The Surveillance Video Retrieval System Based On Moving Target Classification

With the construction of safe city, urban monitoring gradually integrates into various of our lives resulting in the increment of surveillance cameras, which leads to a sharp increase of video data. The problem that how to analyze and retrieve video content of interest fast and efficiently has become urgent.At present, in the field of video monitoring, management and retrieval methods of the surveillance video are usually based on description of the video files (time and place, etc.). Although retrieval application based on the actual content of surveillance video has not gained popularity, but it has attracted more and more attention, which makes it become a hot spot in the field of intelligent video monitoring. In order to resolve the problems existing in the traditional video information retrieval, surveillance video retrieval based on moving target classification is deeply studied in this dissertation. The work is covering moving target detection, classification and pre-processing in the road monitoring scene. The corresponding video analysis algorithm is designed. Furthermore, a surveillance video retrieval system based on moving target classification is constructed. The main work is as follows:(1) Three kinds of target detection methods in video surveillance are introduced, and the detection principle and implementation steps are demonstrated in detail. Contrast experiments are carried out to compare the frame differential method with background subtraction method. Besides, three methods of background subtraction are emphatically analyzed, including SAM, SGM, GMM. Finally, the thesis analyzes and discusses the advantages and disadvantages of each target detection algorithm and their application ranges. Based on road monitoring scene of the fixed system, the single Gauss background modeling is used to detect target in the thesis.(2) Combined with the actual demand, mainly for pedestrians, three kinds of road targets about bicycles/electric vehicles and cars are classified and recognized. Morphological processing is carried out on the target area based on the output of moving target detection module, furthermore, the feature vector is constructed based on the target shape parameters and SVM (Support Vector Machine, SVM) classification is selected to train the study sample, which leads to the realization of classification and recognition of the above three kinds moving targets.(3) A video surveillance retrieval system based on moving target classification is designed and implemented. The system consists of two modules:video analysis (VA) and video retrieval (VR). In VA module, moving objects detection, tracking and classification algorithms are employed to analyze the input videos. Then, each video clips containing different type of moving targets are labeled and the corresponding description files of target categories are generated for subsequent VR applications. In the video retrieval process, users can fast query the corresponding video clips through inputting the query conditions (the date, time and target categories of the moving target). Experimental results show that the system enables users to effectively find out video clips of interest in a large number of monitoring video, and reduce the workload of manual browsing and searching, hence improve the efficiency of the surveillance video retrieval.