重庆理工大学学报（自然科学版）2025，Vol.39Issue(3)：11-18,8.DOI:10.3969/j.issn.1674-8425(z).2025.02.002
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
摘要
Abstract
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Key words
point clouds/3D object detection/adaptive feature pooling/attention mechanism分类
交通工程引用本文复制引用
汪世豪,邓涛..一种面向智能驾驶的单阶段点云三维目标检测算法[J].重庆理工大学学报（自然科学版）,2025,39(3):11-18,8.基金项目
国家自然科学基金项目(52275051) （52275051）
重庆交通大学自然科学类揭榜挂帅项目(XJ2023000701) （XJ2023000701）
重庆市研究生导师团队建设项目(JDDSTD2022007) （JDDSTD2022007）
重庆市研究生联合培养基地建设项目(JDLHPYJD2022001) （JDLHPYJD2022001）
