Deformable object tracking with spatiotemporal segmentation in big vision surveillance

Rapid development of worldwide networks have changed the traditional challenges in vision surveillance to a big data level. Accordingly, the video processing technologies also need to focus more on the new coming big vision problems such as efficient content understanding. As a fundamental and indispensable pre-step for high-level video analysis, e.g. behavior recognition for social security, accurate and robust object tracking can play an essential role because of its capability in extracting the salient information from the captured video dataset. Due to the complexity of the realistic application environments, accurate and robust tracking is not easy because the object appearance may continually change during its moving, especially for the deformable objects, it is difficult for the designed appearance model being adaptive to the heavy shape variations as rotation or distortion. In this paper, a novel object tracking based on spatial segmentation is proposed to handle the problem of drastic appearance changes of the deformable object. By using the motion information between the consecutive frames, the irregular areas of the deformable object can be segmented more accurately by energy function optimization with boundary convergence. In succession, the segmentation areas are modeled by a structural SVM as learning samples to achieve more effective online tracking. Based on the evaluation of the proposed tracking on the standard benchmark database containing the challenges of heavy intrinsic variations and occlusions, the experiment results have demonstrated a significant improvement in accuracy and robustness when compared with other state-of-art tracking approaches.