基于DeepSORT的水下目标声学图像跟踪方法  

Underwater target acoustic image tracking method based on DeepSORT

作　　者：张铁栋 李仁哲[2] 郎硕 曾文静 ZHANG Tiedong;LI Renzhe;LANG Shuo;ZENG Wenjing(School of Ocean Engineering and Technology,Sun Yat-sen University,Zhuhai 519082,Guangdong China;College of Shipbuilding Engineering,Harbin Engineering University,Harbin 150001,China;703 Research Institute of China State Shipbuilding Corporation,Harbin 150001,China)

机构地区：[1]中山大学海洋工程与技术学院,广东珠海519082 [2]哈尔滨工程大学船舶工程学院,黑龙江哈尔滨150001 [3]中船重工703研究所,黑龙江哈尔滨150001

出　　处：《华中科技大学学报（自然科学版）》2023年第10期44-50,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(51879061,U22A2012);黑龙江省开发计划资助项目(GA20A402)。

摘　　要：针对水下运动目标的声学影像模糊、区域特征非稳定,易导致跟踪过程中出现轨迹中断和目标身份(ID)变更等问题,提出基于扩展目标框的DeepSORT水下目标改进跟踪方法.该方法采用改进Faster RCNN作为检测器,扩展了检测器输出的目标框,提取了目标周围散射噪声的特征,增大了目标区域感受野,解决了声学图像中目标特征稀疏的问题.结果表明:本文方法能够抑制轨迹中断和ID变更现象,在复杂场景下相较于传统DeepSORT方法的跟踪结果,ID变更次数下降了80%,轨迹中断占比下降了2.17%,有效提升跟踪网络的稳定性.In view of the fuzzy acoustic image and unstable regional features of underwater moving targets,which were easy to cause track interruption and target identity(ID)change in the tracking process,a DeepSORT improved tracking method based on expanded target bounding box was proposed.Based on the DeepSORT method,the improved Faster RCNN(region-convolutional neural networks)was used as the detector to expand the target bounding box output of the detector,extract the features of scattered noise around the target,and increase the receptive field of the target area,which solved the problem of sparse target features in acoustic images.Results show that the proposed method can suppress the phenomenon of trajectory interruption and ID change.In complex scenarios,compared with the tracking results of traditional DeepSORT method,the number of ID changes decreases by 80%,and the proportion of track interruption decreases by 2.17%,effectively improving the stability of the tracking network.

关 键 词：水下感知 前视声纳 声纳成像 目标跟踪 深度学习 

分 类 号：TB56[交通运输工程—水声工程] TP391.41[理学—物理]