Abstract:
Visual object tracking is defined as the task of locating an object as it moves around in a video sequence. It has widespread applications in the area of human-computer interaction, security and surveillance, video communication and compression, augmented reality, traffic control and medical imaging. Amongst all the trackers, kernel based trackers have gained popularity in the recent past because of their simplicity and robustness to track a variety of objects. However, such trackers usually encode only single view of the object and face problems due to changing appearance patterns of the object, non-rigid object structures, object-to-object and object-to-scene occlusions, and camera motion. In this research, a new kernel-based method for real-time tracking of objects seen from a moving or static camera is proposed with an object to resolve these problems. In contrast to brute-force search, this method uses Powell‟s gradient ascent method to optimally find the most likely target position in every frame. Moreover, a template adaption module has also been proposed which accounts for the changes in shape, size, orientation and shading conditions of the target object over time. The proposed algorithm also handles short-term partial and full occlusion by using Kalman filter for trajectory prediction and Proximity Search for relocking object once it reappears in the scene after occlusion. The performance of proposed algorithm has been evaluated on a number of publicly available real-world sequences. Experimental results show robust performance of tracker for objects with changing appearance and undergoing short-term and long-term full occlusion. The computational complexity of the tracker is exceptionally low, thus making it suitable for real-time applications.
