# ESA-YOLO: An efficient scale-aware traffic sign detection algorithm based on YOLOv11 under adverse weather conditions

## TL;DR

This paper introduces ESA-YOLO, an improved traffic sign detection algorithm that performs better in complex and adverse weather conditions while maintaining efficiency.

## Contribution

The paper proposes three novel components: DMFPN, CAGB, and ASPH to enhance multi-scale detection and robustness in traffic sign recognition.

## Key findings

- ESA-YOLO outperforms YOLOv11n by 3.8% in mAP@50 and 3.9% in mAP@50-95 on the TT100K dataset.
- The model achieves 2.3% and 1.8% improvements in mAP@50 and mAP@50-95 on the CCTSDB2021 dataset.
- ESA-YOLO shows superior small-object detection and robustness in adverse weather conditions.

## Abstract

Traffic sign detection is a critical component of autonomous driving and advanced driver assistance systems, yet challenges persist in achieving high accuracy while maintaining efficiency, particularly for multi-scale and small objects in complex scenes. This paper proposes an improved YOLOv11-based traffic sign detection algorithm that tackles above challenges through three key innovations: (1) A Dense Multi-path Feature Pyramid Network (DMFPN) that boosts multi-scale feature fusion by enabling comprehensive bidirectional interaction between high-level semantic and low-level spatial information, augmented by a dynamic weighted fusion mechanism. (2) A Context-Aware Gating Block (CAGB) that efficiently integrates local and global contextual information through lightweight token and channel mixer, enhancing the small-object detection ability without excessive computational overhead. (3) An Adaptive Scene Perception Head (ASPH) that synergistically combines multi-scale feature extraction with attention mechanisms to improve robustness in adverse weather condition. Extensive experiments on the TT100K and CCTSDB2021 datasets demonstrate the model’s superior performance. On the TT100K dataset, our model outperforms the state-of-the-art YOLOv11n model, achieving improvements of 3.8% in mAP@50 and 3.9% in mAP@50-95 while maintaining comparable computational complexity and reducing parameters by 20%. Similar gains are observed on the CCTSDB2021 dataset, with enhancements of 2.3% in mAP@50 and 1.8% in mAP@50-95. Furthermore, experimental results also demonstrate that our proposed model exhibits superior performance in small object detection and robustness in complex environments compared to mainstream competitors.