基于Transformer的体素化激光点云目标检测算法  

Transformer-based algorithm for voxelized laser point cloud target detection

作　　者：康自祥 王升哲[1] 崔雨勇 高欣仪 陈旺成 KANG Zi-xiang;WANG Sheng-zhe;CUI Yu-yong;GAO Xin-yi;CHEN Wang-cheng(Southwest Institute of Technical Physics,Chengdu 610041,China)

机构地区：[1]西南技术物理研究所,四川成都610041

出　　处：《激光与红外》2023年第2期202-207,共6页Laser & Infrared

摘　　要：3D点云目标检测是计算机3D视觉中的一个关键技术,本文针对激光雷达点云数据的稀疏性、无序性和数据量大,导致神经网络运算效率慢、检测精度低等问题,开展了基于激光雷达点云的目标检测算法研究。在激光雷达点云数据处理阶段,我们将原始点云数据体素化,解决了点云稀疏性和无序性问题,然后使用多层特征下采样层构建特征金字塔,实验验证了该方法使网络在训练阶段更快收敛,有效减少点云数据量大导致的网络运算开销,网络运算效率提升~39%;同时,我们通过引入Transformer注意力模块,提高网络对点云目标关键特征的学习能力,使目标检测的准确率达到88.5%。总体实验结果表明,本文算法在确保检测精度的前提下,提升了网络运算效率。3D point cloud target detection is a key technology in computer 3D vision. To address the problems of sparsity, disorder, and large amount of lidar point cloud data, resulting in slow neural network operation efficiency and low detection accuracy, the research on target detection algorithm of lidar point cloud is carried out in this paper. Firstly, in the LiDAR point cloud data processing stage, the original point cloud data is voxelized to solve the problem of point cloud sparsity and disorder. Secondly, a multi-layer feature down-sampling layer is utilized to build a feature pyramid. Finally, according to the experiment results, the proposed method enables the network to converge faster in the training phase, effectively reducing the network computing overhead caused by the large amount of point cloud data, and improving the network computing efficiency by 39 %. Meanwhile, by introducing Transformer attention module, the network′s ability to learn the key features of point cloud targets is improved, and the accuracy of target detection is improved to 88.5 %. The overall experimental results show that the proposed algorithm improves the network operation efficiency on the premise of ensuring the detection accuracy.

关 键 词：深度学习 TRANSFORMER 体素 点云 目标检测 

分 类 号：TN958.98[电子电信—信号与信息处理]