T-Car: Traffic Light & Sign Detection

1. Overview

This repository documents the traffic perception pipeline used in T-CAR experiments. The detection system is implemented as a 2-stage pipeline:

Stage 1 – Traffic Object Detection (YOLOv11s)

Model: Yolo11s
Classes:

• 0: traffic_sign
• 1: traffic_light Output: bounding boxes for traffic signs and traffic lights

---

Stage 2 – Traffic Light State Classification (MobileNet)

Model: MobileNet classifier
Output: traffic light signal state (e.g., red / yellow / green)

This repository focuses on:

• Stage 1 model comparison (baseline vs large-scale training)
• Visualization results
• Training configuration and reproducibility

2. Pipeline Architecture

Input Image
     │
     ▼
[ Stage 1: YOLOv11s ]
Traffic Object Detection
(traffic_sign / traffic_light)
     │
     ├── traffic_sign → (handled separately)
     │
     └── traffic_light bbox
             │
             ▼
[ Crop ROI ]
             │
             ▼
[ Stage 2: MobileNet ]
Traffic Light Signal Classification
(red / yellow / green)

3. Dataset

3.1 천안(Cheonan) T-CAR Sample Dataset

This repository includes a 260-image dataset:

data/images/

---

3.2 Large-Scale Training Data

The after.pt model was trained with:
~10,000 images (baseline)
~550,000 additional images (expanded dataset)

Total scale: ~560,000 images

Raw dataset (AIHub-based) is not included in this repository.
See train/README.md for training details.

4. Models

Located in:

weights/

before.pt

• Trained on ~10k images
• Baseline YOLOv11s detector

after.pt

• Trained with +550k additional images
• Improved robustness and small-object detection

5. Visualization Results

Stored in:

runs/viz/
  ├── before/
  ├── after/
  ├── compare/
  └── compare.mp4

• before/: detection results using baseline model
• after/: detection results using large-scale trained model
• compare/: side-by-side merged results (left: before, right: after)
• compare.mp4: video comparison

6. Training

Training configuration and conversion scripts are located in:

train/

Includes:

• YOLO dataset conversion script
• data.yaml
• multi-GPU training command
  Refer to train/README.md for details.