融合密集连接和高斯距离的三维目标检测算法  

3D object detection algorithm fusing dense connectivity and Gaussian distance

作　　者：程鑫[1,2] 刘升贤 周经美 周洲 赵祥模[1] CHENG Xin;LIU Sheng-xian;ZHOU Jing-mei;ZHOU Zhou;ZHAO Xiang-mo(School of Information Engineering,Chang'an University,Xi'an 710018,China;Traffic Management Research Institute of the Ministry of Public Security,Wuxi 214151,China;School of Electronics and Control Engineering,Chang'an University,Xi'an 710018,China)

机构地区：[1]长安大学信息工程学院,西安710018 [2]公安部交通管理科学研究所,江苏无锡214151 [3]长安大学电子与控制工程学院,西安710018

出　　处：《吉林大学学报(工学版)》2024年第12期3589-3600,共12页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(52472337,52102452);中国博士后科学基金项目(2023T160129);交通运输部重点科技项目(2022-ZD6-079);陕西省重点研发计划项目(2022QCY-LL-29);云南省交科院科技研发项目(JKYZLX-2023-12)。

摘　　要：为加强对小目标的感知,在F-PointNet网络的基础上,结合密集连接方法和高斯距离特征,提出了FDG-PointNet三维目标检测模型,融合高斯距离特征作为附加注意力特征,有效改善了F-PointNet网络实例分割准确率不高的问题,增强了对点云视锥体中的噪声的过滤;基于密集连接可以加强特征提取的特点,改进主干特征提取PointNet++网络,加强点云特征重用,缓解特征提取过程中对小目标的特征提取程度过低与梯度消失问题,提高三维目标边界框回归的准确性。研究结果表明:本文算法在简单、中等、困难三个难度等级下对汽车、行人、骑行人3种类别的检测整体优于基准方法F-PointNet,在中等难度下对汽车、行人、骑行人的检测分别取得71.12%、61.23%、55.71%的平均检测精度,其中对行人检测提升最明显,在简单和中等难度下提升幅度分别达5.5%和3.1%。综上所述,本文的FDG-PointNet算法有效解决了F-PointNet中小物体检测的低准确性问题,具有较强的适用性。To enhance the perception of small objects,based on the F-PointNet network,the FDGPointNet 3D object detection model is proposed by combining dense connection and Gaussian distance features.Gaussian distance features is fused as additional attention features,and it effectively solves the low accuracy of instance segmentation in the F-PointNet network and enhances the noise filtering in the point cloud view cone.Based on the characteristics that dense connection can enhance feature extraction,the dense connection is used to improve PointNet++network and enhance point cloud feature reuse.It alleviates low degree of feature extraction and gradient disappearance for small objects in the feature extraction process,and improves the accuracy of 3D object bounding box regression.The experimental results show that the proposed algorithm outperforms the benchmark method F-PointNet in three levels(easy,moderate,and hard)for the detection of car,pedestrian,and cyclist,which can achieve the average detection accuracy of 71.12%,61.23%,and 55.71%for car,pedestrian,and cyclist at moderate level.It has the most significant improvement for pedestrian detection,and can increase 5.5%and 3.1%at easy and moderate levels,respectively.In summary,compared to F-PointNet algorithm,the proposed FDG-PointNet algorithm effectively solves the low accuracy of small objects detection and has strong applicability.

关 键 词：计算机科学与技术 三维目标检测 激光雷达 密集连接 高斯距离 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]