Abstract
<title>Abstract</title> <p>Vehicle detection using Unmanned Aerial Vehicles (UAVs) is crucial for applications such as traffic monitoring, smart parking, and search and rescue. However, it remains challenging due to weak discriminative features of small or low-resolution vehicles, information imbalance caused by scale variations and category distribution, and the demand for real-time processing. To address these challenges, this paper introduces a novel global-local feature aggregation network with hybrid attention (GLFA-Net). The framework achieves real-time, high-performance detection via two core components: a bidirectional global-local feature aggregation network (BGLA-Net), which integrates global and local features through top-down and bottom-up pathways to enhance the representation of small targets while balancing spatial and semantic information across layers; and a parallel hybrid attention module (PHAM), which refines features in both spatial and channel dimensions to suppress background interference and accentuate foreground regions, thereby producing more discriminative features for small-vehicle detection. Experimental results demonstrate that GLFA-Net achieves state-of-the-art performance on three public benchmarks, with AP scores of 67.2% on XDUAV, 71.2% on UAVDT, and 62.5% on Stanford Drone, while running at a real-time speed of 52 FPS.</p>