ASTRA-Net: Adaptive Spatiotemporal Track-Before-Detect Recurrent Attention Net for UAV

Stable tracking of maneuvering targets, such as Unmanned Aerial Vehicles (UAVs), in low Signal-to-Noise Ratio (SNR) environments remains a significant challenge. Weak target signals, often obscured by noise, frequently lead to failures in multi-frame energy accumulation, substantial state estimation errors, and degraded tracking performance. To address these limitations, this paper proposes the Adaptive Spatio-temporal Tracking Recurrent Attention Network (ASTRA-Net), an adaptive dual-task learning extension of the Track-Before-Detect (TBD) framework. ASTRA-Net integrates spatiotemporal feature fusion embedding with joint pre-detection and tracking. It captures temporal and spatial dependencies within noisy radar measurements and employs a multi-head self-attention mechanism to dynamically recalibrate feature weights based on varying motion patterns. This framework shares underlying features without stringent prior dynamic assumptions, enhancing overall robustness. It simultaneously learns trajectory prediction and state recognition, discriminates between model states using an information entropy threshold, and adaptively adjusts TBD model parameters during abrupt maneuvers. Experimental results demonstrate that ASTRA-Net excels in low-SNR, noise dominated environments, better than conventional methods in tracking accuracy. This approach offers a novel technical solution for the perception and tracking of low observable targets, particularly suitable for security surveillance and airspace management systems operating in noisy conditions.