Robust approach for people detection and tracking by stereo vision

Abstract

In this paper, a novel method for people detection and tracking is proposed, based on stereo vision. Each person is represented by a group of the feature points. In this method feature point extraction and 2D space construction of projected points on the ground plane is performed in order to provide top view. Occlusion, as a main challenge in tracking systems, can be addressed by top view scene. A robust kernel density estimation method is employed to categorize points. Then Kalman filter is applied to reduce the detection computation complexity from second frame by predicting center of the groups in the next frame. Our method is more practical than existing methods since it has lower computation cost of detection, because of using feature extraction instead of depth map. This low computational complexity makes our method suitable to be used in real time applications.