Research On Methods For Pedestrian Association Across Non-overlapping Camera Views

Object tracking in non-overlapping cameras is an important direction in the research of intelligent monitoring system, and it refers to tracking the object trajectory with the integration of information in multi-cameras in surveillance system. Target association is a core issue in object tracking. This thesis studies target association problem in the non-overlapping surveillance cameras and makes an in-depth study on target feature extraction and matching. By research on the current status of target association, the problems existed and method is analysis. Proper pre-processing and post-processing are designed and add to the process of target matching and association, and proposed an approach of target matching based on target appearance character description. The main contents of the paper include:1. By analysis of the current status of motion detection in static camera, we point out the shortcomings of object extraction with background Subtraction method, and we designed a basic framework of object extraction based on color clustering and human zone model. Through the introduction of color-based segmentation method, we effectively solved the target incomplete problem in background Subtraction method, and improved the target feature extraction results.2. In the step of target feature extraction and matching, an approach based on the color space distribution model is proposed for the shortcomings of the traditional MCSHR method, which is more conform with the appearance differences between different objects and gain better result. Meanwhile, an object matching method based on texture pattern reconstruction is proposed, more in line with the characteristics of human cognition, and designed the specific matching algorithm, which can effectively improve the accuracy of object matching, and enriches the method of object appearance description.3. For the shortcoming of the traditional method is not well associated with the use of target continuous image feature of the monitoring video, we propose a target association approach based on Adaboost continuous target image sequence. Features of the continuous target image sequence is used for training the Adaboost combination classifier, which is used for the target association in non-overlapping surveillance camera views, and it improves the accuracy of the target association.Using the above methods, the moving object in the surveillance videos is associated. The objects in non-overlapping cameras are selected, matching the objects by extracted the features, and the association result is presented. The examples sustain that the method proposed by this thesis is rational and effective.