Article Details

Abstract

To address the deployment requirements of traffic sign detection tasks in resource-constrained scenarios, this paper proposes a lightweight traffic sign detection algorithm based on YOLOv26. Using YOLOv26n as the base model, the CWD knowledge distillation technique is employed, with YOLOv26m serving as the teacher model to train the student model, achieving favorable results. To further enhance model performance, the paper explores the introduction of CBAM and SimAM attention mechanisms, including direct training of attention modules and the adoption of homogeneous distillation strategies. However, experimental results indicate that these improvement schemes fail to meet expectations. Analysis reveals that YOLOv26n, as a lightweight model, has limited capacity and struggles to effectively learn attention mechanisms, while homogeneous distillation cannot provide effective guidance for the student model due to the suboptimal performance of the attention-equipped teacher model. Ultimately, the paper adopts a pure CWD distillation approach. Experiments on the TT100K dataset demonstrate that the YOLOv26n-CWD model achieves an mAP@0.5 of 68.98% and an mAP@0.5:0.95 of 51.74%, representing improvements of 0.95% and 0.20% over the baseline model, respectively, meeting the dual requirements of accuracy and speed in practical applications.

Keywords

traffic sign detection; YOLO26; Attention Mechanism; knowledge distillation; TT100K dataset

References

1. [1] Yang Feifan, Li Jun A Review of YOLO Object Detection Algorithm for Autonomous Driving [J]. Automotive Engineer, 2023 (11): 1-11
2. [2] Guo Fan, Zhang Yongxiang, Tang Wan, etc YOLOv3-A: Traffic Sign Detection Network Based on Attention Mechanism [J]. Journal of Communications, 2021, 42 (1): 87-99
3. [3] Liang Xin, Dai Qian Improved YOLOv3 and Traffic Sign Detection Based on Attention Mechanism [J]. Modern Computer, 2022, 28 (15): 9-16
4. [4] Huang Shang'an Traffic sign detection based on improved YOLO model [J]. Science and Technology Innovation, 2021 (18): 194-196
5. [5] Zhou Zhuofeng Research on Traffic Sign Detection Method Based on Lightweight YOLO [D]. Xi'an: Chang'an University, 2021
6. [6] Liu Liwei, Du Lei Traffic sign detection and recognition algorithm based on improved YOLOv8 [J]. Journal of Changchun University of Technology, 2025, 46 (5): 1-6
7. [7] Sheng Jingwen Research on Lightweight Traffic Sign Recognition Algorithm Based on YOLOv11 [D]. Hefei: Anhui Jianzhu University, 2025
8. [8] Fang Tianrui, Cheng Guang, Liu Hailin, etc YOLOv8n Lightweight Collaborative Improvement Model for Traffic Sign Detection [J]. Computer Engineering and Applications, 2025, 61 (23): 233-247
9. [9] Chen Zhiwei, Liu Gang, Wang Mengjiao Traffic target detection based on improved YOLO network [J]. Computer and Digital Engineering, 2025, 53 (11): 3133-3138
10. [10] Peng Jie, Yu Huijun, Xia Meng, etc Improving the Small Target Traffic Sign Detection Method of YOLO11n [J]. Journal of Chongqing University of Technology (Natural Science), 2025, 39 (11): 145-153
11. [11] Hua Haojun, NAJI Alhusaini. Lightweight Fall Detection Model Based on Unstructured Pruning and Feature Distillation [J]. Journal of Yancheng Institute of Technology (Natural Science Edition), 2025, 38 (3): 1-5
12. [12] Zhang Haiqing, Zhang Lin, Li Daiwei, etc Small target traffic sign detection algorithm integrating progressive scale modeling and deformable attention [J/OL]. Microelectronics and Computer, 2025:1-12 https://link.cnki.net/urlid/61.1123.tn.20251203.1928.004.

About the authors

Kong Qin (1983-), female, Han nationality, from Taixing, Jiangsu, associate professor, master of engineering, research direction: application of artificial intelligence technology, data analysis; Wu Haoyang (2003-), male, Han ethnicity, from Suzhou, Jiangsu, is currently pursuing a bachelor's degree with a research focus on the application of artificial intelligence technology.