基于稀疏自注意力图神经网络的三维目标检测  

3D Object Detection Based on Sparse Self-Attention Graph Neural Network

作　　者：彭志辰 封岸松 王天柱 邵鑫喆 库涛[2,3,4] PENG Zhichen;FENG Ansong;WANG Tianzhu;SHAO Xinzhe;KU Tao(College of Information Engineering,Shenyang University of Chemical Technology,Shenyang 110142,China;Key Laboratory of Networked Control Systems,Chinese Academy of Sciences,Shenyang 110016,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]沈阳化工大学信息工程学院,沈阳110142 [2]中国科学院网络化控制系统重点实验室,沈阳110016 [3]中国科学院沈阳自动化研究所,沈阳110016 [4]中国科学院机器人与智能制造创新研究院,沈阳110169 [5]中国科学院大学,北京100049

出　　处：《计算机工程与应用》2025年第3期295-305,共11页Computer Engineering and Applications

基　　金：国家重点研发计划(2020YFB1708503)。

摘　　要：三维目标检测是自动驾驶环境感知中最重要的技术之一。为了解决远距离漏检问题,提升三维目标检测的效果,提出一种基于稀疏自注意力图神经网络的三维目标检测方法(SSA-GNN),在采样关键点阶段,提出动态区域并行采样法,通过采样区域过滤,场景划分为扇区,融合动态最远体素采样的方式,以保持关键点均匀分布、加速采样同时提升前景点比例。在细化建议框阶段,利用图神经网络在点之间建立联系,通过迭代的消息传递来更好地建模上下文信息和聚合领域信息,并改进多头自注意机制来更好地关注特征聚合后领域中的重要关系,从而提高算法检测性能。SSA-GNN在KITTI公开数据集上进行测试,与基线网络PointPillars、SECOND和PointRCNN相比,在困难等级指标下,Car类平均精度分别提升了7.95、5.50、6.94个百分点,结果表明SSA-GNN可有效提升三维目标检测性能。3D object detection is one of the most important technologies in autonomous driving environment perception.To address the issue of long-range missed detections and enhance the effectiveness of 3D object detection,a 3D object detection method based on a sparse self-attention graph neural network(SSA-GNN)is proposed.In the stage of sampling key points,a dynamic area parallel sampling method is proposed.Through sampling area filtering,the scene is divided into sectors,and the way of dynamic farthest voxel sampling is integrated to maintain the uniform distribution of key points,accelerate sampling,and increase the proportion of foreground points.In the stage of refining the suggestion box,the graph neural network is used to establish connections between points,and the context information is better modeled and the domain information is aggregated through iterative message passing.Moreover,the multi-head self-attention mechanism is improved to pay more attention to the important relationships in the domain after feature aggregation,so as to improve the detection performance of the algorithm.SSA-GNN is tested on the KITTI public data set.Compared with the baseline network PointPillars,SECOND,and PointRCNN,the average accuracy of the Car category is increased by 7.95,5.50,and 6.94 percentage points respectively under the difficulty level index.The results show that SSA-GNN can effectively improve the performance of 3D object detection.

关 键 词：三维目标检测 关键点采样 图神经网络 稀疏自注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]