Abstract
As the increasing amount of surveillance videos, people are more interested in issues like checking out the specific events. However, the vast amount of video data let video analysis become too complicated. Thus, developing a low-complexity and efficient indexing method is very important. In this thesis, based on the characteristic that the object motion varies with time, we extract object information from surveillance video sequences. Then, we proposed an object-based motion summary scheme by analyzing the object motion and generating indices with a proposed object-based motion spectrum. In our scheme, we first extract motion-transition points to get initial object segments using the motion intensity spectrum in [10]. By subtracting the motion-transition frames with the background, we can extract the actual object segments. With these object segments, we further utilize the background subtraction along with the motion tracking algorithm to construct and maintain an object map frame-by-frame. We introduce an object-based motion intensity spectrum to denote the variations of object groups. Finally, several event rules are made based on the object-based motion intensity spectrum, and key frames with anomalous events are summarized. Our experimental results show that the proposed scheme could find out specific events, such as, enter a scene, leave a scene, stop from continuous movement, and move from continuous static, in surveillance videos. We not only propose an effective summary scheme, meanwhile, we also inspire the idea of summarizing spatial location information with temporal relations. Through co-operating the spatial and temporal information, summary schemes can be more powerful and can provide more informative summary indices.