Visual Tracking with adaptive layered-optimizing particles in Multifeature Particle Filtering Framework

In this paper, we propose a particle filter algorithm with adaptive layered-optimization and multi-feature, which is used for motion-based tracking of natural object. A novel reliability measure based on the particle's distribution in the state space is designed to evaluate the tracking quality. According to the tracking quality, the particle set is divided into two parts: one is optimized to be concentrative for the tracking precision and the other keeps being original for the tracking robustness. The number of particles in each part is decided adaptively by the function which uses reliability score as parameter. This algorithm is demonstrated using the color and orientation features weighted by reliability score. Experiments over a set of real-world video sequences are done and the result shows that this algorithm achieves better performance when occlusion and object-motion in variable direction happen; the consuming time meets the requirement of real-time.