On-line object tracking with semi-supervised boosting based particle filter
Shuifa Sun, Xian-Bing Ma, Fangmin Dong, Yin-Shi Qin, Heng Luo, Bangjun Lei · 2013
There are two key issues for the particle filtering based object tracking: the proposed distribution p(Xit|Xiiit-I) and the likelihood p(zt|Xit) between the prediction and the actual observation. The kernel color histogram based particle filter (CHPF) has achieved very good tracking performance with respect to partial occlusion, rotation and scale variations. However, it would easily lose an object when the object has similar appearance as the background or when the illumination changes. To address these problems, in this paper we introduce the Semi-Supervised On-line Boosting algorithm (SSOB) and connect the confidence of SSOB with the observation model of particle filter. This new method, Semi-Supervised Boosting On-line Object Tracking based Particle Filter (SBPF), can better distinguish objects from the background. It also has faster while more robust adaptation to the change of objects' appearance and the environment illumination condition. The core to the enhanced characteristics is implementing the likelihood as the Semi-Supervised On-line Boosting operator. Extensive experiments show the superior performance of this novel method under aforementioned difficult scenarios.