Slashdot Mirror
Algorithms for Motion Tracking?
Keith Handy asks: "I seem to be unable to find algorithms and/or open source programs that will do accurate motion tracking, i.e. you mark a point on an object in frame 36, and the program can follow that point on that object through all the frames following it. This is useful not just for analyzing motion, but also for interpolating/extrapolating frames of video -- so if you had something at only 15 fps, you could generate inbetween frames (which are not just crossfades between the frames) and actually smooth the effect of the motion. Not something so complicated as to get into actual physics -- just something that will indicate where (in 2D only) that part of the object has moved from one frame to the next, for any given point in the whole picture. And for that matter it doesn't have to be 100% accurate, just any means of generating a reasonable motion-flow map." This doesn't strike me as an easy algorithm to develop, but are there any papers online or offline, that might describe an algorithm that can at least track objects in an image?
"In other words, I want something that does this,
in order to write code that will do things like this and this. I already know how to write code to blur and warp images, so to be able to track motion would give me (and you) the same capabilities as these expensive plug-ins.
Anyone know any other resources, directions, or existing code I could look into to find out more about how this works, so I can incorporate it into my own programming instead of paying hundreds or thousands of dollars for limited, proprietary use of the technology?"
Hmm, not sure about any resources on the net, but I did similar stuff at uni so I can recommend a book. Try Image processing, analysis and machine vision by hlavac et al. It's a very good book with plenty of code-neutral algorithms. Good luck.
J-aims
--
Yo, whatever happened to peas? Join T( H)GS
I think the best use of this would be in video compression -- if you can recognize the movement of objects between frames, you can encode how much things have moved instead of re-encoding the entire image.
Which is exactly what MPEG does... very crudely. The MPEG solution seems to be to compare a block (8x8?) of pixels with every block in the previous frame.
The fact that MPEG doesn't use anything more sophisticated than this suggests to me that there probably aren't any algorithms which consistently work better.
Tarsnap: Online backups for the truly paranoid
have a look at openCV (stands for open computer vision), it was originally developed by intel, but was later open souced. runs on both linux and windows and is mainly used for real time motion tracking of live video sources. i'm sure there are some pretty nice algorithms in the source there somewhere. They have their stuff on sourceforge
and a yahoo groups support forum thing here
the original intel pages are here
cheers,
bjpirt
Compute the 2D FFT of each frame (in grayscale), then get
the intercorrelation function of two neighbouring
frames. The maxima are more or less where
the objects have moved.
I only used this method on artificially generated
frames, ie 1 frame with translation and noise
added. Still, the intercorrelation sinks quite
fast. On natural images, there must be a lot
of fiddling to do.
Google passes Turing test : see my journal
using a marker of a known colour (e.g. yellow), then read the raw video stream (scan line at a time) looking for the largest instance of your colour (e.g. the longest instance of yellow on a given scan line), and note where you started from on that scan line and the length of the line. From this you can work out the middle of the marker. This should give you the X, Y coords w.r.t. the camera position.
CJC
Have a look at the KLT tracker - that will probably do what you want.
An implementation can be found here:
http://vision.stanford.edu/~birch/klt/
One way might be to identify "features" in the image e.g. by colour, brightness or changes and build an association tree.
Basically, identify all "peaks" (whatever feature you're interested in) and sort them. Start with the most outstanding feature and associate its nearest neighbours with it. Repeat many times. You will have data structure of references which will produce a map of islands and isthmuses depending on how far down you look.
Attach a "label" (unique ID) to each significant feature in the frame.
Repeat for the next frame.
Compare significant features. Using some sort of threshold, you can attach a confidence level that you're looking at the safe feature in the previous frame.
That's a simplistic overview, but I did it many years ago for looking at the output of stellar formation simulations.
I'm out of my tree just now but please feel free to leave a banana.
http://motion.technolust.cx
there are some examples and a sample video which demonstrate tracking "motion."
http://motion.technolust.cx/
"Motion uses a video4linux device and detects changes in the image. If a change is detected a snapshot will be taken. "
For other applications (e.g., colorization), you need somewhat better segmentation. Doing this well in the general case is still a research topic; but that's good: you can get lots of research software from around the net that does this sort of thing. Look for keywords like "computer vision", "motion", "segmentation", and "tracking" on Google.
why don't you try this far-fetched possibility:
:-)
break up the iimage into N x N submatrices, and do a fourier transform on each subsection of the image. then do this for the next frame, and calculate the phase differences between each frame, and use linear/cubic/etc interpolation to generate the frames in between. not too difficult, and I think there is even a 2-D FFT library located somwhere on download.com. this, however might introduce a couple of artifacts, but if you're doing high framerate video, it shouldn't be too noticeable.
or even more far-fetched:
assuming that the translation of the objects in the image plane between frames are small and uniform enough, you might also be able to pull this off with a properly trained neural network on subsections of the image (so each individual feature fits approximately in each subsection). neural networks can do non-linear regression, but thier outputs are continuous, so I figure if you train it right, it'll give you what you want.
good luck
Once you have done this for every block in the original frame, you have a set of motion vectors from which you can construct an intermediate frame.
Check this:
http://robotics.stanford.edu/~birch/klt/
I think that different MPEG compression schemes track motion differently - some using a brute force method. This method treats your image like a linear function so that it can search for the region of interest in the next image by using a "newton's method" like scheme - Much more efficient than brute force pixel comparison. I could be wrong though - I wasn't really paying attention in class
Posting anonymously, due to broken slash code.
I picked up a book a while ago titled "Image Analysis and Processing", and it's part of the "Lecture Notes in Computer Science" series... it contains tons of information on image segmentation and has a few sections dealing specifically with object recognition and motion prediction. You could probably adapt many of the processes in there to suit your needs.
I picked up this book (and many other computer and math books) at my local Coles bookstore for $2-$5 CDN$ each... I guess they were trying to get rid of them. I don't know if you'll be able to find a copy, but here's the info anyways:
Lecture Notes in Computer Science
Volume 1310
Image Analysis and Processing
Alberto Del Bimbo (Editor)
Published by Springer
ISSN: 0302-9743
ISBN: 3-540-63507-6
The editor's email address is listed in the cover page: delbimbo@aguirre.ing.unifi.it, so you might be able to contact him to see where you could find a copy... Good luck!
... of the CGI industry. Many have tried and many have failed. I've known a couple of people who have been involved in development of systems to do this, and I've seen a lot of companies come and go who have promised such systems at industry shows (quite a few have claimed to be defence funded which is a little scary). None that I know of have borrowed too heavily from public algorithms to do this (although one did claim their system generated some splendid comical morphing effects when mis-applied!). There are quite a few commercial systems out there (have a look at highend3d.com's lists etc), but I guess this isn't what you're after. As I understand the state of the art, simply following a feature on an image frame (using fairly simple algorithms) is not a difficult problem in itself. The tricksy part comes when you need to follow a feature (e.g edge / point) which is changing itself during the sequence, or even in the worst case being eclipsed by another feature (person walking in front of camera). Know it doesn't help, and I hate to be negative, but I don't think you're gonna see a sourceforge motion-tracking project any time soon.
I found the audio commentry on the SG1 dvd for Small Victories facinating, with how they used lasers as points (they later brushed out some), so they could sync the CGI bugs with the moving camera.
h tm l
Also, the BBC have something camera based in the works
http://www.bbc.co.uk/rd/tour/virtualproduction.
I was considering using my webcam as a motion sensor. I do not know if it will work or not, but I was considering having it sound the alert if the .jpg changes by more than so many bytes.
.jpg size will change, therefore meaning something has happened.
That way, if the structure of the picture changes, with more or less pixels of the same color, the
Will this work?
checking the size of a jpg file would not work. maybe an uncompressed jpg file would contain the information. A very small change in the original picture could lead to very much change in the resulting bytes of the jpg file.
but try seaching for webcam motion detector on google and you will find some useful stuff.
This is a well known imageprocessing problem, which is (among others) successfully implemented in the Philips 100 hz Natural Motion TV. This TV converts the traditional 50 hz or 60 hz to 100 hz by estimating the true motion in the video sequence and interpolating the missing fields (=half of an interlaced frame) accordingly. This works extremely well, making tv (especially action movies) a lot smoother.
;)
The best technique I know of is called 3DRS, which is patented by Philips. (So it's NOT open source.) One book which explains it really well. (together with lots of other image processing algorithms) is "Video processing for multimedia systems" by G. de Haan
The only reason I know about all this because I recently started working at research into motion estimation
Hope this helps.
Hi.
Our lab is doing very similar work. We've interpolated frames of video from an 8fps image sequence (taken with a wearable computer) into a smooth 30fps video sequence, using VideoOrbits. Theres a short video example available somewhere on my homepage. Perhaps this would be of interest to you. VideoOrbits is freely available at http://wearcam.org/orbits.
Video Orbits runs at over 11 fps on
a 700 MHz dual processor machine. Its also a featureless tracking algorithm so no point correspondences need to be identified.