In the first section, according to the assumption that only the silhouette could be extracted to describe the motion of a moving object, a three-frame differencing model is constructed based on the variance ratio of successive blocks of different frames, simultaneously the Gaussian kernel based density estimation model in \ is used as a filter to suppress false positives, and the SVC algorithm in \ is also employed to realize real-time multi-object motion detection. A novel contour tracing algorithm, namely the RW (roller wheel) method, is proposed in the next section. Edge information detected by the forgoing model can be fully used by this algorithm, while the interior details will be skipped. This algorithm works in a similar way that human vision system works. Experimental results are given in the end.