遥感卫星视频中的目标小,场景变化多样,在遥感卫星视频上的进行多目标跟踪存在一定的困难。针对卫星视频中目标小的特点,提出了一种高分辨率遥感卫星视频的多目标跟踪算法,该算法先检测出疑似目标再进行轨迹关联。检测阶段构建小目标检测器,首先在骨干网络中通过Transformer捕获全局的上下文信息,然后利用注意力机制增强目标特征,最后添加了一个用于预测小目标的分支；为使检测出的小目标与已有轨迹匹配,轨迹关联阶段,采用关注低置信度检测的弱关联算法。本文选取高分辨率遥感卫星视频(Gaofen Challenge,2021)进行实验,实验结果表明,本文提出的方法在遥感卫星视频中的多目标跟踪数据集上的MOTA指标达到63.1%,相较于基准(baseline)模型提升13.5%,能够显著提升遥感卫星视频中多目标跟踪的性能。

Multi-object tracking in high-resolution remote sensing satellite video, aiming at estimating the position and identities of objects, has many applications such as security monitoring, automatic driving, and intelligent transportation. However, compared to the surveillance video in nature scene, the objects in remote sensing satellite video are quite small and have few features, which makes it difficult for multi-object tracking. In addition, the size of remote sensing image is huge, which puts forward higher requirements on computation and storage. Multi-object tracking in high-resolution remote sensing satellite video faces higher real-time requirements.