YOLO-D2: YOLO Model for Drone Object Detection

Abstract

With the increasing use of drones in our current era, it has become diffcult to detect them in light of the tough and diffcult conditions, especially if they are small in size, due to their height and distance from the perspective of regular cameras whose mission is limited to capturing shapes, colors and fine-texture over limited distances, and thus they occupy a very small area of the image, which makes it hard to detect them and distinguish them from airplanes and even birds. For this purpose, this work propose a system named YOLO-D2 capable of detecting drones even the small ones, with the help of deep learning models combined with event cameras concept. This latter can capture and track the path of the drone and monitor their motion alongside regular cameras. The experimental results showed the superiority of the YOLO-D2 system over traditional detection systems based on RGB only, such as the YOLOv8 model. The method of merging the extracted features from RGB images with the extracted ones from event camera contributed to improving detection accuracy and increasing performance effciency, the YOLO-D2 system achieved a precision of 0.789, with a recorded recall rate of 0.687, mAP@0.5 value of 0.737 on the validation set and a value of 0.629 on the test set, and achieved the best frame rate per second (FPS) among the fusion models of 29.95. However, the system relies on synthetic event data rather than true event camera recordings, and was developed and evaluated on a single dataset (SafeShore), limiting the extent to which generalization to other domains and real event-sensor data could be assessed