YOLOXpress: A lightweight real-time unmanned aerial vehicle detection algorithm

Abstract

The widespread use of drones has made drone detection a critical factor in various fields, particularly in security and defense. However, this task presents unique challenges due to the high speed, small size, and ability of drones to blend into their surroundings, which can hinder detection effectiveness. This paper introduces enhancements to the You Only Look Once (YOLO)-v8 model to improve real-time drone detection capabilities, especially when deployed on resource-constrained devices. We propose an improved model called YOLOXpress, which optimizes both processing speed and model size while maintaining an acceptable level of accuracy. By replacing the Cross-Stage Feature Fusion modules in the Backbone and Neck with Re-parameterization Convolution and RepC3 modules, we significantly reduced the number of computations, achieving a 12.25% increase in processing speed (frames per second) and a 69.96% reduction in model size. Although there was a 6% decrease in average accuracy compared to the original YOLO-v8 model, YOLOXpress remained effective for real-time drone detection. Experiments conducted on the TIB-Net dataset confirmed that this model is highly suitable for deployment on resource-limited devices, such as compact embedded systems.