Object Tracking Under large Occlusions based on Spatio-temporal Context Learning

This paper presents a visual tracking framework which has been designed based on spatio-temporal context learning with robustness against large occlusions. The proposed method is a quick and robust algorithm which extracts the spatio-temporal context information. In this method, initially, a local context model between the target object and its surrounding local background is learned in terms of spatial relationships at a tracking scene using deconvolution problem solving. Then, the learned spatial context model is used for updating the spatio-temporal context of next frame. In the next frame, tracking procedure is performed by computing confidence mapping as a convolution problem which completes the spatio-temporal context and can formulate the best object’s location using the estimated confidence mapping maximum. Also, this algorithm uses the quick Fourier transform for quick detection and learning. To avoid over-fitting problem and to reduce the noise which has been defined using the estimation error, the estimated object scale is computed through filtering and filter’s fixed parameter (< 0). Finally, the number of successfully tracked frames using the proposed tracker compared with other trackers indicates that it has considerably improved the large occlusions problem.

Keywords:

Large Occlusions , Tracking , Quick Fourier Transform , Confidence Mapping , Spatio-temporal Context Learning