Real-Time Tracking Based on Rotation-Invariant Descriptors

Common tracking algorithms based on descriptors usually use a bounding box containing a target for extracting of its features. Disjoint background noise inside of the box strongly affects target descriptors. We propose to compute the histograms of oriented gradients in several circular windows within the actual region of support of a target. Such descriptors are background noise-free and rotation-invariant. The suggested tracking algorithm additionally utilizes depth information from a Kinect camera for better tracking when partial occlusions of the target are faced. The performance of the proposed algorithm is tested in terms of recognition rate using the Princeton Tracking Benchmark scenarios and compared with that of the state-of-the-art tracking algorithms. Finally, in order to achieve high rate of processing, the algorithm was implemented with GPU parallel processing technologies.