Improved SORT for vehicles tracking in satellite videos

Abstract

Multi-object tracking in satellite videos has been widely used in civilian and military fields. Among them, the tracking of vehicles has important applications in the field of traffic monitoring. However, the tracking of vehicles in satellite videos still remains challenging and unsolved due to the extremely small size and the lack of appearance and geometric features. In this paper, we propose an improved SORT to tackle the tracking of vehicles in satellite videos by introducing C3D to CenterNet to improve the detection performance and promote the overall tracking performance. Specifically, we use C3D as the backbone of CenterNet to extract spatio-temporal information and use a 3D channel attention mechanism to fuse the information extracted by C3D to improve the detection performance, thereby improving the tracking results. The qualitative and quantitative results of experiments on videos of Jilin-1 satellite constellation show that our method can efficiently improve the tracking performance of vehicles in satellite videos.