| With the continuous increase of car ownership,the complex traffic environment leads to tremendous traffic pressure.Issue of traffic safety cannot be ignored.Trajectories of road users with high precision and instantaneity are critical for traffic congestion reduction,traffic safety enhancement,and traffic efficiency ensurance.Present methods focus more on multiple target tracking utilizing one sensor,which could cause issues of low accuracy and poor robustness.This paper conducts a study on the temporal and spatial registration of roadside LiDAR and camera,and investigated the effect of different deployment postures on the registration results,which provide basic for the trajectory fusion of roadside LiDAR and camera.Based on multiple evaluation indicators,the performance impact of four attention mechanisms on 2D and 3D target detection algorithms is systematically investigated,algorithms of target detection with higher accuracy and precision are compared and selected as the support of trajectory tracking.Considering distance decay characteristic of point cloud quality(quantity,density,etc.),trajectory fusion algorithm is proposed based on adaptive weight coefficient,realizing the fusion of multi-scale trajectories in the same space-time and making up for the deficiency of single sensor in trajectory tracking.According to the research results,the main conclusions obtained are as follows:(1)The multi-source data-collecting platform was designed and established,and the time registration method based on point cloud and image data timestamp was proposed.After time registration,the time difference between lidar and camera data at the same time was less than 0.1s.Based on the plane calibration method of Zhang Zhengyou and the plane target model,the internal and external parameters of roadside lidar and camera can be acquired.By comparing and analyzing the reprojection error and effect of different spacing between horizontal and vertical layout,the appropriate spacing is obtained.(2)It can be concluded from the experiment of target detection algorithm that adding attention mechanism can effectively improve the performance of 2D or 3D target detection model.Tests on multi-scale target detection algorithm with four attention mechanism show that accuracy of detection was improved by 4.48%through adding SELayer module compared with original algorithm.In terms of 3D target detection algorithm,the improvement effect of the addition of attention mechanism is not good when the target is "car".As for target "pedestrian",the detection algorithm improves in the four performance evaluation indicators after adding attention module,the average increase is 1%as the Ecalayer module is added.Detection performance of"non-motor vehicle" is improved by 8%in the four evaluation indicators when the addition is Ecalayer module and SELayer module.(3)By introducing the distance decay property of point cloud mass,a target trajectory tracking method based on adaptive weight coefficient is proposed.Tests on trajectory tracking algorithm based on adaptive weight coefficient indicate this algorithm can accurately identify the speed of target with calculation error no more than 10%in the situation of car moving at constant speeds.As for robustness,the proposed tracking algorithm perform well under 5 different real scenarios.Compared with the single 2D or 3D trajectory tracking algorithm,the proposed method can effectively improve the range of trajectory tracking.For vehicle,non-motor vehicle and pedestrian,this method can effectively alleviate the phenomenon of target loss,the repair rate of disconnection track is 12.9%、40%and 38.24%respectively. |
