Research On Description And Recognition Of Events In Surveillance Video

Nowadays the technique of video surveillance is playing a more and more important role in the maintenance of public security. Chinese cities are carrying out the "Safe City" project, in which a great number of surveillance cameras needs to be mounted. To deal with the large number of cameras, it is unpractical to barely depend on detecting the suspicious events manually. Therefore, an intelligent video surveillance system, which is a combination of computer vision and AI, is really in need. This system can auto-recognize video events with a low error rate, bring a great improvement to video surveillance systems and can effectively solve the problems in video surveillance of different places and situations.This paper mainly discussed the design method of the structure of video event description and recognition system and the method of high-level semantic representations of video events and the technology of video events recognition. The key problems below are solved in the process of finishing the study:(1) A general, extensible and a user-centered system architecture for the description and recognition of events in surveillance videos is proposed, which is composed of description subsystem and recognition subsystem.(2) Based on the knowledge of ontology, the structure of video events is defined and a method of the knowledge description of video events is built.(3) By using the fuzzy petri net, the modeling of video events is realized, which transforms the ontology of video events to fuzzy petri net model, and provides the rules for reasoning in recognition of video events.(4) Based on the reasoning arithmetic of fuzzy petri net, a viable method of recognition of video events is founded with the events' certainty factor calculated, which makes it possible to define the alert level of video events.(5) Combining the methods mentioned above, an event of "stealing bag in public" is detected, and a multi-sensor cooperation is realized. The examples sustain that the method of description and recognition of video events proposed in this paper is effective.