Research On Cascaded 3D Vehicle Detection Method Based On Local Aware Space Aggregation

Given the swift advancement of autonomous driving technology,lidar is widely used in auxiliary automatic driving systems,Therefore,the 3D detection technology based on laser point cloud is particularly important for the development of autonomous driving technology.However,in complex environments,due to the existence of problems such as laser point cloud sparsity,disorder,background clutter,and occlusion,the performance of 3D detection based on laser point cloud is seriously affected.After studying the development history and recent achievements of 3D detection,this paper designs a cascaded 3D Vehicle Detection Framework(C3DVD).The main research content of this paper is as follows:(1)In view of the inherent characteristics of the detection equipment and complex environments that easily cause point cloud noise and redundant point clouds,and the traditional point cloud simplification algorithm ignores the inherent characteristics of the point cloud,a combination of DBSCAN and point cloud information entropy is proposed.point cloud reduction algorithm.A KNN-based point cloud cluster density estimation method is designed,and Shannon entropy is used to quantify point cloud cluster information.Based on the global information quantification results,noise data and redundant point clouds are filtered out by setting an adaptive threshold.The global error and local error of the simplified results are evaluated by the average Euclidean distance and Hausdorff distance before and after simplification.In the comparative experiment with other types of simplification methods,it is found that the point cloud simplification algorithm combined with DBSCAN and point cloud information entropy can retain the details and shape features of the original point cloud on the basis of effectively filtering out redundant point clouds.(2)Aiming at the problems of laser point cloud sparsity,disorder,and unbalanced number of foreground and background points,a class agnostic instance segmentation method based on foreground points(Class Agnostic Instance Segmentation Method Based on Foreground Points,CAISM)is proposed..First,with Point Net++ as the backbone network,a global attention module is embedded in the point cloud encoder and decoder modules,point cloud features are extracted and feature masks at different feature scales are constructed to realize the weight adjustment of global features and global attention The mechanism works in two parts of point cloud encoding and decoding respectively.The former is used to strengthen the weight of interest points,and the latter guides feature restoration according to the adjusted weight,and finally obtains a point-by-point feature map.Second,in order to accomplish the tasks of foreground segmentation and coarse object localization simultaneously,a "classification +regression" dual-task head structure is designed,and the point-wise feature maps will be shared between the foreground segmentation and instance-centered regression tasks.Then,the instance center coordinates of the regression output are used as the initial cluster center,combined with the improved K-means clustering to realize the instance segmentation of the foreground point.Finally,an experiment is designed to compare and analyze the segmentation performance of CAISM,Point RCNN and Squeeze Seg.CAISM shows high segmentation and positioning performance,and is better than Point RCNN and Squeeze Seg algorithms.(3)Aiming at the problems of high computing power and poor real-time performance of existing 3D bounding box generation algorithms,a 3D bounding box generation method based on dynamic adaptive Sailfish optimizer is proposed on the basis of instance segmentation completed by CAISM.First,aiming at the slow convergence speed of the original sailfish optimizer,a dynamic adaptive sailfish optimizer is proposed in combination with the dynamic population mixing strategy,which improves the convergence speed of the original algorithm;secondly,according to the instance center and foreground point space under the BEV view Features generate a large number of 2D pre-selected bounding boxes;then optimize the generated pre-selected boxes through the dynamic adaptive sailfish optimizer,output 2D bounding boxes,restore the 2D bounding boxes under the BEV view to 3D according to prior knowledge,and generate the final 3D Bounding box;finally,combine the instance point cloud to output the 3D detection result.The comparison experiment with various 3D object detection algorithms shows that C3 DVD can further improve the accuracy of 3D vehicle detection under the premise of satisfying real-time performance.