Tracking an object in a recorded video

[henrykay]

Hi guys,

i stumble across this forum searching for a way to tag an object in a recorder video and track it as it goes.

Can i check is this possible to do ?

for example i had a video of my 2 dogs playing fetch. is it possible to tag both my dogs and the stick and track them throughout the video?

is it hard to achieve this if its possible to do it?


Thanks a lot for the advise given.

 

[gendou2]

It is possible but a very difficult task.
It can be made easier your black dog is running on a grey background fetching a white stick.
A computer algorithm called "blob detection" can identify blobs of color in an image.
For example, you might track the black and white blobs frame be frame, and interpolate the motion.
For simple experiments, I recommend experimenting with this library:

http://v3ga.net/processing/BlobDetection/

 

[maverick_starstrider]

This is called image segmentation and image registration and is a big field of computer science

 

[CylonMath]

How much time is needed to do such a task ?

 

[maverick_starstrider]

well that entirely depends on the specific task. It could be an easy thing you could write in matlab, it could take a supercomputer, it could not be doable at all. It could involve simple integral transformations on an image or it could involve developing a machine learning algorithm. It entirely depends on the specifics. (I used to tinker with this kinda stuff when I worked in medical physics in undergrad, they were trying to write algorithms to detect cancer in MRI image automatically, it was an extraordinarly difficult undertaking).

 

[henrykay]

thanks all for your valuable opinion.

gendou2 thank you for the link, i'll take a look at the information there .

hopefully its not too hard to achieve this and will not require a super computer to do so.

I assume the more object that needs to be track in a video, the harder it will be to do so. is there a chance that the tracking can be done even if the object went out of the video and re-enter ?

Appreciate any advise given .

 

[minger]

I remember seeing features of a new video editing software that was still unreleased. In it, the user drew a box on the side of a car in the first frame. The person then drew inside that box (text IIRC). Then they simply hit play and the code was able to track the side of the car through the entire stage as it both translated and rotated.

I really wish I could remember the name of it; perhaps it was just a beta version of Premeire or something like that?

 

[maverick_starstrider]

the difficulty is not in the number of objects. Once you have it working for one, tracking multiple identical objects is trivial. The difficult is dependent on how much help the computer gets before hand in determining what it's tracking and how difficult it is to distinguish the object you're tracking from the background (not so much a camoflage object on a plain background but busy backgrounds are a big problem). So if you want the algorithm to track a ball of unknown dimensions, shape and color across a very 'busy' image without you giving it any hints to start with that could be very difficult.

 

[mgb_phys]

As maverick says, it's not hard to do the simple case - the hard cases can be impossible.
A popular image processing library for this sort of stuff is OpenCV http://opencv.willowgarage.com/wiki/ there are probably even some examples of this topic - it's a common application of image processing

 

[henrykay]

interesting.

more reading to be done :D

 

[CylonMath]

Digital Image Processing - Rafael Gonzales, Richard Woods
Introductory Techniques for 3-D Computer Vision - Trucco, Alessandro Verri
Computer Vision - Linda G. Shapiro, George Stockman
Learning OpenCV : Computer Vision with the OpenCV - Gary Bradski, Adrian Kaehler

That books are worth having a look if youre interested in doing that task.

 

[henrykay]

cool. Time to hit the library for some reading up :D Hopefully those books you've mentioned are available at my region :D