融合点云和体素信息的目标检测网络  

Object detection network fusing point cloud and voxel information

作　　者：刘慧 董振阳[1] 田帅华 LIU Hui;DONG Zhen-yang;TIAN Shuai-hua(School of Electrical and Information Engineering,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Beijing Key Laboratory of Intelligent Processing for Building Big Data,Beijing University of Civil Engineering and Architecture,Beijing 102616,China)

机构地区：[1]北京建筑大学电气与信息工程学院,北京102616 [2]北京建筑大学建筑大数据智能处理方法研究北京市重点实验室,北京102616

出　　处：《计算机工程与设计》2024年第9期2771-2778,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62176018)。

摘　　要：为解决复杂自动驾驶场景下小目标检测效果不佳和漏检的问题,提出一种融合点云和体素信息的高性能网络架构。通过预处理模块、空间语义特征关联模块、坐标注意力机制模块等改进PV-RCNN网络的检测性能,构建网络架构PSC-RCNN。在KITTI上进行验证,实验结果表明,PSC-RCNN在简单、中等、困难3种检测难度的类别下,对于自行车这种形状复杂的小物体识别精度分别为82.99%、67.03%、59.88%,相对原有的PV-RCNN网络,识别精度分别提高了4.39%、3.32%、2.23%;相对于现有3D目标检测网络,识别精度分别提高了0.51%、2.93%、2.23%。To solve the problem of poor and missed detection of small objects in complex autonomous driving scenarios,a high-performance network architecture that fused point cloud and voxel information was proposed.The object detection performance of the PV-RCNN network was improved through the preprocessing module,the spatial semantic feature concatenate module,and the coordinate attention mechanism module,and a network architecture PSC-RCNN was constructed.Validated on the KITTI,experimental results show that the recognition accuracy of PSC-RCNN for small objects with complex shapes like bicycle is 82.99%,67.03%,and 59.88%under three categories of detection difficulty(easy,medium,and difficult)respectively,the recognition accuracy is improved by 4.39%,3.32%,and 2.23%respectively.Compared with the existing 3D point cloud object detection network,the recognition accuracy is improved by 0.51%,2.93%,and 2.23%,respectively.

关 键 词：机器视觉 三维点云 三维体素 目标检测 空间语义特征关联 坐标注意力 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]