Study On Video Image Processing Techniques In Community Security System

Along with the development of economy and society,social demand for the security of living environment becomes higher and higher.However,traditional surveillance system can not fully support the intelligence requirement.On the other hand,the massive community monitoring video acquired in real time is capable of recording individual behaviors and interactions in the crowd.Such video data make it possible for the quantitative perception of the characteristics of the crowd,opening up a new way for community intelligent security research.Therefore,this thesis focuses on the following techniques,including person re-identification,crowd aggregation detection,violent behavior detection and fire detection,as well as their applications in the community security system.Due to the environmental factors such as monitoring angle of view,illumination,occlusion and camera parameters,etc.,it is difficult to detect and track the moving object consecutively in video.In addition,the differences in individual postures and movements have brought great challenges to the existing person re-identification algorithms.This research has proposed a pedestrian image feature extraction method based on the deep learning and triplet neural network.The proposed method combines the global pedestrian image feature with the local pedestrian attribute feature,improving the identification accuracy of pedestrian.And this method can be applied to the continuous track of specific people in surveillance video.In addition,the data enhancement of the pedestrian image is performed by using the Generative Adversarial Networks(GAN)to improve the generalization ability of the pedestrian feature extraction model.The experimental results on the Market1501 dataset show that this method can achieve better pedestrian recognition than the traditional methods.To solve the problem of crowd density estimation in parse and dense populations,this thesis introduces the attention mechanism to capture the global pedestrian feature and utilizes the image local classification to acquire the perspective view.Then,by combining the pedestrian category response map and the image block regression network,the proposed method can effectively solve the problem of pedestrian density estimation under the situation of small pedestrian pixel area and large pedestrian occlusion.A satisfactory effect has been obtained in the real street scene and community environment application.Furthermore,aiming at the problem that the existing violence recognition algorithms can not support the real-time calculation,an online violent behavior recognition algorithm based on deep learning has been established,which also improves the detection accuracy of abnormal behaviors in the complex environment.The proposed algorithm combines the precision of 3D convolution and the high efficiency of 2D convolution to realize fast and accurate violent behavior recognition in real-time video data.The result of actual running demonstrates that the speed and accuracy of the proposed algorithm satisfy the practical application in real-time security system.At last,an efficient end-to-end fire detection algorithm framework based on deep learning has been constructed,which provides technical support for fire detection in security systems.Generally,the fire detection in real surveillance video has some problems.For example,the fire area in the security camera is relatively small,the light and other fire-like objects affect the fire recognition,and the smoke movement characteristics are more than the color characteristics.To solve these problems,the proposed algorithm framework fuses the flame category response areas,image static feature and smoke optical flow feature,and its application on real-world data illustrates the validity in the security system.