应用于地面战场目标的点云检测算法研究  

Research on Point Cloud Detection Algorithm for Ground Battlefield Targets

作　　者：吴奕霆 吴新建[1] 黄涛[1] WU Yi-ting;WU Xin-jian;HUANG Tao(Huazhong Institute of Electro-Optics-Wuhan National Laboratory for Optoelectronics,Wuhan 430233,China)

机构地区：[1]华中光电技术研究所-武汉光电国家研究中心,湖北武汉430223

出　　处：《光学与光电技术》2022年第2期77-83,共7页Optics & Optoelectronic Technology

摘　　要：地面战场上目标检测是精准跟踪以及准确打击的前提,在现代无人化陆战中起着至关重要的作用。传统的图像检测方法受到光照,天气等条件制约,利用激光雷达技术进行3D目标检测能够显著改善车体的感知能力。针对应用于陆战无人车的检测任务,提出了一种基于卷积神经网络的3D点云检测算法,通过优化VoxelNet的体素化和特征融合模块设计了一组端对端的高效网络,并改进了一种基于距离的非极大值抑制策略。实验表明原始VoxelNet在自建数据集上车辆目标AP值为78.53%,而改进后的网络性能达84.11%,对未来军事领域的三维目标检测任务具有参考价值。Object detection on the ground battlefield is the basis of precise strike,playing a vital role in modern unmanned warfare. The traditional image algorithm is restricted by lighting,weather and other conditions,which can be solved by 3D detection algorithm using lidar. For unmanned vehicles’ detection task on land battlefield,a 3D detection algorithm based on convolutional neural network is proposed in this paper. By optimizing the feature fusion module of VoxelNet,a group of end-to-end efficient networks are designed,and a non-maximum suppression strategy based on distance is improved.Experiments show that on the self-built dataset,original VoxelNet’s AP of vehicle target is 78.53%,while our network performance is 84.11%,which has great value for 3D detection task in the feature military field.

关 键 词：无人驾驶 激光雷达 点云 目标检测 深度学习 

分 类 号：TN958.98[电子电信—信号与信息处理]