In this paper, we construct a visual surveillance system,in which each monitoring device works like the eyes of a chameleon. Particularly, each device is built based on embedded systems, with the motion tracking and simple stereo matching functions. As the embedded systems lack strong computing capability for image processing, we introduce some methods to reduce the load of the device. More importantly, in our system, all monitoring devices are operated based on a distributed collaborative mechanism to automatically and efficiently assign the proper number of cameras to track an object. In summary, the technologies of computer vision, distributed computation and embedded systems are elaborately integrated into our work.