STLTrack: Dual-Memory Cooperative Perception for Robust UAV Object Tracking

As a pivotal technology within aerial intelligent perception systems, Unmanned Aerial Vehicle (UAV) object tracking confronts significant challenges in complex dynamic environments. Targets frequently undergo drastic appearance variations and prolonged occlusion, rendering existing methods incapable of simultaneously reconciling the conflicting requirements of rapid responsiveness and cross-temporal stability. Drawing inspiration from the synergistic mechanism of short-term and long-term memory in the biological visual system, we propose a Dual-Memory Cooperative Perception framework for robust UAV object tracking, termed STLTrack. By explicitly modeling the functional differentiation and complementarity of short-term and long-term memory across different temporal scales, STLTrack effectively alleviates cross-temporal instability caused by appearance fluctuations and insufficient utilization of historical information. We design a Short-term Memory Perception (SMP) module to rapidly focus on instantaneous salient regions in the current frame. The Long-term Memory Perception (LMP) module maintains high quality historical feature representations to provide stable semantic support over extended temporal horizons. Furthermore, a Dynamic Memory Gated Fusion (DMGF) module is designed to adaptively balance the contributions of short-term and long-term memory according to the current perceptual state, enabling flexible adaptation to diverse tracking scenarios. Extensive experimental results on multiple UAV tracking benchmarks demonstrate that STLTrack significantly outperforms existing state-of-the-art methods in terms of precision and success rate.