Research On LiDAR-Camera Based Perception Fusion And System Development And Vehicle Field Testing

With the advancement of science and technology,autonomous driving technology has gradually become one of the important directions of future transportation development.Autonomous vehicles can improve traffic safety,alleviate traffic congestion,and reduce carbon emissions.In the autonomous driving technology,environmental perception is the key to achieving the autonomous driving of the vehicle.The vehicle perceives the road conditions and the surrounding environment through various sensors to make corresponding driving decisions.However,the traditional single-modal perception system uses a single sensor for perception,and the information obtained is limited,and it is difficult to fully obtain real road condition information,thus affecting the safety of autonomous vehicles.In order to solve this problem,perception fusion technology has been proposed and widely used.This technology fuses the information obtained by multiple sensors to achieve information complementation and correction,reduce the errors and deficiencies in single sensor,and obtain a more comprehensive environmental perception information,thereby improving the driving safety and efficiency of autonomous vehicles.For the closed park scene of medium and low-speed roads,this paper proposes an obstacle recognition algorithm based on Li DAR and camera perception fusion,aiming to improve the environment perception ability of autonomous vehicles.The algorithm recognizes the 2D image information obtained by the camera and the 3D point cloud information obtained by the Li DAR,and outputs the single-modal perception obstacle list and the candidate frame list before postprocessing,and then uses the CLOCs post-fusion algorithm to fuse the 2D and the list of candidate boxes output by 3D object detection can output more accurate 3D prediction boxes.In this paper,the accuracy of object detection is improved by improving the object detection algorithm.First,by improving the loss function of the YOLOv3 algorithm to achieve a better regression performance,the accuracy of 2D object detection is improved,and the BN layer is merged into the convolutional layer to improve the convergence speed and inference speed of training.Secondly,the detection accuracy and detection speed are balanced by improving the point cloud sampling method of the CenterPoint algorithm,and a 3D CIoU loss function is proposed based on the CIoU idea to obtain a better regression performance.Finally,the dimensionality of the input feature tensor of the CLOCs algorithm is extended,and the convolutional layer is replaced by the residual block so that the network can better learn the probability correlation of fusion,and the accuracy of the post-perception fusion algorithm is improved.In this paper,an object detection and perception fusion system based on Li DAR and monocular camera is designed and implemented.The Li DAR and camera are installed and calibrated by retrofitting the real vehicle,and the perception system is deployed on the real vehicle and tested on the real vehicle.The perception system takes vehicles,pedestrians and cyclists as detection objects,obtains their pose and category information in three-dimensional space,and generates a fused obstacle list,which can improve the perception ability of a single vehicle.The experimental results show that compared with the original YOLOv3 algorithm,the average precision of the improved YOLOv3 is increased by 4.54%,and the detection speed is increased by8%.Compared with the original CenterPoint algorithm,the average precision of the improved CenterPoint algorithm is increased by 3.42%,and the detection speed is improved by 12.82%.Compared with the original CLOCs algorithm,the average accuracy of the improved CLOCs algorithm has increased by 3.5%.Compared with the single-modal perception algorithm,the perception fusion algorithm proposed in this paper effectively utilizes the advantages of different perception modalities,and the average precision is improved by 6.79%.The improved method proposed in this paper can be applied to the perception fusion of other object detection algorithms,which has certain universality and is of great significance for promoting the development of autonomous driving technology.