The Research On Object Tracking And Recognition In Intelligent Visual Surveillance

Intelligent visual surveillance is one of new arising high-tech application fields. It spans many subjects including image processing, image analysis, machine vision, pattern recognition, artificial intelligence, etc. The key technologies of intelligent visual surveillance include as follows: object detection and classification, object tracking, object correspondence, object recognition and video content understanding, etc. Intelligent visual surveillance is in the stage of research and development now.In this paper, we focus our work on some key technologies in intelligent vusual surveillance systems: object tracking, object correspondence and object recognition. We discuss some key problems of these technologies and try to provide some solutions. These problems are as follows: how to track objects in frame-skipping videos, how to perform object correspondence between multiple cameras with the results of object tracking, how to identify roles of tracked or detected people, how to verify the identity of a person after role identification. The contributions of this paper are as follows:1. To solve the tracking problem in frame-skipping videos, we propose a method using a particle filter with erratic motion detection. Wireless internet cameras have been becoming in popular in recent years, but the retrieved videos with wireless internet cameras have the frame skipping problems. Frame-skipping data brings the difficulties in calculating the posterior distributions of the states, viz. how to design a better transition model to describe objects moving between frames. We combine the traditional motion detection with object tracking to solve the problem of object tracking in frame-skipping videos by acceptable calculation.2. To solve the problem of object correspondence between multiple cameras, we propose a region-SIFT descriptor based target matching method. Object correspondence is finding the correspondences between objects in the different image sequences at the same time. It is difficult for object correspondene to obtain many a priori knowledge such as camera intrinsic and extrinsic parameters. Our method is a region based method and the region is represented by SIFT descriptor instead of traditional color features. The proposed method can be used for many categories of objects; neither camera calibrations nor the constraint that objects stand in the same plane is required in our method; and it is robust to changes of the light intensity.3. To solve the problem of role identification using multiple cameras, we propose a method by using a causal network. The difficulty of role identification is how to distinguish the facts which effect the identification and how to analyze the dependencies between the facts and the roles. After the analysis of persons with special roles and the extraction of visual features, spatio-temporal features and additional features, our method performs the role identification by using a causal network to fuse the features. It is beneficial to use multi-view videos in locating the position of the moving object and overcoming occlusions.4. To solve the problem of identity verification in visual surveillance, we propose a method for decreasing the influence under variable illumination intensity by using the line-based singular value feature vector instead of image gray-level value to calculate "distance" between two lines. We prove the robostness of our method to changes of the light intensity. Moreover, we suggest a distributed solution to increase the speed of face recognition.