Research On 3D Point Cloud Object Detection Algorithm For Driverless Vehicle

With the continuous development of driverless technology,the market size of the self-driving car industry is growing rapidly.The system of unmanned driving mainly consists of three core technologies:environment perception,behavior decision and motion control,of which the perception system enables unmanned vehicles to understand the road environment like a driver through on-board sensors,so as to make prediction and planning.On-board sensors mainly include cameras,millimeter wave radar,LIDAR and ultrasonic radar,etc.Meanwhile,high-speed chips such as TPU and GPU also play an important role in the driverless field.Unmanned driving is the ultimate manifestation of intelligence,which concentrates on the use of computer,artificial intelligence and automatic control technologies.In this paper,based on the point cloud data collected by 3D Li DAR,the target recognition task of obstacles is accomplished,which mainly includes ground segmentation,after getting the non-ground points,the obstacle targets are extracted by clustering algorithm,and finally the obstacle recognition task is realized by using the point cloud classification network of deep learning.The main work and contributions of this paper are as follows:For the phenomenon of under-segmentation in existing ground segmentation methods,this paper combines the index information of each scan line in the LIDAR collected point cloud data,the index information of each scan line is processed separately for each scan line according to the geometric features of the road surface,and the extraction and filtering of the ground point cloud is carried out.After removing the ground points,we use the DBSCAN algorithm combined with KD-Tree to cluster the non-ground After removing the ground points,we use the DBSCAN algorithm combined with KD-Tree to cluster the non-ground points,so that we can get the target clustering results more quickly and accurately while removing the noise points.In response to the existing methods that do not consider the importance difference of feature channels,we propose the ADGCNN network combining attention mechanism,introduce channel attention module in Edge Conv structure,assign different weights to the channels to improve the feature expression ability of the network,and train on the actual dataset to obtain the point cloud classifier.