DETECTION AND RECOGNITION OF ROAD SIGNS USING YOLOv5

Ettazi, Haitam; Rafalia, Najat; Abouchabaka, Jaafar · 2023 · DOAJ

DOI: 10.24874/PES05.03.017

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper evaluates the performance of the YOLOv5s deep learning model for the detection and recognition of traffic signs, a critical component for Advanced Driver Assistance Systems (ADAS) and autonomous vehicles. The study is motivated by the need for real-time object detection algorithms that balance speed and accuracy, addressing the limitations of slower two-stage detectors like R-CNN. The authors specifically aim to assess the inference capabilities and accuracy of YOLOv5s across various environmental conditions and compare its effectiveness against previous iterations like YOLOv4. The experimental design utilized a subset of the Mapillary Traffic Sign Dataset, selecting 5,772 annotated images from five specific classes: keep-right, no-entry, yield, pedestrians-crossing, and stop. The YOLOv5s model was trained on Google Colab using a Tesla T4 GPU. The authors conducted experiments varying batch sizes (32, 64, and 80) and training epochs (up to 200) to determine optimal hyperparameters. The model architecture leverages a CSPNet backbone for feature extraction and a PANet neck for feature aggregation, designed to improve processing speed and reduce gradient duplication. Performance was evaluated using standard metrics including precision, recall, and mean Average Precision (mAP). The results indicate that the YOLOv5s model achieved a mean Average Precision (mAP) of 63.7% across the five tested classes. The study found that a batch size of 64 yielded superior performance compared to a batch size of 80, achieving precision rates above 80% for the categories, whereas the larger batch size resulted in precision below 80%, likely due to insufficient training data and time. Specific precision rates for the five classes ranged from 63.2% to 88.3%. The authors note that the model exhibited slight overfitting, with optimal convergence occurring around the 76th epoch. When compared to a prior study using YOLOv4 on the Indonesia Traffic Signs dataset, which reported an mAP of 74.91% but lower precision (74%), the current study suggests YOLOv5 offers competitive precision advantages despite the different datasets. The significance of this work lies in demonstrating the viability of YOLOv5s for real-time traffic sign recognition, highlighting its speed and efficiency on low-power devices. The authors conclude that while YOLOv5 is robust and capable of handling diverse lighting and weather conditions, it remains susceptible to errors from reflections or visually similar non-sign objects. The study underscores the importance of hyperparameter tuning, specifically batch size, and identifies the reliance on large, labeled datasets as a primary limitation. Future research is recommended to improve robustness in challenging conditions such as low lighting or cluttered backgrounds.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• sign visibility legibility