基于局部信息融合的点云3D目标检测算法  

Point cloud 3D object detection algorithm based on local information fusion

作　　者：张林杰 柴志雷 王宁 ZHANG Linjie;CHAI Zhilei;WANG Ning(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214122,China;Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence,Wuxi 214122,China)

机构地区：[1]江南大学人工智能与计算机学院,江苏无锡214122 [2]江苏省模式识别与计算智能工程实验室,江苏无锡214122

出　　处：《浙江大学学报（工学版）》2024年第11期2219-2229,共11页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(61972180);江苏省模式识别与计算智能工程实验室资助项目。

摘　　要：针对当前基于点云的三维目标检测算法缺乏目标准确的空间位置信息,提出局部信息编码模块和后期交叉融合模块的三维目标检测算法.在特征提取阶段,模型通过三维稀疏卷积高效地编码全局特征.局部信息编码模块利用目标内部的原始点云信息,构建目标的细粒度语义信息,通过自注意力机制对这些信息进行重新加权,增强局部特征的表达能力.提出交叉融合模块,用于局部特征与全局特征的信息交互,产生表达能力更强的目标检测特征.使用KITTI和Waymo公开数据集,验证所提出的方法.在KITTI数据集的简单、中等和困难任务上,本文方法的平均准确率AP0.7分别达到了91.60%、82.53%和77.83%,在Waymo数据集上的平均准确率AP0.7达到74.92%.A three-dimensional object detection algorithm with a local information encoding module and a subsequent cross-fusion module was proposed aiming at the current lack of accurate spatial position information for three-dimensional object detection algorithms based on point clouds.Global features were efficiently encoded using 3D sparse convolution during the feature extraction phase.The local information encoding module leveraged the intrinsic information within the object’s point cloud,constructing fine-grained semantic details.The information was reweighted to enhance the representation of local features through a self-attention mechanism.A cross-fusion module was introduced to facilitate interaction between local and global features,resulting in enhanced object detection features.The proposed method was validated using the KITTI and Waymo datasets.The average precision at IoU 0.7 for easy,moderate and hard tasks achieved 91.60%,82.53%,and 77.83%,respectively on the KITTI dataset.The average precision at IoU 0.7 reached 74.92%on the Waymo dataset.

关 键 词：点云 稀疏卷积 局部信息 注意力机制 交叉融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]