EMA-WIoU-YOLO11: Efficient Multi-Scale Attention and Geometry-Aware Regression for Public-Space Obstacle Detection

Reliable detection of diverse obstacles in unstructured public spaces is critical for assistive navigation systems serving visually impaired individuals and autonomous platforms. Existing approaches face three key limitations: (1) vanilla single-scale attention mechanisms fail to discriminate visually similar obstacle pairs under varying illumination and occlusion; (2) standard IoU-based regression losses assign equal gradient weight to geometrically inferior proposals, degrading localisation in crowded scenes; (3) prior benchmark models are validated on narrow-category datasets that do not reflect the fine-grained 22-class diversity encountered in real-world urban environments. To address these challenges, we propose EMA-WIoU-YOLO11, an enhanced detection framework built upon the YOLO11n backbone. Our key innovations are threefold: (1) we embed an Efficient Multi-Scale Attention (EMA) module after the final C3k2 backbone stage to jointly capture global channel statistics and local spatial context with negligible parameter overhead; (2) we replace the standard box regression loss with the Wise-IoU (WIoU) formulation, which introduces a sample-wise focusing weight to suppress geometrically inferior anchors; (3) we provide the first systematic evaluation on the 22-class Dist-YOLO Public-Space Obstacle dataset comprising 3,350 annotated images. Experiments demonstrate that our model achieves mAP@0.5=0.849, precision=0.90, and recall=0.80, with peak F1=0.81 at a confidence threshold of 0.374, confirming the effectiveness and practical applicability of the proposed framework.