A SVM embedded particle filter for multi-object detection and tracking

Tracking a vary number of moving objects from an unsteady aerial platform has three major challenges. There are fast camera motion, changing scene and lighting, and targets entering and leaving the field of view at arbitrary position. In order to deal with these difficulties, a vision system that is capable of learning, detecting and tracking multiple objects of interest is developed. Our approach combines two important algorithms: SVM and particle filter. The particle filter using color histogram as observation feature is a powerful technique for tacking multi-target. However, the design of proposal distribution and the treatment for objects entering and leaving the scene are two crucial issues. In this paper, we incorporate information from HOG-based SVM detector. The offline learned SVM detector has two major functions in the tracking system. One is to initialize the starting states of particle filter automatically and add new objects entering the scene quickly. The other is to improve the proposal distribution for the particle filter. Meanwhile, the color-based particle filtering process enable us to efficiently and reliably track the individual objects. The experiments demonstrate that our SVM embedded particle filter is an effective and fully automatic method which can be used in multi-object tracking system.