UAV-Mamba: Efficient Mamba for Real-time Infrared Small Target Detection from UAV Perspective

The proliferation of unmanned aerial vehicles (UAVs) for surveillance and reconnaissance has heightened the demand for effective detection of small infrared targets. While previous efforts have achieved high accuracy, they fail to adequately address the imperative for real-time processing requirements in UAV operations. In this paper, we introduce the UAV-Mamba framework, designed to address this challenge from three perspectives: the efficient Mamba block, lightweight DS backbone and multi-scale feature learning architecture. Additionally, we present the UAV-SIRST dataset, tailored for UAV-based Single-frame InfraRed Small Target (SIRST) detection, to further support research in this area. Our experimental results on several benchmarks demonstrate that UAV-Mamba achieves real-time performance without sacrificing accuracy, outperforming existing methods. Moreover, we have successfully deployed the proposed framework on the Jetson AGX Orin platform, further confirming its computational efficiency and practicality for real-world UAV applications. The code and dataset are available at https://github.com/ram-hash/UAV-MAMBA.