CSE 499B Research Project

Real-Time Bangladeshi Traffic Sign Detection Using Deep Learning

A comparative analysis of YOLOv11 and SSD architectures achieving 99.45% mAP@50 with a lightweight 5.2MB model running at 22 FPS on CPU.

99.45%
mAP@50
5.2MB
Model Size
22.2
FPS on CPU
8,953
Dataset Images

Research Highlights

Key contributions and achievements of this study

Rank #2 Accuracy

Achieved 99.45% mAP@50, ranking 2nd among 10 state-of-the-art traffic sign detection studies (2012-2024).

97% Smaller Model

5.2MB model size compared to 182.9MB industry average—enabling deployment on edge devices and mobile.

Real-Time on CPU

22.2 FPS inference on CPU without GPU acceleration, making it practical for real-world deployment.

First BD Dataset

BRSDD: First comprehensive Bangladeshi Road Sign Detection Dataset with 8,953 images across 29 classes.

Production Ready

Includes Android app, web demo, and deployment guides for immediate real-world application.

Comprehensive Study

Detailed comparison of YOLOv11 vs SSD with analysis of training dynamics, efficiency metrics, and failure cases.

Experimental Results

Performance metrics, comparisons, and visualizations

Benchmark comparison with state-of-the-art methods
Performance comparison with 10 state-of-the-art traffic sign detection methods (2012-2024)
Model mAP@50 Size (MB) FPS Year
YOLOv11n (Ours) 99.45% 5.2 22.2 2024
Tabernik et al. 99.89% ~200 15 2020
Dewi et al. 99.30% ~250 45 2021
SSD-MobileNet 89.20% 23.5 18 2024
Training metrics over 50 epochs
Training analysis: loss convergence, accuracy progression, and learning rate schedule over 50 epochs
YOLOv11 vs SSD comparison
YOLOv11n vs SSD-MobileNet: Accuracy, speed, and model size comparison
Complete results dashboard
6-panel results dashboard: radar chart, convergence, comparison table, speed, efficiency, and loss curves

Interactive Demo

Upload an image or use your camera to detect Bangladeshi traffic signs in real-time

Traffic Sign Detection — YOLOv11n Open in new tab

Demo powered by Gradio on Hugging Face Spaces. Model: YOLOv11n (5.2MB)

BRSDD Dataset

Bangladeshi Road Sign Detection Dataset — the first comprehensive collection for BD traffic signs

8,953
Total Images
29
Sign Classes
12,847
Bounding Boxes
94.2%
Annotation Agreement
Class distribution across 29 categories
Class distribution across 29 traffic sign categories (Regulatory, Warning, Mandatory)

Dataset Access

The BRSDD dataset is available for research purposes. Please cite our paper if you use this dataset.

Research Paper

Full methodology, experiments, and analysis

Real-Time Bangladeshi Traffic Sign Detection Using Deep Learning: A Comparative Analysis of YOLOv11 and SSD Architectures

Mohammad Mansib Newaz

CSE 499B Senior Project, North South University, 2024

Citation

If you use our work, please cite:

@article{newaz2024bdtrafficsigns,
  title={Real-Time Bangladeshi Traffic Sign Detection Using Deep Learning: 
         A Comparative Analysis of YOLOv11 and SSD Architectures},
  author={Newaz, Mohammad Mansib},
  journal={CSE 499B Senior Project, North South University},
  year={2024},
  note={Available at: https://github.com/mnxtr/bd-traffic-signs}
}