YOLO-VSSBlock: A YOLOv8m-Based UAV Object Detection Method with Visual State Space Modelling
Abstract
Object detection in Uncrewed Aerial Vehicle (UAV) imagery remains challenging due to densely distributed small objects, complex backgrounds, substantial scale variations, and frequent occlusions. Although You Only Look Once (YOLO) based detectors achieve a favourable balance between detection accuracy and computational efficiency, their convolution-dominated feature extraction models long-range spatial dependencies less explicitly. This paper presents YOLO-VSSBlock, a YOLOv8m-based detector that selectively integrates visual state-space modules into the P5 path. Specifically, the backbone P5 C2f module and the final P5 fusion C2f module in the neck are replaced with VSSBlock and XSSBlock, respectively, while the original P3 and P4 structures are retained. This placement enhances contextual modelling at the low-resolution semantic stage while preserving fine-grained local features in the high-resolution branches. Experiments on VisDrone2019 show that YOLO-VSSBlock improves mAP@0.5 from 0.426 to 0.446 and mAP@0.5-0.95 from 0.258 to 0.272. Theoretical complexity increases from 79.1 to 81.2 GFLOPs (approximately 2.7%), indicating an improved accuracy-complexity trade-off without implying device-level acceleration.