Deep feature based correlation filter for single object tracking in satellite videos

Abstract

In computer vision, single object tracking in satellite video is a trending research topic with challenging problems. The high-resolution remote sensing camera captures spatiotemporal information from Earth, where tracking of interested targets is possible. Due to nadir view, single object tracking in satellite videos faces model drift problem due to limited features of tiny objects. Rotation of object, background clutter and long period occlusion of moving objects lead to model drift problem. This study proposes a kernelized correlation filter framework-based tracker using modified ResNet50 with spatial pyramid pooling as feature extractor which solve the short rotation problem and get the shape, textures and patterns features of small size object. To solve long period occlusion problem, this paper proposes an average of average based motion estimation method with Kalman filtering. The proposed method was implemented, tested and quantitative experiments were conducted on satellite video SatSOT dataset. From the obtained results, it is found that the proposed tracker produces better performance in comparison to state-of-the-art tracker methods available in literature in terms of success score, precision score and speed. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.