Traffic Sign Detection Model (YOLOv11)

Model Description

This model is an object detection model trained using the Ultralytics YOLOv11 framework. The model detects and classifies multiple types of traffic signs in road images by predicting bounding boxes and class labels.

The model was trained by fine-tuning the pretrained YOLOv11n architecture on a dataset of annotated traffic signs derived from the LISA Traffic Sign Dataset.

Training Approach

Base model: YOLOv11n
Framework: Ultralytics YOLO
Training method: transfer learning / fine-tuning
Task: object detection

Transfer learning allows the model to start with pretrained visual features learned from large datasets and then specialize those features for traffic sign detection.

Intended Use

This model is designed for:

traffic sign detection research
computer vision experimentation
academic coursework projects
demonstrations of object detection systems

Possible applications include:

driver assistance research
automated traffic sign recognition
road scene analysis

This model should not be used in safety-critical systems such as autonomous vehicles without extensive additional testing and validation.

Training Data

Dataset Source

The model was trained using images derived from the LISA Traffic Sign Dataset.

The LISA Traffic Sign Dataset is a publicly available dataset created for traffic sign detection and classification research. The dataset contains traffic sign images captured from real driving environments in the United States.

Dataset link:
https://cvrr.ucsd.edu/LISA/lisa-traffic-sign-dataset.html

Dataset Citation

Mogelmose, A., Trivedi, M. M., & Moeslund, T. B. (2012).
Vision-based traffic sign detection and analysis for intelligent driver assistance systems: Perspectives and survey.
IEEE Transactions on Intelligent Transportation Systems.

Classes

The model detects the following traffic sign classes:

doNotEnter
pedestrianCrossing
speedLimit15
speedLimit25
speedLimit30
speedLimit35
speedLimit40
speedLimit45
speedLimit50
speedLimit65
stop
yield

Each object instance is annotated with bounding boxes.

Data Collection Methodology

Images in the dataset were collected using vehicle-mounted cameras capturing real road scenes. These images include a variety of:

lighting conditions
road environments
viewing angles
traffic sign scales

This diversity helps improve the model’s ability to generalize to new images.

Annotation Process

Bounding box annotations were created using Roboflow annotation tools.

Annotation workflow:

Images were uploaded to Roboflow
Bounding boxes were drawn around each traffic sign
Each object was labeled with its correct class
Annotations were reviewed and corrected
Dataset exported in YOLO format for training

Train / Validation / Test Split

The dataset was divided into three sets:

Dataset Split	Percentage
Training	~70%
Validation	~20%
Test	~10%

Data Augmentation

During training, several augmentation techniques were applied to improve generalization:

horizontal flipping
mosaic augmentation
image scaling
color adjustments (HSV)

These augmentations help the model learn to detect objects under different visual conditions.

Training Procedure

Framework

Training was performed using the Ultralytics YOLO training framework.

Hardware

Training environment:

GPU: Tesla T4
Platform: Google Colab
Training time: ~1 hour

Hyperparameters

Parameter	Value
Epochs	300
Batch Size	16
Image Size	640
Learning Rate	0.01
Weight Decay	0.0005

Evaluation Results

Overall Model Performance

Metric	Score
Precision	0.99
Recall	~0.99
mAP@0.5	0.994
mAP@0.5–0.95	~0.89
Best F1 Score	0.99

These results indicate that the model performs extremely well on the validation dataset, detecting traffic signs with high accuracy and minimal false positives.

Per-Class Performance

Class	Average Precision
doNotEnter	0.995
pedestrianCrossing	0.985
speedLimit15	0.995
speedLimit25	0.995
speedLimit30	0.995
speedLimit35	0.994
speedLimit40	0.995
speedLimit45	0.995
speedLimit50	0.995
speedLimit65	0.995
stop	0.995
yield	0.995

Most classes achieved extremely high detection accuracy.

The slightly lower performance for pedestrianCrossing may be due to higher variation in appearance and background conditions.

Confusion Matrix Analysis

The confusion matrix shows that most predictions fall along the diagonal, indicating that the model correctly classifies the majority of traffic sign instances.

Examples of strong performance include:

pedestrianCrossing: 146 correct detections
speedLimit35: 76 correct detections
speedLimit25: 57 correct detections
stop: 118 correct detections
yield: 41 correct detections

Misclassifications are rare and usually occur between visually similar traffic signs.

Key Visualizations

Precision-Recall Curve

The precision-recall curve demonstrates that the model maintains high precision across most recall values. This indicates that the model produces very few false positives while still detecting most objects.

F1-Confidence Curve

The F1-confidence curve shows that the optimal detection confidence threshold is approximately 0.73, where the model achieves an F1 score of about 0.99.

This threshold provides the best balance between precision and recall.

Performance Analysis

The model performs extremely well on the validation dataset due to several factors:

Transfer learning from a pretrained YOLO model.
Consistent visual characteristics of traffic signs.
Data augmentation during training.
Clear visual differences between most traffic sign classes.

However, these results reflect performance on the validation dataset and may not fully represent real-world performance in different environments.

Limitations and Biases

Visually Similar Classes

Speed limit signs such as 25 mph, 30 mph, and 35 mph have similar shapes and layouts. If the number on the sign is partially obscured or blurred, the model may confuse these classes.

Environmental Limitations

Model performance may degrade under certain conditions:

poor lighting
nighttime driving scenes
motion blur
heavy shadows
extreme viewing angles

Dataset Bias

The dataset primarily contains traffic signs captured in specific geographic and environmental conditions. This may introduce bias related to:

geographic location
road environment
weather conditions

Performance may vary in unfamiliar environments.

Ethical Considerations

This model should be used responsibly and should not be deployed in safety-critical systems without rigorous real-world testing and validation.

Reproducibility

Training command used:

from ultralytics import YOLO

model = YOLO("yolo11n.pt")

model.train(
    data="/content/dataset/data.yaml",
    epochs=300,
    imgsz=640,
    batch=16,
    device=0
)

Downloads last month: 27