MBD-YOLO: A Multi-Path Coordinate Attention and Boundary-Aware Dual-Stream Fusion YOLO for UAV Small Object Detection

Abstract

UAV imagery is particularly difficult for object detection because targets occupy few pixels, occur in dense or occluded groups, and must be separated from extensive background texture. Adding a high-resolution detection layer improves the visibility of tiny instances, but it also carries road markings, building contours, shadows, and other distracting structures into the feature pyramid. We address this trade-off by developing MBD-YOLO from the P2-based SOD-YOLO detector. A Dual-Stream Attention Module (DSAM) transfers deep semantic cues to the high-resolution path through complementary foreground- and background-priority streams. A Multi-Path Coordinate Attention (MPCA) module extends horizontal and vertical coordinate encoding with a two-dimensional spatial-response path. A training-only Boundary-Aware Branch (BAB) further supervises the enhanced high-resolution feature with fused multi-scale Laplacian edges and is removed for inference. Experiments on VisDrone2019-DET, TinyPerson, and AI-TOD show consistent improvements over YOLOv8n and the direct SOD-YOLO baseline under the common evaluation setting. On VisDrone2019, MBD-YOLO achieves 0.397 Recall, 0.485 mAP50, and 0.308 mAP50:95, exceeding SOD-YOLO by 0.053, 0.094, and 0.068, respectively. Its measured pipeline speed is 76.44 FPS, compared with 98.13 FPS for SOD-YOLO, at $640\times 640$ resolution with batch size 1 on an RTX 3090.