基于改进PointPillars的点云车辆目标检测

Point cloud vehicle target detection based on improved PointPillars

针对自动驾驶领域中,激光雷达获取的点云数据存在稀疏性以及边缘噪点误检等问题,提出一种基于改进PointPil-lars的点云车辆目标检测方法。首先,基于SimAM注意力机制改进体素化特征输入,使得网络特征提取阶段能更加关注关键信息,提高特征学习的全局性。其次,基于卷积块注意力模块(CBAM)改进骨干网络结构,提出全新的轻量化通道注意力模块Tiny-CAM和可变形空间注意力模块Deformable-SAM,构建Multi-CBAM骨干网络,提升网络特征提取及特征融合能力。在KITTI数据集以及非公开车库点云数据集上进行验证,实验结果表明,与原网络相比,改进PointPillars方法具有更高的检测精度,平均检测精度提升2.98%,针对遮挡小于30%的点云车辆目标检测精度提升6.51%,证明了该方法的有效性。

In the field of autonomous driving,the point cloud data obtained by lidar has problems such as sparsity and edge noise false detection.This paper proposes a point cloud vehicle target detection method based on improved PointPillars.Firstly,the voxelized feature input is improved based on the SimAM attention mechanism,so that the network feature extraction stage can pay more attention to the key information and improve the globality of feature learning.Secondly,based on the improved backbone network structure of CBAM,a new lightweight channel attention module Tiny-CAM and a deformable spatial attention module Deformable-SAM are proposed to construct the Multi-CBAM backbone network and improve the network feature extraction and feature fusion ability.The KITTI data set and the non-public garage point cloud data set are verified.The experimental results show that the method adopted in this paper has higher detection accuracy.Compared with the original network,the average detection accuracy is improved by 2.98%,and the detection accuracy of point cloud vehicle targets with occlusion less than 30%is improved by 6.51%,which proves the effectiveness of the method.