Fast and Robust Tracker in Distance Learning Applications Using Uncalibrated Stereo Cameras

This paper presents a novel and fast homography- based tracking method to segment and track the instructor placed in front of a class board using two synchronized views. In the proposed method we estimate the instructor as a planar object and track its planar parameters. At first, the homography of board and fundamental geometry of two views are extracted using six point correspondences. The homography transformation matrix of object is calculated by obtaining three point correspondences in parallax region introduced with homography of the board. The parameters to be tracked by Kalman filter are obtained frame by frame by object's homography decomposition. Our proposed method is robust against improper illumination conditions and can effectively segment and track the instructor. We evaluated our method subjectively and objectively using alternative background subtraction and methods; namely, temporal averaging model, mixture of Gaussian model, CamShift, and our previous work. Experimental results show that our proposed method tracks the instructor in an educational video faster than previous works with higher accuracy while effectively handling the shadow.