I'm trying to make my OpenCV-based fiducial marker detection more robust when the user moves the camera (phone) violently. Markers are ArTag-style with a Hamming code embedded within a black border. Borders are detected by thresholding the image, then looking for quads based on the found contours, then checking the internals of the quads.
In general, decoding of the marker is fairly robust if the black border is recognized. I've tried the most obvious thing, which is downsampling the image twice, and also performing quad-detection on those levels. This helps with camera defocus on extreme nearground markers, and also with very small levels of image blur, but doesn't hugely help the general case of camera motion blur
Is there available research on ways to make detection more robust? Ideas I'm wondering about include:
Can you do some sort of optical flow tracking to "guess" the positions of the marker in the next frame, then some sort of corner detection in the region of those guesses, rather than treating the rectangle search as a full-frame thresholding?
On PCs, is it possible to derive blur coeffiients (perhaps by registration with recent video frames where the marker was detected) and deblur the image prior to processing?
On smartphones, is it possible to use the gyroscope and/or accelerometers to get deblurring coefficients and pre-process the image? (I'm assuming not, simply because if it were, the market would be flooded with shake-correcting camera apps.)
Links to failed ideas would also be appreciated if it saves me trying them.
Yes, you can use optical flow to estimate where the marker might be and localise your search, but it's just relocalisation, your tracking will have broken for the blurred frames.
I don't know enough about deblurring except to say it's very computationally intensive, so real-time might be difficult
You can use the sensors to guess the sort of blur you're faced with, but I would guess deblurring is too computational for mobile devices in real time.
Then some other approaches:
There is some really smart stuff in here: http://www.robots.ox.ac.uk/~gk/publications/KleinDrummond2004IVC.pdf where they're doing edge detection (which could be used to find your marker borders, even though you're looking for quads right now), modelling the camera movements from the sensors, and using those values to estimate how an edge in the direction of blur should appear given the frame-rate, and searching for that. Very elegant.
Similarly here http://www.eecis.udel.edu/~jye/lab_research/11/BLUT_iccv_11.pdf they just pre-blur the tracking targets and try to match the blurred targets that are appropriate given the direction of blur. They use Gaussian filters to model blur, which are symmetrical, so you need half as many pre-blurred targets as you might initially expect.
If you do try implementing any of these, I'd be really interested to hear how you get on!
From some related work (attempting to use sensors/gyroscope to predict likely location of features from one frame to another in video) I'd say that 3 is likely to be difficult if not impossible. I think at best you could get an indication of the approximate direction and angle of motion which may help you model blur using the approaches referenced by dabhaid but I think it unlikely you'd get sufficient precision to be much more help.
Related
I am trying to build a solution where I could differentiate between a 3D textured surface with the height of around 200 micron and a regular text print.
The following image is a textured surface. The black color here is the base surface.
Regular text print will be the 2D print of the same 3D textured surface.
[EDIT]
Initial thought about solving this problem, could look like this:
General idea here would be, images shot at different angles of a 3D object would be less related to each other than the images shot for a 2D object in the similar condition.
One of the possible way to verify could be: 1. Take 2 images, with enough light around (flash of the camera). These images should be shot at as far angle from the object plane as possible. Say, one taken at camera making 45 degree at left side and other with the same angle on the right side.
Extract the ROI, perspective correct them.
Find GLCM of the composite of these 2 images. If the contrast of the GLCM is low, then it would be a 3D image, else a 2D.
Please pardon the language, open for edit suggestion.
General idea here would be, images shot at different angles of a 3D object would be less related to each other than the images shot for a 2D object in the similar condition.
One of the possible way to verify could be:
1. Take 2 images, with enough light around (flash of the camera). These images should be shot at as far angle from the object plane as possible. Say, one taken at camera making 45 degree at left side and other with the same angle on the right side.
Extract the ROI, perspective correct them.
Find GLCM of composite of these 2 images. If contrast of the GLCM is low, then it would be a 3D image, else a 2D.
Please pardon the language, open for edit suggestion.
If you can get another image which
different angle or
sharper angle or
different lighting condition
you may get result. However, using two image with different angle with calibrate camera can get stereo vision image which solve your problem easily.
This is a pretty complex problem and there is no plug-in-and-go solution for this. Using light (structured or laser) or shadow to detect a height of 0.2 mm will almost surely not work with an acceptable degree of confidence, no matter of how much "photos" you take. (This is just my personal intuition, in computer vision we verify if something works by actually testing).
GLCM is a nice feature to describe texture, but it is, as far as I know, used to verify if there is a pattern in the texture, so, I believe it would output a positive value for 2D print text if there is some kind of repeating pattern.
I would let the computer learn what is text, what is texture. Just extract a large amount of 3D and 2D data, and use a machine learning engine to learn which is what. If the feature space is rich enough, it may be able to find a way to differentiate one from another, in a way our human mind wouldn't be able to. The feature space should consist of edge and colour features.
If the system environment is stable and controlled, this approach will work specially well, since the training data will be so similar to the testing data.
For this problem, I'd start by computing colour and edge features (local image pixel sums over different edge and colour channels) and try a boosted classifier. Boosted classifiers aren't the state of the art when it comes to machine learning, but they are good at not overfitting (meaning you can just insert as much data as you want), and will most likely work in a stable environment.
Hope this helps,
Good luck.
I'm currently working on a project where I need to be able to very reliable get the positions of the balls on a pool table.
I'm using a Kinect v2 above the table as the source.
Initial image looks like this (after converting it to 8-bit from 16-bit by throwing away pixels which is not around table level):
Then a I subtract a reference image with the empty table from the current image.
After thresholding and equalization it looks like this: image
It's fairly easy to detect the individual balls on a single image, the problem is that I have to do it constantly with 30fps.
Difficulties:
Low resolution image (512*424), a ball is around 4-5 pixel in diameter
Kinect depth image has a lot of noise from this distance (2 meters)
Balls look different on the depth image, for example the black ball is kind of inverted compared to the others
If they touch each other then they can become one blob on the image, if I try to separate them with depth thresholding (only using the top of the balls) then some of the balls can disappear from the image
It's really important that anything other than balls should not be detected e.g.: cue, hands etc...
My process which kind of works but not reliable enough:
16bit to 8bit by thresholding
Subtracting sample image with empty table
Cropping
Thresholding
Equalizing
Eroding
Dilating
Binary threshold
Contour finder
Some further algorithms on the output coordinates
The problem is that a pool cue or hand can be detected as a ball and also if two ball touches then it can cause issues. Also tried with hough circles but with even less success. (Works nicely if the Kinect is closer but then it cant cover the whole table)
Any clues would be much appreciated.
Expanding comments above:
I recommend improving the IRL setup as much as possible.
Most of the time it's easier to ensure a reliable setup than to try to "fix" that user computer vision before even getting to detecting/tracking anything.
My suggestions are:
Move the camera closer to the table. (the image you posted can be 117% bigger and still cover the pockets)
Align the camera to be perfectly perpendicular to the table (and ensure the sensor stand is sturdy and well fixed): it will be easier to process a perfect top down view than a slightly tilted view (which is what the depth gradient shows). (sure the data can be rotated, but why waste CPU cycles when you can simply keep the sensor straight)
With a more reliable setup you should be able to threshold based on depth.
You can possible threshold to the centre of balls since the information bellow is occluded anyway. The balls do not deform, so it the radius decreases fast the ball probably went in a pocket.
One you have a clear threshold image, you can findContours() and minEnclosingCircle(). Additionally you should contrain the result based on min and max radius values to avoid other objects that may be in the view (hands, pool cues, etc.). Also have a look at moments() and be sure to read Adrian's excellent Ball Tracking with OpenCV article
It's using Python, but you should be able to find OpenCV equivalent call for the language you use.
In terms tracking
If you use OpenCV 2.4 you should look into OpenCV 2.4's tracking algorithms (such as Lucas-Kanade).
If you already use OpenCV 3.0, it has it's own list of contributed tracking algorithms (such as TLD).
I recommend starting with Moments first: use the simplest and least computationally expensive setup initially and see how robuts the results are before going into the more complex algorithms (which will take to understand and get the parameters right to get expected results out of)
I have to make a bot which has to overcome obstacles autonomously in an arena that will be filled with rocks. The bot has to find its way through this area and reach the end point. I am thinking of using edge detector operators like canny and sobel for this problem.
I want to know whether those will be suitable options for this problem. If so, then after detecting the edges, how can I make the bot find the path, overcoming the rock obstacles?
I am using QT IDE and opencv library.
Since you will be analyzing frames of video, and the robot will be moving most of the time, image differences and optical flow too will be helpful. Edge detection alone might not help a lot, unless the surroundings and obstacles are simple and have known properties. Posting a photo of the scene can help those who want to answer the question.
Yes, canny is a very good edge detector. In fact the opencv implementation uses sobel to get the gradient estimate. You may need to apply a Gaussian filter to the image before edge detection. Edges are good features to look for rocks, but depending on the background other features such as color may also be useful. It probably would be easier if you gather 3D scene information via stereo, or laser scanner, or kinect like sensor. Also consider detecting when you bump into rocks and building up a map of where they are.
You can use contours to detect any object. You can estimate its size by finding the area of the contours. Then you can use moments to find the center of the object.
I would like to locate a car (front center point x,y) using a high resolution single camera. The camera setup is fixed at 1-2m high, and tilted around 25 degrees. The camera can provide images in where the front side of the car is visible. The intrinsic and extrinsic parameters are known.
So far, I tried to detect the headlights and number plates. Issues... Headlights are not detected as blobs all the time. The shape of the headlights are changing depending on the distance. Also, the number plate is not visible in the dark.
Is there a robust algorithm to detect a car? or to detect headlights? or detect number plate?How could I proceed?
Thanks in advance,
Are you detecting the same car everytime? If yes, then presumably the appearance remains consistent. Rather than detect and recognise blobs and shapes, you may be better off using scale and rotation invariant features combined with a machine learning algorithm. Look into the SIFT and SURF feature descriptors. For easy experimentation, use OpenCV's implementation of feature description and matching. Take a look at this example.
This is not an easy problem because of the change in the scale and point of view. Ideally, you would need a collection of training images with the car seen from different points of view to match later some of them to your input image. Then, you need local features (SIFT, SURF) or some classifier to decide on the match.
On the other hand, if you are tracking the same car all the time, check out the MeanShift algorithm. The problem is you need an initial position to carry on with the tracking.
I've 2 pictures of the same scene from an uncalibrated camera. The pics are from a slightly different angle and scale(zoom) and I'd like to superpose them, rejecting any kind of shake. In other words, I should transform them so the shake becomes imperceptible, do a Motion Compensation.
I've already tried using a simple SURF (feature) detector along with Homography but sometimes the result isn't satisfactory. So I am thinking about trying Image Rectification to compensate the motion.
- Would it work with slight changes, such as user shake?
- Would it really work to reject shake for these 2 frames? And for a bigger buffer of pictures (10 maybe)?
- Anyone knows if it would fix scale disparity (different zoom in the images)?
- What the algorithm really do? Will it transform both pictures into a third orientation?
If there is a better solution, I would be glad to know =)
EDIT
I don't aim to compensate blur motion but the displacement itself. For example, in this file the author compensates the angle difference between two cameras by Image Rectification. How does it actually work? Does it always create an intermediate picture orientation or can I specify that one of the pictures shall remains still??
Also, would I be able to apply this to many frames or it would always find an intermediate orientation for each two frames I put in?
Cheers,
I'm not sure how well superimposing the images would work. Another way to remove blur (including motion blur which should dominate in handheld camera devices) from an image is by blind deconvolution. It is basically a method of finding the inverse of the blur filter that was physically applied (camera shaken) to the real image. There's plenty of techniques out on the web. I've specifically had good results using a modified version of the algorithm in this paper: http://www.cse.cuhk.edu.hk/~leojia/all_final_papers/motion_deblur_cvpr07.pdf
It also comes with an executable file somewhere around the web so you can see if it's fit for your purpose.
Good luck out there!