Research On The Key Technologies Of Distributed Intelligent Visual Surveillance Systems

With the development of the network and sensor technologies, automatic and intelligent visual surveillance systems have become more and more popular. Existing visual surveillance systems provide the infrastructure only to capture, store and distribute video, while leaving the task of threat detection exclusively to human operators. Intelligent visual surveillance systems should create a shift in the security paradigm from "investigation of incidents" to "prevention of potentially catastrophic incidents". There exist some open problems and new challenges including software platform, event analysis, and etc., which are different from those in the existing visual surveillance systems. This dissertation addresses some key technologies in intelligent visual surveillance systems and has made innovative progress in the following respects:1) A Distributed Multi-Agent Software Platform based on Quality of ServiceA distributed multi-agent software platform based on quality of service is proposed. Each agent in the software architecture is an independent process, and can consistently communicate each other. The loose-coupled publish-subscribe model based on group messages is adopted, which caters for the spontaneous interaction between modules. A data transmitting mechanism based on quality of service is proposed according to the transmitting data types and features in the distributed intelligent visual surveillance systems. Experiment tests show that this mechanism has less delay for data transmitting.2) Consistent Foreground Object Detection and Tracking for the Indoor ScenesAn algorithm of consistent foreground object detection for the indoor scenes is presented for foreground object detection tracking and eliminating ghost effect. The basic idea of this method is as follows: both color and gradient information are fused, and two background models, the original background and the runtime background, are created and dynamically updated in whole detection process. This method can efficiently solve the problem of consistent foreground object detection for the objects in a "move and stop" way. The tracking method based on the particle filter method can robustly track multiple objects. Experiment results show that this tracking method can simultaneously track the whole body, head and hand if person stop or move.3) Face Recognition Based on the Probability Outputs of Multi-class Support Vector MachinesStandard Support Vector Machines (SVMs) does not provide probabilities output, a directly solving posterior probability method is presented for the probability outputs of' one against one' multi-class SVMs, which improves the classification ability. Face recognition based on the probability outputs of multi-class SVMs is proposed combining the advantages both SVM and probability modeling. Considering the special situation for face recognition in intelligent meeting scene, the front face is chosen for face recognition to reduce the pose effect, which is the ratio of the lengths or areas between the head and the face. Experiment results show that this method not only improves the precision of face recognition, but also provides the reliability of the classification.4) Real-time Analysis of Situation Events Based on Hierarchical Events FusionThe definition of semantic event is given. Its three characteristics are analyzed. (1) The event is the spatio-temporal object; (2) The event is related with the dynamic environment; (3) The event has hierarchical structure. Hierarchical dynamic Bayesian network based on hierarchical events fusion is modeled for situation event analysis, and a real-time recognition method based on RBPF method is proposed. Based on the hierarchical features for the event and these relations among events at different levels, situation event is decomposed into a sequence of sub-events at different levels. The corresponding RBPF method are constructed for the inference of the posterior probability of each node in a hierarchical dynamic Bayesian network in order to recognize situation events in real time. Simulation experiments results show that this method can analyze situation event in real time and achieve the better recognition precision and the less computation time than the PF method.In a word, it can be drawn that the achievements and innovations summarized above effectively solve these key problems in the field of intelligent visual surveillances, and have broad application perspective and potential economic value. This dissertation is supported by National Natural Science Foundation of China under Grant 90304018, 60672137, the Doctoral Program Foundation of Ministry of Education of China under Grant 20060497015 and the Natural Science Foundation of Hubei Province (No. 2004 ABA043).