Association of Vehicle Object Detection and the Time-space Trajectory Matching from Aerial Videos

High resolution track data contains rich information about vehicle travel and traffic flow. The fusion method of cross-frame vehicle detection association and trajectory matching is developed to extract the vehicle trajectories from the aerial video. The convolutional neural network, YOLOv5, is used to obtain video-wide vehicle object detection. Base on the result of detection, a correlation algorithm of a cross-frame target under the constraints of vehicle dynamics and trajectory confidence is proposed. Then, broken track matching and constructing algorithms based on the maximum correlation are established for identifying unique vehicles. The trajectory is converted from image coordinates to Freenet coordinates under lane reference, and the ensemble empirical mode decomposition(EEMD)model has been constructed to eliminate data noise. Two sets of open-source aerial videos, coving congestion and free-flow traffic status, are taken by a drone on the Nanjing expressway to test the effect of the trajectory extraction algorithm. The results show that the trajectory accuracies are 98.86 and 98.83% under the free flow and congested conditions, respectively. Besides, the track recall rates are 93.00 and 86.69%. The trajectory extraction speed of the algorithm is 0.07 s/vehicle/m. The vehicle trajectory dataset processed by this method can provide extensive data support for traffic flow, traffic safety, and traffic control research. The dataset is published at http://seutraffic.com/.