At this moment, I am able to use NN to identify an object, such as a human, when given a frame from the camera. Once locate the object, then I can feed the human object image to either NN that's designed to classify male or female.

Let's say I get 1 frame per second from the camera and perform detection, the objective is to track the number of male and female walk pass the camera within the given hours.

My question is, the same person in multiple frames will be overcounted. I couldn't wrap my head around how can I train a NN to understand that this is the same person without dive into facial recognition? I'm sure there is some tracking technique that I just don't know.

One little constraint, if the person left the camera frame and come back into it later, it is fine to treat it as two people.


1 Answer 1


One way to solve the issue may be a tracker (like Kalman). It would be faster and easier approach than neural nets.

If you insist about solving this issue through neural nets, then some magic and creativity is needed. Based on the nature of tracking, you need to feed multiple frames to predict next location of the object and check if there is any object close by. So, you may consider combining RNNs with CNNs to predict next bounding box to track and replace Kalman filter's prediction with RNNs. (check)

  • $\begingroup$ Thanks for the quick reply. So, is there no real tracking? I believe Kalman filter is using a predictive motion model. So, if I only have 2 frames, I cannot track the object. Also, what if the object is human? I cannot really predict the motion as human may stop and turn around at random time. Any insight to my situations? $\endgroup$
    – WorldWind
    Dec 13, 2017 at 7:32
  • $\begingroup$ It would be better for you to check how tracking algorithms work. Tracking algorithms are designed to solve those kinds of problems. $\endgroup$ Dec 13, 2017 at 13:04

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .