An Abnormal Behavior Detection Algorithm And System Based On Multi-target Tracking

With the continuous development of smart cities and smart transportation in China,intelligent video surveillance has attracted more and more attention.Intelligent video surveillance mainly includes the identification of people or objects,object tracking,individual state or scene state analysis and other tasks.Multi-object tracking(MOT)algorithm is the key to intelligent video surveillance.Therefore,this paper proposes a multi-object tracking algorithm based on deep hash feature.In addition,we propose an abnormal behavior detection algorithm based on multi-object tracking for Unmanned Aerial Vehicle(UAV)road monitoring scenarios,and develop a UAV road monitoring system.The main research work of this paper is described as follows:1.In order to solve the problem of ID Switch and tracking interruption caused by object occlusion or object interaction in MOT,this paper proposes a novel multi-object tracking algorithm by fusing deep hash appearance features and motion features for more robust feature representation.In addition,this paper designs a tracking interrupt recovery mechanism to improve the accuracy of data association.Inspired by the idea of person re-identification and hash representation for image retrieval,this paper connects the classic person re-identification network Res Net with a fully connected hash layer to construct a deep hash network,and then extracts deep hash appearance features of pedestrians through this network.In terms of data association,this paper improves the traditional multi-object tracking mechanism,and realizes the ID recovery after tracking interruption by saving multi-frame trajectory features.Evaluated by the multi-object tracking dataset MOT Challenge,the proposed multi-object tracking algorithm based on deep hash feature effectively reduces the number of ID Switches,and improves the tracking accuracy without changing object detection algorithm.2.In the scene of UAV road monitoring,this paper takes vehicles as the research target,and proposes an abnormal behavior detection algorithm based on sparse representation.Firstly,we utilize multi-object tracking algorithm to obtain trajectories of vehicles.Then we segment the trajectories,and calculate the speed and angle characteristics of vehicles based on the trajectory coordinates.After that,a Gaussian filtering algorithm is used to remove noise of speed features caused by external factors such as detection error or drone shake.Finally,we fuse the speed and angle features,and then construct a dictionary by the normal trajectory features.The dictionary reconstruction error is employed to determine whether the test trajectory sample is abnormal.Evaluated by the UAV video data set,the proposed abnormal behavior detection algorithm can effectively detect the abnormal behavior of vehicles with a high detection accuracy.3.Combining multi-object tracking and abnormal behavior detection algorithms,this paper designs and develops a UAV road monitoring system.The system adopts C / S architecture,including a visualization module,an object detection module,a multi-object tracking module,and a data query module.The modules cooperate with each other to realize the real-time object detection and tracking,vehicle flow statistic,real-time abnormal behavior detection and data query.