A New Framework for Enhancing Anti-Drone Detection Performance

Abstract

The spread of unmanned aerial vehicles (UAVs), termed drones, challenges modern security systems. Their small size complicates reliable detection in diverse environments. High maneuverability allows evasion of conventional radar-based methods. Low observability reduces detectability in cluttered settings. Conventional approaches struggle under variable weather and terrain conditions. Advances in computer vision address these limitations through the use of real-time strategies. Deep learning frameworks like YOLO excel in rapid object localization. These frameworks classify objects accurately in complex scenes. This paper integrates YOLOv8 with the Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA-Net), and Squeeze-and-Excitation (SE-Net). The multi-stage configuration combines these advanced attention modules. The framework specifically targets anti-drone system requirements. It enhances the detection of small, fast-moving UAVs reliably. Experimental results validate the framework's effectiveness clearly. This paper systematically embeds CBAM, ECA-Net, and SE-Net into YOLOv8. Multiple model configurations test different attention mechanism placements. Experiments confirm enhanced UAV detection capability and reliability. The dataset comprises over 1,300 UAV images captured across diverse environments and altitudes. It includes challenging scenarios such as cluttered backgrounds, low visibility, and varying drone scales. The framework improves both recall and overall accuracy. Optimal configurations for YOLOv8 + CBAM <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(2,3,5)+$</tex> ECA <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(7,9)$</tex> improved recall to 92.9%. Another model YOLOv8 + CBAM (1) +ECA (5)+SE (18) achieved a precision of 98.6%. The framework maintains real-time processing speeds despite added complexity.