Multi-level human tracking

In this paper, we describe a system on automatically tracking people in indoor environment. Parametric Estimation methods are adopted in training images to obtain means and covariance of the background pixels, They are processed in pixel classification step to extract moving people and eliminate their shadows. Our tracking algorithm consists of two levels: region level and blob level. The region level tracks a whole human by utilizing the position information of the detected persons, The EM algorithm is adopted at the blob level to segment human image into different parts. When region level tracking fails, the system switches to blob-level tracking, Average Bhattacharyya distances between corresponding blobs of consecutive frames are calculated to get the correct match between different regions. Potential applications of the proposed algorithm include HCI and visual surveillance, Experimental results are given to demonstrate the robustness and efficiency of our algorithm.