航拍视频车辆检测目标关联与时空轨迹匹配

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

高解析度轨迹数据蕴含丰富车辆行驶与交通流时空信息。为从航拍视频中提取车辆轨迹,构建了车辆检测目标跨帧关联与轨迹匹配融合方法。采用卷积神经网络YOLOv5构建视频全域车辆目标检测,提出车辆动力学与轨迹置信度约束下跨帧目标关联算法,建立了基于最大相关性的断续轨迹匹配与融合构建算法,实现轨迹车辆唯一编号。将轨迹从图像坐标转换为车道基准下Frenet坐标,构建集合经验模态分解(EEMD)模型进行轨迹数据噪声消除。采用南京市快速路无人机拍摄的2组开源航拍视频,涵盖拥堵与自由流交通状态,对轨迹提取算法进行效果测试。结果表明,在自由流和拥挤条件下轨迹准确率分别为98.86%和98.83%,轨迹召回率为93.00%和86.69%,构建算法的轨迹提取速度为0.07 s/辆/m。该方法处理得到的详细车辆时空轨迹信息能为交通流、交通安全、交通管控研究提供广泛的数据支撑,数据公开于http://seutraffic.com/。

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/.