基于注意力机制的三维点云车辆目标检测
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

基金项目:

长沙理工大学青年教师成长计划(2019QJCZ014)


3D Point Cloud Vehicle Target Detection Based on Attention Mechanism
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 增强出版
  • |
  • 文章评论
    摘要:

    针对自动驾驶场景下三维点云车辆的识别和定位问题, 提出了一种基于注意力机制的三维点云车辆目标检测算法. 算法将稀疏无序的点云空间划分成等距规则的体素表示, 用三维稀疏卷积和辅助网络同步从所有体素中提取内部点云特征, 进而生成鸟瞰图. 但在将内部三维的点云特征转化为二维的鸟瞰图后, 通常会造成目标空间特征信息丢失, 使得最终检测结果以及方向性预估差. 为进一步提取鸟瞰图中特征信息, 提出了一种注意力机制模块, 其中包含两种注意力模型, 并对其采用首、中、尾的“立体式”布局结构, 实现对鸟瞰图中特征信息的放大和抑制, 最后使用卷积神经网络和PS-Warp变换机制对处理过后的鸟瞰图进行三维目标检测. 实验表明, 该算法在保证实时检测效率的前提下, 与现有算法相比, 具有更好的方向预估性以及更高的检测精度.

    Abstract:

    In this study, a 3D point-cloud target detection algorithm for vehicles based on attention mechanism is proposed for the recognition and positioning of the targets in autonomous driving scenarios. The algorithm first divides the sparse and disordered point cloud space into equidistant and regular voxel representations. Then, 3D sparse convolution and auxiliary network are used to synchronously extract the internal point cloud features from all voxels. Afterward, a bird’s-eye view is generated. After the internal 3D point cloud features are converted into a 2D bird’s-eye view, the spatial feature information of the target will be lost generally, which makes the final detection result and the direction prediction unsatisfactory. To further extract the feature information of the bird’s-eye view, this study also proposes an attention mechanism module, which contains two attention models and adopts a three-dimensional layout structure (front, middle, and back) to realize amplification and suppression of the feature information of the bird’s-eye view. The convolutional neural network and PS-Warp transformation mechanism are employed to perform 3D target detection on the processed bird’s-eye view. Experiments show that, under the premise of ensuring real-time detection efficiency, this algorithm has better direction prediction and higher detection accuracy than existing algorithms.

    参考文献
    相似文献
    引证文献
引用本文

彭玉旭,董胜超.基于注意力机制的三维点云车辆目标检测.计算机系统应用,2021,30(12):211-217

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-03-22
  • 最后修改日期:2021-04-19
  • 录用日期:
  • 在线发布日期: 2021-12-10
  • 出版日期:
您是第位访问者
版权所有:中国科学院软件研究所 京ICP备05046678号-3
地址:北京海淀区中关村南四街4号 中科院软件园区 7号楼305房间,邮政编码:100190
电话:010-62661041 传真: Email:csa (a) iscas.ac.cn
技术支持:北京勤云科技发展有限公司

京公网安备 11040202500063号