Sum-of-Squared Differences Enhanced The Performance Of Kernel Particle Filter In Non-rigid Object Tracking

We studies the effects of the tracking system caused by object's sudden variable motion in order to improve the system's performance and efficiency. Our approach combines the strengths of two successful algorithms: sum-of-squared differences(SSD) and kernel particle filters(KPF) which was developed from usual particle filters. In cascaded algorithm SKPF, both of them can exert their own characteristic.First, KPF inherit good tracking performance from traditional PF. It was observed that a PF does not perform well when the dynamic system has a very small system noise or if the observation noise has very small variance. But KPF can overcome this problem. When the object tolerance of sudden variable motion, KPF and PF often would lost the tracked object. In addition, the iteration times of mean-shift in KPF may raise up dramatically and result in decreasing real-time performance. So we introduce the SSD tracker to improve tracking system's performance. The SSD tracker is ideally suited to handling large displacements of object. It can locate object's position quickly , avoid lost object and reduce the iteration times of mean-shift. Second, the added SSD tracker made particle set quickly collapse to one point in the state space, and the model changes rapidly in SSD tracker. While all the side-effects accompanied with SSD tracker are circumvented by the KPF module. In order to prevent degeneration , a Gauss perturbation is added to the particles at each iterations. And KPF also has a robust model to improve SSD's stability. So they complement one another. Finally , we developed traditional SSD algorithm to integrated with SKPF more well. With this improving, we can reduce the calculation capacity. The SKPF can update object model synchronously and evaluate credibility of particle set by making use of SSD's intermediate result. It makes our SKPF tracking system more robust and effective.The paper proved the effectiveness of SKPF tracker with some experiments finally. The experiment results show us introduced SSD reduces iteration times of mean-shift in KPF and improve performance of SKPF to dealing with large displacement movement. For the attachment of particle credibility evaluation, SKPF tracker also performs robustly under occlusion phenomenon. We can see SKPF tracker is a simple but powerful object tracking system.