远距离和遮挡下三维目标检测算法研究  

Long-distance and occluded 3D target detection algorithm

作　　者：陆军[1] 李杨[1] 鲁林超 LU Jun;LI Yang;LU Linchao(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001

出　　处：《智能系统学报》2024年第2期259-266,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(52171332);黑龙江省自然科学基金项目(F201123)。

摘　　要：针对现有三维目标检测算法对存在遮挡及距离较远目标检测效果差的问题,以基于点云的三维目标检测算法(3D object proposal generation and detection from point cloud,PointRCNN)为基础,对网络进行改进,提高三维目标检测精度。对区域生成网络(region proposal network,RPN)获取的提议区域(region of interest,ROI)体素化处理,同时构建不同尺度的区域金字塔来捕获更加广泛的兴趣点;加入点云Transformer模块来增强对网格中心点局部特征的学习;在网络中加入球查询半径预测模块,使得模型可以根据点云密度自适应调整球查询的范围。最后,对所提算法的有效性进行了试验验证,在KITTI数据集下对模型的性能进行评估测试,同时设计相应的消融试验验证模型中各模块的有效性。To address the limitations of existing 3D target detection algorithms,particularly their poor detection performance for occluded and long-distance objects,we have implemented an enhancement to the PointRCNN network,a 3D object detection algorithm based on point cloud.We began by voxelizing the region of interest obtained from the region proposal network and constructing region pyramids of different scales to capture a wider range of points of interest.Simultaneously,we introduced a point cloud transformer module to enhance the learning of the local features of grid center points.Moreover,we incorporated a sphere query radius prediction module into the network.This addition allows the model to adaptively adjust the sphere query range according to the density of the point cloud.Finally,the effectiveness of the proposed algorithm was validated through rigorous experimental testing.We evaluated the performance of the model using the KITTI data set and designed corresponding ablation experiments to verify the effectiveness of each module in the model.

关 键 词：目标检测 深度学习 激光雷达点云 远距离目标 遮挡下目标 自动驾驶 区域金字塔 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]