自航船模点云数据集的海上船舶检测  

Marine ship detection on the point cloud dataset of autonomous navigation ship models

作　　者：何芸倩 夏桂华 冯鸿超 向晶 胡乃元 HE Yunqian;XIA Guihua;FENG Hongchao;XIANG Jing;HU Naiyuan(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China;Heilongjiang Provincial Key Laboratory of Environment Intelligent Perception,Harbin 150001,China;Key Laboratory of Intelligent Technology and Application of Marine Equipment(Harbin Engineering University),Ministry of Education,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 [2]黑龙江省环境智能感知重点实验室,黑龙江哈尔滨150001 [3]“船海装备智能化技术与应用”教育部重点实验室(哈尔滨工程大学),黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2022年第8期1156-1162,1168,共8页Journal of Harbin Engineering University

基　　金：国家重点研发计划(2019YFE0105400).

摘　　要：为了进行激光雷达海上目标检测的算法研究,本文利用自主航行船模与激光雷达等效采集海上场景点云数据,制作了船舶点云数据集。利用深度学习方法,提出了一种适用于船舶点云目标检测的点结构轻量型目标检测网络LASSD,并通过网络剪枝的方式提升了速度并缩减了所需的计算资源。提出一种基于候选目标的高阶点云特征局部注意力模块,弥补网络剪枝带来的精度损失。实验表明:本文的LASSD网络仅使用5.3×106的参数量在船舶数据集中达到79.42%的精度,在检测中单幅场景仅花费13.5 ms,检测精度以及运行速度能够在实际应用中提供实时有效的检测结果。In order to start the research work of LiDAR maritime object detection,a ship point cloud dataset was established by equivalently collecting the marine point cloud data using autonomous navigation ship models and LIDAR.Using the deep learning method,a point-structured lightweight object detection network LASSD was proposed for ship point cloud object detection,and the network pruning method improved the speed and reduced the required computing resources.A local attention module for high-level point cloud features based on candidate targets was proposed to compensate for the accuracy loss caused by network pruning.Experiments show that the LASSD network in this paper uses only 5.3×106 parameters to achieve 79.42%accuracy in the ship dataset,and single scene detection only takes 13.5 ms.Detection accuracy and operation speed can provide valid results in practical applications.

关 键 词：激光雷达 船舶 海上环境 点云 计算机视觉 目标检测 数据集 深度学习 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]