Research And Implementation Of Point Cloud Data Fusion Method Based On Edge Computing Of Internet Of Vehicles

With the rapid development of deep learning,Internet of Vehicles edge computing and 5G communication technologies,autonomous driving is becoming a key technology affecting the automotive industry in the next decade.Sensors are the medium through which the autonomous driving system perceives the external environment.The synergy and complementarity between sensors are the key to further improving the safety of autonomous driving.In recent years,multi-sensor fusion technology has received more and more attention,and the application of point cloud fusion methods in three-dimensional target detection and map construction has also increased.The Internet of Vehicles edge computing technology aims to give more powerful information processing capabilities and content delivery capabilities to vehicle edge nodes,and to provide a more efficient and low-latency service support platform for in-vehicle applications.With the help of Io V edge computing technology,on the one hand,it can make up for the long time period and low accuracy of point cloud fusion caused by insufficient computing power on the vehicle side;on the other hand,by interacting with the nearby edge server,the data of the vehicle node itself can be further integrated.And processing to expand the perception range of a single vehicle.The current mainstream 3D laser point cloud target detection methods are mainly divided into deep learning methods and traditional laser point cloud methods.Aiming at the problems of traditional laser sensing methods such as poor clustering generalization and loss of surrounding environment information,this paper combines two-dimensional image target detection based on laser point cloud perception,and designs a three-dimensional target detection scheme and process based on point cloud fusion.The overall process is mainly divided into point cloud filtering and segmentation,point cloud projection clustering,and target tracking estimation.Compared with traditional laser perception methods,this paper uses target detection to obtain the position of the target object in the image,and project the point cloud into the picture through coordinate transformation to obtain the point cloud in the corresponding ROI area,so as to obtain the point cloud information of the corresponding target object.To achieve the purpose of clustering.Finally,the target is tracked and motion estimation is performed by means of image key point extraction and matching and extended Kalman filtering.Aiming at the problems of large errors and low accuracy of pure laser odometer mapping,this paper proposes a SLAM map construction scheme based on point cloud fusion,which integrates sensor information such as IMU and GPS,and combines graph optimization algorithms and closed-loop detection of Scan Context environmental descriptors.The technology constrains and optimizes map construction to achieve the purpose of reducing errors and improving accuracy.The overall process mainly includes data preprocessing,frontend odometer estimation,back-end graph optimization,and closed-loop detection.Finally,in view of the current situation of long single-vehicle lidar mapping time period and large computing power requirements,this method proposes a multi-vehicle mapping positioning scheme based on edge-cloud collaboration,which realizes lidar positioning services based on edge computing,and has many verifications.The feasibility of combining sub-maps between cars in the cloud.