Research On Key Technologies Of Vehicle Detection And Tracking And Behavior Analysis In Complex Scenes

Intelligent vehicle detection,tracking and behavior analysis are of great significance to improve the comprehensive management level of urban intelligent transportation,solving traffic safety problems,and implementing traffic violation control.At this stage,due to the influence of occlusion,deformation,congestion,and complicated weather,video-based automatic vehicle driving state detection technology has many problems,such as poor detection robustness,broken tracking trajectory,slow behavior recognition and high error rate.In response to this problem,this thesis focuses on abnormal event detection and fine-grained vehicle behavior recognition in complex traffic scenes,studies the vehicle detection using pyramidal multi-scale calculation method and multi-dimensional vehicle tracking using dynamic and static feature matching method.Then,calculates the relevant traffic parameters based on the results of vehicle detection and tracking,and test the effectiveness,accuracy and real-time performance of our proposed methods in complex scenes.The main research content includes the following aspects:(1)Constructed a roadside camera-based vehicle detection data set and a vehicle-mounted camera-based vehicle detection data set in complex traffic scene.Aiming at the defects of the existing labeling tools with single category labeling,this thesis designs a set of multi-functional labeling software with the function of batch label modification,target quantity statistics,label format conversion and automatic annotation.Meanwhile,a standardized annotation rule is formulated for vehicle annotation in complex scenes,that is,multiple rules,no interoperability,and no reuse.After that,relying on roadside cameras and vehicle-mounted cameras,we have completed a large number of sample collections and annotation for complex scenes such as rain,fog,congestion,and night,and constructed a data set based on roadside cameras and a data set based on vehicle-mounted cameras.The two data sets we constructed can solve the problems of sparse sample size of complex scenes and inconsistent annotation rules in the existing vehicle detection data sets,provide data support for different traffic application requirements,and provide effective data guarantee for testing the impact of complex scenes on vehicle detection performance.(2)Proposed a V-SSD vehicle detection method based on the improvement of Single Shot Multi-Box Detector(SSD)using multi-feature fusion and pyramid multi-scale calculation method.Aiming at the defects of the SSD method,such as feature dispersion,loss of feature information,and poor detection robustness,a multi-layer feature fusion method is proposed to improve the network structure of SSD,which greatly improves the feature learning ability of SSD for small targets and reduces the feature loss in the process of deep learning.On the other hand,a pyramid multi-scale calculation method based on enumeration is proposed to improve the detection algorithm of SSD,which improves the robustness of the SSD detection algorithm.The experimental results show that the V-SSD method based on the improvement of SSD’s network structure and detection algorithm greatly improves the defects of SSD,and has excellent comprehensive performance under multiple evaluation indicators such as accuracy and speed.Furthermore,experiments in complex traffic scenes also show that the detection speed of V-SSD method is slightly lower than SSD and YOLOv3(YOLOv3: an incremental improvement)methods.However,the detection accuracy of V-SSD method is more than 93%,and it can still ensure the real-time requirements in practical application,and can maintain a stable vehicle detection result in complex traffic scenes.(3)Proposed a multi-dimensional vehicle tracking algorithm using dynamic and static feature matching method.Aiming at the problems of trajectory loss,fracture and jump caused by occlusion,deformation and other factors,a multi-dimensional vehicle tracking used the dynamic and static feature matching method is proposed,which based on three dimensions of dynamic displacement,static category and dynamic rotation angle during vehicle driving.The algorithm mainly designs position correction method of the key point for vehicle tracking,the vehicle re-identification method based on multi-dimensional constraints,and the linear optimization method for trajectory.Among them,the position correction method of the key point for vehicle tracking fully considers the expression ability of the vehicle’s bounding box to the vehicle position,and proposes an offset correction algorithm for the center point of the bottom edge of the vehicle’s bounding box to obtain the best vehicle tracking point position.The vehicle re-identification algorithm based on multi-dimensional constraints puts forward the minimum Euclidean distance,the maximum matching area and trajectory angle constraints that meet the threshold,which successfully filters the false trajectory points,and effectively suppresses the problem of vehicle missing and error tracking caused by occlusion and deformation.The linear optimization of the tracking trajectory focuses on the smoothing of the trajectory in order to describe the driving state of the vehicle better.Finally,the experiments in complex traffic scenarios show that our vehicle tracking algorithm has the advantages of less calculation,high accuracy,and stable tracking performance.(4)From two levels of precise classification and rapid detection,a fine-grained identification method and a multi-factor comprehensive decision model for vehicle abnormal behavior are proposed.Combined with the information of vehicle position,category and trajectory obtained by detection and tracking,the fine-grained identification of parking,speeding,retrograde,lane change,and congestion is designed and realized.Aiming at the slow recognition speed caused by the large amount of calculation in the fine-grained recognition method,two different multifactor comprehensive decision-making models of K-nearest neighbor and random forest are designed to realize the rapid detection of whether the vehicle behavior is normal or not.The experimental results show that the fine-grained vehicle abnormal behavior recognition method is suitable for precise classification that does not pay attention to the speed,and the multi-factor comprehensive decision model only determines whether the vehicle behavior is abnormal,so it is more valuable for traffic applications that focus on detecting speed without emphasizing abnormal categories.In summary,this thesis focuses on the key technologies of vehicle detection and tracking and behavior analysis in complex traffic scenes,and puts forward solutions to the false warning problems caused by occlusion,illumination,congestion and complex weather conditions,which has a certain reference value for intelligent traffic management.