Tracking video object based on central macroblocks

Since the approaches suggested so far for video moving object tracking are sensitive to the accuracy of object segmentation, we propose a new video object tracking algorithm that provides the strength of robustness to the problems of both under-segmentation and over-segmentation. The proposed tracking algorithm introduces a concept of central macroblock, which is used to establish the correspondences between objects inside neighboring frames via two levels of similarity measurement and observations. Furthermore, MPEG motion estimation and compensation and Kalman filtering techniques are also exploited to enhance the tracking performances. While the first level of similarity measurements is limited to local texture matching via SAD (Sum of Absolute Differences) values between central macroblocks, the second level is established upon objects via directional vectors, characterizing the internal structure of the segmented objects. And both levels of similarity measurements are integrated by Kalman filtering. Experimental results carried out on PETS2001 and PETS2006 database show that the proposed algorithm achieves successful tracking performances robust to inaccuracy of object segmentation as well as other distracting factors such as occlusion, deformation, lighting effect, object disappearances and appearances etc.