用于激光雷达目标检测的单阶段无锚框优化网络  

An Optimized Network for One-Stage Anchor-Free LiDAR Object Detection

作　　者：朱望江 郭建伟[1,2] 张吉光 孟维亮 张晓鹏[1,2] Zhu Wangjiang;Guo Jianwei;Zhang Jiguang;Meng Weiliang;Zhang Xiaopeng(State Key Laboratory of Multimodal Artificial Intelligence Systems,Institute of Automation,Chinese Academy of Sciences,Beijing 100190;School of Artificial Intelligence,University of Chinese Academy of Sciences,Beijing 100049;Zhejiang Lab,Hangzhou 311121)

机构地区：[1]中国科学院自动化研究所多模态人工智能系统全国重点实验室,北京100190 [2]中国科学院大学人工智能学院,北京100049 [3]之江实验室,杭州311121

出　　处：《计算机辅助设计与图形学学报》2025年第3期457-464,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(U21A20515,U22B2034,61972459,61971418,62172416,32271983);之江实验室开放课题(2021KE0AB07).

摘　　要：激光雷达目标检测近年来开始借鉴图像目标检测的网络设计,但依然存在计算低效无法满足实时应用以及网络结构简单导致性能不足的问题.为此,提出的网络采用了单阶段无锚框的简洁设计;优化了激光点云体素化表达,在提升计算效率的同时保留了一部分点云高程特征;基于残差网络的思想,设计了更深的主干网络结构用于提取深度特征;引入特征金字塔来提升小目标的检测效果.在公开数据集KITTI上,所提网络的mAP指标在各类别目标的检测中均取得了领先的性能(提高了1%~3%).在自动驾驶计算平台上的运行时间测试表明,所提网络能够达到43 ms/帧的处理速度,满足实时性需求.In recent years,the realm of laser radar target detection has increasingly drawn inspiration from network designs employed in image target detection.Nevertheless,persistent challenges encompass inefficient computations impeding real-time applications and suboptimal performance resulting from overly simplistic network structures.The proposed network embraces a refined design characterized by a single-stage,anchor-free methodology.It strategically enhances the voxelization representation of laser point clouds,thereby bolstering computational efficiency while preserving crucial elevation features.Leveraging the principles of residual networks,a deeper backbone network structure is crafted to adeptly extract deep features.Additionally,the incorporation of a feature pyramid enhances the detection capabilities,especially for smaller targets.Across various object categories,the proposed network demonstrates outstanding performance on the KITTI public dataset,as reflected in the mAP indicator with 1%to 3%improvement.Runtime assessments on an autonomous driving computing platform have a processing speed of 43 ms per frame.validating the network’s ability to meet real-time requirements.

关 键 词：目标检测 激光雷达 自动驾驶 单阶段 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]