Drone Detection and Tracking: An Edge-Deployable Efficient Algorithm Based on Vision Sensor

The widespread use of drones has intensified the demand for effective counter-drone systems to mitigate security and privacy threats. Detection methods based on vision sensors offer economic and practical advantages but often struggle with the accurate identification of small, fast-moving aerial targets, especially under real-time constraints on edge devices. In order to overcome these challenges, You Only Look Once (YOLO)-DroneDet is proposed, which is an efficient drone detection framework built upon YOLOv8 and specifically tailored for vision sensor-based ground-to-air applications. The framework integrates two key innovations: a multiscale dense connection (MDC) module to improve feature propagation for small-object detection and a multilayer shared head (MSH) to reduce detection head parameters, enhancing deployment efficiency on edge platforms. Furthermore, model pruning and knowledge distillation are employed to compress the model while maintaining high accuracy. A gimbal-mounted vision sensor is also incorporated to dynamically expand the field of view, addressing the limitations of static sensing setups. The experimental results show that YOLO-DroneDet improves the detection performance, achieving an mAP50:95 of 54.9%—a 6.7% increase over the baseline YOLOv8—while maintaining real-time inference at 14.7 ms/frame on the NVIDIA Jetson AGX Orin. This work presents a practical solution for real-time drone detection and tracking using vision sensors. Dataset and demo: https://github.com/ballballubsmart/YOLO-DroneDet