Satellite-borne Vehicles Tracking: A Motion Position Graph Approach and Dataset.

Video satellites have opened a new era for city-scale traffic monitoring. However, in the broad and aerial field-of-view of satellite video, vehicles appear extremely dim and dense, often comprising only several pixels. Even worse, the noise signals, illumination fluctuations, and cloud interferences are ubiquitous in the video frames. Considering the limited spatial information of vehicles in satellite videos, this paper fully exploits the inter-frame "spatio-temporal consistency" inherent in videos for multi-object tracking in dim and dense environments. Our contributions are as follows: (1) We build a Velocity-labeled Dim and Dense Vehicles (VDD-VEH) dataset. This dataset comprises 70,665 images and 15,461 instances, and provides velocity annotations for accurate motion flow estimation. (2) We propose a Motion Position Graph (MPG) that jointly represents both temporal and spatial position information of all vehicles across the entire satellite video. Specifically, we map object detection positions and motion flows into a graph, and then use Integer Linear Programming (ILP) and our proposed Trajectory Refinement Module (TRM) to conduct global optimization and association within this graph. Experimental results demonstrate that our MPG tracker achieves state-of-the-art accuracy, with a MOTA of 71.4% and an IDF1 of 70.5%. Besides, our dataset and method are fully available at: https://github.com/yinz-creator/MPGTracker.