前视声呐图像小目标智能感知与跟踪算法  

Intelligent sensing and tracking algorithm for small targets in forward-looking sonar images

作　　者：贾昊明 于晓阳 周天[1,2,3,4] JIA Haoming;YU Xiaoyang;ZHOU Tian(National Key Laboratory of Underwater Acoustic Technology,Harbin Engineering University,Harbin 150001,China;Key Laboratory of Marine Information Acquisition and Security(Harbin Engineering University),Ministry of Industry and Information Technology,Harbin 150001,China;College of Underwater Acoustic Engineering,Harbin Engineering University,Harbin 150001,China;Key Laboratory for Polar Acoustics and Application of Ministry of Education,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学水声技术全国重点实验室,黑龙江哈尔滨150001 [2]海洋信息获取与安全工业和信息化部重点实验室(哈尔滨工程大学)工业和信息化部,黑龙江哈尔滨150001 [3]哈尔滨工程大学水声工程学院,黑龙江哈尔滨150001 [4]极地海洋声学与技术应用教育部重点实验室,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2025年第1期129-137,共9页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(42306212,62076215,42176188);黑龙江省自然科学基金项目(LH2023D018);海南省自然科学基金项目(421CXTD442).

摘　　要：针对传统深度学习模型对小目标感知能力有限的难题,本研究首先提出基于注意力机制的YOLOv5_cs检测模型,在此基础上通过声呐图像公开数据集开展模型预训练,利用迁移学习来增强网络对目标的特征提取能力。联合DeepSORT应用于多目标跟踪任务中,通过对真实水池实验中采集的多目标跟踪数据集的测试结果分析表明:本研究提出的算法相比于传统YOLOv5联合DeepSORT算法,多项评价指标得到提升,其中多目标跟踪准确度指标提升了4.85%,多目标跟踪精度指标提升了0.95%,身份识别精度得分指标提升了2.66%。同时提出的算法较好地解决了目标形态发生变化条件下目标检测效果不佳导致的错跟、漏跟等问题,具有较高实际应用潜力。Forward-looking sonar images have low resolution,complex noise,and blurred target edges,which limit the application of the YOLOv5(You Only Look Once)combined with the DeepSORT(Simple Online and Realtime Tracking with a Deep Association Metric)algorithm for target detection and tracking tasks.This study first proposes the YOLOv5_cs detection model based on the attention mechanism to address the limited ability of the traditional deep learning model to perceive small targets.The model is pre-trained using a public dataset of sonar images,and transfer learning is utilized to enhance the feature extraction ability of the network on the target.When applied jointly with DeepSORT in multiobject tracking tasks,the analysis of test results from a real pool experiment dataset shows that the algorithm proposed in this study outperforms the traditional YOLOv5 combined with the DeepSORT algorithm across several evaluation metrics.Specifically,MOTA(Multiple Object Tracking Accuracy)is improved by 4.85%,MOTP(Multiple Object Tracking Precision)is enhanced by 0.95%,and IDF1 is increased by 2.66%.The proposed algorithm effectively solves problems such as tracking errors and omissions due to poor target detection under the conditions of changing target shapes.Thus,it has promising potential for practical applications.

关 键 词：前视声呐图像 多目标跟踪 深度学习 注意力机制模块 空间转深度 非跨步卷积 迁移学习 YOLOv5_cs 

分 类 号：P754[天文地球—海洋科学]