Robust Object Detection and Tracking Using a Space-Temporal Mutual Feedback Scheme

A statistical object detection and tracking mutual feedback scheme, combining Gaussian mixture model (GMM) based on principal component analysis (PCA) and expectation maximization (EM) Kalman filter algorithm, is proposed in this paper. In space object detection stage, PCA provides compact and decorrelated feature space, the tracked object feature is statistically represented as GMM in RGB color space, objects are detected by maximum a posteriori (MAP) estimation. In temporal tracking stage, the tracked object is determined by the Bhattacharyya similarity measurement, the object position of consecutive frame is predicted by EM Kalman filter algorithm. The integration of object detection and tracking spatio-temporal mutual feedback scheme can decrease the accumulation error. We have applied the proposed method to object detection and tracking under the partial occlusion and the changes of moving speed with encouraging results.