A Fast and Adaptive Object Tracking Method

Robust object tracking is still a challenging research topic due to dynamical changes of illumination, viewpoints, camera jitter and partial occlusion. To handle such problems, this paper presents an adaptive object tracking method based on co-training and particle filtering algorithms. An online learning scheme with cooperation of two visual classifiers is designed to alleviate the drifting problem. The histogram of oriented gradients (HOG) and local binary pattern (LBP) are used to describe the object appearance. Two SVM classifiers are constructed separately based on the above two visual features, and online updated by the co-training framework. This updating strategy considers the same classification problem from two complementary viewpoints which successfully overcome the error accumulation problem. In order to reduce the searching state space, we introduce a dynamical model and importance sampling from the ICONDENSATION framework to improve the precision and efficiency of the sampling procedure. A correction term is introduced to decrease the weights of false positive samples, and hereby improve the performance of object tracking and classifier updating. Experimental results on four datasets verify the effectiveness and robustness of the proposed adaptive tracking method. Without decreasing the tracking performance, our method also achieves a speedup of 25 times for object state estimation compared with the traditional sliding window algorithm.