Research On Traffic Subject Detection Method Based On Fusion Of Image And Point Cloud

With the development and application of target detection technology,the detection methods of traffic subject such as vehicles and pedestrians are playing a crucial role in the field of autonomous driving.Aiming at the problems of slow speed,low accuracy and distance measurement precision of current traffic subject target detection,a traffic subject detection method based on deep learning image and point cloud fusion is proposed based on the project "Holographic Traffic State Reconstruction and Vehicle Cluster Cooperative Control Testing and Verification"(2018YFB1600605)of the National Key Research and Development Program.The main contents are as follows:(1)Aiming at the problem that the resolution of the sparse depth map obtained from the conversion of the original point cloud of Li DAR is much lower than that of the color image,a depth-completion method relying only on the point cloud information is proposed.Firstly,the sparse depth map is obtained by constructing a spatial alignment model of the laser point cloud and the color image.Then,considering that the depth-completion method based on image guidance will bring in noise points in the case of poor image imaging quality,this paper adopts an improved bilateral filter to complete the sparse depth map into a dense depth map while relying only on the point cloud depth information.The experimental results show that the method can significantly improve the density of sparse point clouds and achieve data matching of images and point clouds.(2)Aiming at the problem of slow speed and low accuracy of traffic subject detection in autonomous driving,YOLOv5 real-time target detection algorithm is used to detect vehicles,pedestrians,etc.Firstly,the basic principle of YOLOv5 model is analyzed.Then,it is applied to color images and dense depth maps for vehicle,pedestrian and other traffic subject detection respectively.The experimental results show that the YOLOv5 algorithm meets the requirements of fast speed and high accuracy for traffic subject detection.(3)Aiming at the problem of low confidence rate of traffic subject detection and large distance measurement error in some scenarios,a decision layer fusion scheme is proposed.Firstly,by analyzing the low confidence rate in the scenario of insufficient image detection,DS evidence theory is used to achieve the fusion of color image confidence rate and dense depth map confidence rate.Then,a joint ranging method is proposed by analyzing the shortage of point cloud ranging and monocular ranging.The experimental results show that the confidence rate fusion method in this paper can effectively improve the confidence rate of the traffic subject target;the average error of the distance of the joint ranging method is 0.26 m within 20 m from the measured target to the sensor,and the absolute errors are all lower than 0.5m,which makes up for the shortage of point cloud ranging in the cases of no point cloud in the detection frame and interference point cloud in the detection frame.