Detection Of Vehicle Abnormal Behavior Based On Multi-object Tracking

In order to reduce the probability of traffic accidents and improve the efficiency of traffic flow and law enforcement,vehicle abnormal behavior detection technology has been applied to the traffic video surveillance system.However,the ambiguous definition,unpredictable occurrence,and data sparsity make it difficult or low accuracy to detect vehicle abnormal behaviors.Based on the intelligent demand of traffic video surveillance system,the thesis designs a complete vehicle abnormal behavior detection system to solve the existing problems of low vehicle multi-target tracking accuracy,difficulty in vehicle abnormal behavior detection,and few realistic vehicle abnormal behavior detection applications.Researches are carried out around the design of vehicle multi-target tracking algorithm,the design of vehicle abnormal trajectory discrimination algorithm and the application and deployment of vehicle abnormal behavior detection algorithm.The thesis is mainly expanded from the following three aspects:1.An improved vehicle multi-target tracking algorithm based on CenterTrack is proposed.First,a small sample amplification method is used to increase the number of distant small vehicles to training samples.Then the multi-target tracking algorithm,CenterTrack,is retrained to locate the position of the vehicle in the image sequence and learns the center displacement of vehicles between adjacent frames.When the to-be-associated trajectories fail to match with the detection,the to-be-associated trajectories use the short-term trajectory memory method to predict its position information in the next frame through its historical motion information.At the same time,the to-be-associated trajectories associating with the to-be-associated detection are classified according to the association failure time to maintain the vehicle trajectory ID with less vehicle ID switch.Finally,in the five test scenarios,extracted from the traffic monitoring dataset UA-DETRAC,the improved method maintains the advantages of the CenterTrack algorithm,and achieves a nearly 20% improvement in the multi-target tracking accuracy where CenterTrack does not perform well.Compared with the YOLOv4-Deep Sort algorithm,the multi-target tracking accuracy is improved nearly10% in four test scenarios,and the effect is significant.2.A vehicle abnormal trajectory discrimination algorithm based on reconstruction error is proposed.Firstly,the normal vehicle trajectory in the traffic surveillance video is extracted,and a large amount of normal sub-trajectories data is obtained through trajectory data enhancement and trajectory segmentation.Then a vehicle trajectory reconstruction model combining long-short-term-memory network and variational autoencoder is constructed.The abnormal trajectory of the vehicle is discriminated by the difference of the reconstruction error generated by the normal and abnormal trajectory inputted of the trained reconstruction model.The experiment shows that the proposed algorithm has an accuracy of 90% above in the test dataset including illegal parking,road occupation and wrong-way driving and the false positive rate does not exceed 9%.The performance in the test scene is significantly better than the traditional isolation forest and support vector machine methods,and it also has great advantages compared with the variational encoder method based on full-connection network.3.A vehicle abnormal behavior detection algorithm is designed and deployed in the embedded platform Jetson TX2,and a vehicle abnormal behavior detection system for traffic video surveillance is implemented.The experiments tested by realistic traffic surveillance video shows that the algorithm deployed on the Jetson TX2 platform meets the needs of the accuracy in practical application scenarios,and has the potential to process traffic video in real time.In the realistic needs of the intelligent traffic video surveillance system,the thesis designs an improved vehicle multi-target tracking algorithm based on CenterTrack for trajectory extraction and an abnormality discrimination algorithm based on vehicle trajectory.The two algorithms are integrated to realize the vehicle abnormal behavior detection algorithm.Then the vehicle abnormal behavior detection algorithm is deployed in the embedded hardware platform Jetson TX2,and a vehicle abnormal behavior detection system for traffic video surveillance is built.Overall,the work of the thesis is in line with the expected research objectives,and has academic research significance and practical application value.