低检测置信度下轻量化水下多目标跟踪算法  

A light underwater multi-object tracking algorithm with low detection confidence

作　　者：张文凯 余敏 刘浩煜 叶颢 冯琳琳 汤奇荣[1] ZHANG Wenkai;YU Min;LIU Haoyu;YE Hao;FENG Linlin;TANG Qirong(Laboratory of Robotics and Multibody System,School of Mechanical Engineering,Tongji University,Shanghai 201804,China;Department of Physical Education,Tongji University,Shanghai 200092,China)

机构地区：[1]同济大学机械与能源工程学院机器人技术与多体系统实验室,上海201804 [2]同济大学体育教学部,上海200092

出　　处：《舰船科学技术》2025年第6期128-133,共6页Ship Science and Technology

基　　金：国家自然科学基金资助项目(62373285);上海市产业协同创新项目(HCXBCY-2022-051);机器人技术与系统全国重点实验室开放基金(SKLRS-2024-KF-04);某部基础科研计划项目(XXXX2022YYYC133)。

摘　　要：水下声呐图像存在背景噪声严重等问题,导致水下分类器输出大量检测低置信度对象。而现有的水下多目标跟踪框架大多简单排除低置信度目标,导致跟踪轨迹中断。本文提出一种低检测置信度下水下多目标跟踪算法YOLO-Fair MOT;引入多通道随机混合注意力模块,抑制背景噪声的影响;采用深度可分离卷积降低模型复杂性,提升跟踪过程的整体速度;结合低置信度数据匹配算法与广义交并比匹配算法,改善跟踪轨迹的中断问题。实验结果表明,YOLO-Fair MOT算法具有更好的跟踪准确度、跟踪精确度、轨迹保持性以及检测速度。Underwater forward-looking sonar images face problems like severe background noise,leading to a large number of low confidence detection targets from the underwater classifier.Most existing underwater multi-object tracking frameworks simply exclude those targets,which cause trajectory interruption problem.A light underwater multi-target tracking algorithm YOLO-Fair MOT is proposed.The channel-spatial random fusion attention module is introduced to suppress the influence of image background noise.Depthwise convolutions are used to reduce model complexity and improves the overall speed of the tracking process.And the Byte association algorithm and the generalized intersection over union association algorithm are combined to improve the fragmentation of tracking trajectories.The experimental results show that YOLO-Fair MOT has higher value of tracking accuracy,tracking accuracy,trajectory retention,and detection speed.

关 键 词：水下多目标跟踪 前视声呐图像 轻量化 YOLOv5 Fair MOT 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]