自动驾驶场景中,通常会用基于体素化的算法来完成点云3D目标检测任务,因为该类方法拥有计算量少、耗时少等方面的优势。但是当下常用的方法往往会带来双重信息损失,其一是体素化带来的量化误差造成的,其二则是对体素化后的点云信息利用不充分造成的。设计一个三阶段的网络结构来解决信息损失大的问题。第一阶段使用基于体素化的优秀算法完成输出边界框的任务;第二阶段利用一阶段特征图上的信息精修边界框,以解决一阶段对输入信息利用不充分的问题;第三阶段利用了原始点的精确位置信息再次精修边界框,以弥补体素化带来的点云信息损失。在Waymo Open Dataset上,所提多阶段3D目标检测算法的检测精度超过了CenterPoint等受工业界青睐的优秀算法,且满足自动驾驶落地的时间要求。
Three-Dimensional Object Detection Based on Multistage Information Enhancement in Point Clouds
Voxel-based method is usually used in autonomous driving when conducting three-dimensional(3D)object detection based on a point cloud.This method is associated with small computational complexity and small latency.However,the current algorithms used in the industry often result in double information loss.Voxelization can bring information loss of point cloud.In addition,these algorithms do not entirely utilize the point cloud information after voxelization.Thus,this study designs a three-stage network to solve the problem of large information loss.In the first stage,an excellent voxel-based algorithm is used to output the proposal bounding box.In the second stage,the information on the feature map associated with the proposal is used to refine the bounding box,which aims to solve the problem of insufficient information utilization.The third stage uses the precise location of the original points,which make up for the information loss caused by voxelization.On the Waymo Open Dataset,the detection accuracy of the proposed multistage 3D object detection method is better than CenterPoint and other excellent algorithms favored by the industry.Meanwhile,it meets the requirement of latency for autonomous driving.
machine visionthree-dimensional object detectionlaser point cloudmultistageinformation enhancement