基于强语义关键点采样的三维目标检测方法  

3D Object Detection Based on Strong Semantic Key Point Sampling

作　　者：车运龙 袁亮 孙丽慧 CHE Yunlong;YUAN Liang;SUN Lihui(College of Information Science and Technology,Beijing University of Chemical Technology,Beijing 100029,China;Non-Commissioned Officer Academy,Space Engineering University,Beijing 100000,China)

机构地区：[1]北京化工大学信息科学与技术学院,北京100029 [2]航天工程大学士官学校,北京100000

出　　处：《计算机工程与应用》2024年第9期254-260,共7页Computer Engineering and Applications

基　　金：国家自然科学基金(52275003);中央高校基本科研业务费资助项目(buctrc202105)。

摘　　要：目标检测算法中关键信息的特征提取是影响算法精度的重要因素。针对当前三维目标检测算法中存在的关键点采样困难、特征提取不充分等问题,借鉴PV-RCNN三维目标检测网络,提出一种基于强语义关键点采样的三维目标检测方法SSPS-RCNN(strong semantic point sampling RCNN)。在关键点采样阶段,该算法采用语义加权点采样和提案区域点过滤相融合的方法,获得更具特征代表性的采样关键点,以提升采样点中前景点的比例;在不增加网络结构的基础上,将语义信息重新加权关键点特征,以进一步细化关键点的特征贡献提升算法精度。在KITTI数据集上的实验结果表明该算法与现有主流算法相比,对减少物体检测中的漏检与错检问题和整体检测精度提升,表现出良好的稳定性和鲁棒性。Feature extraction of key information in the target detection algorithm is an important factor affecting the accu�racy of the algorithm.Aiming at the problems of key point sampling difficulty and insufficient feature extraction in the current 3D target detection algorithm,used for reference of PV-RCNN 3D target detection network,a 3D target detection algorithm SSPS-RCNN(strong semantic point sampling RCNN)based on strong semantic key point sampling is pro�posed.In the key point sampling stage,the algorithm used the fusion method of semantic weighted point sampling and proposal area point filtering to obtain more characteristic representative key points and increase the proportion of fore�ground points in the sampling points.Without adding the network structure,the point by semantic information is reweighted to further refine the feature contribution of the key points to improve the algorithm accuracy.Experiments show that this algorithm can reduce the problem of missing detection and wrong detection than the existing mainstream algorithms,and show good stability and robustness.Experimental results on KITTI dataset show that the proposed algorithm has good sta�bility and robustness compared with the existing mainstream algorithms,which can reduce the problem of missing detec�tion and wrong detection and improve the overall detection accuracy.

关 键 词：深度学习 激光雷达 三维目标检测 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]