Multiframe spatio-temporal attention motion-adaptive network for moving space target detection

Space target detection based on optical observations is a fundamental approach for space situational awareness. However, in ground-based optical images, space targets often resemble stars, making single-frame differentiation challenging. The dense stellar background with high dynamic range further complicates object extraction. In real-world scenarios, varying relative motion between objects and the observation platform induces diverse apparent speeds, scales, and shapes, posing challenges for existing methods to exploit multiframe motion cues effectively, thereby limiting detection performance. To address these challenges, we propose the Multiframe Spatio-temporal Attention Motion-adaptive Network (MSAMNet) for robust small moving target detection in complex backgrounds. MSAMNet integrates Adaptive Attention Feature Enhancement (AAFE) and Spatio-Temporal Dynamic Motion-Aware (STDMA) modules to enhance spatio-temporal feature representation. AAFE generates attention maps based on channel-wise differences and local feature distributions to suppress noise and highlight target details. STDMA captures motion features across consecutive frames through exponential decay weighting and multi-scale heterogeneous kernel convolution, improving sensitivity to diverse motion patterns. Furthermore, we introduce BUAA-MSOD, the first open-source dataset of multiple moving space objects in real-world scenarios, covering diverse motion patterns and target morphologies with both mask and point-level annotations. Experimental results demonstrate that MSAMNet significantly outperforms state-of-the-art methods on BUAA-MSOD, achieving higher detection accuracy and lower false alarm rates across various space targets.