Research On Vehicle Detection Method Based On Fusion Of LiDAR And Camera

Autonomous vehicles are the main direction of automobile development in the future,and vehicle detection is the key technology to ensure their safe and stable driving.Multi-sensor fusion technology is an important way to achieve stable,fast,and accurate vehicle detection.However,the existing detection methods have the problems of poor robustness,incompatible detection speed and detection accuracy.In this context,based on the national key research and development plan(SQ2018YFB16003105),this paper studies the decision-level fusion strategy of Li DAR and camera,and proposes a multi-adaptive,high real-time,and robust vehicle detection method,which consists of three parts: data matching,real-time detection and decision-level fusion.The main research contents are as follows:(1)Deep fusion of information at the data level of Li DAR and camera is realized.Based on the analysis of the working principle and geometric relationship of the two sensors,an alignment fusion model of the three-dimensional laser point clouds and the two-dimensional image is established,and a sparse depth map is obtained.Then through the depth completion algorithm,the two-dimensional sparse depth map is converted into a two-dimensional dense depth map,which solves the sparsity problem of the laser point cloud data,so that the laser point cloud data and color image have the same resolution.Aiming at the single adaptability problem of the existing deep completion methods,a multi-adaptive depth completion method is proposed,which can change the completion strategy according to the day and night environment,and provide higher precision data for the subsequent stage.(2)The vehicle detection algorithm based on YOLOv3 is studied.Based on the analysis of the typical structure of convolutional neural network,the YOLO series real-time detection algorithm is studied.Combined with the high real-time requirement of autonomous vehicles,YOLOv3 algorithm with consideration of detection speed and detection accuracy is selected for real-time detection of color images and dense depth maps.(3)A decision-level fusion method is proposed to achieve robust fusion of Li DAR and camera.First,the bounding box fusion method is designed to fuse the bounding boxes of the vehicle detected by the color image and the dense depth map,and obtain the candidate regions of the vehicle.Then through the improved D-S evidence theory based on the evidence distance,the confidence score fusion is further realized,and the accuracy of candidate areas is analyzed,thus the final detection results are obtained.Finally,according to the vehicle position information provided by the fusion results and the depth information provided by Li DAR,the distance between the vehicles is obtained.Through the KITTI dataset and the Waymo Open dataset,the proposed method is tested in daytime and nighttime driving environments respectively.The experimental results show that the fusion detection method proposed in this paper can accurately obtain the location and distance of the target vehicle,and has high real-time and robustness.In addition,it is verified that the depth completion method proposed in this paper has high accuracy and strong practicability.