Method And Realization Of Vehicle Trajectory Features Recognition From Aerial Videos

Urban population and car ownership are increasing year by year,and traffic safety problems such as road accidents and vehicle congestion remain serious.The urban traffic congestion index increases year by year,and the possibility of traffic accidents increases under the condition of traffic congestion.The characteristics of urban road traffic flow and vehicle movement are rich in information,and the collection of real and reliable urban road vehicle trajectory data plays an important role in the study of traffic safety and traffic flow theory in road sections.With the rapid development of machine vision and deep learning technology in recent years,how to identify vehicle trajectory and traffic flow feature data more efficiently and accurately has become a research hotspot and technical difficulty.Most existing vehicle trajectory data collection schemes use fixed mounted cameras to collect video information.Due to the limitations of shooting Angle and position,the accuracy of vehicle track data extracted by this scheme is limited.However,unmanned aerial photography equipment can collect vehicle driving picture data from the perspective of altitude,and vehicle trajectory extracted based on this data can greatly improve accuracy.In this paper,UAV is firstly used to collect road vehicle driving videos from a high altitude perspective,and then initial track data of road vehicles are extracted based on video stabilization,deep learning target detection and multi-target tracking technologies,including various motion characteristic data of vehicles,such as: Vehicle speed,acceleration,lane number,front vehicle ID,rear vehicle ID,space headway,headway and TTC have eight vehicle motion feature data,which have important supporting role in the study of driving behavior and traffic flow theory,and can be applied in road traffic safety and traffic flow theory related research.The main contents of this paper are as follows:(1)Unmanned aerial photography equipment was used to take pictures of vehicle driving in different sections of urban roads,and video data of vehicle driving in three typical sections of straight section,curved section and confluent section were collected respectively.In order to eliminate the jitter problem in video shooting,image feature point feature vector is extracted based on SURF feature point detection algorithm,and RANSAC matching algorithm is used to match the reference image and the image to be matched,so as to realize image stabilization of aerial video.(2)Deep learning target detection and multi-target tracking technology are used to identify and track vehicle targets in aerial video.On the basis of Vis Drone image data set,high-altitude image training set is made by using video frame images shot in this paper,and YOLOv5 model is trained.The results show that the recognition accuracy of the model m AP@0.5 index is 99.1% after training,which is 49.1% higher than that of the model without adding self-made image set.Then,based on Kalman filter and Hungarian matching algorithm,the vehicle target in video continuous frame image is matched and associated,and the initial vehicle trajectory data of video continuous frame is obtained.(3)The initial trajectory data of vehicles in different sections are preprocessed to reduce the system error.The vehicle speed,acceleration,lane number,front vehicle ID,rear vehicle ID,headway,headway and TTC motion characteristic data were extracted under different road sections.The proposed automatic lane number detection algorithm can extract vehicle lane number data,and it is verified that the detection accuracy of lane line position in straight sections is 94.94%,and that in curved sections is 90.25%.Then,the traffic flow data of different sections were analyzed based on the vehicle movement characteristic data,and the lane changing and following behavior characteristics of different sections were analyzed.Finally,the collected vehicle trajectory data is compared with the existing vehicle trajectory data,and the results show that the vehicle trajectory data based on the proposed algorithm has higher accuracy.To sum up,the research on urban road vehicle track data collection and motion feature data extraction carried out in this paper can provide data basis for road traffic safety and traffic flow theory and other studies,and has important practical significance for studying vehicle characteristics of urban road traffic and solving urban road traffic problems.