RAD-YOLO: a lightweight edge-deployed visual measurement method for small-object detection in UAV imagery

Abstract

Small-object detection in unmanned aerial vehicle (UAV) remote-sensing imagery remains difficult because targets often have low spatial resolution, dense spatial distribution, partial occlusion and strong background clutter. These factors weaken discriminative features and restrict real-time inference on embedded edge platforms. To address the coupled requirements of detection accuracy and deployability, we propose RAD-YOLO, a YOLOv8s-based small-object detector, and develops an edge-deployable variant named RAD-YOLO-Slim. An adaptive-field convolutional block attention module is first introduced into the backbone to enhance multi-scale receptive-field modelling and joint spatial-channel representation. A gated bidirectional asymptotic feature pyramid network is then designed to improve cross-scale semantic interaction while suppressing the propagation of shallow background responses. A distance-normalised Wise-IoU (WIoU) loss and soft non-maximum suppression are further combined to refine bounding-box localisation and retain adjacent detections in dense small-object scenes. Finally, group-shuffle convolution reconstruction, layer-adaptive magnitude-based pruning and 8-bit integer (INT8) post-training quantisation are used to construct RAD-YOLO-Slim for resource-constrained edge deployment. Experiments on VisDrone2019-DET show that RAD-YOLO achieves 47.6% mean average precision at an IoU threshold of 0.5 (mAP@0.5), an improvement of 8.6 percentage points over YOLOv8s under the same protocol. RAD-YOLO-Slim maintains 44.1% mAP@0.5 with 5.42 M parameters and 6.2 giga floating-point operations (GFLOPs). After INT8 deployment on an RK3576 platform, RAD-YOLO-Slim achieves 43.1% mAP@0.5, 51 frames per second (FPS) and 11.81 FPS/W. On the RK3576 platform, RAD-YOLO-Slim provides a practical balance of accuracy, speed and energy efficiency.