STDNet: A spatio-temporal dual-stream network for detecting faint space objects from ground-based optical imagery

The accurate detection of faint, small, and moving space objects in ground-based optical imagery is critical for space situational awareness. However, this task faces significant challenges, including low signal-to-noise ratios, complex target dynamics, background motion, the scarcity of diverse and high-quality annotated datasets. Existing methods, including traditional and deep learning-based, often struggle with these complexities and lack robust end-to-end frameworks for diverse motion scenarios and enhanced representation of faint targets. To address these limitations, this paper introduces STDNet, a novel Spatio-temporal Dual-stream Network that adaptively processes sequence inputs of varying lengths. STDNet effectively decouples and integrates morphological and spatio-temporal features through a parallel architecture. The Static Stream employs a tree-structured hierarchical feature aggregation network to integrate cross-level and multi-scale information, enhancing the spatial feature modeling for faint targets. Meanwhile, the Dynamic Stream introduces a unified 3D spatio-temporal window attention mechanism, processing volumetric data from image sequences to directly and precisely characterize complex target motion patterns. Furthermore, a parallel fusion module using spatio-temporal attention integrates spatial and temporal cues, significantly improving detection accuracy and efficiency. To overcome the lack of extensive real-world data, we innovatively develop a high-fidelity data simulation methodology incorporating focal length inversion and a full-spectrum image degradation pipeline, generating thousands of space objects across varying signal-to-noise ratios for robust training and comprehensive evaluation. Extensive experiments demonstrate that STDNet achieves an F1 of 0.953 on real-world datasets, outperforming baselines particularly in detecting dim objects under low-SNR constraints. With a rapid inference time of 32.24 ms per frame, the model effectively balances high-precision detection with real-time survey requirements.