Multi-Object Tracking in Satellite Videos with Multi-Perception Fusion and Gaussian Process Regression

As an emerging topic in remote sensing, multi-object tracking in satellite video can provide important support for dynamic earth observation. The major challenges stem from the diverse sizes of objects and sparse pixel occupancy in wide satellite observation areas, leading to a high frequency of missing detection and severe false alarms, and thus fragmented trajectories. In this study, a novel multi-object tracking technique (MP-GRMOT) is proposed, under the framework of detection-based multi-object tracking (DBT). Firstly, for objects of diverse sizes, task perception, scale perception, and spatial perception are combined into a single framework via a self-attention mechanism based on a dynamic detection head. Additionally, for sparse pixel occupancy of the objects, a P2 layer is introduced into the shallower layers of the detection network to obtain high-resolution feature maps. Finally, for more accurate and complete prediction of the trajectories, Gaussian process regression is applied after Kalman filtering to reduce errors caused by the lack of observations while updating the trajectory state. MP-GRMOT provides competitive performance, as verified by thorough experimental analyses utilizing the typical satellite video dataset (VISO). MP-GRMOT can enhance the tracking accuracy (MOTA) and ID F1 score by roughly 3% in comparison to state-of-the-art techniques.