Advancing Multiobject Tracking for Small Vehicles in Satellite Videos: A More Focused and Continuous Approach

With the development of satellite videos, multiobject tracking (MOT) in satellite video is possible and has become a new challenging task. The challenges of the task are mainly from the characteristics of objects in satellite videos: 1) small objects; 2) similarity of objects and background; 3) sparse object distribution; and 4) various data quality. These characteristics make it difficult to apply MOT methods for ground-based video directly to satellite video. To tackle the problem, this article proposes a more focused and continuous two-stage satellite video MOT method named FoCos. In the detection part, a multiscale center-point-based small object detector assisted by multitype dynamic data is introduced. This detector uses the dynamic characteristics of vehicles and their correlation with the background to better localize the vehicles. Meanwhile, considering that vehicles are sparsely distributed, a feature de-redundancy convolution (FDC) module is applied within the detector. In the tracking part, considering the problem of poor generalization caused by an uneven quality between data, a quality-aware confidence hierarchy (QACH) module, as well as a two-stage data association module, is introduced to perceive the quality of the data and separate confidence into hierarchies for the association, improving the performance of the method under the different quality of videos. Meanwhile, a trajectory smoothing (TS) strategy is introduced to obtain more continuous trajectories. Finally, the proposed method is evaluated in three satellite video datasets and its superiority is demonstrated.