Pedestrian Retrieval System Based On Multi-camera Association

The association and retrieval of multi-camera pedestrian are hot topics in the field of intelligent video surveillance. they aim to accomplish the pedestrian tracking and retrieval of multiple camera scenes. The system of multi-camera pedestrian retrieval incorporates theories and techniques of several image processing fields. The system improves the efficiency of security monitoring and saves a lot of manpower and resources, so it has a very important practical value.The thesis will introduce the basic theory of object detection and tracking in video processing, and then a fast pedestrian detection algorithm is proposed. The multi-camera pedestrian retrieval system is designed at last. Algorithmic process of the system includes object detection, pedestrian detection, multi-objective tracking in single camera, object correlation between several cameras and suspects retrieval. The main works and innovations in this thesis is as follows:(1) To improve the detection rate of pedestrian detection, decrease the large dimensions of its features and solve its time-consuming problem, the fast pedestrian detection based on combinatory feature of improved HOG and Sobel-LBP is proposed. The algorithm integrates multi-scale and multi-block mode to extract the weighted histogram of oriented gradient and Sobel-LBP as detecting. Experimental results show that our algorithm has better performance in real-time and detection accuracy.(2) The pedestrian retrieval system of multi-camera association is designed in the thesis. Object detection based on clustering background modeling method is used first, and then pedestrian detection is used to correct the object detection results, while the adjacent frames feature matching method is used to achieve multi-target tracking, the camera network topology and the similarity match of key objects is combined to achieve the association of multi-camera targets. At last, according to user request the suspect's picture is retrieved in the key objects.The system software is divided into two parts, the client is responsible for interaction with the user and the server processing multi-channel video. Experimental results show that the real-time system is able to retrieve pedestrians accurately.