Active Tracking Using Kernel-Based Vision Processor and Robust Fuzzy Control

In this paper we introduce a practicable system by combining a vision processing algorithm and a fuzzy controller to obtain an efficient active tracker. Target tracking performance is heavily dependent on a good blend of vision algorithm and control. Because of performance and computational complexity, in practice, many visual tracking algorithms cannot be linked with control systems to track objects in realtime. Robustness and speed are the two major bottlenecks of current visual tracking algorithms. In this paper, the target's visual model is used along with a kernel-based searching algorithm to predict the target position. A model update is also incorporated to recognize when the target’s appearance is changing due to its pose change. In case of target track loss, a search algorithm sweeps the space in vicinity of last target position to recover the lost target. A motion detection module used in our tracking system not only helps to initiate the tracking process automatically, it also finds the target’s presence after track loss. A fuzzy control is also synthesized to reach our control performance objectives. The idea is implemented in an active camera system to track moving targets. In addition, we've used the parallel processing technique for vision and control to reach acceptable speed and accuracy in real-time tracking.