Research On Vehicle Trajectory Extraction Method Based On UAV Aerial Video

Vehicle trajectory data is a sequence of positions formed by vehicles traveling in a certain time period or road section,which contains a wealth of information such as traffic data and vehicle driving behavior.It has played a supporting role in the establishment of traffic flow theory models,vehicle driving behavior,traffic conflict points and other models.In recent years,UAV aerial video has replaced ordinary camera video in traffic video collection,which has the advantages of convenience,efficiency,and wide shooting angle of view.UAV aerial video can record a large number of vehicle trajectories on the road.If all vehicles are manually marked to extract their trajectories,it is not only inefficient,but also prone to missing vehicles.This paper aims to use computer vision instead of human vision to extract vehicle trajectories from aerial video.The main research results are as follows.(1)Aiming at the problem of the image shifting due to the jitter caused by the wind and its own PTZ during the aerial video process of the UAV.This paper adopted SIFT features and RANSAC algorithm for image registration,took the first frame as the reference image and aligned the rest of the video sequence frames,gave the singleresponse matrix for each frame.(2)Using bilateral filters and morphology to solved the problems of blurred contours and noise pollution that can exist in traditional motion target detection using background elimination method.For the problem of pseudo-target in the images,this paper proposed the eight-neighborhood contour tracking method to obtain the foreground object contour and removed the pseudo-target according to the physical characteristics of the model.(3)Applying the principle of homography matrix transformation to the reference image,the problem of deploying detectors in the video image is studied.The center point of the fitted contour of the vehicle falls in the detector to determine the detected vehicle,and the vehicle contour coordinates obtained by it are added to the tracking queue list.(4)In order to get the motion trajectories of all vehicles,the KCF algorithm that fuses FHOG and CN features was used to track the trajectories of the vehicles in the tracking queue list.And the coordinates of the fitted contour center points of the tracked vehicles in each frame were connected to obtain the current tracked vehicle trajectory,and this process was repeated to obtain all vehicle trajectories.Finally,for the repeated vehicle trajectories,the LCSS algorithm was used to delete the duplicates.(5)Using the above set of process methods to conduct experiments on the flat and peak videos taken by UAV aerial photography,the effectiveness of the algorithm in this paper is verified.Successfully extracted the trajectories of vehicles on the roads at various times and obtained three kinds of traffic information such as traffic volume,speed,and acceleration.Experiments show that the success rate of vehicle trajectory extraction during flat periods reaches 98.8%,and the success rate of vehicle trajectory extraction during peak periods reaches 98.3%.At the same time,the experiments also prove that the vehicle trajectories of different vehicle models can be extracted well.In summary,the above-mentioned vehicle trajectory extraction method based on UAV aerial photography and computer vision recognition proposed in this paper can extract vehicle trajectories and traffic parameters such as traffic volume,speed and acceleration well,also can provide technical support for other value-added traffic applications in the follow-up.