YOLOv8 Traffic Sign Detection

A computer vision project for traffic sign detection using YOLOv8 and custom lightweight architecture variants.

This project compares a standard YOLOv8 baseline with several attention and feature-fusion extensions, including CBAM, LCFE, IMCMD, and YOLO-TS-style fusion. The goal is to improve small-object traffic sign detection while evaluating robustness across day and night driving conditions.

---

Overview

Traffic sign detection is an important perception task for autonomous driving and driver-assistance systems. Road signs are often small, visually similar, partially occluded, or affected by lighting changes, which makes detection challenging.

This project investigates:

• Whether lightweight feature-fusion modules can improve YOLOv8 traffic sign detection
• Which custom modules contribute most to detection performance
• How model performance changes between day and night conditions
• Whether higher accuracy can be achieved with fewer parameters

---

Project Workflow

Traffic sign dataset
        ↓
Dataset configuration
        ↓
YOLOv8 baseline training
        ↓
Custom model variant training
        ↓
Evaluation and comparison
        ↓
Day/night robustness analysis
        ↓
Result visualization

---

Model Variants

Variant	Description
baseline	Standard YOLOv8 training baseline
cbam	YOLOv8 with CBAM attention
lcfe	YOLOv8 with lightweight context feature enhancement
lcfe_v2	Stability-improved LCFE variant
imcmd	Custom IMCMD feature-fusion design
imcmd_ts	IMCMD with YOLO-TS-style AGRFM fusion
ts	YOLO-TS-style ablation variant

---

Key Features

• YOLOv8-based traffic sign detection pipeline
• Multiple custom architecture variants for comparison
• Lightweight attention and feature-fusion experiments
• Day vs. night robustness analysis
• Ablation-style model comparison
• CLI wrappers for training, evaluation, and prediction
• Benchmark summaries and visualization charts

---

Dataset

The project is designed around the LISA traffic sign dataset setup used in the experiments.

Expected class set:

• go
• goForward
• goLeft
• stop
• stopLeft
• warning
• warningLeft

The full dataset is not included in this repository. Dataset paths should be updated in the YAML files before training or evaluation.

Dataset configuration files:

configs/datasets/
lisa_day.yaml
lisa_night.yaml

More details are available in:

docs/dataset_card.md

---

Methodology

1. Baseline Training

The project first trains a standard YOLOv8 baseline to establish a reference point for traffic sign detection performance.

The baseline is used to evaluate whether custom modules improve accuracy, robustness, or parameter efficiency.

---

2. Custom Architecture Variants

Several model variants are tested around the YOLOv8 framework.

The custom modules focus on:

• Strengthening feature fusion
• Improving small-object representation
• Adding lightweight attention
• Comparing accuracy vs. parameter efficiency
• Testing robustness under lighting changes

---

3. Evaluation

Models are evaluated using object detection metrics, with emphasis on:

• mAP@0.5
• Parameter count
• Overall model comparison
• Day vs. night performance
• Ablation-style component contribution

Results are summarized in:

docs/results.md
reports/metrics_summary.csv
reports/day_night_summary.csv

---

Results

The strongest reported variant in this project is IMCMD.

Model	mAP@0.5	Params	Notes
IMCMD	42.74	1.98M	Best overall accuracy
IMCMD-TS	40.57	2.21M	Strong multi-scale fusion variant
YOLOv8s Baseline	39.14	11.14M	Reference model
YOLO-TS	38.10	13.71M	Ablation comparison
CCA_Light	29.91	-	Lightweight context baseline
CBAM	11.73	-	Attention-only comparison

Key observations:

• IMCMD achieved the best reported mAP@0.5.
• IMCMD-TS remained competitive and showed strong robustness behavior.
• The custom variants outperformed the standard YOLOv8 baseline in this setup.
• IMCMD achieved better accuracy with substantially fewer parameters than the baseline.

---

Visual Results

The repository includes generated charts for:

• Overall model comparison
• Day vs. night robustness
• Ablation study
• Training convergence

Chart files are stored in:

charts/

Detailed result notes are available in:

docs/results.md

---

Repository Structure

YOLOv8-Traffic-Sign-Detection/
├── .github/                    # CI workflow
├── assets/
│   └── sample_predictions/     # Sample prediction assets
├── charts/                     # Result figures used in README
├── configs/
│   ├── datasets/               # Dataset YAML files
│   └── experiments/            # Experiment presets
├── data/
│   ├── external/               # Raw datasets, not tracked
│   ├── interim/                # Intermediate assets
│   └── processed/              # Processed dataset assets
├── docs/
│   ├── dataset_card.md
│   ├── project_overview.md
│   ├── results.md
│   └── roadmap.md
├── outputs/                    # Prediction outputs
├── reports/
│   ├── day_night_summary.csv
│   ├── metrics_summary.csv
│   └── metrics_summary_template.csv
├── scripts/
│   ├── train.py
│   ├── evaluate.py
│   ├── predict.py
│   └── prepare_folders.py
├── src/
│   └── yolo_traffic_sign/
│       ├── models/
│       ├── cli.py
│       ├── inference.py
│       ├── legacy.py
│       └── paths.py
├── tests/
├── pyproject.toml
├── requirements.txt
└── README.md

---

Installation

1. Clone the repository

git clone https://github.com/zehuanyu/YOLOv8-Traffic-Sign-Detection.git
cd YOLOv8-Traffic-Sign-Detection

2. Create environment

py -m venv .venv
.venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt
pip install -e .

If using GPU training, install the correct CUDA-enabled PyTorch build before installing the remaining dependencies.

---

Quick Start

Prepare project folders

py scripts/prepare_folders.py

Train YOLOv8 baseline

py scripts/train.py train --variant baseline --data configs/datasets/lisa_day.yaml --epochs 100 --batch 16 --name baseline_yolov8s_100epochs

Train IMCMD variant

py scripts/train.py train --variant imcmd --model-type small --data configs/datasets/lisa_night.yaml --epochs 100 --batch 16 --name imcmd_small_100epochs

Evaluate a trained model

py scripts/evaluate.py --weights runs/traffic_sign/imcmd_small_100epochs/weights/best.pt --data configs/datasets/lisa_day.yaml

Export prediction images

py scripts/predict.py predict --weights runs/traffic_sign/imcmd_small_100epochs/weights/best.pt --source assets/sample_predictions --variant imcmd --name imcmd_demo

Prediction outputs are saved under:

outputs/predict/

---

Engineering Notes

• Root-level model scripts are preserved for compatibility with the original experiment code.
• Reusable wrapper logic is organized under src/yolo_traffic_sign/.
• Result summaries in reports/ are aligned with the charts used in the README.
• Dataset YAML paths must be updated locally before training.

---

Future Improvements

• Move remaining legacy modules into src/yolo_traffic_sign/models/
• Add automated metrics export after training
• Add qualitative prediction examples to the README
• Add stricter evaluation report generation
• Release pretrained model artifacts
• Add more detailed error analysis for failure cases

---

Tech Stack

Python, PyTorch, Ultralytics YOLOv8, OpenCV, Object Detection, Computer Vision, Traffic Sign Detection, Model Evaluation, Data Visualization

---

Author

Zehuan Yu