I am trying to apply kalman filter for video processing , i am studying about it from different sources but it take me towards question that if i don't know that where my object come in frame mean i don't know about its angle (location in frame) then is it possible that Kalman filter automatically detect its measurements e.g location (as i already applied algorithms for specific object detection , so i only want that detected object tracking) and then track it in other frames ?

Other thing which come to my mind is that two objects (which i want to detect) come in the same frame and in the same measurements to is kalman filter take them as two different objects for tracking ?

  • $\begingroup$ The first sentence is oft-repeated and not well suited for the engineering process. Start with what you want to do. For the sake of the argument, if what you want to do lends itself to a Kalman filter solution (I saw track, so it could be a valid choice), then you've stumbled upon the problem of "Data Association". That is, once you have a detect (which typically comes first), associating it with an existing track or the birth of a new track. The multiple target or clutter scenarios you alluded to also fall into this category. Nevertheless, I would rephrase the question. $\endgroup$ – Bryan Feb 21 '14 at 18:27
  • $\begingroup$ @Bryan Isn't it computationally expensive if i track 2/3 vehicles at the same time in one frame and i have to make different track for them ? $\endgroup$ – ARG Feb 21 '14 at 18:38
  • $\begingroup$ again it goes back to what you want to do which you've yet to define. If you want to do multiple target tracking, there are algorithms tailored for that. If you're treating some detections as clutter, then there are algorithms suited for that assumption. $\endgroup$ – Bryan Feb 21 '14 at 18:51
  • $\begingroup$ I want to track the vehicle as i already applied detection algorithms for it and there may be two different vehicles in one frame or 3 or my be one $\endgroup$ – ARG Feb 21 '14 at 18:56
  • $\begingroup$ I don't know how I can be any more clear; you need to fully define your problem. Until then I can only speculate. $\endgroup$ – Bryan Feb 21 '14 at 19:46

Main thing is that in the first frame it will be required to select the object of detection because it is obvious that the algorithm will not know automatically which object you want to track if a number of things will be moving in the video scene.Lucas-Kannade method is one of the methods which can detect moving objects in a given video frame . If you do not specify object it will show movement of all moving objects.Thats why you have to tell the algorithm which thing you want to track. Second thing i want to tell you is about Kalman filter which i think you are interpreting wrong , what Kalman filter actually do is that it uses the system model and input signal and generates its own output which is actually a predicted value . Now using the observation from the system it calculate the correction to be applied to the predicted output and hence final results are better than observations obtained from the system. Hence very first step is that you need a system model and the measurements. In tracking algorithms generally, Image processing algorithm is applied to detect the object of interest and then Kalman filter is applied to make the detected location more precise. Do not think that it will predict on its own. I hope it will help you.

  • $\begingroup$ I already implement the algorithms for object detection , now i only want to track it $\endgroup$ – ARG Feb 21 '14 at 19:05
  • $\begingroup$ then in each frame you will have to detect the object using your algorithm ,determine its location and keep on passing them to kalman filter.According to system model kalman filter will provide more accurate location of your object in frame under processing $\endgroup$ – Amit_DSP Feb 21 '14 at 19:10
  • $\begingroup$ I don't want its location , the only thing i want is the first frame of that object until the last frame of that object , so that i can save only one frame of that particular object $\endgroup$ – ARG Feb 21 '14 at 19:11
  • $\begingroup$ i did not get it? $\endgroup$ – Amit_DSP Feb 21 '14 at 19:16
  • $\begingroup$ I have nothing to do with locations , the only thing i want is the first frame of that object until the last frame of that object either even move its location in every frame , i want that i can save only one frame of that particular object which i am detecting because that object take 60/70 frames to pass from the eye of camera $\endgroup$ – ARG Feb 21 '14 at 19:18

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.