Research And Implement Of Video Analysis Technology On Community Surveillance

As the video surveillance system grows,the contents of the surveillance video are growing at an exponential level,how to efficiently manage,view and retrieve these video contents becomes an urgent need to be solved.Community surveillance video has lots of features,such as poor image quality,large number of illumination changes,economic and practical design etc.,this thesis mainly focuses on the research and improvement of foreground detection,multi-target tracking and image retrieval technology which required in surveillance video analysis.All these improved technologies are experimented and Comparative analyzed on real community surveillance video.The main research contents of this thesis are as follows:(1)As lots of noise in the surveillance video which influences video analysis,this thesis carries out theoretical analysis and experimental comparison on some commonly image preprocess techniques and picked out the best fit algorithm for real surveillance video.In the process of video dynamic and static segmentation,this thesis proposes a block multi-threshold inter-frame difference algorithm based on the law of pedestrian movement.The results of experiments on the real community surveillance video show that the algorithm has faster segmentation speed and lower false detection rate than the classic inter-frame difference algorithm.(2)Due to lots of sudden illumination changes and illumination irregularities in community surveillance video,inspired of dual background model method and an adaptive lighting compensation strategy,this thesis proposes a fast robust foreground detection algorithm which can cope with the illumination change.Some experiments are conducted in Wallflower data set and real community surveillance video.The results show that this algorithm has higher detection accuracy than the mixed Gaussian model(GMM),the VIBE algorithm and adaptive self-organizing network algorithm(LBASOM)in the event of sudden illumination change,and has faster process speed than GMM,LBASOM and multi-layer background model.(3)Because of the high complexity of the existing multi-target tracking algorithms,this thesis proposes an online multi-target tracking algorithm which combined with Meanshift algorithm and target feature matching.The algorithm does not need training model and basically does not need a priori knowledge.The results of experiments on the real community surveillance video show that this algorithm has a lower ratio of mostly lost trajectories(ML)and faster processing speedcompared with the CMOT algorithm.(4)A theoretical analysis is conduct on four kinds of target features,including color histogram,color and edge direction feature descriptor,image hash and SURF feature.Then,this thesis improved different feature similarity measurements in order to obtain a similarity value ranging from 0 to 1.Experiments show that the improved similarity measurements are more suitable for image retrieval.(5)This thesis divide the community surveillance video analysis system into three part: dynamic and static segmentation and information extraction layer,multi-target tracking analysis layer and video target retrieval layer.Combined with the actual needs of community surveillance,this thesis realizes a community surveillance video analysis system layer by layer.