一种面向智能驾驶的单阶段点云三维目标检测算法  

A single-stage point cloud 3D object detection algorithm for intelligent driving

作　　者：汪世豪 邓涛 WANG Shihao;DENG Tao(School of Mechanotronics&Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China;School of Aeronautics,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing Key Laboratory of Green Aviation Energy and Power,Chongqing 400074,China;The Green Aerotechnics Research Institute of Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学机电与车辆工程学院,重庆400074 [2]重庆交通大学航空学院,重庆400074 [3]绿色航空能源动力重庆市重点实验室,重庆400074 [4]重庆交通大学绿色航空技术研究院,重庆400074

出　　处：《重庆理工大学学报(自然科学)》2025年第2期11-18,共8页Journal of Chongqing University of Technology:Natural Science

基　　金：国家自然科学基金项目(52275051);重庆交通大学自然科学类揭榜挂帅项目(XJ2023000701);重庆市研究生导师团队建设项目(JDDSTD2022007);重庆市研究生联合培养基地建设项目(JDLHPYJD2022001)。

摘　　要：为解决点云三维目标检测算法中单阶段网络与双阶段网络在检测精度和检测效率之间难以平衡的问题,提出一种基于点云体素的伪双阶段三维目标检测网络—AFPNet。构建轻量化的中心热图模块预测目标中心点,去除传统区域建议网络中的锚框设置与非极大值抑制操作。为更好地利用多尺度体素特征,设计了AFP-Cross-Attention模块,通过设计的自适应特征池化(AFP)方法,在多尺度体素中提取高价值特征并进行交叉注意力计算,与全局注意力相比有效减少了计算量。基于AFP方法搭建了AFP-Transformer检测头,建模查询特征与高价值特征的依赖关系,提升网络精度。实验结果表明:在KITTI数据集上,所提方法在3种主要目标类别的平均精度均值(mAP)方面相较基准方法分别提升了1.43%、5.23%和4.41%,平均每帧推理时间为34.19 ms,有效缩短了单阶段和双阶段网络之间的精度差距,保留了良好的检测效率。To well balance the detection accuracy and efficiency between single-stage and two-stage networks in current 3D object detection algorithms based on point cloud data,we propose a pseudo two-stage 3D object detection framework utilizing point cloud voxels.This framework includes a lightweight Center Heatmap Module for predicting object center points,eliminating anchor box settings and non-maximum suppression operations in traditional region proposal networks.To better utilize multi-scale voxel features,an AFP-Cross-Attention module is designed to extract high-value features in multi-scale voxels and perform cross-attention computations,reducing computational complexity compared to global attention.The AFP-Transformer detection head,built on the AFP method,effectively models the dependency between query features and high-value features,enhancing network accuracy.Experimental results on the KITTI dataset show our method improves the average precision for three main object categories by 1.43%,5.23%,and 4.41%respectively compared to the baseline method,with an average inference time of 34.19 ms per frame.Our approach effectively narrows the accuracy gap between single-stage and two-stage networks while maintaining high detection efficiency.

关 键 词：点云 三维目标检测 自适应特征池化 注意力机制 

分 类 号：U495[交通运输工程—交通运输规划与管理]