Lin Xiuxiu, Niu Yusu, Yu Xinran, Fan Zhun, Zhuang Jiafan, Zou An-Min
College of Engineering, Shantou University, Shantou, 515063, China.
University of Electronic Science and Technology of China, Chengdu, 611731, China.
Neural Netw. 2025 May;185:107182. doi: 10.1016/j.neunet.2025.107182. Epub 2025 Jan 21.
Under the advancement of artificial intelligence, Unmanned Aerial Vehicles (UAVs) exhibit efficient flexibility in military reconnaissance, traffic monitoring, and crop analysis. However, the UAV detection faces unique challenges due to the UAV's small size in images, high flight speeds, and limited computational resources. This paper introduces a novel Background-centric Attention Module (BAM) to address these challenges. Unlike traditional methods relying on UAV visual features, the BAM utilizes complex background information to identify UAV presence. The BAM seamlessly integrates into existing UAV detection frameworks, improving accuracy with no significant increase in the computation time. Extensive experiments on challenging datasets, Naval Postgraduate School Drones (NPS), and Flying drones (FLDrones) using mainstream detectors YOLOv5 and TphPlus demonstrate the effectiveness of the BAM in significantly enhancing detection accuracy. This research emphasizes the importance of background information in the UAV detection and proposes a method aligning with human perceptual processes, paving the way for further advancements in the field.
