A self-attention driven multi-scale object detection framework for adverse weather in smart cities

一种基于自注意力机制的智能城市恶劣天气多尺度目标检测框架

阅读:1

Abstract

Detecting abandoned or small-scale objects under adverse weather conditions remains a significant challenge for maintaining road safety and urban security in smart cities. Existing detection systems often suffer performance degradation in the presence of fog, rain, snow, or sand, where reduced visibility and background noise obscure object boundaries. To address this limitation, this study proposes a Self-Attention Driven Multi-Scale Object Detection Framework, which establishes a tightly integrated pipeline combining the Adaptive Dual-Background Model (ADBM), the Pixel-based Finite State Machine (PFSM), and the Attention-based Scale Module (ASM) with the Self-Attention Optimized You Only Look Once (SAO-YOLO) network. Unlike conventional YOLO-based approaches that treat enhancement modules independently, the proposed framework enables holistic interaction among background refinement, feature selection, and attention-guided detection, leading to more stable and context-aware predictions. Furthermore, a thermal-visual feature fusion mechanism aligns thermal and visual representations at the P3 level of the Feature Pyramid Network (FPN), enhancing visibility robustness under poor illumination. Experimental evaluation on the Detection in Adverse Weather Nature dataset demonstrates superior performance, achieving 97.25% Accuracy, 95.33% Precision, 96.37% Recall, and 96.55% F1-score, outperforming recent state-of-the-art models such as YOLOv5, YOLOv8, and DEtection TRansformer. These results validate the effectiveness of the integrated ADBM-PFSM-ASM-SAO-YOLO design in maintaining high detection reliability under complex weather variations. The proposed model offers a practical solution for real-time traffic monitoring, urban surveillance, and public safety applications where detection consistency under environmental uncertainty is critical.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。