Vehicle Micro-Trajectory Automatic Acquisition Method Based on Multisensor Fusion

A vehicle micro-trajectory automatic acquisition method based on multisensor fusion is proposed to enhance the accuracy of trajectory sensed by traffic monitoring equipment while minimizing equipment expenditures. First, the YOLOv5 + DeepSORT vehicle recognition and tracking model is applied to obtain the video’s pixel trajectories of moving vehicles. Second, calibrate the camera’s internal parameters by calibration method with a checkerboard grid and obtain the conversion formula between pixel and camera coordinate system. Third, calibrate the camera’s external parameters by matching the feature point information of the diffuse calibration plate both in light detection and ranging (LiDAR) and camera, and determine the positional relationship between the LiDAR and camera coordinate system. Then, to convert pixel trajectory to real trajectory, we construct the conversion formula from the pixel coordinate system to the LiDAR coordinate system by combining the internal and external parameters. Finally, the kernel regression method based on the Gaussian function reduces inaccuracies caused by camera vibration and coordinate transformation. Also, a trajectory fusion method based on Kalman filtering with dynamically assigned weight coefficients is proposed. It uses multiple cameras to collect data simultaneously and effectively fuse them to improve the trajectory integrity and continuity further while improving the accuracy and reliability of vehicle tracking. Experiments are conducted on three scenarios to validate the proposed method: a straight road section, a curving road section, and an intersection. The experimental results were compared with the vehicle trajectories retrieved from LiDAR data. The results indicate that the average absolute error of the optimized trajectory under monocular vision is reduced by about 40%, and the average absolute error of the velocity is reduced by about 45% compared with that before the optimization. Compared with the real trajectory, the error of the calibrated trajectory is within 0.2 m. The fused trajectory under binocular vision can fill in more than 85% of the missing trajectory points. Compared with monocular vision, the average absolute error of the trajectory is reduced by more than 0.03 m, and the velocity error index is further improved. This method effectively improves the accuracy of acquiring vehicle trajectories through traffic monitoring videos.