In recent years, Transformer-based target tracking algorithms have performed well in ordinary scenarios. However, when applied to satellite video scenarios, the tracking effect of the algorithms is not satisfactory due to the small size of satellite video targets, blurred features, and complex background interference. To address this issue, this paper proposes an algorithm for Enhancing Spatial Perception for Satellite Video Target Tracking (ESPTrack). This algorithm, through the spatial collaborative attention module, integrates local and global spatial information to enhance the multi-level representation of the target’s detailed features and overall structure. Meanwhile, a Gaussian prior cross-attention module is constructed. The Gaussian distribution weighting is utilized to enhance the key context information, improving the model’s ability to perceive the target’s spatial position and reducing the impact of background interference. To verify the effectiveness of the algorithm proposed in this paper, experiments were conducted on the satellite video datasets SatSOT and OOTB. The results show that the proposed algorithm has better performance compared with the existing target tracking algorithms, and it is verified that enhancing spatial perception in complex satellite video scenarios can effectively improve tracking performance.
Perpendicular-orbit circular scanning satellites overcome the conflict between ground resolution and width observed in traditional optical satellites by using a perpendicular-orbit circular scanning imaging method and splicing along the orbit, achieving a balance between an ultra-large width and a high resolution. However, laboratory calibrations of perpendicular-orbit circular scanning satellites exhibit large errors due to the influence of factors such as the thermal and mechanical environment of space during the launch and operation of satellites, and thus, they cannot be applied. In this paper, we start by analysing the in-camera azimuth element errors of perpendicular-orbit circular scanning satellites, then derive a probe element pointing angle calibration model from the physical in-camera calibration model and carry out in-camera parameter calibration based on simulated image data from an ultra-wide perpendicular-orbit circular scanning satellite. Edge and centre strips were selected for the experiment, and a certain number of control points were placed uniformly near the middle column (perpendicular orbit) of the image in each strip and covering all row directions (along orbit). Checkpoints were uniformly selected across a range of widths. The results show that in-orbit geometric calibration can significantly improve the direct-to-ground positioning accuracy of perpendicular-orbit circular scanning satellites, with the positioning accuracy error shown to be better than 30 m within a width of 300 km, 30 m within a width of 1000 km, and 50 m within a width of 2000 km.
11 December 2025 in Astrophysics and Space Science
To address the challenges of low signal-to-noise ratio and poor localization in detecting faint celestial objects from ground-based optical images, we propose CRFusion-Det, a plug-and-play probabilistic detection head. It introduces innovations at both the feature representation and inference levels. First, dilated convolutions and the CBAM attention module are integrated into the heatmap and width-height regression branches to enhance multi-scale contextual perception. Second, for offset estimation, the keypoint coordinate regression is innovatively reformulated as a probability distribution modeling problem. This is achieved via a learnable Prospect/Background Probability Estimation Module (PBPEM) and a Spatial-Appearance message Transmission Module (SATM), which explicitly capture inter-target geometric constraints and appearance consistency. A mean-field iterative algorithm is employed for structured inference, enabling progressive distribution refinement and sub-pixel localization. Extensive experiments on real-world datasets demonstrate CRFusion-Det’s effectiveness and generalization. When integrated into five different baseline networks, it consistently improved the recall by 1.68%–5.24% and reduced the normalized mean error to as low as 0.05 (0.73 pixels). The proposed CRFusion-Det significantly enhances the detection and localization accuracy of baseline models for faint targets, validating its superiority as a solution for astronomical image processing.