Occlusion Resistant Transformer Based Network for Efficient Object Tracking in Satellite Videos

With the advancement of high-resolution remote sensing satellites, object tracking in satellite videos has emerged as a new area of study. However, problems like occlusion and background clutter render object tracking in these videos difficult. To address these issues, the paper proposes ORMET - Occlusion Resistance with Motion Estimation Transformer which brings together the best of currently prevalent object-tracking technologies to track minute objects in satellite videos. The prediction head introduced by this study extends the work done by SMAT and exploits both the appearance and motion features of the target. It presents a self and mixed attention-based transformer structure that extracts the appearance features from the target and the search patch and then combines them with the motion features of the target using an adaptive fusion strategy. Further, the accuracy is improved by taking the help of a trajectory estimator Kalman filter for cases when the target is completely occluded and no satisfactory results are obtained by the prediction head. Comprehensive experiments conducted on the SatSOT dataset conclusively show how well the suggested approach works when compared to the most advanced tracking techniques.