基于多尺度候选融合与优化的三维目标检测算法  

Three-dimensional object detection algorithm based on multi-scale candidate fusion and optimization

作　　者：才华[1] 郑延阳 付强[2] 王晟宇 王伟刚[3] 马智勇[3] CAI Hua;ZHENG Yan-yang;FU Qiang;WANG Sheng-yu;WANG Wei-gang;MA Zhi-yong(School of Electronic Information Engineer,Changchun University of Science and Technology,Changchun 130022,China;School of Opto-Electronic Engineer,Changchun University of Science and Technology,Changchun 130022,China;No.2 Department of Urology,The First Hospital of Jilin University,Changchun 130061,China)

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]长春理工大学空间光电技术研究所,长春130022 [3]吉林大学第一医院泌尿外二科,长春130061

出　　处：《吉林大学学报(工学版)》2025年第2期709-721,共13页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金重大项目(61890963);吉林省科技发展计划项目(20210204099YY);吉林省医疗卫生人才专项项目(JLSWSRCZX2023-70)。

摘　　要：为了改善点云场景下的检测任务中,基于单一低分辨特征图生成的候选框容易造成目标丢失和关键点采样过程中引入大量背景点的问题,本文提出了一种基于PV-RCNN网络的改进算法。通过区域候选融合网络和加权非极大值抑制融合不同尺度下的候选框并消除冗余。利用分割网络对原始点云进行前景点分割,并根据候选框确定目标中心点位置,利用高斯密度函数进行区域密度估计得到不同的采样权重以解决稀疏区域采样困难的问题。本文使用KITTI数据集进行实验评估,在汽车、行人和骑行者中等难度下的平均精度分别较基线算法提升0.39%、1.31%和0.63%,并同样在Waymo open数据集上进行泛化实验。实验结果证明本文算法与目前大部分三维目标检测算法相比取得更高的检测精度。To address the issues of target omission and the inclusion of a large number of background pointsin keypoint sampling for point cloud-based object detection,an improved algorithm based on the PV-RCNN network is introduced.This approach employs both a regional proposal fusion network andweighted non-maximum suppression(NMS)to merge proposals generated at various scales whileeliminating redundancy.A segmentation network is utilized to segment foreground points from the originalpoint cloud,and object center points are identified based on these proposals.Gaussian density functions areemployed for regional density estimation,which assigns different sampling weights to solve the problem ofdifficult sampling in sparse areas.Experimental evaluations on the KITTI dataset indicate that thealgorithm enhances the average precision at medium difficulty levels by 0.39%,1.31%,and 0.63%forcars,pedestrians,and cyclists,respectively.Generalization experiments were also conducted on theWaymo open dataset.The results suggest that the introduced algorithm achieves higher accuracy comparedto most of the existing 3D object detection networks.

关 键 词：计算机视觉 三维目标检测 区域候选融合 加权非极大值抑制 关键点采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]