面向船舶智能航行的多目标实时跟踪方法  

Real-time multi-target tracking method for ship intelligent navigation

作　　者：徐海祥[1,2] 卜瑞波 冯辉[1,2] XU Haixiang;BU Ruibo;FENG Hui(Key Laboratory of High Performance Ship Technology,Ministry of Education,Wuhan University of Technology,Wuhan 430064,China;School of Transportation,Wuhan University of Technology,Wuhan 430064,China)

机构地区：[1]武汉理工大学高性能船舶技术教育部重点实验室,湖北武汉430064 [2]武汉理工大学交通学院,湖北武汉430064

出　　处：《华中科技大学学报（自然科学版）》2022年第1期138-143,共6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(51879210,51979210);中央高校基本科研业务费专项资金资助项目(2019Ⅲ040,2019III132CG)。

摘　　要：为能够实时跟踪周围船舶目标,提出一种船舶多目标实时跟踪方法.利用训练好的检测器来检测航行中会遇到的各种船舶,然后通过改进的DeepSort跟踪算法将检测结果进行关联匹配,从而完成多个船舶目标的实时跟踪.通过生成船舶重识别数据集来训练DeeSort算法中的特征提取网络,并改进了表观匹配中特征向量集的更新方式,使得空间有限的集合存储更多种表观特征.实验结果表明:提出的算法能够显著提升跟踪性能,其中轨迹切换身份的次数降低11%,轨迹被打断的次数降低5.8%,且不会增加计算时间,能够满足海上船舶目标感知的准确性和实时性要求.In order to track the surrounding ship targets in real time,a ship multi-target real-time tracking method was proposed.The trained detector was used to detect all kinds of ships encountered during navigation,and then the detection results were correlated and matched by the improved deep sort tracking algorithm,so as to complete the real-time tracking of multiple ship targets.The feature extraction network in deeport algorithm was trained by generating ship re identification data set,and the update mode of feature vector set in apparent matching is improved,so that the set with limited space could store more kinds of apparent features. The experimental results show that the proposed algorithm can significantly improve the tracking performance,in which the number of track switching identity is reduced by 11%,and the number of track interruptions is reduced by 5.8%.It can meet the accuracy and real-time requirements of marine ship target perception.

关 键 词：智能船舶 水面图像 多目标跟踪 特征提取 检测器 

分 类 号：U664.82[交通运输工程—船舶及航道工程]