中国空间科学技术

中国空间科学技术 ›› 2024, Vol. 44 ›› Issue (1): 144-153.doi: 10.16708/j.cnki.1000-758X.2024.0016

• 论文 • 上一篇    

面向视频卫星的多目标跟踪技术

陈海涛，马骏，李峰，鹿明，鲁啸天，张南   

  1. 1.河南大学 软件学院，开封475100
    2.中国空间技术研究院 钱学森空间技术试验室，北京100094
    3.中国航天科技创新研究院 先进智能算法中心，北京100163
  • 出版日期:2024-02-25 发布日期:2024-02-01

A new multi-target tracking method for video satellite data

CHEN Haitao，MA Jun，LI Feng，LU Ming，LU Xiaotian，ZHANG Nan   

  1. 1.College of Software，Henan University，Kaifeng 475100，China
    2.Qian Xuesen Space Technology Laboratory，China Academy of Space Technology，Beijing 100094，China
    3.Advanced Intelligent Algorithm Center，China Academy of Aerospace Science and Technology Innovation，
    Beijing 100163，China
  • Published:2024-02-25 Online:2024-02-01

PDF (PC)

238

摘要： 随着面阵探测器的广泛使用，面向视频卫星的多目标跟踪具有重要意义，但基于图结构的多目标跟踪方法，在图的构建中，大多数从相邻帧提取线索，而忽略了以往帧的线索。针对这个问题，提出了一个端到端的图网络框架，利用从多帧中提取的运动特征、外观特征、拓扑信息等多种线索，对图的节点、边和全局变量进行构建。实现这个统一框架的一个关键原则是为不同的线索和不同的来源（轨迹和检测目标）设计兼容的特征表示和图网络更新机制。该框架以前馈的方式运行，并以在线的方式进行训练。在公共数据集VISO、MOT16、MOT17基准上评测，取得了99.8%、48.8%、51.8%的多目标跟踪精度，优于其他相关多目标跟踪算法，并通过消融试验验证了各个线索对多目标跟踪性能提高的有效性，未来在智慧交通、智慧城市、军事战争等诸多领域具有广泛应用场景。

关键词: 多目标跟踪, 视频卫星, 图结构, 时空信息, 运动特征, 外观特征

Abstract:  With the widespread use of area array detectors，multitarget tracking for video satellites has become of great significance.However，for multi-target tracking methods based on graph structure，in the construction of graphs，most of them extract clues from adjacent frames，ignoring the previous frame clues.In response to the above problems，an end-to-end graph network framework was proposed to construct the nodes，edges and global variables of the graph，using various clues such as motion features，appearance features，and topology information extracted from multiple frames.A key principle to realize this unified framework is to design compatible feature representations and graph network update mechanisms for different clues and different sources（trajectories and detection targets）.The framework operated in a feed-forward fashion and trained on line.Being evaluated on the public datasets VISO，MOT16，MOT17 benchmarks，the multi-target tracking accuracy of 99.8%，48.8% and 51.8% was achieved respectively，which was better than other related multi-target tracking algorithms.And the ablation experiments were used to verify the improvement that each clue tracks multiple targets.The effectiveness of performance improvement will have a wide range of application scenarios in many fields such as smart transportation，smart cities，and military warfare in the future.

Key words: multi-target tracking, video satellite, graph structure, spatiotemporal information, motion feature, appearance feature