WGS84 Geoid Height Altitude Offset for external GPS data on IOS - ios

For an application I'm writing we are interfacing IOS devices with an external sensor which outputs GPS data over a local wifi network. This data comes across in a "raw" format with respect to altitude. In general all GPS altitude needs to have a correction factor applied related to the WGS84 geoid height based on the current location.
For example, in the following Geo Control Point (http://www.ngs.noaa.gov/cgi-bin/ds_mark.prl?PidBox=HV9830) which resides at Lat 38 56 36.77159 and a Lon 077 01 08.34929
HV9830* NAD 83(2011) POSITION- 38 56 36.77159(N) 077 01 08.34929(W) ADJUSTED
HV9830* NAD 83(2011) ELLIP HT- 42.624 (meters) (06/27/12) ADJUSTED
HV9830* NAD 83(2011) EPOCH - 2010.00
HV9830* NAVD 88 ORTHO HEIGHT - 74.7 (meters) 245. (feet) VERTCON
HV9830 ______________________________________________________________________
HV9830 GEOID HEIGHT - -32.02 (meters) GEOID12A
HV9830 NAD 83(2011) X - 1,115,795.966 (meters) COMP
HV9830 NAD 83(2011) Y - -4,840,360.447 (meters) COMP
HV9830 NAD 83(2011) Z - 3,987,471.457 (meters) COMP
HV9830 LAPLACE CORR - -2.38 (seconds) DEFLEC12A
You can see that the Geoid Height is -32 meters. So given a RAW GPS reading near this point one would have to apply a correction of -32 meters in order to calculate the correct altitude. (Note:corrections are negative so you would actually be subtracting a negative and thus shifting the reading up 32 meters).
As opposed to Android it is our understanding that with regards to coreLocation this GeoidHeight information is automagically calculated internally by IOS. Where we are running into difficulty is that we are using a local wifi network with a sensor that calculates uncorrected GPS and collecting both the external sensor data as well as coreLocation readings for GPS. I was wondering if anybody was aware of a library (C/Objective-C) which has the Geoid information and can help me do these calculations on the fly when I'm reading the raw GPS signal from our sensor package.
Thank you for your help.
Side note: Please don't suggest I look at the following post: Get altitude by longitude and latitude in Android This si a good solution however we do not have a live internet connection so we cannot make a live query to Goole or USGS.

I've gone ahead and solved my problems here. What I did was create an ObjectiveC implementation of a c implementation of fortran code to do what I needed. The original c can be found here: http://sourceforge.net/projects/egm96-f477-c/
You would need to download the project from source forge in order to access the input files required for this code: CORCOEF and EGM96
My objective-c implementation is as follows:
GeoidCalculator.h
#import <Foundation/Foundation.h>
#interface GeoidCalculator : NSObject
+ (GeoidCalculator *)instance;
-(double) getHeightFromLat:(double)lat andLon:(double)lon;
-(double) getCurrentHeightOffset;
-(void) updatePositionWithLatitude:(double)lat andLongitude:(double)lon;
#end
GeoidCalculator.m
#import "GeoidCalculator.h"
#import <stdio.h>
#import <math.h>
#define l_value (65341)
#define _361 (361)
#implementation GeoidCalculator
static int nmax;
static double currentHeight;
static double cc[l_value+ 1], cs[l_value+ 1], hc[l_value+ 1], hs[l_value+ 1],
p[l_value+ 1], sinml[_361+ 1], cosml[_361+ 1], rleg[_361+ 1];
+ (GeoidCalculator *)instance {
static GeoidCalculator *_instance = nil;
#synchronized (self) {
if (_instance == nil) {
_instance = [[self alloc] init];
init_arrays();
currentHeight = -9999;
}
}
return _instance;
}
- (double)getHeightFromLat:(double)lat andLon:(double)lon {
[self updatePositionWithLatitude:lat andLongitude:lon];
return [self getCurrentHeightOffset];
}
- (double)getCurrentHeightOffset {
return currentHeight;
}
- (void)updatePositionWithLatitude:(double)lat andLongitude:(double)lon {
const double rad = 180 / M_PI;
double flat, flon, u;
flat = lat; flon = lon;
/*compute the geocentric latitude,geocentric radius,normal gravity*/
u = undulation(flat / rad, flon / rad, nmax, nmax + 1);
/*u is the geoid undulation from the egm96 potential coefficient model
including the height anomaly to geoid undulation correction term
and a correction term to have the undulations refer to the
wgs84 ellipsoid. the geoid undulation unit is meters.*/
currentHeight = u;
}
double hundu(unsigned nmax, double p[l_value+ 1],
double hc[l_value+ 1], double hs[l_value+ 1],
double sinml[_361+ 1], double cosml[_361+ 1], double gr, double re,
double cc[l_value+ 1], double cs[l_value+ 1]) {/*constants for wgs84(g873);gm in units of m**3/s**2*/
const double gm = .3986004418e15, ae = 6378137.;
double arn, ar, ac, a, b, sum, sumc, sum2, tempc, temp;
int k, n, m;
ar = ae / re;
arn = ar;
ac = a = b = 0;
k = 3;
for (n = 2; n <= nmax; n++) {
arn *= ar;
k++;
sum = p[k] * hc[k];
sumc = p[k] * cc[k];
sum2 = 0;
for (m = 1; m <= n; m++) {
k++;
tempc = cc[k] * cosml[m] + cs[k] * sinml[m];
temp = hc[k] * cosml[m] + hs[k] * sinml[m];
sumc += p[k] * tempc;
sum += p[k] * temp;
}
ac += sumc;
a += sum * arn;
}
ac += cc[1] + p[2] * cc[2] + p[3] * (cc[3] * cosml[1] + cs[3] * sinml[1]);
/*add haco=ac/100 to convert height anomaly on the ellipsoid to the undulation
add -0.53m to make undulation refer to the wgs84 ellipsoid.*/
return a * gm / (gr * re) + ac / 100 - .53;
}
void dscml(double rlon, unsigned nmax, double sinml[_361+ 1], double cosml[_361+ 1]) {
double a, b;
int m;
a = sin(rlon);
b = cos(rlon);
sinml[1] = a;
cosml[1] = b;
sinml[2] = 2 * b * a;
cosml[2] = 2 * b * b - 1;
for (m = 3; m <= nmax; m++) {
sinml[m] = 2 * b * sinml[m - 1] - sinml[m - 2];
cosml[m] = 2 * b * cosml[m - 1] - cosml[m - 2];
}
}
void dhcsin(unsigned nmax, double hc[l_value+ 1], double hs[l_value+ 1]) {
// potential coefficient file
//f_12 = fopen("EGM96", "rb");
NSString* path2 = [[NSBundle mainBundle] pathForResource:#"EGM96" ofType:#""];
FILE* f_12 = fopen(path2.UTF8String, "rb");
if (f_12 == NULL) {
NSLog([path2 stringByAppendingString:#" not found"]);
}
int n, m;
double j2, j4, j6, j8, j10, c, s, ec, es;
/*the even degree zonal coefficients given below were computed for the
wgs84(g873) system of constants and are identical to those values
used in the NIMA gridding procedure. computed using subroutine
grs written by N.K. PAVLIS*/
j2 = 0.108262982131e-2;
j4 = -.237091120053e-05;
j6 = 0.608346498882e-8;
j8 = -0.142681087920e-10;
j10 = 0.121439275882e-13;
m = ((nmax + 1) * (nmax + 2)) / 2;
for (n = 1; n <= m; n++)hc[n] = hs[n] = 0;
while (6 == fscanf(f_12, "%i %i %lf %lf %lf %lf", &n, &m, &c, &s, &ec, &es)) {
if (n > nmax)continue;
n = (n * (n + 1)) / 2 + m + 1;
hc[n] = c;
hs[n] = s;
}
hc[4] += j2 / sqrt(5);
hc[11] += j4 / 3;
hc[22] += j6 / sqrt(13);
hc[37] += j8 / sqrt(17);
hc[56] += j10 / sqrt(21);
fclose(f_12);
}
void legfdn(unsigned m, double theta, double rleg[_361+ 1], unsigned nmx)
/*this subroutine computes all normalized legendre function
in "rleg". order is always
m, and colatitude is always theta (radians). maximum deg
is nmx. all calculations in double precision.
ir must be set to zero before the first call to this sub.
the dimensions of arrays rleg must be at least equal to nmx+1.
Original programmer :Oscar L. Colombo, Dept. of Geodetic Science
the Ohio State University, August 1980
ineiev: I removed the derivatives, for they are never computed here*/
{
static double drts[1301], dirt[1301], cothet, sithet, rlnn[_361+ 1];
static int ir;
int nmx1 = nmx + 1, nmx2p = 2 * nmx + 1, m1 = m + 1, m2 = m + 2, m3 = m + 3, n, n1, n2;
if (!ir) {
ir = 1;
for (n = 1; n <= nmx2p; n++) {
drts[n] = sqrt(n);
dirt[n] = 1 / drts[n];
}
}
cothet = cos(theta);
sithet = sin(theta);
/*compute the legendre functions*/
rlnn[1] = 1;
rlnn[2] = sithet * drts[3];
for (n1 = 3; n1 <= m1; n1++) {
n = n1 - 1;
n2 = 2 * n;
rlnn[n1] = drts[n2 + 1] * dirt[n2] * sithet * rlnn[n];
}
switch (m) {
case 1:
rleg[2] = rlnn[2];
rleg[3] = drts[5] * cothet * rleg[2];
break;
case 0:
rleg[1] = 1;
rleg[2] = cothet * drts[3];
break;
}
rleg[m1] = rlnn[m1];
if (m2 <= nmx1) {
rleg[m2] = drts[m1 * 2 + 1] * cothet * rleg[m1];
if (m3 <= nmx1)
for (n1 = m3; n1 <= nmx1; n1++) {
n = n1 - 1;
if ((!m && n < 2) || (m == 1 && n < 3))continue;
n2 = 2 * n;
rleg[n1] = drts[n2 + 1] * dirt[n + m] * dirt[n - m] *
(drts[n2 - 1] * cothet * rleg[n1 - 1] - drts[n + m - 1] * drts[n - m - 1] * dirt[n2 - 3] * rleg[n1 - 2]);
}
}
}
void radgra(double lat, double lon, double *rlat, double *gr, double *re)
/*this subroutine computes geocentric distance to the point,
the geocentric latitude,and
an approximate value of normal gravity at the point based
the constants of the wgs84(g873) system are used*/
{
const double a = 6378137., e2 = .00669437999013, geqt = 9.7803253359, k = .00193185265246;
double n, t1 = sin(lat) * sin(lat), t2, x, y, z;
n = a / sqrt(1 - e2 * t1);
t2 = n * cos(lat);
x = t2 * cos(lon);
y = t2 * sin(lon);
z = (n * (1 - e2)) * sin(lat);
*re = sqrt(x * x + y * y + z * z);/*compute the geocentric radius*/
*rlat = atan(z / sqrt(x * x + y * y));/*compute the geocentric latitude*/
*gr = geqt * (1 + k * t1) / sqrt(1 - e2 * t1);/*compute normal gravity:units are m/sec**2*/
}
double undulation(double lat, double lon, int nmax, int k) {
double rlat, gr, re;
int i, j, m;
radgra(lat, lon, &rlat, &gr, &re);
rlat = M_PI / 2 - rlat;
for (j = 1; j <= k; j++) {
m = j - 1;
legfdn(m, rlat, rleg, nmax);
for (i = j; i <= k; i++)p[(i - 1) * i / 2 + m + 1] = rleg[i];
}
dscml(lon, nmax, sinml, cosml);
return hundu(nmax, p, hc, hs, sinml, cosml, gr, re, cc, cs);
}
void init_arrays(void) {
int ig, i, n, m;
double t1, t2;
NSString* path1 = [[NSBundle mainBundle] pathForResource:#"CORCOEF" ofType:#""];
//correction coefficient file: modified with 'sed -e"s/D/e/g"' to be read with fscanf
FILE* f_1 = fopen([path1 cStringUsingEncoding:1], "rb");
if (f_1 == NULL) {
NSLog([path1 stringByAppendingString:#" not found"]);
}
nmax = 360;
for (i = 1; i <= l_value; i++)cc[i] = cs[i] = 0;
while (4 == fscanf(f_1, "%i %i %lg %lg", &n, &m, &t1, &t2)) {
ig = (n * (n + 1)) / 2 + m + 1;
cc[ig] = t1;
cs[ig] = t2;
}
/*the correction coefficients are now read in*/
/*the potential coefficients are now read in and the reference
even degree zonal harmonic coefficients removed to degree 6*/
dhcsin(nmax, hc, hs);
fclose(f_1);
}
#end
I've done some limited testing against the Geoid Height Calculator (http://www.unavco.org/community_science/science-support/geoid/geoid.html) and looks like everything is a match
UPDATE iOS8 or Greater
As of IOS8 This code might not work correctly. You may need to change how the bundle is loaded:
[[NSBundle mainBundle] pathForResource:#"EGM96" ofType:#""];
Do some googling or add a comment here.

Impressive stuff Jeef! I just used your code to create this sqlite which may be easier to add/use in a project, assuming integer precision for lat/lon is good enough:
https://github.com/vectorstofinal/geoid_heights

You could use GeoTrans.
Provided by http://earth-info.nga.mil/GandG/geotrans/index.html
The keyword is "vertical datum". So you want to convert from WGS84 to e.g EGM96 vertical datum. Make sure which Geoid modell you want to use. EGM96 is one of that.
Maybe these answer help you, too:
How to calculate the altitude above from mean sea level
Next read the ios Open Source License Text: Available in
Settings -> General -> About -> Legal -> License ...
There you get a list of all libs that ios uses. One of them I found was the calculation of magnetic decilination usung a sw of USGS. Chances are verry high that the Geoid height calculation is listed there too.

Related

How to get XY value from ct in Philips Hue?

How to get XY value from ct.
Ex: ct = 217, I want to get x="0.3127569", y= "0.32908".
I'm able to convert XY value into ct value using this below code.
float R1 = [hue[0] floatValue];
float S1 = [hue[1] floatValue];
float result = ((R1-0.332)/(S1-0.1858));
NSString *ctString = [NSString stringWithFormat:#"%f", ((-449*result*result*result)+(3525*result*result)-(6823.3*result)+(5520.33))];
float micro2 = (float) (1 / [ctString floatValue] * 1000000);
NSString *ctValue = [NSString stringWithFormat:#"%f", micro2];
ctValue = [NSString stringWithFormat:#"%d", [ctValue intValue]];
if ([ctValue integerValue] < 153) {
ctValue = [NSString stringWithFormat:#"%d", 153];
}
Now I want reverse value, which is from ct to XY.
On Phillips HUE
2000K maps to 500 and 6500K maps to 153 given in ct as color temperature but can be thought as actually being Mired.
Mired means micro reciprocal degree wikipedia.
ct is possibly used because it is not 100% Mired. Quite sure Phillips uses a lookup table as a lot CIE algorithms do because there are just 347 indexes in this range from 153 to 500.
The following is not a solution, it's just simple concept of a lookup table.
And as the CIE 1931 xy to CCT Formula by McCamy suggests found here it is possible to use a lookup table to find x and y as well.
A table can be found here but i am not sure if that is the right lookup table.
reminder so the following is not a solution, but to find an reverse algo the code may help.
typedef int Kelvin;
typedef float Mired;
Mired linearMiredByKelvin(Kelvin k) {
if (k==0) return 0;
return 1000000.0/k;
}
-(void)mired {
Mired miredMin = 2000.0/13.0; // 153,84 = reciprocal 6500K
Mired miredMax = 500.0; // 500,00 = reciprocal 2000K
Mired lookupMiredByKelvin[6501]; //max 6500 Kelvin + 1 safe index
//Kelvin lookupKelvinByMired[501]; //max 500 Mired + 1 safe index
// dummy stuff, empty unused table space
for (Kelvin k = 0; k < 2000; k++) {
lookupMiredByKelvin[k] = 0;
}
//for (Mired m = 0.0; m < 154.0; m++) {
// lookupKelvinByMired[(int)m] = 0;
//}
for (Kelvin k=2000; k<6501; k++) {
Mired linearMired = linearMiredByKelvin(k);
float dimm = (linearMired - miredMin) / ( miredMax - miredMin);
Kelvin ct = (Kelvin)(1000000.0/(dimm*miredMax - dimm*miredMin + miredMin));
lookupMiredByKelvin[k] = linearMiredByKelvin(ct);
if (k==2000 || k==2250 || k==2500 || k==2750 ||
k==3000 || k==3250 || k==3500 || k==3750 ||
k==4000 || k==4250 || k==4500 || k==4750 ||
k==5000 || k==5250 || k==5500 || k==5750 ||
k==6000 || k==6250 || k==6500 || k==6501 )
fprintf(stderr,"%d %f %f\n",ct, dimm, lookupMiredByKelvin[k]);
}
}
at least this is proof that x and y will not sit on a simple vector.
CCT means correlated colour temperature and like the implementation in the question shows can be calculated via n= (x-0.3320)/(0.1858-y); CCT = 437*n^3 + 3601*n^2 + 6861*n + 5517. (after McCamy)
but a cct=217 is out of range of above link'ed lookup table.
following the idea in this git-repo from colour-science
and ported to C it could look like..
void CCT_to_xy_CIE_D(float cct) {
//if (CCT < 4000 || CCT > 25000) fprintf(stderr, "Correlated colour temperature must be in domain, unpredictable results may occur! \n");
float x = calculateXviaCCT(cct);
float y = calculateYviaX(x);
NSLog(#"cct=%f x%f y%f",cct,x,y);
}
float calculateXviaCCT(float cct) {
float cct_3 = pow(cct, 3); //(cct*cct*cct);
float cct_2 = pow(cct, 2); //(cct*cct);
if (cct<=7000)
return -4.607 * pow(10, 9) / cct_3 + 2.9678 * pow(10, 6) / cct_2 + 0.09911 * pow(10, 3) / cct + 0.244063;
return -2.0064 * pow(10, 9) / cct_3 + 1.9018 * pow(10, 6) / cct_2 + 0.24748 * pow(10, 3) / cct + 0.23704;
}
float calculateYviaX(float x) {
return -3.000 * pow(x, 2) + 2.870 * x - 0.275;
}
CCT_to_xy_CIE_D(6504.38938305); //proof of concept
//cct=6504.389160 x0.312708 y0.329113
CCT_to_xy_CIE_D(217.0);
//cct=217.000000 x-387.131073 y-450722.750000
// so for sure Phillips hue temperature given in ct between 153-500 is not a good starting point
//but
CCT_to_xy_CIE_D(2000.0);
//cct=2000.000000 x0.459693 y0.410366
this seems to work fine with CCT between 2000 and 25000, but maybe confusing is CCT is given in Kelvin here.
EDIT
This has been through so many revisions and ideas. To keep it simple I edited most of that out and just give you the final result.
This fits your function perfectly except for a region in the middle (temp from 256 to 316) where it deviates a bit.
The problem with your function is that it has approximately infinite solutions, so to solve it nicely you need more constraints, but what? Ol Sen's reference https://www.waveformlighting.com/tech/calculate-color-temperature-cct-from-cie-1931-xy-coordinates discusses it in some detail and then mentions that you want a Duv to be zero. It also gives a way to calculate Duv and so I added that to my optimiser and voila!
Nice and smooth. The optimiser now solves for x and y that both satisfies your function and also minimises Duv.
To get it to work nicely I had to scale Duv quite a bit. That page mentions that Duv should be very small so I think this is a good thing. Also, as the temp increases the scaling should to help the optimiser.
Below prints from 153 to 500.
#import <Foundation/Foundation.h>
// Function taken from your code
// Simplified a bit
int ctFuncI ( float x, float y )
{
// float R1 = [hue[0] floatValue];
// float S1 = [hue[1] floatValue];
float result = (x-0.332)/(y-0.1858);
float cubic = - 449 * result * result * result + 3525 * result * result - 6823.3 * result + 5520.33;
float micro2 = 1 / cubic * 1000000;
int ct = ( int )( micro2 + 0.5 );
if ( ct < 153 )
{
ct = 153;
}
return ct;
}
// Need this
// Float version of your code
float ctFuncF ( float x, float y )
{
// float R1 = [hue[0] floatValue];
// float S1 = [hue[1] floatValue];
float result = (x-0.332)/(y-0.1858);
float cubic = - 449 * result * result * result + 3525 * result * result - 6823.3 * result + 5520.33;
return 1000000 / cubic;
}
// We need an additional constraint
// https://www.waveformlighting.com/tech/calculate-duv-from-cie-1931-xy-coordinates
// Given x, y calculate Duv
// We want this to be 0
float duv ( float x, float y )
{
float f = 1 / ( - 2 * x + 12 * y + 3 );
float u = 4 * x * f;
float v = 6 * y * f;
// I'm typing float but my heart yells double
float k6 = -0.00616793;
float k5 = 0.0893944;
float k4 = -0.5179722;
float k3 = 1.5317403;
float k2 = -2.4243787;
float k1 = 1.925865;
float k0 = -0.471106;
float du = u - 0.292;
float dv = v - 0.24;
float Lfp = sqrt ( du * du + dv * dv );
float a = acos( du / Lfp );
float Lbb = k6 * pow ( a, 6 ) + k5 * pow( a, 5 ) + k4 * pow( a, 4 ) + k3 * pow( a, 3 ) + k2 * pow(a,2) + k1 * a + k0;
return Lfp - Lbb;
}
// Solver!
// Returns iterations
int ctSolve ( int ct, float * x, float * y )
{
int iter = 0;
float dx = 0.001;
float dy = 0.001;
// Error
// Note we scale duv a bit
// Seems the higher the temp, the higher scale we require
// Also note the jump at 255 ...
float s = 1000 * ( ct > 255 ? 10 : 1 );
float d = fabs( ctFuncF ( * x, * y ) - ct ) + s * fabs( duv ( * x, * y ) );
// Approx
while ( d > 0.5 && iter < 250 )
{
iter ++;
dx *= fabs( ctFuncF ( * x + dx, * y ) - ct ) + s * fabs( duv ( * x + dx, * y ) ) < d ? 1.2 : - 0.5;
dy *= fabs( ctFuncF ( * x, * y + dy ) - ct ) + s * fabs( duv ( * x, * y + dy ) ) < d ? 1.2 : - 0.5;
* x += dx;
* y += dy;
d = fabs( ctFuncF ( * x, * y ) - ct ) + s * fabs( duv ( * x, * y ) );
}
return iter;
}
// Tester
int main(int argc, const char * argv[]) {
#autoreleasepool
{
// insert code here...
NSLog(#"Hello, World!");
float x, y;
int sume = 0;
int sumi = 0;
for ( int ct = 153; ct <= 500; ct ++ )
{
// Initial guess
x = 0.4;
y = 0.4;
// Approx
int iter = ctSolve ( ct, & x, & y );
// CT and error
int ctEst = ctFuncI ( x, y );
int e = ct - ctEst;
// Diagnostics
sume += abs ( e );
sumi += iter;
// Print out results
NSLog ( #"want ct = %d x = %f y = %f got ct %d in %d iter error %d", ct, x, y, ctEst, iter, e );
}
NSLog ( #"Sum of abs errors %d iterations %d", sume, sumi );
}
return 0;
}
To use it, do as below.
// To call it, init x and y to some guess
float x = 0.4;
float y = 0.4;
// Then call solver with your temp
int ct = 217;
ctSolve( ct, & x, & y ); // Note you pass references to x and y
// Done, answer now in x and y
a bit more compact answer and functions to convert back and forth..
beware there are rounding issues because McCamy's formula relies and mathematical assumptions. And so the backward calculation does also.
if you want to find more results search directly for "n= (x-0.3320)/(0.1858-y); CCT = 437*n^3 + 3601*n^2 + 6861*n + 5517" there are plenty of different methods to convert back and forth.
so here Phillips-Hue #[#x,#y] to Phillips-ct,Phillips-ct to CCT, CCT to x,y
void CCT_to_xy_CIE_D(float cct) {
//if (CCT < 4000 || CCT > 25000) fprintf(stderr, "Correlated colour temperature must be in domain, unpredictable results may occur! \n");
float x = calculateXviaCCT(cct);
float y = calculateYviaX(x);
fprintf(stderr,"cct=%f x%f y%f",cct,x,y);
}
float calculateXviaCCT(float cct) {
float cct_3 = pow(cct, 3); //(cct*cct*cct);
float cct_2 = pow(cct, 2); //(cct*cct);
if (cct<=7000.0)
return -4.607 * pow(10, 9) / cct_3 + 2.9678 * pow(10, 6) / cct_2 + 0.09911 * pow(10, 3) / cct + 0.244063;
return -2.0064 * pow(10, 9) / cct_3 + 1.9018 * pow(10, 6) / cct_2 + 0.24748 * pow(10, 3) / cct + 0.23704;
}
float calculateYviaX(float x) {
return -3.000 * x*x + 2.870 * x - 0.275;
}
int calculate_PhillipsHueCT_withCCT(float cct) {
if (cct>6500.0) return 2000.0/13.0;
if (cct<2000.0) return 500.0;
//return (float) (1 / cct * 1000000); // same as..
return 1000000 / cct;
}
float calculate_CCT_withPhillipsHueCT(float ct) {
if (ct == 0.0) return 0.0;
return 1000000 / ct;
}
float calculate_CCT_withHueXY(NSArray *hue) {
float x = [hue[0] floatValue]; //R1
float y = [hue[1] floatValue]; //S1
//x = 0.312708; y = 0.329113;
float n = (x-0.3320)/(0.1858-y);
float cct = 437.0*n*n*n + 3601.0*n*n + 6861.0*n + 5517.0;
return cct;
}
// MC Camy formula n=(x-0.3320)/(0.1858-y); cct = 437*n^3 + 3601*n^2 + 6861*n + 5517;
-(void)testPhillipsHueCt_backAndForth {
NSArray *hue = #[#(0.312708),#(0.329113)];
float cct = calculate_CCT_withHueXY(hue);
float ct = calculate_PhillipsHueCT_withCCT(cct);
NSLog(#"ct %f",ct);
CCT_to_xy_CIE_D(cct); // check
CCT_to_xy_CIE_D(6504.38938305); //proof of concept
CCT_to_xy_CIE_D(2000.0);
CCT_to_xy_CIE_D(calculate_CCT_withPhillipsHueCT(217.0));
}

Google Maps heat map color by average weight

The Google Maps iOS SDK's heat map (more specifically the Google-Maps-iOS-Utils framework) decides the color to render an area in essentially by calculating the density of the points in that area.
However, I would like to instead select the color based on the average weight or intensity of the points in that area.
From what I understand, this behavior is not built in (but who knows––the documentation sort of sucks). The file where the color-picking is decided is I think in /src/Heatmap/GMUHeatmapTileLayer.mThis is a relatively short file, but I am not very well versed in Objective-C, so I am having some difficulty figuring out what does what. I think -tileForX:y:zoom: in GMUHeatmapTileLayer.m is the important function, but I'm not sure and even if it is, I don't quite know how to modify it. Towards the end of this method, the data is 'convolved' first horizontally and then vertically. I think this is where the intensities are actually calculated. Unfortunately, I do not know exactly what it's doing, and I am afraid of changing things because I suck at obj-c. This is what the convolve parts of this method look like:
- (UIImage *)tileForX:(NSUInteger)x y:(NSUInteger)y zoom:(NSUInteger)zoom {
// ...
// Convolve data.
int lowerLimit = (int)data->_radius;
int upperLimit = paddedTileSize - (int)data->_radius - 1;
// Convolve horizontally first.
float *intermediate = calloc(paddedTileSize * paddedTileSize, sizeof(float));
for (int y = 0; y < paddedTileSize; y++) {
for (int x = 0; x < paddedTileSize; x++) {
float value = intensity[y * paddedTileSize + x];
if (value != 0) {
// convolve to x +/- radius bounded by the limit we care about.
int start = MAX(lowerLimit, x - (int)data->_radius);
int end = MIN(upperLimit, x + (int)data->_radius);
for (int x2 = start; x2 <= end; x2++) {
float scaledKernel = value * [data->_kernel[x2 - x + data->_radius] floatValue];
// I THINK THIS IS WHERE I NEED TO MAKE THE CHANGE
intermediate[y * paddedTileSize + x2] += scaledKernel;
// ^
}
}
}
}
free(intensity);
// Convole vertically to get final intensity.
float *finalIntensity = calloc(kGMUTileSize * kGMUTileSize, sizeof(float));
for (int x = lowerLimit; x <= upperLimit; x++) {
for (int y = 0; y < paddedTileSize; y++) {
float value = intermediate[y * paddedTileSize + x];
if (value != 0) {
int start = MAX(lowerLimit, y - (int)data->_radius);
int end = MIN(upperLimit, y + (int)data->_radius);
for (int y2 = start; y2 <= end; y2++) {
float scaledKernel = value * [data->_kernel[y2 - y + data->_radius] floatValue];
// I THINK THIS IS WHERE I NEED TO MAKE THE CHANGE
finalIntensity[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit] += scaledKernel;
// ^
}
}
}
}
free(intermediate);
// ...
}
This is the method where the intensities are calculated for each iteration, right? If so, how can I change this to achieve my desired effect (average, not summative colors, which I think are proportional to intensity).
So: How can I have averaged instead of summed intensities by modifying the framework?
I think you are on the right track. To calculate average you divide the point sum by the point count. Since you already have the sums calculated, I think an easy solution would be to also save the count for each point. If I understand it correctly, this it what you have to do.
When allocating memory for the sums also allocate memory for the counts
// At this place
float *intermediate = calloc(paddedTileSize * paddedTileSize, sizeof(float));
// Add this line, calloc will initialize them to zero
int *counts = calloc(paddedTileSize * paddedTileSize, sizeof(int));
Then increase the count in each loop.
// Below this line (first loop)
intermediate[y * paddedTileSize + x2] += scaledKernel;
// Add this
counts[y * paddedTileSize + x2]++;
// And below this line (second loop)
finalIntensity[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit] += scaledKernel;
// Add this
counts[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit]++;
After the two loops you should have two arrays, one with your sums finalIntensity and one with your counts counts. Now go through the values and calculate the averages.
for (int y = 0; y < paddedTileSize; y++) {
for (int x = 0; x < paddedTileSize; x++) {
int n = y * paddedTileSize + x;
if (counts[n] != 0)
finalIntensity[n] = finalIntensity[n] / counts[n];
}
}
free(counts);
The finalIntensity should now contain your averages.
If you prefer, and the rest of the code makes it possible, you can skip the last loop and instead do the division when using the final intensity values. Just change any subsequent finalIntensity[n] to counts[n] == 0 ? finalIntensity[n] : finalIntensity[n] / counts[n].
I may have just solved the same issue for the java version.
My problem was having a custom gradient with 12 different values.
But my actual weighted data does not necessarily contain all intensity values from 1 to 12.
The problem is, the highest intensity value gets mapped to the highest color.
Also 10 datapoints with intensity 1 that are close by will get the same color as a single point with intensity 12.
So the function where the tile gets created is a good starting point:
Java:
public Tile getTile(int x, int y, int zoom) {
// ...
// Quantize points
int dim = TILE_DIM + mRadius * 2;
double[][] intensity = new double[dim][dim];
int[][] count = new int[dim][dim];
for (WeightedLatLng w : points) {
Point p = w.getPoint();
int bucketX = (int) ((p.x - minX) / bucketWidth);
int bucketY = (int) ((p.y - minY) / bucketWidth);
intensity[bucketX][bucketY] += w.getIntensity();
count[bucketX][bucketY]++;
}
// Quantize wraparound points (taking xOffset into account)
for (WeightedLatLng w : wrappedPoints) {
Point p = w.getPoint();
int bucketX = (int) ((p.x + xOffset - minX) / bucketWidth);
int bucketY = (int) ((p.y - minY) / bucketWidth);
intensity[bucketX][bucketY] += w.getIntensity();
count[bucketX][bucketY]++;
}
for(int bx = 0; bx < dim; bx++)
for (int by = 0; by < dim; by++)
if (count[bx][by] != 0)
intensity[bx][by] /= count[bx][by];
//...
I added a counter and count every addition to the intensities, after that I go through every intensity and calculate the average.
For C:
- (UIImage *)tileForX:(NSUInteger)x y:(NSUInteger)y zoom:(NSUInteger)zoom {
//...
// Quantize points.
int paddedTileSize = kGMUTileSize + 2 * (int)data->_radius;
float *intensity = calloc(paddedTileSize * paddedTileSize, sizeof(float));
int *count = calloc(paddedTileSize * paddedTileSize, sizeof(int));
for (GMUWeightedLatLng *item in points) {
GQTPoint p = [item point];
int x = (int)((p.x - minX) / bucketWidth);
// Flip y axis as world space goes south to north, but tile content goes north to south.
int y = (int)((maxY - p.y) / bucketWidth);
// If the point is just on the edge of the query area, the bucketing could put it outside
// bounds.
if (x >= paddedTileSize) x = paddedTileSize - 1;
if (y >= paddedTileSize) y = paddedTileSize - 1;
intensity[y * paddedTileSize + x] += item.intensity;
count[y * paddedTileSize + x] ++;
}
for (GMUWeightedLatLng *item in wrappedPoints) {
GQTPoint p = [item point];
int x = (int)((p.x + wrappedPointsOffset - minX) / bucketWidth);
// Flip y axis as world space goes south to north, but tile content goes north to south.
int y = (int)((maxY - p.y) / bucketWidth);
// If the point is just on the edge of the query area, the bucketing could put it outside
// bounds.
if (x >= paddedTileSize) x = paddedTileSize - 1;
if (y >= paddedTileSize) y = paddedTileSize - 1;
// For wrapped points, additional shifting risks bucketing slipping just outside due to
// numerical instability.
if (x < 0) x = 0;
intensity[y * paddedTileSize + x] += item.intensity;
count[y * paddedTileSize + x] ++;
}
for(int i=0; i < paddedTileSize * paddedTileSize; i++)
if (count[i] != 0)
intensity[i] /= count[i];
Next is the convolving.
What I did there, is to make sure that the calculated value does not go over the maximum in my data.
Java:
// Convolve it ("smoothen" it out)
double[][] convolved = convolve(intensity, mKernel, mMaxAverage);
// the mMaxAverage gets set here:
public void setWeightedData(Collection<WeightedLatLng> data) {
// ...
// Add points to quad tree
for (WeightedLatLng l : mData) {
mTree.add(l);
mMaxAverage = Math.max(l.getIntensity(), mMaxAverage);
}
// ...
// And finally the convolve method:
static double[][] convolve(double[][] grid, double[] kernel, double max) {
// ...
intermediate[x2][y] += val * kernel[x2 - (x - radius)];
if (intermediate[x2][y] > max) intermediate[x2][y] = max;
// ...
outputGrid[x - radius][y2 - radius] += val * kernel[y2 - (y - radius)];
if (outputGrid[x - radius][y2 - radius] > max ) outputGrid[x - radius][y2 - radius] = max;
For C:
// To get the maximum average you could do that here:
- (void)setWeightedData:(NSArray<GMUWeightedLatLng *> *)weightedData {
_weightedData = [weightedData copy];
for (GMUWeightedLatLng *dataPoint in _weightedData)
_maxAverage = Math.max(dataPoint.intensity, _maxAverage)
// ...
// And then simply in the convolve section
intermediate[y * paddedTileSize + x2] += scaledKernel;
if (intermediate[y * paddedTileSize + x2] > _maxAverage)
intermediate[y * paddedTileSize + x2] = _maxAverage;
// ...
finalIntensity[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit] += scaledKernel;
if (finalIntensity[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit] > _maxAverage)
finalIntensity[(y2 - lowerLimit) * kGMUTileSize + x - lowerLimit] = _maxAverage;
And finally the coloring
Java:
// The maximum intensity is simply the size of my gradient colors array (or the starting points)
Bitmap bitmap = colorize(convolved, mColorMap, mGradient.mStartPoints.length);
For C:
// Generate coloring
// ...
float max = [data->_maxIntensities[zoom] floatValue];
max = _gradient.startPoints.count;
I did this in Java and it worked for me, not sure about the C-code though.
You have to play around with the radius and you could even edit the kernel. Because I found that when I have a lot of homogeneous data (i.e. little variation in the intensities, or a lot of data in general) the heat map will degenerate to a one-colored overlay, because the gradient on the edges will get smaller and smaller.
But hope this helps anyway.
// Erik

Two drawn lines affecting each other [duplicate]

Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
This question does not appear to be about programming within the scope defined in the help center.
Closed 6 years ago.
Locked. This question and its answers are locked because the question is off-topic but has historical significance. It is not currently accepting new answers or interactions.
How do I determine whether or not two lines intersect, and if they do, at what x,y point?
There’s a nice approach to this problem that uses vector cross products. Define the 2-dimensional vector cross product v × w to be vx wy − vy wx.
Suppose the two line segments run from p to p + r and from q to q + s. Then any point on the first line is representable as p + t r (for a scalar parameter t) and any point on the second line as q + u s (for a scalar parameter u).
The two lines intersect if we can find t and u such that:
p + t r = q + u s
Cross both sides with s, getting
(p + t r) × s = (q + u s) × s
And since s × s = 0, this means
t (r × s) = (q − p) × s
And therefore, solving for t:
t = (q − p) × s / (r × s)
In the same way, we can solve for u:
(p + t r) × r = (q + u s) × r
u (s × r) = (p − q) × r
u = (p − q) × r / (s × r)
To reduce the number of computation steps, it's convenient to rewrite this as follows (remembering that s × r = − r × s):
u = (q − p) × r / (r × s)
Now there are four cases:
If r × s = 0 and (q − p) × r = 0, then the two lines are collinear.
In this case, express the endpoints of the second segment (q and q + s) in terms of the equation of the first line segment (p + t r):
t0 = (q − p) · r / (r · r)
t1 = (q + s − p) · r / (r · r) = t0 + s · r / (r · r)
If the interval between t0 and t1 intersects the interval [0, 1] then the line segments are collinear and overlapping; otherwise they are collinear and disjoint.
Note that if s and r point in opposite directions, then s · r < 0 and so the interval to be checked is [t1, t0] rather than [t0, t1].
If r × s = 0 and (q − p) × r ≠ 0, then the two lines are parallel and non-intersecting.
If r × s ≠ 0 and 0 ≤ t ≤ 1 and 0 ≤ u ≤ 1, the two line segments meet at the point p + t r = q + u s.
Otherwise, the two line segments are not parallel but do not intersect.
Credit: this method is the 2-dimensional specialization of the 3D line intersection algorithm from the article "Intersection of two lines in three-space" by Ronald Goldman, published in Graphics Gems, page 304. In three dimensions, the usual case is that the lines are skew (neither parallel nor intersecting) in which case the method gives the points of closest approach of the two lines.
FWIW, the following function (in C) both detects line intersections and determines the intersection point. It is based on an algorithm in Andre LeMothe's "Tricks of the Windows Game Programming Gurus". It's not dissimilar to some of the algorithm's in other answers (e.g. Gareth's). LeMothe then uses Cramer's Rule (don't ask me) to solve the equations themselves.
I can attest that it works in my feeble asteroids clone, and seems to deal correctly with the edge cases described in other answers by Elemental, Dan and Wodzu. It's also probably faster than the code posted by KingNestor because it's all multiplication and division, no square roots!
I guess there's some potential for divide by zero in there, though it hasn't been an issue in my case. Easy enough to modify to avoid the crash anyway.
// Returns 1 if the lines intersect, otherwise 0. In addition, if the lines
// intersect the intersection point may be stored in the floats i_x and i_y.
char get_line_intersection(float p0_x, float p0_y, float p1_x, float p1_y,
float p2_x, float p2_y, float p3_x, float p3_y, float *i_x, float *i_y)
{
float s1_x, s1_y, s2_x, s2_y;
s1_x = p1_x - p0_x; s1_y = p1_y - p0_y;
s2_x = p3_x - p2_x; s2_y = p3_y - p2_y;
float s, t;
s = (-s1_y * (p0_x - p2_x) + s1_x * (p0_y - p2_y)) / (-s2_x * s1_y + s1_x * s2_y);
t = ( s2_x * (p0_y - p2_y) - s2_y * (p0_x - p2_x)) / (-s2_x * s1_y + s1_x * s2_y);
if (s >= 0 && s <= 1 && t >= 0 && t <= 1)
{
// Collision detected
if (i_x != NULL)
*i_x = p0_x + (t * s1_x);
if (i_y != NULL)
*i_y = p0_y + (t * s1_y);
return 1;
}
return 0; // No collision
}
BTW, I must say that in LeMothe's book, though he apparently gets the algorithm right, the concrete example he shows plugs in the wrong numbers and does calculations wrong. For example:
(4 * (4 - 1) + 12 * (7 - 1)) / (17 * 4 + 12 * 10)
= 844/0.88
= 0.44
That confused me for hours. :(
The problem reduces to this question: Do two lines from A to B and from C to D intersect? Then you can ask it four times (between the line and each of the four sides of the rectangle).
Here's the vector math for doing it. I'm assuming the line from A to B is the line in question and the line from C to D is one of the rectangle lines. My notation is that Ax is the "x-coordinate of A" and Cy is the "y-coordinate of C." And "*" means dot-product, so e.g. A*B = Ax*Bx + Ay*By.
E = B-A = ( Bx-Ax, By-Ay )
F = D-C = ( Dx-Cx, Dy-Cy )
P = ( -Ey, Ex )
h = ( (A-C) * P ) / ( F * P )
This h number is the key. If h is between 0 and 1, the lines intersect, otherwise they don't. If F*P is zero, of course you cannot make the calculation, but in this case the lines are parallel and therefore only intersect in the obvious cases.
The exact point of intersection is C + F*h.
More Fun:
If h is exactly 0 or 1 the lines touch at an end-point. You can consider this an "intersection" or not as you see fit.
Specifically, h is how much you have to multiply the length of the line in order to exactly touch the other line.
Therefore, If h<0, it means the rectangle line is "behind" the given line (with "direction" being "from A to B"), and if h>1 the rectangle line is "in front" of the given line.
Derivation:
A and C are vectors that point to the start of the line; E and F are the vectors from the ends of A and C that form the line.
For any two non-parallel lines in the plane, there must be exactly one pair of scalar g and h such that this equation holds:
A + E*g = C + F*h
Why? Because two non-parallel lines must intersect, which means you can scale both lines by some amount each and touch each other.
(At first this looks like a single equation with two unknowns! But it isn't when you consider that this is a 2D vector equation, which means this is really a pair of equations in x and y.)
We have to eliminate one of these variables. An easy way is to make the E term zero. To do that, take the dot-product of both sides of the equation using a vector that will dot to zero with E. That vector I called P above, and I did the obvious transformation of E.
You now have:
A*P = C*P + F*P*h
(A-C)*P = (F*P)*h
( (A-C)*P ) / (F*P) = h
I have tried to implement the algorithm so elegantly described by Jason above; unfortunately while working though the mathematics in the debugging I found many cases for which it doesn't work.
For example consider the points A(10,10) B(20,20) C(10,1) D(1,10) gives h=.5 and yet it is clear by examination that these segments are no-where near each other.
Graphing this makes it clear that 0 < h < 1 criteria only indicates that the intercept point would lie on CD if it existed but tells one nothing of whether that point lies on AB.
To ensure that there is a cross point you must do the symmetrical calculation for the variable g and the requirement for interception is:
0 < g < 1 AND 0 < h < 1
Here's an improvement to Gavin's answer. marcp's solution is similar also, but neither postpone the division.
This actually turns out to be a practical application of Gareth Rees' answer as well, because the cross-product's equivalent in 2D is the perp-dot-product, which is what this code uses three of. Switching to 3D and using the cross-product, interpolating both s and t at the end, results in the two closest points between the lines in 3D.
Anyway, the 2D solution:
int get_line_intersection(float p0_x, float p0_y, float p1_x, float p1_y,
float p2_x, float p2_y, float p3_x, float p3_y, float *i_x, float *i_y)
{
float s02_x, s02_y, s10_x, s10_y, s32_x, s32_y, s_numer, t_numer, denom, t;
s10_x = p1_x - p0_x;
s10_y = p1_y - p0_y;
s32_x = p3_x - p2_x;
s32_y = p3_y - p2_y;
denom = s10_x * s32_y - s32_x * s10_y;
if (denom == 0)
return 0; // Collinear
bool denomPositive = denom > 0;
s02_x = p0_x - p2_x;
s02_y = p0_y - p2_y;
s_numer = s10_x * s02_y - s10_y * s02_x;
if ((s_numer < 0) == denomPositive)
return 0; // No collision
t_numer = s32_x * s02_y - s32_y * s02_x;
if ((t_numer < 0) == denomPositive)
return 0; // No collision
if (((s_numer > denom) == denomPositive) || ((t_numer > denom) == denomPositive))
return 0; // No collision
// Collision detected
t = t_numer / denom;
if (i_x != NULL)
*i_x = p0_x + (t * s10_x);
if (i_y != NULL)
*i_y = p0_y + (t * s10_y);
return 1;
}
Basically it postpones the division until the last moment, and moves most of the tests until before certain calculations are done, thereby adding early-outs. Finally, it also avoids the division by zero case which occurs when the lines are parallel.
You also might want to consider using an epsilon test rather than comparison against zero. Lines that are extremely close to parallel can produce results that are slightly off. This is not a bug, it is a limitation with floating point math.
Question C: How do you detect whether or not two line segments intersect?
I have searched for the same topic, and I wasn't happy with the answers. So I have written an article that explains very detailed how to check if two line segments intersect with a lot of images. There is complete (and tested) Java-code.
Here is the article, cropped to the most important parts:
The algorithm, that checks if line segment a intersects with line segment b, looks like this:
What are bounding boxes? Here are two bounding boxes of two line segments:
If both bounding boxes have an intersection, you move line segment a so that one point is at (0|0). Now you have a line through the origin defined by a. Now move line segment b the same way and check if the new points of line segment b are on different sides of line a. If this is the case, check it the other way around. If this is also the case, the line segments intersect. If not, they don't intersect.
Question A: Where do two line segments intersect?
You know that two line segments a and b intersect. If you don't know that, check it with the tools I gave you in "Question C".
Now you can go through some cases and get the solution with 7th grade math (see code and interactive example).
Question B: How do you detect whether or not two lines intersect?
Let's say your point A = (x1, y1), point B = (x2, y2), C = (x_3, y_3), D = (x_4, y_4).
Your first line is defined by AB (with A != B), and your second one by CD (with C != D).
function doLinesIntersect(AB, CD) {
if (x1 == x2) {
return !(x3 == x4 && x1 != x3);
} else if (x3 == x4) {
return true;
} else {
// Both lines are not parallel to the y-axis
m1 = (y1-y2)/(x1-x2);
m2 = (y3-y4)/(x3-x4);
return m1 != m2;
}
}
Question D: Where do two lines intersect?
Check with Question B if they intersect at all.
The lines a and b are defined by two points for each line.
You can basically apply the same logic was used in Question A.
The answer once accepted here is incorrect (it has since been unaccepted, so hooray!). It does not correctly eliminate all non-intersections. Trivially it may appear to work but it can fail, especially in the case that 0 and 1 are considered valid for h.
Consider the following case:
Lines at (4,1)-(5,1) and (0,0)-(0,2)
These are perpendicular lines which clearly do not overlap.
A=(4,1)
B=(5,1)
C=(0,0)
D=(0,2)
E=(5,1)-(4,1)=(-1,0)
F=(0,2)-(0,0)=(0,-2)
P=(0,1)
h=((4,1)-(0,0)) dot (0,1) / ((0,-2) dot (0,1)) = 0
According to the above answer, these two line segments meet at an endpoint (values of 0 and 1). That endpoint would be:
(0,0)+(0,-2)*0=(0,0)
So, apparently the two line segments meet at (0,0), which is on line CD, but not on line AB. So what is going wrong? The answer is that the values of 0 and 1 are not valid and only sometimes HAPPEN to correctly predict endpoint intersection. When the extension of one line (but not the other) would meet the line segment, the algorithm predicts an intersection of line segments, but this is not correct. I imagine that by testing starting with AB vs CD and then also testing with CD vs AB, this problem would be eliminated. Only if both fall between 0 and 1 inclusively can they be said to intersect.
I recommend using the vector cross product method if you must predict end-points.
-Dan
Python version of iMalc's answer:
def find_intersection( p0, p1, p2, p3 ) :
s10_x = p1[0] - p0[0]
s10_y = p1[1] - p0[1]
s32_x = p3[0] - p2[0]
s32_y = p3[1] - p2[1]
denom = s10_x * s32_y - s32_x * s10_y
if denom == 0 : return None # collinear
denom_is_positive = denom > 0
s02_x = p0[0] - p2[0]
s02_y = p0[1] - p2[1]
s_numer = s10_x * s02_y - s10_y * s02_x
if (s_numer < 0) == denom_is_positive : return None # no collision
t_numer = s32_x * s02_y - s32_y * s02_x
if (t_numer < 0) == denom_is_positive : return None # no collision
if (s_numer > denom) == denom_is_positive or (t_numer > denom) == denom_is_positive : return None # no collision
# collision detected
t = t_numer / denom
intersection_point = [ p0[0] + (t * s10_x), p0[1] + (t * s10_y) ]
return intersection_point
Finding the correct intersection of two line segments is a non-trivial task with lots of edge cases. Here's a well documented, working and tested solution in Java.
In essence, there are three things that can happen when finding the intersection of two line segments:
The segments do not intersect
There is a unique intersection point
The intersection is another segment
NOTE: In the code, I assume that a line segment (x1, y1), (x2, y2) with x1 = x2 and y1 = y2 is a valid line segment. Mathematically speaking, a line segment consists of distinct points, but I am allowing segments to be points in this implementation for completeness.
Code is taken from my github repo
/**
* This snippet finds the intersection of two line segments.
* The intersection may either be empty, a single point or the
* intersection is a subsegment there's an overlap.
*/
import static java.lang.Math.abs;
import static java.lang.Math.max;
import static java.lang.Math.min;
import java.util.ArrayList;
import java.util.List;
public class LineSegmentLineSegmentIntersection {
// Small epsilon used for double value comparison.
private static final double EPS = 1e-5;
// 2D Point class.
public static class Pt {
double x, y;
public Pt(double x, double y) {
this.x = x;
this.y = y;
}
public boolean equals(Pt pt) {
return abs(x - pt.x) < EPS && abs(y - pt.y) < EPS;
}
}
// Finds the orientation of point 'c' relative to the line segment (a, b)
// Returns 0 if all three points are collinear.
// Returns -1 if 'c' is clockwise to segment (a, b), i.e right of line formed by the segment.
// Returns +1 if 'c' is counter clockwise to segment (a, b), i.e left of line
// formed by the segment.
public static int orientation(Pt a, Pt b, Pt c) {
double value = (b.y - a.y) * (c.x - b.x) -
(b.x - a.x) * (c.y - b.y);
if (abs(value) < EPS) return 0;
return (value > 0) ? -1 : +1;
}
// Tests whether point 'c' is on the line segment (a, b).
// Ensure first that point c is collinear to segment (a, b) and
// then check whether c is within the rectangle formed by (a, b)
public static boolean pointOnLine(Pt a, Pt b, Pt c) {
return orientation(a, b, c) == 0 &&
min(a.x, b.x) <= c.x && c.x <= max(a.x, b.x) &&
min(a.y, b.y) <= c.y && c.y <= max(a.y, b.y);
}
// Determines whether two segments intersect.
public static boolean segmentsIntersect(Pt p1, Pt p2, Pt p3, Pt p4) {
// Get the orientation of points p3 and p4 in relation
// to the line segment (p1, p2)
int o1 = orientation(p1, p2, p3);
int o2 = orientation(p1, p2, p4);
int o3 = orientation(p3, p4, p1);
int o4 = orientation(p3, p4, p2);
// If the points p1, p2 are on opposite sides of the infinite
// line formed by (p3, p4) and conversly p3, p4 are on opposite
// sides of the infinite line formed by (p1, p2) then there is
// an intersection.
if (o1 != o2 && o3 != o4) return true;
// Collinear special cases (perhaps these if checks can be simplified?)
if (o1 == 0 && pointOnLine(p1, p2, p3)) return true;
if (o2 == 0 && pointOnLine(p1, p2, p4)) return true;
if (o3 == 0 && pointOnLine(p3, p4, p1)) return true;
if (o4 == 0 && pointOnLine(p3, p4, p2)) return true;
return false;
}
public static List<Pt> getCommonEndpoints(Pt p1, Pt p2, Pt p3, Pt p4) {
List<Pt> points = new ArrayList<>();
if (p1.equals(p3)) {
points.add(p1);
if (p2.equals(p4)) points.add(p2);
} else if (p1.equals(p4)) {
points.add(p1);
if (p2.equals(p3)) points.add(p2);
} else if (p2.equals(p3)) {
points.add(p2);
if (p1.equals(p4)) points.add(p1);
} else if (p2.equals(p4)) {
points.add(p2);
if (p1.equals(p3)) points.add(p1);
}
return points;
}
// Finds the intersection point(s) of two line segments. Unlike regular line
// segments, segments which are points (x1 = x2 and y1 = y2) are allowed.
public static Pt[] lineSegmentLineSegmentIntersection(Pt p1, Pt p2, Pt p3, Pt p4) {
// No intersection.
if (!segmentsIntersect(p1, p2, p3, p4)) return new Pt[]{};
// Both segments are a single point.
if (p1.equals(p2) && p2.equals(p3) && p3.equals(p4))
return new Pt[]{p1};
List<Pt> endpoints = getCommonEndpoints(p1, p2, p3, p4);
int n = endpoints.size();
// One of the line segments is an intersecting single point.
// NOTE: checking only n == 1 is insufficient to return early
// because the solution might be a sub segment.
boolean singleton = p1.equals(p2) || p3.equals(p4);
if (n == 1 && singleton) return new Pt[]{endpoints.get(0)};
// Segments are equal.
if (n == 2) return new Pt[]{endpoints.get(0), endpoints.get(1)};
boolean collinearSegments = (orientation(p1, p2, p3) == 0) &&
(orientation(p1, p2, p4) == 0);
// The intersection will be a sub-segment of the two
// segments since they overlap each other.
if (collinearSegments) {
// Segment #2 is enclosed in segment #1
if (pointOnLine(p1, p2, p3) && pointOnLine(p1, p2, p4))
return new Pt[]{p3, p4};
// Segment #1 is enclosed in segment #2
if (pointOnLine(p3, p4, p1) && pointOnLine(p3, p4, p2))
return new Pt[]{p1, p2};
// The subsegment is part of segment #1 and part of segment #2.
// Find the middle points which correspond to this segment.
Pt midPoint1 = pointOnLine(p1, p2, p3) ? p3 : p4;
Pt midPoint2 = pointOnLine(p3, p4, p1) ? p1 : p2;
// There is actually only one middle point!
if (midPoint1.equals(midPoint2)) return new Pt[]{midPoint1};
return new Pt[]{midPoint1, midPoint2};
}
/* Beyond this point there is a unique intersection point. */
// Segment #1 is a vertical line.
if (abs(p1.x - p2.x) < EPS) {
double m = (p4.y - p3.y) / (p4.x - p3.x);
double b = p3.y - m * p3.x;
return new Pt[]{new Pt(p1.x, m * p1.x + b)};
}
// Segment #2 is a vertical line.
if (abs(p3.x - p4.x) < EPS) {
double m = (p2.y - p1.y) / (p2.x - p1.x);
double b = p1.y - m * p1.x;
return new Pt[]{new Pt(p3.x, m * p3.x + b)};
}
double m1 = (p2.y - p1.y) / (p2.x - p1.x);
double m2 = (p4.y - p3.y) / (p4.x - p3.x);
double b1 = p1.y - m1 * p1.x;
double b2 = p3.y - m2 * p3.x;
double x = (b2 - b1) / (m1 - m2);
double y = (m1 * b2 - m2 * b1) / (m1 - m2);
return new Pt[]{new Pt(x, y)};
}
}
Here is a simple usage example:
public static void main(String[] args) {
// Segment #1 is (p1, p2), segment #2 is (p3, p4)
Pt p1, p2, p3, p4;
p1 = new Pt(-2, 4); p2 = new Pt(3, 3);
p3 = new Pt(0, 0); p4 = new Pt(2, 4);
Pt[] points = lineSegmentLineSegmentIntersection(p1, p2, p3, p4);
Pt point = points[0];
// Prints: (1.636, 3.273)
System.out.printf("(%.3f, %.3f)\n", point.x, point.y);
p1 = new Pt(-10, 0); p2 = new Pt(+10, 0);
p3 = new Pt(-5, 0); p4 = new Pt(+5, 0);
points = lineSegmentLineSegmentIntersection(p1, p2, p3, p4);
Pt point1 = points[0], point2 = points[1];
// Prints: (-5.000, 0.000) (5.000, 0.000)
System.out.printf("(%.3f, %.3f) (%.3f, %.3f)\n", point1.x, point1.y, point2.x, point2.y);
}
Just wanted to mention that a good explanation and explicit solution can be found in the Numeric Recipes series. I've got the 3rd edition and the answer is on page 1117, section 21.4. Another solution with a different nomenclature can be found in a paper by Marina Gavrilova Reliable Line Section Intersection Testing. Her solution is, to my mind, a little simpler.
My implementation is below:
bool NuGeometry::IsBetween(const double& x0, const double& x, const double& x1){
return (x >= x0) && (x <= x1);
}
bool NuGeometry::FindIntersection(const double& x0, const double& y0,
const double& x1, const double& y1,
const double& a0, const double& b0,
const double& a1, const double& b1,
double& xy, double& ab) {
// four endpoints are x0, y0 & x1,y1 & a0,b0 & a1,b1
// returned values xy and ab are the fractional distance along xy and ab
// and are only defined when the result is true
bool partial = false;
double denom = (b0 - b1) * (x0 - x1) - (y0 - y1) * (a0 - a1);
if (denom == 0) {
xy = -1;
ab = -1;
} else {
xy = (a0 * (y1 - b1) + a1 * (b0 - y1) + x1 * (b1 - b0)) / denom;
partial = NuGeometry::IsBetween(0, xy, 1);
if (partial) {
// no point calculating this unless xy is between 0 & 1
ab = (y1 * (x0 - a1) + b1 * (x1 - x0) + y0 * (a1 - x1)) / denom;
}
}
if ( partial && NuGeometry::IsBetween(0, ab, 1)) {
ab = 1-ab;
xy = 1-xy;
return true;
} else return false;
}
Plenty of solutions are available above, but I think below solution is pretty simple and easy to understand.
Two segments Vector AB and Vector CD intersect if and only if
The endpoints a and b are on opposite sides of the segment CD.
The endpoints c and d are on opposite side of the segment AB.
More specifically a and b are on opposite side of segment CD if and only if exactly one of the two triples a,c,d and b,c,d is in counterclockwise order.
Intersect(a, b, c, d)
if CCW(a, c, d) == CCW(b, c, d)
return false;
else if CCW(a, b, c) == CCW(a, b, d)
return false;
else
return true;
Here CCW represent counterclockwise which returns true/false based on the orientation of the points.
Source : http://compgeom.cs.uiuc.edu/~jeffe/teaching/373/notes/x06-sweepline.pdf
Page 2
C and Objective-C
Based on Gareth Rees' answer
const AGKLine AGKLineZero = (AGKLine){(CGPoint){0.0, 0.0}, (CGPoint){0.0, 0.0}};
AGKLine AGKLineMake(CGPoint start, CGPoint end)
{
return (AGKLine){start, end};
}
double AGKLineLength(AGKLine l)
{
return CGPointLengthBetween_AGK(l.start, l.end);
}
BOOL AGKLineIntersection(AGKLine l1, AGKLine l2, CGPoint *out_pointOfIntersection)
{
// http://stackoverflow.com/a/565282/202451
CGPoint p = l1.start;
CGPoint q = l2.start;
CGPoint r = CGPointSubtract_AGK(l1.end, l1.start);
CGPoint s = CGPointSubtract_AGK(l2.end, l2.start);
double s_r_crossProduct = CGPointCrossProductZComponent_AGK(r, s);
double t = CGPointCrossProductZComponent_AGK(CGPointSubtract_AGK(q, p), s) / s_r_crossProduct;
double u = CGPointCrossProductZComponent_AGK(CGPointSubtract_AGK(q, p), r) / s_r_crossProduct;
if(t < 0 || t > 1.0 || u < 0 || u > 1.0)
{
if(out_pointOfIntersection != NULL)
{
*out_pointOfIntersection = CGPointZero;
}
return NO;
}
else
{
if(out_pointOfIntersection != NULL)
{
CGPoint i = CGPointAdd_AGK(p, CGPointMultiply_AGK(r, t));
*out_pointOfIntersection = i;
}
return YES;
}
}
CGFloat CGPointCrossProductZComponent_AGK(CGPoint v1, CGPoint v2)
{
return v1.x * v2.y - v1.y * v2.x;
}
CGPoint CGPointSubtract_AGK(CGPoint p1, CGPoint p2)
{
return (CGPoint){p1.x - p2.x, p1.y - p2.y};
}
CGPoint CGPointAdd_AGK(CGPoint p1, CGPoint p2)
{
return (CGPoint){p1.x + p2.x, p1.y + p2.y};
}
CGFloat CGPointCrossProductZComponent_AGK(CGPoint v1, CGPoint v2)
{
return v1.x * v2.y - v1.y * v2.x;
}
CGPoint CGPointMultiply_AGK(CGPoint p1, CGFloat factor)
{
return (CGPoint){p1.x * factor, p1.y * factor};
}
Many of the functions and structs are private, but you should pretty easy be able to know what's going on.
This is public on this repo https://github.com/hfossli/AGGeometryKit/
This is working well for me. Taken from here.
// calculates intersection and checks for parallel lines.
// also checks that the intersection point is actually on
// the line segment p1-p2
Point findIntersection(Point p1,Point p2,
Point p3,Point p4) {
float xD1,yD1,xD2,yD2,xD3,yD3;
float dot,deg,len1,len2;
float segmentLen1,segmentLen2;
float ua,ub,div;
// calculate differences
xD1=p2.x-p1.x;
xD2=p4.x-p3.x;
yD1=p2.y-p1.y;
yD2=p4.y-p3.y;
xD3=p1.x-p3.x;
yD3=p1.y-p3.y;
// calculate the lengths of the two lines
len1=sqrt(xD1*xD1+yD1*yD1);
len2=sqrt(xD2*xD2+yD2*yD2);
// calculate angle between the two lines.
dot=(xD1*xD2+yD1*yD2); // dot product
deg=dot/(len1*len2);
// if abs(angle)==1 then the lines are parallell,
// so no intersection is possible
if(abs(deg)==1) return null;
// find intersection Pt between two lines
Point pt=new Point(0,0);
div=yD2*xD1-xD2*yD1;
ua=(xD2*yD3-yD2*xD3)/div;
ub=(xD1*yD3-yD1*xD3)/div;
pt.x=p1.x+ua*xD1;
pt.y=p1.y+ua*yD1;
// calculate the combined length of the two segments
// between Pt-p1 and Pt-p2
xD1=pt.x-p1.x;
xD2=pt.x-p2.x;
yD1=pt.y-p1.y;
yD2=pt.y-p2.y;
segmentLen1=sqrt(xD1*xD1+yD1*yD1)+sqrt(xD2*xD2+yD2*yD2);
// calculate the combined length of the two segments
// between Pt-p3 and Pt-p4
xD1=pt.x-p3.x;
xD2=pt.x-p4.x;
yD1=pt.y-p3.y;
yD2=pt.y-p4.y;
segmentLen2=sqrt(xD1*xD1+yD1*yD1)+sqrt(xD2*xD2+yD2*yD2);
// if the lengths of both sets of segments are the same as
// the lenghts of the two lines the point is actually
// on the line segment.
// if the point isn’t on the line, return null
if(abs(len1-segmentLen1)>0.01 || abs(len2-segmentLen2)>0.01)
return null;
// return the valid intersection
return pt;
}
class Point{
float x,y;
Point(float x, float y){
this.x = x;
this.y = y;
}
void set(float x, float y){
this.x = x;
this.y = y;
}
}
I tried some of these answers, but they didnt work for me (sorry guys); after some more net searching I found this.
With a little modification to his code I now have this function that will return the point of intersection or if no intersection is found it will return -1,-1.
Public Function intercetion(ByVal ax As Integer, ByVal ay As Integer, ByVal bx As Integer, ByVal by As Integer, ByVal cx As Integer, ByVal cy As Integer, ByVal dx As Integer, ByVal dy As Integer) As Point
'// Determines the intersection point of the line segment defined by points A and B
'// with the line segment defined by points C and D.
'//
'// Returns YES if the intersection point was found, and stores that point in X,Y.
'// Returns NO if there is no determinable intersection point, in which case X,Y will
'// be unmodified.
Dim distAB, theCos, theSin, newX, ABpos As Double
'// Fail if either line segment is zero-length.
If ax = bx And ay = by Or cx = dx And cy = dy Then Return New Point(-1, -1)
'// Fail if the segments share an end-point.
If ax = cx And ay = cy Or bx = cx And by = cy Or ax = dx And ay = dy Or bx = dx And by = dy Then Return New Point(-1, -1)
'// (1) Translate the system so that point A is on the origin.
bx -= ax
by -= ay
cx -= ax
cy -= ay
dx -= ax
dy -= ay
'// Discover the length of segment A-B.
distAB = Math.Sqrt(bx * bx + by * by)
'// (2) Rotate the system so that point B is on the positive X axis.
theCos = bx / distAB
theSin = by / distAB
newX = cx * theCos + cy * theSin
cy = cy * theCos - cx * theSin
cx = newX
newX = dx * theCos + dy * theSin
dy = dy * theCos - dx * theSin
dx = newX
'// Fail if segment C-D doesn't cross line A-B.
If cy < 0 And dy < 0 Or cy >= 0 And dy >= 0 Then Return New Point(-1, -1)
'// (3) Discover the position of the intersection point along line A-B.
ABpos = dx + (cx - dx) * dy / (dy - cy)
'// Fail if segment C-D crosses line A-B outside of segment A-B.
If ABpos < 0 Or ABpos > distAB Then Return New Point(-1, -1)
'// (4) Apply the discovered position to line A-B in the original coordinate system.
'*X=Ax+ABpos*theCos
'*Y=Ay+ABpos*theSin
'// Success.
Return New Point(ax + ABpos * theCos, ay + ABpos * theSin)
End Function
There seems to be some interest in Gavin's answer for which cortijon proposed a javascript version in the comments and iMalc provided a version with slightly fewer computations. Some have pointed out shortcomings with various code proposals and others have commented on the efficiency of some code proposals.
The algorithm provided by iMalc via Gavin's answer is the one that I am currently using in a javascript project and I just wanted to provide a cleaned up version here if it may help anyone.
// Some variables for reuse, others may do this differently
var p0x, p1x, p2x, p3x, ix,
p0y, p1y, p2y, p3y, iy,
collisionDetected;
// do stuff, call other functions, set endpoints...
// note: for my purpose I use |t| < |d| as opposed to
// |t| <= |d| which is equivalent to 0 <= t < 1 rather than
// 0 <= t <= 1 as in Gavin's answer - results may vary
var lineSegmentIntersection = function(){
var d, dx1, dx2, dx3, dy1, dy2, dy3, s, t;
dx1 = p1x - p0x; dy1 = p1y - p0y;
dx2 = p3x - p2x; dy2 = p3y - p2y;
dx3 = p0x - p2x; dy3 = p0y - p2y;
collisionDetected = 0;
d = dx1 * dy2 - dx2 * dy1;
if(d !== 0){
s = dx1 * dy3 - dx3 * dy1;
if((s <= 0 && d < 0 && s >= d) || (s >= 0 && d > 0 && s <= d)){
t = dx2 * dy3 - dx3 * dy2;
if((t <= 0 && d < 0 && t > d) || (t >= 0 && d > 0 && t < d)){
t = t / d;
collisionDetected = 1;
ix = p0x + t * dx1;
iy = p0y + t * dy1;
}
}
}
};
I think there is a much much simpler solution for this problem. I came up with another idea today and it seems to work just fine (at least in 2D for now). All you have to do, is to calculate the intersection between two lines, then check if the calculated intersection point is within the boundig boxes of both line segments. If it is, the line segments intersect. That's it.
EDIT:
This is how I calculate the intersection (I don't know anymore where I found this code snippet)
Point3D
comes from
System.Windows.Media.Media3D
public static Point3D? Intersection(Point3D start1, Point3D end1, Point3D start2, Point3D end2) {
double a1 = end1.Y - start1.Y;
double b1 = start1.X - end1.X;
double c1 = a1 * start1.X + b1 * start1.Y;
double a2 = end2.Y - start2.Y;
double b2 = start2.X - end2.X;
double c2 = a2 * start2.X + b2 * start2.Y;
double det = a1 * b2 - a2 * b1;
if (det == 0) { // lines are parallel
return null;
}
double x = (b2 * c1 - b1 * c2) / det;
double y = (a1 * c2 - a2 * c1) / det;
return new Point3D(x, y, 0.0);
}
and this is my (simplified for the purpose of the answer) BoundingBox class:
public class BoundingBox {
private Point3D min = new Point3D();
private Point3D max = new Point3D();
public BoundingBox(Point3D point) {
min = point;
max = point;
}
public Point3D Min {
get { return min; }
set { min = value; }
}
public Point3D Max {
get { return max; }
set { max = value; }
}
public bool Contains(BoundingBox box) {
bool contains =
min.X <= box.min.X && max.X >= box.max.X &&
min.Y <= box.min.Y && max.Y >= box.max.Y &&
min.Z <= box.min.Z && max.Z >= box.max.Z;
return contains;
}
public bool Contains(Point3D point) {
return Contains(new BoundingBox(point));
}
}
This solution may help
public static float GetLineYIntesept(PointF p, float slope)
{
return p.Y - slope * p.X;
}
public static PointF FindIntersection(PointF line1Start, PointF line1End, PointF line2Start, PointF line2End)
{
float slope1 = (line1End.Y - line1Start.Y) / (line1End.X - line1Start.X);
float slope2 = (line2End.Y - line2Start.Y) / (line2End.X - line2Start.X);
float yinter1 = GetLineYIntesept(line1Start, slope1);
float yinter2 = GetLineYIntesept(line2Start, slope2);
if (slope1 == slope2 && yinter1 != yinter2)
return PointF.Empty;
float x = (yinter2 - yinter1) / (slope1 - slope2);
float y = slope1 * x + yinter1;
return new PointF(x, y);
}
I ported Kris's above answer to JavaScript. After trying numerous different answers, his provided the correct points. I thought I was going crazy that I wasn't getting the points I needed.
function getLineLineCollision(p0, p1, p2, p3) {
var s1, s2;
s1 = {x: p1.x - p0.x, y: p1.y - p0.y};
s2 = {x: p3.x - p2.x, y: p3.y - p2.y};
var s10_x = p1.x - p0.x;
var s10_y = p1.y - p0.y;
var s32_x = p3.x - p2.x;
var s32_y = p3.y - p2.y;
var denom = s10_x * s32_y - s32_x * s10_y;
if(denom == 0) {
return false;
}
var denom_positive = denom > 0;
var s02_x = p0.x - p2.x;
var s02_y = p0.y - p2.y;
var s_numer = s10_x * s02_y - s10_y * s02_x;
if((s_numer < 0) == denom_positive) {
return false;
}
var t_numer = s32_x * s02_y - s32_y * s02_x;
if((t_numer < 0) == denom_positive) {
return false;
}
if((s_numer > denom) == denom_positive || (t_numer > denom) == denom_positive) {
return false;
}
var t = t_numer / denom;
var p = {x: p0.x + (t * s10_x), y: p0.y + (t * s10_y)};
return p;
}
I tried lot of ways and then I decided to write my own. So here it is:
bool IsBetween (float x, float b1, float b2)
{
return ( ((x >= (b1 - 0.1f)) &&
(x <= (b2 + 0.1f))) ||
((x >= (b2 - 0.1f)) &&
(x <= (b1 + 0.1f))));
}
bool IsSegmentsColliding( POINTFLOAT lineA,
POINTFLOAT lineB,
POINTFLOAT line2A,
POINTFLOAT line2B)
{
float deltaX1 = lineB.x - lineA.x;
float deltaX2 = line2B.x - line2A.x;
float deltaY1 = lineB.y - lineA.y;
float deltaY2 = line2B.y - line2A.y;
if (abs(deltaX1) < 0.01f &&
abs(deltaX2) < 0.01f) // Both are vertical lines
return false;
if (abs((deltaY1 / deltaX1) -
(deltaY2 / deltaX2)) < 0.001f) // Two parallel line
return false;
float xCol = ( ( (deltaX1 * deltaX2) *
(line2A.y - lineA.y)) -
(line2A.x * deltaY2 * deltaX1) +
(lineA.x * deltaY1 * deltaX2)) /
((deltaY1 * deltaX2) - (deltaY2 * deltaX1));
float yCol = 0;
if (deltaX1 < 0.01f) // L1 is a vertical line
yCol = ((xCol * deltaY2) +
(line2A.y * deltaX2) -
(line2A.x * deltaY2)) / deltaX2;
else // L1 is acceptable
yCol = ((xCol * deltaY1) +
(lineA.y * deltaX1) -
(lineA.x * deltaY1)) / deltaX1;
bool isCol = IsBetween(xCol, lineA.x, lineB.x) &&
IsBetween(yCol, lineA.y, lineB.y) &&
IsBetween(xCol, line2A.x, line2B.x) &&
IsBetween(yCol, line2A.y, line2B.y);
return isCol;
}
Based on these two formulas: (I simplified them from equation of lines and other formulas)
This based on Gareth Ree's answer. It also returns the overlap of the line segments if they do. Coded in C++, V is a simple vector class. Where the cross product of two vectors in 2D returns a single scalar. It was tested and passed by my schools automatic testing system.
//Required input point must be colinear with the line
bool on_segment(const V& p, const LineSegment& l)
{
//If a point is on the line, the sum of the vectors formed by the point to the line endpoints must be equal
V va = p - l.pa;
V vb = p - l.pb;
R ma = va.magnitude();
R mb = vb.magnitude();
R ml = (l.pb - l.pa).magnitude();
R s = ma + mb;
bool r = s <= ml + epsilon;
return r;
}
//Compute using vector math
// Returns 0 points if the lines do not intersect or overlap
// Returns 1 point if the lines intersect
// Returns 2 points if the lines overlap, contain the points where overlapping start starts and stop
std::vector<V> intersect(const LineSegment& la, const LineSegment& lb)
{
std::vector<V> r;
//http://stackoverflow.com/questions/563198/how-do-you-detect-where-two-line-segments-intersect
V oa, ob, da, db; //Origin and direction vectors
R sa, sb; //Scalar values
oa = la.pa;
da = la.pb - la.pa;
ob = lb.pa;
db = lb.pb - lb.pa;
if (da.cross(db) == 0 && (ob - oa).cross(da) == 0) //If colinear
{
if (on_segment(lb.pa, la) && on_segment(lb.pb, la))
{
r.push_back(lb.pa);
r.push_back(lb.pb);
dprintf("colinear, overlapping\n");
return r;
}
if (on_segment(la.pa, lb) && on_segment(la.pb, lb))
{
r.push_back(la.pa);
r.push_back(la.pb);
dprintf("colinear, overlapping\n");
return r;
}
if (on_segment(la.pa, lb))
r.push_back(la.pa);
if (on_segment(la.pb, lb))
r.push_back(la.pb);
if (on_segment(lb.pa, la))
r.push_back(lb.pa);
if (on_segment(lb.pb, la))
r.push_back(lb.pb);
if (r.size() == 0)
dprintf("colinear, non-overlapping\n");
else
dprintf("colinear, overlapping\n");
return r;
}
if (da.cross(db) == 0 && (ob - oa).cross(da) != 0)
{
dprintf("parallel non-intersecting\n");
return r;
}
//Math trick db cross db == 0, which is a single scalar in 2D.
//Crossing both sides with vector db gives:
sa = (ob - oa).cross(db) / da.cross(db);
//Crossing both sides with vector da gives
sb = (oa - ob).cross(da) / db.cross(da);
if (0 <= sa && sa <= 1 && 0 <= sb && sb <= 1)
{
dprintf("intersecting\n");
r.push_back(oa + da * sa);
return r;
}
dprintf("non-intersecting, non-parallel, non-colinear, non-overlapping\n");
return r;
}
Here's a basic implementation of a line segment in C#, with corresponding intersection detection code. It requires a 2D vector/point struct called Vector2f, though you can replace this with any other type that has X/Y properties. You could also replace float with double if that suits your needs better.
This code is used in my .NET physics library, Boing.
public struct LineSegment2f
{
public Vector2f From { get; }
public Vector2f To { get; }
public LineSegment2f(Vector2f #from, Vector2f to)
{
From = #from;
To = to;
}
public Vector2f Delta => new Vector2f(To.X - From.X, To.Y - From.Y);
/// <summary>
/// Attempt to intersect two line segments.
/// </summary>
/// <remarks>
/// Even if the line segments do not intersect, <paramref name="t"/> and <paramref name="u"/> will be set.
/// If the lines are parallel, <paramref name="t"/> and <paramref name="u"/> are set to <see cref="float.NaN"/>.
/// </remarks>
/// <param name="other">The line to attempt intersection of this line with.</param>
/// <param name="intersectionPoint">The point of intersection if within the line segments, or empty..</param>
/// <param name="t">The distance along this line at which intersection would occur, or NaN if lines are collinear/parallel.</param>
/// <param name="u">The distance along the other line at which intersection would occur, or NaN if lines are collinear/parallel.</param>
/// <returns><c>true</c> if the line segments intersect, otherwise <c>false</c>.</returns>
public bool TryIntersect(LineSegment2f other, out Vector2f intersectionPoint, out float t, out float u)
{
var p = From;
var q = other.From;
var r = Delta;
var s = other.Delta;
// t = (q − p) × s / (r × s)
// u = (q − p) × r / (r × s)
var denom = Fake2DCross(r, s);
if (denom == 0)
{
// lines are collinear or parallel
t = float.NaN;
u = float.NaN;
intersectionPoint = default(Vector2f);
return false;
}
var tNumer = Fake2DCross(q - p, s);
var uNumer = Fake2DCross(q - p, r);
t = tNumer / denom;
u = uNumer / denom;
if (t < 0 || t > 1 || u < 0 || u > 1)
{
// line segments do not intersect within their ranges
intersectionPoint = default(Vector2f);
return false;
}
intersectionPoint = p + r * t;
return true;
}
private static float Fake2DCross(Vector2f a, Vector2f b)
{
return a.X * b.Y - a.Y * b.X;
}
}
A C++ program to check if two given line segments intersect
#include <iostream>
using namespace std;
struct Point
{
int x;
int y;
};
// Given three colinear points p, q, r, the function checks if
// point q lies on line segment 'pr'
bool onSegment(Point p, Point q, Point r)
{
if (q.x <= max(p.x, r.x) && q.x >= min(p.x, r.x) &&
q.y <= max(p.y, r.y) && q.y >= min(p.y, r.y))
return true;
return false;
}
// To find orientation of ordered triplet (p, q, r).
// The function returns following values
// 0 --> p, q and r are colinear
// 1 --> Clockwise
// 2 --> Counterclockwise
int orientation(Point p, Point q, Point r)
{
// See 10th slides from following link for derivation of the formula
// http://www.dcs.gla.ac.uk/~pat/52233/slides/Geometry1x1.pdf
int val = (q.y - p.y) * (r.x - q.x) -
(q.x - p.x) * (r.y - q.y);
if (val == 0) return 0; // colinear
return (val > 0)? 1: 2; // clock or counterclock wise
}
// The main function that returns true if line segment 'p1q1'
// and 'p2q2' intersect.
bool doIntersect(Point p1, Point q1, Point p2, Point q2)
{
// Find the four orientations needed for general and
// special cases
int o1 = orientation(p1, q1, p2);
int o2 = orientation(p1, q1, q2);
int o3 = orientation(p2, q2, p1);
int o4 = orientation(p2, q2, q1);
// General case
if (o1 != o2 && o3 != o4)
return true;
// Special Cases
// p1, q1 and p2 are colinear and p2 lies on segment p1q1
if (o1 == 0 && onSegment(p1, p2, q1)) return true;
// p1, q1 and p2 are colinear and q2 lies on segment p1q1
if (o2 == 0 && onSegment(p1, q2, q1)) return true;
// p2, q2 and p1 are colinear and p1 lies on segment p2q2
if (o3 == 0 && onSegment(p2, p1, q2)) return true;
// p2, q2 and q1 are colinear and q1 lies on segment p2q2
if (o4 == 0 && onSegment(p2, q1, q2)) return true;
return false; // Doesn't fall in any of the above cases
}
// Driver program to test above functions
int main()
{
struct Point p1 = {1, 1}, q1 = {10, 1};
struct Point p2 = {1, 2}, q2 = {10, 2};
doIntersect(p1, q1, p2, q2)? cout << "Yes\n": cout << "No\n";
p1 = {10, 0}, q1 = {0, 10};
p2 = {0, 0}, q2 = {10, 10};
doIntersect(p1, q1, p2, q2)? cout << "Yes\n": cout << "No\n";
p1 = {-5, -5}, q1 = {0, 0};
p2 = {1, 1}, q2 = {10, 10};
doIntersect(p1, q1, p2, q2)? cout << "Yes\n": cout << "No\n";
return 0;
}
Based on #Gareth Rees answer, version for Python:
import numpy as np
def np_perp( a ) :
b = np.empty_like(a)
b[0] = a[1]
b[1] = -a[0]
return b
def np_cross_product(a, b):
return np.dot(a, np_perp(b))
def np_seg_intersect(a, b, considerCollinearOverlapAsIntersect = False):
# https://stackoverflow.com/questions/563198/how-do-you-detect-where-two-line-segments-intersect/565282#565282
# http://www.codeproject.com/Tips/862988/Find-the-intersection-point-of-two-line-segments
r = a[1] - a[0]
s = b[1] - b[0]
v = b[0] - a[0]
num = np_cross_product(v, r)
denom = np_cross_product(r, s)
# If r x s = 0 and (q - p) x r = 0, then the two lines are collinear.
if np.isclose(denom, 0) and np.isclose(num, 0):
# 1. If either 0 <= (q - p) * r <= r * r or 0 <= (p - q) * s <= * s
# then the two lines are overlapping,
if(considerCollinearOverlapAsIntersect):
vDotR = np.dot(v, r)
aDotS = np.dot(-v, s)
if (0 <= vDotR and vDotR <= np.dot(r,r)) or (0 <= aDotS and aDotS <= np.dot(s,s)):
return True
# 2. If neither 0 <= (q - p) * r = r * r nor 0 <= (p - q) * s <= s * s
# then the two lines are collinear but disjoint.
# No need to implement this expression, as it follows from the expression above.
return None
if np.isclose(denom, 0) and not np.isclose(num, 0):
# Parallel and non intersecting
return None
u = num / denom
t = np_cross_product(v, s) / denom
if u >= 0 and u <= 1 and t >= 0 and t <= 1:
res = b[0] + (s*u)
return res
# Otherwise, the two line segments are not parallel but do not intersect.
return None
If each side of the rectangle is a line segment, and the user drawn portion is a line segment, then you need to just check the user drawn segment for intersection with the four side line segments. This should be a fairly simple exercise given the start and end points of each segment.
Based on t3chb0t's answer:
int intersezione_linee(int x1, int y1, int x2, int y2, int x3, int y3, int x4, int y4, int& p_x, int& p_y)
{
//L1: estremi (x1,y1)(x2,y2) L2: estremi (x3,y3)(x3,y3)
int d;
d = (x1-x2)*(y3-y4) - (y1-y2)*(x3-x4);
if(!d)
return 0;
p_x = ((x1*y2-y1*x2)*(x3-x4) - (x1-x2)*(x3*y4-y3*x4))/d;
p_y = ((x1*y2-y1*x2)*(y3-y4) - (y1-y2)*(x3*y4-y3*x4))/d;
return 1;
}
int in_bounding_box(int x1, int y1, int x2, int y2, int p_x, int p_y)
{
return p_x>=x1 && p_x<=x2 && p_y>=y1 && p_y<=y2;
}
int intersezione_segmenti(int x1, int y1, int x2, int y2, int x3, int y3, int x4, int y4, int& p_x, int& p_y)
{
if (!intersezione_linee(x1,y1,x2,y2,x3,y3,x4,y4,p_x,p_y))
return 0;
return in_bounding_box(x1,y1,x2,y2,p_x,p_y) && in_bounding_box(x3,y3,x4,y4,p_x,p_y);
}
I read these algorithm from the book "multiple view geometry"
following text using
' as transpose sign
* as dot product
x as cross product, when using as operator
1. line definition
a point x_vec = (x, y)' lies on the line ax + by + c = 0
we denote L = (a, b, c)', the point as (x, y, 1)' as homogeneous coordinates
the line equation can be written as
(x, y, 1)(a, b, c)' = 0 or x' * L = 0
2. intersection of lines
we have two lines L1=(a1, b1, c1)', L2=(a2, b2, c2)'
assume x is a point, a vector, and x = L1 x L2 (L1 cross product L2).
be careful, x is always a 2D point, please read homogeneous coordinates if you are confused about (L1xL2) is a three elements vector, and x is a 2D coordinates.
according to triple product, we know that
L1 * ( L1 x L2 ) = 0, and L2 * (L1 x L2) = 0, because of L1,L2 co-plane
we substitute (L1xL2) with vector x, then we have L1*x=0, L2*x=0, which means x lie on both L1 and L2, x is the intersection point.
be careful, here x is homogeneous coordinates, if the last element of x is zero, it means L1 and L2 are parallel.
Many answers have wrapped up all the calculations into a single function. If you need to calculate the line slopes, y-intercepts, or x-intercepts for use elsewhere in your code, you'll be making those calculations redundantly. I have separated out the respective functions, used obvious variable names, and commented my code to make it easier to follow. I needed to know if lines intersect infinitely beyond their endpoints, so in JavaScript:
http://jsfiddle.net/skibulk/evmqq00u/
var point_a = {x:0, y:10},
point_b = {x:12, y:12},
point_c = {x:10, y:0},
point_d = {x:0, y:0},
slope_ab = slope(point_a, point_b),
slope_bc = slope(point_b, point_c),
slope_cd = slope(point_c, point_d),
slope_da = slope(point_d, point_a),
yint_ab = y_intercept(point_a, slope_ab),
yint_bc = y_intercept(point_b, slope_bc),
yint_cd = y_intercept(point_c, slope_cd),
yint_da = y_intercept(point_d, slope_da),
xint_ab = x_intercept(point_a, slope_ab, yint_ab),
xint_bc = x_intercept(point_b, slope_bc, yint_bc),
xint_cd = x_intercept(point_c, slope_cd, yint_cd),
xint_da = x_intercept(point_d, slope_da, yint_da),
point_aa = intersect(slope_da, yint_da, xint_da, slope_ab, yint_ab, xint_ab),
point_bb = intersect(slope_ab, yint_ab, xint_ab, slope_bc, yint_bc, xint_bc),
point_cc = intersect(slope_bc, yint_bc, xint_bc, slope_cd, yint_cd, xint_cd),
point_dd = intersect(slope_cd, yint_cd, xint_cd, slope_da, yint_da, xint_da);
console.log(point_a, point_b, point_c, point_d);
console.log(slope_ab, slope_bc, slope_cd, slope_da);
console.log(yint_ab, yint_bc, yint_cd, yint_da);
console.log(xint_ab, xint_bc, xint_cd, xint_da);
console.log(point_aa, point_bb, point_cc, point_dd);
function slope(point_a, point_b) {
var i = (point_b.y - point_a.y) / (point_b.x - point_a.x);
if (i === -Infinity) return Infinity;
if (i === -0) return 0;
return i;
}
function y_intercept(point, slope) {
// Horizontal Line
if (slope == 0) return point.y;
// Vertical Line
if (slope == Infinity)
{
// THE Y-Axis
if (point.x == 0) return Infinity;
// No Intercept
return null;
}
// Angled Line
return point.y - (slope * point.x);
}
function x_intercept(point, slope, yint) {
// Vertical Line
if (slope == Infinity) return point.x;
// Horizontal Line
if (slope == 0)
{
// THE X-Axis
if (point.y == 0) return Infinity;
// No Intercept
return null;
}
// Angled Line
return -yint / slope;
}
// Intersection of two infinite lines
function intersect(slope_a, yint_a, xint_a, slope_b, yint_b, xint_b) {
if (slope_a == slope_b)
{
// Equal Lines
if (yint_a == yint_b && xint_a == xint_b) return Infinity;
// Parallel Lines
return null;
}
// First Line Vertical
if (slope_a == Infinity)
{
return {
x: xint_a,
y: (slope_b * xint_a) + yint_b
};
}
// Second Line Vertical
if (slope_b == Infinity)
{
return {
x: xint_b,
y: (slope_a * xint_b) + yint_a
};
}
// Not Equal, Not Parallel, Not Vertical
var i = (yint_b - yint_a) / (slope_a - slope_b);
return {
x: i,
y: (slope_a * i) + yint_a
};
}

Multi otsu(multi-thresholding) with openCV

I am trying to carry out multi-thresholding with otsu. The method I am using currently is actually via maximising the between class variance, I have managed to get the same threshold value given as that by the OpenCV library. However, that is just via running otsu method once.
Documentation on how to do multi-level thresholding or rather recursive thresholding is rather limited. Where do I do after obtaining the original otsu's value? Would appreciate some hints, I been playing around with the code, adding one external for loop, but the next value calculated is always 254 for any given image:(
My code if need be:
//compute histogram first
cv::Mat imageh; //image edited to grayscale for histogram purpose
//imageh=image; //to delete and uncomment below;
cv::cvtColor(image, imageh, CV_BGR2GRAY);
int histSize[1] = {256}; // number of bins
float hranges[2] = {0.0, 256.0}; // min andax pixel value
const float* ranges[1] = {hranges};
int channels[1] = {0}; // only 1 channel used
cv::MatND hist;
// Compute histogram
calcHist(&imageh, 1, channels, cv::Mat(), hist, 1, histSize, ranges);
IplImage* im = new IplImage(imageh);//assign the image to an IplImage pointer
IplImage* finalIm = cvCreateImage(cvSize(im->width, im->height), IPL_DEPTH_8U, 1);
double otsuThreshold= cvThreshold(im, finalIm, 0, 255, cv::THRESH_BINARY | cv::THRESH_OTSU );
cout<<"opencv otsu gives "<<otsuThreshold<<endl;
int totalNumberOfPixels= imageh.total();
cout<<"total number of Pixels is " <<totalNumberOfPixels<< endl;
float sum = 0;
for (int t=0 ; t<256 ; t++)
{
sum += t * hist.at<float>(t);
}
cout<<"sum is "<<sum<<endl;
float sumB = 0; //sum of background
int wB = 0; // weight of background
int wF = 0; //weight of foreground
float varMax = 0;
int threshold = 0;
//run an iteration to find the maximum value of the between class variance(as between class variance shld be maximise)
for (int t=0 ; t<256 ; t++)
{
wB += hist.at<float>(t); // Weight Background
if (wB == 0) continue;
wF = totalNumberOfPixels - wB; // Weight Foreground
if (wF == 0) break;
sumB += (float) (t * hist.at<float>(t));
float mB = sumB / wB; // Mean Background
float mF = (sum - sumB) / wF; // Mean Foreground
// Calculate Between Class Variance
float varBetween = (float)wB * (float)wF * (mB - mF) * (mB - mF);
// Check if new maximum found
if (varBetween > varMax) {
varMax = varBetween;
threshold = t;
}
}
cout<<"threshold value is: "<<threshold;
To extend Otsu's thresholding method to multi-level thresholding the between class variance equation becomes:
Please check out Deng-Yuan Huang, Ta-Wei Lin, Wu-Chih Hu, Automatic
Multilevel Thresholding Based on Two-Stage Otsu's Method with Cluster
Determination by Valley Estimation, Int. Journal of Innovative
Computing, 2011, 7:5631-5644 for more information.
http://www.ijicic.org/ijicic-10-05033.pdf
Here is my C# implementation of Otsu Multi for 2 thresholds:
/* Otsu (1979) - multi */
Tuple < int, int > otsuMulti(object sender, EventArgs e) {
//image histogram
int[] histogram = new int[256];
//total number of pixels
int N = 0;
//accumulate image histogram and total number of pixels
foreach(int intensity in image.Data) {
if (intensity != 0) {
histogram[intensity] += 1;
N++;
}
}
double W0K, W1K, W2K, M0, M1, M2, currVarB, optimalThresh1, optimalThresh2, maxBetweenVar, M0K, M1K, M2K, MT;
optimalThresh1 = 0;
optimalThresh2 = 0;
W0K = 0;
W1K = 0;
M0K = 0;
M1K = 0;
MT = 0;
maxBetweenVar = 0;
for (int k = 0; k <= 255; k++) {
MT += k * (histogram[k] / (double) N);
}
for (int t1 = 0; t1 <= 255; t1++) {
W0K += histogram[t1] / (double) N; //Pi
M0K += t1 * (histogram[t1] / (double) N); //i * Pi
M0 = M0K / W0K; //(i * Pi)/Pi
W1K = 0;
M1K = 0;
for (int t2 = t1 + 1; t2 <= 255; t2++) {
W1K += histogram[t2] / (double) N; //Pi
M1K += t2 * (histogram[t2] / (double) N); //i * Pi
M1 = M1K / W1K; //(i * Pi)/Pi
W2K = 1 - (W0K + W1K);
M2K = MT - (M0K + M1K);
if (W2K <= 0) break;
M2 = M2K / W2K;
currVarB = W0K * (M0 - MT) * (M0 - MT) + W1K * (M1 - MT) * (M1 - MT) + W2K * (M2 - MT) * (M2 - MT);
if (maxBetweenVar < currVarB) {
maxBetweenVar = currVarB;
optimalThresh1 = t1;
optimalThresh2 = t2;
}
}
}
return new Tuple(optimalThresh1, optimalThresh2);
}
And this is the result I got by thresholding an image scan of soil with the above code:
(T1 = 110, T2 = 147).
Otsu's original paper: "Nobuyuki Otsu, A Threshold Selection Method
from Gray-Level Histogram, IEEE Transactions on Systems, Man, and
Cybernetics, 1979, 9:62-66" also briefly mentions the extension to
Multithresholding.
https://engineering.purdue.edu/kak/computervision/ECE661.08/OTSU_paper.pdf
Hope this helps.
Here is a simple general approach for 'n' thresholds in python (>3.0) :
# developed by- SUJOY KUMAR GOSWAMI
# source paper- https://people.ece.cornell.edu/acharya/papers/mlt_thr_img.pdf
import cv2
import numpy as np
import math
img = cv2.imread('path-to-image')
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
a = 0
b = 255
n = 6 # number of thresholds (better choose even value)
k = 0.7 # free variable to take any positive value
T = [] # list which will contain 'n' thresholds
def sujoy(img, a, b):
if a>b:
s=-1
m=-1
return m,s
img = np.array(img)
t1 = (img>=a)
t2 = (img<=b)
X = np.multiply(t1,t2)
Y = np.multiply(img,X)
s = np.sum(X)
m = np.sum(Y)/s
return m,s
for i in range(int(n/2-1)):
img = np.array(img)
t1 = (img>=a)
t2 = (img<=b)
X = np.multiply(t1,t2)
Y = np.multiply(img,X)
mu = np.sum(Y)/np.sum(X)
Z = Y - mu
Z = np.multiply(Z,X)
W = np.multiply(Z,Z)
sigma = math.sqrt(np.sum(W)/np.sum(X))
T1 = mu - k*sigma
T2 = mu + k*sigma
x, y = sujoy(img, a, T1)
w, z = sujoy(img, T2, b)
T.append(x)
T.append(w)
a = T1+1
b = T2-1
k = k*(i+1)
T1 = mu
T2 = mu+1
x, y = sujoy(img, a, T1)
w, z = sujoy(img, T2, b)
T.append(x)
T.append(w)
T.sort()
print(T)
For full paper and more informations visit this link.
I've written an example on how otsu thresholding work in python before. You can see the source code here: https://github.com/subokita/Sandbox/blob/master/otsu.py
In the example there's 2 variants, otsu2() which is the optimised version, as seen on Wikipedia page, and otsu() which is more naive implementation based on the algorithm description itself.
If you are okay in reading python codes (in this case, they are pretty simple, almost pseudo code like), you might want to look at otsu() in the example and modify it. Porting it to C++ code is not hard either.
#Antoni4 gives the best answer in my opinion and it's very straight forward to increase the number of levels.
This is for three-level thresholding:
#include "Shadow01-1.cuh"
void multiThresh(double &optimalThresh1, double &optimalThresh2, double &optimalThresh3, cv::Mat &imgHist, cv::Mat &src)
{
double W0K, W1K, W2K, W3K, M0, M1, M2, M3, currVarB, maxBetweenVar, M0K, M1K, M2K, M3K, MT;
unsigned char *histogram = (unsigned char*)(imgHist.data);
int N = src.rows*src.cols;
W0K = 0;
W1K = 0;
M0K = 0;
M1K = 0;
MT = 0;
maxBetweenVar = 0;
for (int k = 0; k <= 255; k++) {
MT += k * (histogram[k] / (double) N);
}
for (int t1 = 0; t1 <= 255; t1++)
{
W0K += histogram[t1] / (double) N; //Pi
M0K += t1 * (histogram[t1] / (double) N); //i * Pi
M0 = M0K / W0K; //(i * Pi)/Pi
W1K = 0;
M1K = 0;
for (int t2 = t1 + 1; t2 <= 255; t2++)
{
W1K += histogram[t2] / (double) N; //Pi
M1K += t2 * (histogram[t2] / (double) N); //i * Pi
M1 = M1K / W1K; //(i * Pi)/Pi
W2K = 1 - (W0K + W1K);
M2K = MT - (M0K + M1K);
if (W2K <= 0) break;
M2 = M2K / W2K;
W3K = 0;
M3K = 0;
for (int t3 = t2 + 1; t3 <= 255; t3++)
{
W2K += histogram[t3] / (double) N; //Pi
M2K += t3 * (histogram[t3] / (double) N); // i*Pi
M2 = M2K / W2K; //(i*Pi)/Pi
W3K = 1 - (W1K + W2K);
M3K = MT - (M1K + M2K);
M3 = M3K / W3K;
currVarB = W0K * (M0 - MT) * (M0 - MT) + W1K * (M1 - MT) * (M1 - MT) + W2K * (M2 - MT) * (M2 - MT) + W3K * (M3 - MT) * (M3 - MT);
if (maxBetweenVar < currVarB)
{
maxBetweenVar = currVarB;
optimalThresh1 = t1;
optimalThresh2 = t2;
optimalThresh3 = t3;
}
}
}
}
}
#Guilherme Silva
Your code has a BUG
You Must Replace:
W3K = 0;
M3K = 0;
with
W2K = 0;
M2K = 0;
and
W3K = 1 - (W1K + W2K);
M3K = MT - (M1K + M2K);
with
W3K = 1 - (W0K + W1K + W2K);
M3K = MT - (M0K + M1K + M2K);
;-)
Regards
EDIT(1): [Toby Speight]
I discovered this bug by applying the effect to the same picture at different resoultions(Sizes) and seeing that the output results were to much different from each others (Even changing resolution a little bit)
W3K and M3K must be the totals minus the Previous WKs and MKs.
(I thought about this for Code-similarity with the one with one level less)
At the moment due to my lacks of English I cannot explain Better How and Why
To be honest I'm still not 100% sure that this way is correct, even thought from my outputs I could tell that it gives better results. (Even with 1 Level more (5 shades of gray))
You could try yourself ;-)
Sorry
My Outputs:
3 Thresholds
4 Thresholds
I found a useful piece of code in this thread. I was looking for a multi-level Otsu implementation for double/float images. So, I tried to generalize example for N-levels with double/float matrix as input. In my code below I am using armadillo library as dependency. But this code can be easily adapted for standard C++ arrays, just replace vec, uvec objects with single dimensional double and integer arrays, mat and umat with two-dimensional. Two other functions from armadillo used here are: vectorise and hist.
// Input parameters:
// map - input image (double matrix)
// mask - region of interest to be thresholded
// nBins - number of bins
// nLevels - number of Otsu thresholds
#include <armadillo>
#include <algorithm>
#include <vector>
mat OtsuFilterMulti(mat map, int nBins, int nLevels) {
mat mapr; // output thresholded image
mapr = zeros<mat>(map.n_rows, map.n_cols);
unsigned int numElem = 0;
vec threshold = zeros<vec>(nLevels);
vec q = zeros<vec>(nLevels + 1);
vec mu = zeros<vec>(nLevels + 1);
vec muk = zeros<vec>(nLevels + 1);
uvec binv = zeros<uvec>(nLevels);
if (nLevels <= 1) return mapr;
numElem = map.n_rows*map.n_cols;
uvec histogram = hist(vectorise(map), nBins);
double maxval = map.max();
double minval = map.min();
double odelta = (maxval - abs(minval)) / nBins; // distance between histogram bins
vec oval = zeros<vec>(nBins);
double mt = 0, variance = 0.0, bestVariance = 0.0;
for (int ii = 0; ii < nBins; ii++) {
oval(ii) = (double)odelta*ii + (double)odelta*0.5; // centers of histogram bins
mt += (double)ii*((double)histogram(ii)) / (double)numElem;
}
for (int ii = 0; ii < nLevels; ii++) {
binv(ii) = ii;
}
double sq, smuk;
int nComb;
nComb = nCombinations(nBins,nLevels);
std::vector<bool> v(nBins);
std::fill(v.begin(), v.begin() + nLevels, true);
umat ibin = zeros<umat>(nComb, nLevels); // indices from combinations will be stored here
int cc = 0;
int ci = 0;
do {
for (int i = 0; i < nBins; ++i) {
if(ci==nLevels) ci=0;
if (v[i]) {
ibin(cc,ci) = i;
ci++;
}
}
cc++;
} while (std::prev_permutation(v.begin(), v.end()));
uvec lastIndex = zeros<uvec>(nLevels);
// Perform operations on pre-calculated indices
for (int ii = 0; ii < nComb; ii++) {
for (int jj = 0; jj < nLevels; jj++) {
smuk = 0;
sq = 0;
if (lastIndex(jj) != ibin(ii, jj) || ii == 0) {
q(jj) += double(histogram(ibin(ii, jj))) / (double)numElem;
muk(jj) += ibin(ii, jj)*(double(histogram(ibin(ii, jj)))) / (double)numElem;
mu(jj) = muk(jj) / q(jj);
q(jj + 1) = 0.0;
muk(jj + 1) = 0.0;
if (jj>0) {
for (int kk = 0; kk <= jj; kk++) {
sq += q(kk);
smuk += muk(kk);
}
q(jj + 1) = 1 - sq;
muk(jj + 1) = mt - smuk;
mu(jj + 1) = muk(jj + 1) / q(jj + 1);
}
if (jj>0 && jj<(nLevels - 1)) {
q(jj + 1) = 0.0;
muk(jj + 1) = 0.0;
}
lastIndex(jj) = ibin(ii, jj);
}
}
variance = 0.0;
for (int jj = 0; jj <= nLevels; jj++) {
variance += q(jj)*(mu(jj) - mt)*(mu(jj) - mt);
}
if (variance > bestVariance) {
bestVariance = variance;
for (int jj = 0; jj<nLevels; jj++) {
threshold(jj) = oval(ibin(ii, jj));
}
}
}
cout << "Optimized thresholds: ";
for (int jj = 0; jj<nLevels; jj++) {
cout << threshold(jj) << " ";
}
cout << endl;
for (unsigned int jj = 0; jj<map.n_rows; jj++) {
for (unsigned int kk = 0; kk<map.n_cols; kk++) {
for (int ll = 0; ll<nLevels; ll++) {
if (map(jj, kk) >= threshold(ll)) {
mapr(jj, kk) = ll+1;
}
}
}
}
return mapr;
}
int nCombinations(int n, int r) {
if (r>n) return 0;
if (r*2 > n) r = n-r;
if (r == 0) return 1;
int ret = n;
for( int i = 2; i <= r; ++i ) {
ret *= (n-i+1);
ret /= i;
}
return ret;
}

Fast bilinear interpolation on old iOS devices

I've got the following code to do a biliner interpolation from a matrix of 2D vectors, each cell has x and y values of the vector, and the function receives k and l indices telling the bottom-left nearest position in the matrix
// p[1] returns the interpolated values
// fieldLinePointsVerts the raw data array of fieldNumHorizontalPoints x fieldNumVerticalPoints
// only fieldNumHorizontalPoints matters to determine the index to access the raw data
// k and l horizontal and vertical indices of the point just bellow p[0] in the raw data
void interpolate( vertex2d* p, vertex2d* fieldLinePointsVerts, int fieldNumHorizontalPoints, int k, int l ) {
int index = (l * fieldNumHorizontalPoints + k) * 2;
vertex2d p11;
p11.x = fieldLinePointsVerts[index].x;
p11.y = fieldLinePointsVerts[index].y;
vertex2d q11;
q11.x = fieldLinePointsVerts[index+1].x;
q11.y = fieldLinePointsVerts[index+1].y;
index = (l * fieldNumHorizontalPoints + k + 1) * 2;
vertex2d q21;
q21.x = fieldLinePointsVerts[index+1].x;
q21.y = fieldLinePointsVerts[index+1].y;
index = ( (l + 1) * fieldNumHorizontalPoints + k) * 2;
vertex2d q12;
q12.x = fieldLinePointsVerts[index+1].x;
q12.y = fieldLinePointsVerts[index+1].y;
index = ( (l + 1) * fieldNumHorizontalPoints + k + 1 ) * 2;
vertex2d p22;
p22.x = fieldLinePointsVerts[index].x;
p22.y = fieldLinePointsVerts[index].y;
vertex2d q22;
q22.x = fieldLinePointsVerts[index+1].x;
q22.y = fieldLinePointsVerts[index+1].y;
float fx = 1.0 / (p22.x - p11.x);
float fx1 = (p22.x - p[0].x) * fx;
float fx2 = (p[0].x - p11.x) * fx;
vertex2d r1;
r1.x = fx1 * q11.x + fx2 * q21.x;
r1.y = fx1 * q11.y + fx2 * q21.y;
vertex2d r2;
r2.x = fx1 * q12.x + fx2 * q22.x;
r2.y = fx1 * q12.y + fx2 * q22.y;
float fy = 1.0 / (p22.y - p11.y);
float fy1 = (p22.y - p[0].y) * fy;
float fy2 = (p[0].y - p11.y) * fy;
p[1].x = fy1 * r1.x + fy2 * r2.x;
p[1].y = fy1 * r1.y + fy2 * r2.y;
}
Currently this code needs to be run every single frame in old iOS devices, say devices with arm6 processors
I've taken the numeric sub-indices from the wikipedia's equations http://en.wikipedia.org/wiki/Bilinear_interpolation
I'd accreciate any comments on optimization for performance, even plain asm code
This code should not be causing your slowdown if it's only run once per frame. However, if it's run multiple times per frame, it easily could be.
I'd run your app with a profiler to see where the true performance problem lies.
There is some room for optimization here: a) Certain index calculations could be factored out and re-used in subsequent calculations), b) You could dereference your fieldLinePointsVerts array to a pointer once and re-use that, instead of indexing it twice per index...
but in general those things won't help a great deal, unless this function is being called many, many times per frame. In which case every little thing will help.

Resources