Would a recurrent neural network be appropriate for object tracking tasks? Mainly I will have 3D feature vectors $(x, y, t)$ where $x$ and $y$ are the positions of an object in the image and $t$ is either the current timecode or a number indexing the current frame being processed.

Since $x$, $y$ always have noise and are just "approximately" located in an image the network would have to do some de-noising. Finally the object should be tracked and the next feature vector $(x, y, t_{t+1}$) estimated.

I guess a Kalman filter would perform quite well for this problem but since the image sequences will have highly different noises (based on lighting and the object to be tracked) I assume an architecture which can learn those additional features would perform better?!

My first idea was using an Elman network to keep it simple...however I would highly appreciate any input on this topic because I do not want to "waste time" on solutions just to notice that they barely work later on.

I also read about recurrent neural networks mimic a Kalman Filter's behaviour but I have no experience whatsoever in which scenarios this would be helpful or producing good results.

Any experience, advice, literature or somehow related post is welcome :)

  • $\begingroup$ Nobody has any interesting ideas / sources? $\endgroup$ – daniel451 Dec 5 '15 at 17:28
  • 1
    $\begingroup$ I am not sure if you still need advice, but check out: robots.ox.ac.uk/~mobile/Papers/2016AAAI_ondruska.pdf $\endgroup$ – user101306 Jan 20 '16 at 16:07
  • $\begingroup$ Yeah, I'm always open minded for feedback or ideas. Thank you! $\endgroup$ – daniel451 Jan 22 '16 at 0:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.