首页|远距离和遮挡下三维目标检测算法研究

远距离和遮挡下三维目标检测算法研究

扫码查看
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.
Long-distance and occluded 3D target detection algorithm
To address the limitations of existing 3D target detection algorithms,particularly their poor detection per-formance 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 re-gion 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 al-lows the model to adaptively adjust the sphere query range according to the density of the point cloud.Finally,the ef-fectiveness of the proposed algorithm was validated through rigorous experimental testing.We evaluated the perform-ance of the model using the KITTI data set and designed corresponding ablation experiments to verify the effectiveness of each module in the model.

target detectiondeep learningLidar point cloudlong-distance targetoccluded targetautopilotregional pyramidfeature extraction

陆军、李杨、鲁林超

展开 >

哈尔滨工程大学 智能科学与工程学院, 黑龙江 哈尔滨 150001

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

国家自然科学基金黑龙江省自然科学基金

52171332F201123

2024

智能系统学报
中国人工智能学会 哈尔滨工程大学

智能系统学报

CSTPCD北大核心
影响因子:0.672
ISSN:1673-4785
年,卷(期):2024.19(2)
  • 31