Mob-YOLO: A Lightweight UAV Object Detection Method

Abstract

With the increasing use of Unmanned Aerial Vehicles (UAVs) in various fields, the coordinated execution of tasks by multiple UAVs has become an important development trend in the future. To avoid the collision of multiple UAVs with each other during flight and ensure flight safety, it is essential to be able to achieve high-precision, real-time airborne UAV object detection. In this work, a UAV object detection method called Mob-YOLO is proposed. Based on the high-performance model YOLOv4, MobileNetv2, a lightweight convolutional neural network, is used to replace the original YOLOv4 backbone CSPDarknet53 for model size reduction and computing operation simplification. Meanwhile, to solve the issue of poor accuracy for small UAV objects after network replacement, this work also designs a multi-scale feature extraction and fusion branch to expand the receptive field of the object detector by multi-scale feature fusion. The proposed method is evaluated using a self-built UAV dataset. The results demonstrate that Mob-YOLO can satisfy accurate real-time monitoring of UAV objects, and the model size is tiny, which can be used for deployment on airborne embedded processors.