Deep Learning-Based Real-Time Detection and Localization for Targets in UAV Remote Sensing Images

Abstract

The rapid and accurate recognition and positioning of targets in UAV remote sensing imagery is a critical challenge in fields such as precision agriculture, disaster monitoring, and urban planning. This paper presents a deep learning-based algorithm research for the fast recognition and precise positioning of targets in UAV remote sensing images. We propose an optimized single-shot multibox detector (SSD) architecture integrated with an attention mechanism to enhance feature representation for small and densely distributed targets in complex backgrounds. The algorithm incorporates a feature pyramid network (FPN) to leverage multi-scale features, improving detection accuracy across varying target sizes. Additionally, we design a lightweight backbone network to ensure computational efficiency, enabling real-time processing on embedded platforms commonly deployed on UAVs. The proposed method is trained and validated on a custom dataset comprising diverse UAV-captured scenes, demonstrating a significant improvement in both inference speed and detection precision compared to existing approaches. Experimental results show that our algorithm achieves a mean average precision (mAP) of 89.7% with a processing speed of 32 frames per second on a single GPU, striking an effective balance between accuracy and efficiency. This research provides a practical solution for real-time UAV remote sensing applications, offering substantial potential for autonomous monitoring and rapid response systems.