系统工程与电子技术 ›› 2025, Vol. 47 ›› Issue (1): 41-51.doi: 10.12305/j.issn.1001-506X.2025.01.05

• 电子技术 • 上一篇    下一篇

基于惯性预测的遥感场景舰船多目标跟踪模型

潘超凡1, 李润生1,*, 胡庆2, 包全福3, 保永强4   

  1. 1. 信息工程大学数据与目标工程学院, 河南 郑州 450001
    2. 中国人民解放军61191部队, 浙江 杭州 310000
    3. 中国人民解放军95806部队, 北京 100076
    4. 北京航天遥感国际科技发展有限公司, 北京 100076
  • 收稿日期:2023-03-07 出版日期:2025-01-21 发布日期:2025-01-25
  • 通讯作者: 李润生
  • 作者简介:潘超凡(1997—), 男, 硕士, 主要研究方向为基于深度学习的目标检测
    李润生(1985—), 男, 副教授, 博士, 主要研究方向为遥感目标检测
    胡庆(1997—), 男, 助理研究员, 硕士, 主要研究方向为遥感影像目标检测
    包全福(1986—), 男, 工程师, 硕士, 主要研究方向为图像识别
    保永强(1986—), 男, 工程师, 硕士, 主要研究方向为摄影测量与遥感
  • 基金资助:
    青年科学基金(41901378)

Ship multiple-object tracking model in remote sensing scene based on inertial prediction

Chaofan PAN1, Runsheng LI1,*, Qing HU2, Quanfu BAO3, Yongqiang BAO4   

  1. 1. School of Data and Target Engineering, University of Information Engineering, Zhengzhou 450001, China
    2. Unit 61191 of the PLA, Hangzhou 310000, China
    3. Unit 95806 of the PLA, Beijing 100076, China
    4. Beijing Aerospace Remote Sensing International Technology Development Co. LTD, Beijing 100076, China
  • Received:2023-03-07 Online:2025-01-21 Published:2025-01-25
  • Contact: Runsheng LI

摘要:

遥感场景下的舰船目标跟踪具有重要的战略意义和经济价值, 如何克服遥感视角下舰船朝向任意性、近岸舰船密集排列等问题对跟踪性能的影响是一项具有挑战性的任务。针对遥感场景下舰船等大长宽比目标的多目标跟踪(multiple-object tracking, MOT)任务, 提出一种基于惯性预测的多目标跟踪器(inertial predicting multiple-object tracker, IPMOT)。首先, 利用检测-跟踪(tracking-by-detection, TBD)范式级联检测器和跟踪器有效避免训练过程对时序关系的依赖, 通过公开的目标检测数据集实现对检测器的训练, 解决跟踪数据集缺乏的问题。其次, 针对TBD范式在检测阶段存在的漏检严重影响跟踪性能的问题, 构建惯性跟踪模型(inertial tracking model, ITM), 通过多步预测来实现检测器漏检时的跟踪保持, 并利用角度修正消除边界处角度突变的影响。最后, 为实现所提算法的模型训练和性能测试, 制作舰船MOT (ship MOT, SMOT)数据集。实验结果表明, 所提模型在MOT精度(MOT accuracy, MOTA)和识别F1分数(identity F1 score, IDF1)指标上分别提升3.9%和7.2%, 在IDs和Frag指标上的表现有明显改善, 具有较好的跟踪精度和稳定性。

关键词: 遥感场景, 舰船目标, 多目标跟踪, 惯性预测, 检测-跟踪范式

Abstract:

Ship object tracking in remote sensing scenario is of great strategic significance and economic value. How to overcome the influence of ship orientation arbitrariness, close arrangement of offshore ship targets on tracking performance from remote sensing perspective is a challenging task. Multiple-object tracking (MOT) tasks with large aspect ratio such as ships in remote sensing scenarios are studied. An inertial predicting multiple object tracker (IPMOT) is proposed. Firstly, the tracing-by-detection (TBD) paradigm is used to cascade the detector and tracker to effectively avoid the dependence of the training process on the time sequence relationship, and the detector is trained through the open target detection data set to solve the problem of lack of tracking data set. Secondly, addressing the problem of missed detections in the detection phase of the TBD paradigm, which significantly affects tracking performance, an inertial tracking model (ITM) is built. It uses multi-step prediction to track the errors of detector, and uses angle correction to eliminate the impact of angle mutation at the edge. Finally, in order to realize the model training and performance testing of the proposed algorithm, the ship MOT (SMOT) dataset is constructed. The experimental results show that the proposed model improves the performance of MOT accuracy (MOTA) and inverse document frequency (IDF1) indexes by 3.9% and 7.2%, respectively, and the performance of IDs and Frag indexes is significantly improved, with good tracking accuracy and stability.

Key words: remote sensing scenario, ship target, multiple-object tracking (MOT), inertial predicting, tracing-by-detection (TBD) paradigm

中图分类号: