Towards lightweight and accurate deep neural network for low-altitude flying unmanned aerial vehicle detection

Abstract With the increasingly wide use of unmanned aerial vehicles (UAVs) in low-altitude airspace, the detection of low-altitude flying UAVs is significant to avoid collisions between UAVs. At present, deep learning has made great progress in the field of object detection. However, existing approaches still cannot meet the practical needs of low-altitude UAV detection due to the various scales, complex backgrounds and model complexity. To address these challenges, we propose an enhanced version of YOLOv7, incorporating multiple receptive fields, parameter-free attention, and lightweight down-sampling modules (LDM) for effective UAV detection at low altitudes. Firstly, a multiple receptive fields module is constructed to enrich the receptive fields of feature maps at different levels. Secondly, a simple, parameter-free attention module is introduced to highlight the key features for UAV detection and suppress the influence of interference factors in the background. Thirdly, an LDM based on depthwise separable convolution is adopted to reduce the model complexity. Finally, a low-altitude flying UAV dataset is constructed to verify the detection performance. Compared with other state-of-the-art methods, the proposed method has better performance in terms of average precision (AP), network parameters, model size and floating-point operations. The experimental results demonstrate the efficiency and superiority of the proposed method in detecting low-altitude UAVs.