Street Sign Sense

Real-Time Traffic Sign Detection

Machine learning object detection system based on YOLO12, running client-side directly in the browser with WebGL hardware acceleration.

View Report.pdf

Author: Alessandro Ferrante Visite: ...

Project & Hugging Face Live Statistics

Taxonomy

1 Sign Classes

Curated

1 Dataset Images

YOLO Box

1 Total Annotations

Dataset

1 Dataset Downloads

Nano

1 YOLO12n Downloads

Small

1 YOLO12s Downloads

Medium

1 YOLO12m Downloads

1. Context & Safety-Critical Relevance

In ADAS and autonomous driving, rapid perception of traffic signs is vital for passenger safety. Real-time recognition demands sub-frame latency (>30 FPS), robust handling of adverse weather, motion blur, and occlusions, and zero margin for misclassification on mandatory or regulatory signs.

2. The YOLO Family & Ultralytics Evolution

Pioneered as a single-stage end-to-end regression paradigm and industrially perfected by Ultralytics (from YOLOv5 and YOLOv8 through to YOLO11 and YOLO12), the YOLO ecosystem has defined modern real-time computer vision. Ultralytics revolutionized edge AI by delivering lightweight, anchor-free, attention-enhanced models engineered for extreme efficiency across PyTorch, ONNX, and in-browser WebGL.

3. Object Detection: Spatial Localization & Classification

Unlike pure image classification which only answers what is in an image, Object Detection solves two intertwined vision challenges at once: classification (identifying the road sign category) and spatial localization (predicting exact bounding box coordinates [x, y, w, h]). This dual capability allows ADAS vehicles and edge devices to simultaneously detect and track multiple street signs across cluttered urban scenes and variable distances.

Object Detection: Bounding Box Localization and Multi-Class Classification

4. YOLO12 Architecture: Attention-Centric Design

YOLO12 integrates attention mechanisms to boost detection accuracy without sacrificing real-time inference speed:

R-ELAN Backbone: Combines residual connections with advanced feature aggregation. Uses Stacked Blocks Reduction to streamline computation.
Area Attention Module Neck: Partitions feature maps into segments ($l=4$), turning global attention into local operations and cutting computational costs by 75%.
Dynamic Head + FlashAttention: Optimizes memory access during attention operations and eliminates positional encoding for small signs.
Optimized MLP Expansion: Reduced to ~1.2 to prevent bottlenecks, combined with depthwise convolutions and batch normalization.

3. Evaluation Metrics & Benchmark

Official validation set metrics obtained from the best model checkpoint (best.pt) as documented in Chapter 4 of Report.pdf:

Model	Precision	Recall	F1-Score	mAP@50	mAP@50-95	Inference (ms)*	Size
YOLO12n (Nano)	0.849	0.819	0.825	0.870	0.760	5.4 ms (~185 FPS)	5.32 MB
YOLO12s (Small)	0.887	0.846	0.853	0.904	0.796	12.6 ms (~79 FPS)	18.1 MB
YOLO12m (Medium)	0.882	0.859	0.856	0.916	0.811	31.7 ms (~31 FPS)	77.6 MB