Adaptive Traffic Signal Management (ATSM) — Vehicle Counting with YOLOv8 + SORT

A lightweight research prototype that counts vehicles in a region-of-interest and adapts a simulated red-light timer based on real-time traffic density. Built with Ultralytics YOLOv8 for detection and SORT for object tracking.

This repository includes:

• Real-time vehicle detection and tracking using YOLOv8 + SORT
• Line-based vehicle counting and unique ID handling
• Dynamic signal timer reduction when congestion exceeds a threshold
• Example media and mask for quick local runs

Note: This is an unofficial research artifact and has not been peer reviewed. The source code is GPL-3.0; the accompanying paper is licensed under CC BY-SA 4.0 (see CC_LICENCE).

Quickstart

• Supported OS: Windows, Linux, macOS
• Recommended Python: 3.12+

1) Create a virtual environment (Windows PowerShell)

py -3.12 -m venv .venv ; .\.venv\Scripts\Activate.ps1

2) Install dependencies

pip install --upgrade pip ; pip install -r requirements.txt

Tip: Installing PyTorch can be slow. If you have issues, install the recommended build from the official site for your CUDA/CPU, then re-run pip install -r requirements.txt.

3) Run the demo

python car_counter.py

• The script reads Media/cars2.mp4 and applies Media/mask.png as the region of interest.
• Press q to exit.

How it works

• Detector: Ultralytics YOLOv8 (Weights/yolov8n.pt by default)
• Tracker: SORT (Kalman Filter + IOU-based association)
• Counting: A horizontal line (count_line = [199, 363, 1208, 377]) registers a vehicle when the tracked center crosses it.
• Signal logic: If the number of unique vehicles above the threshold (car_count_threshold = 10) is exceeded, the simulated red-light timer is reduced (60s -> 30s), with a simple cooldown to prevent oscillation.

You can change:

• Input video: update video_path in car_counter.py
• Model weights: swap to Weights/yolov8l.pt for higher accuracy
• Region mask or counting line: replace Media/mask.png and edit count_line

File structure

• car_counter.py — main demo script (detection, tracking, counting, timer)
• sort.py — SORT tracker (GPLv3 by Alex Bewley)
• Media/ — example video(s) and mask.png
• Weights/ — YOLOv8 weights (small yolov8n.pt, large yolov8l.pt)
• CC_LICENCE — paper license (CC BY-SA 4.0)
• LICENSE — code license (GPL-3.0)

Results preview

Below are sample outputs from the demo run:

Advanced usage

Use the YAML config or override via CLI:

# Use config.yaml (default)
python car_counter.py

# Point to a different config
python car_counter.py --config my_config.yaml

# Override key params
python car_counter.py --video Media/cars3.mp4 --weights Weights/yolov8l.pt --line 150 360 1200 360 --threshold 15 --normal 70 --reduced 40 --cooldown 15 --conf 0.4

Paper and citation

• PDF: Unofficial Research Paper
• If you build on this work, please cite the repository (see CITATION.cff) and reference the accompanying unpublished paper:
  • Jambhale, Anuj. "Adaptive Traffic Signal Management (ATSM): Vehicle Counting with YOLOv8 + SORT." Unofficial research manuscript, 2025.

License

• Source code: GPL-3.0 (see LICENSE). Includes sort.py from Alex Bewley (GPLv3).
• Paper and text in CC_LICENCE: CC BY-SA 4.0.
• Ultralytics YOLOv8 is licensed under AGPL-3.0 — see the Ultralytics License for details.

Acknowledgements

• Ultralytics YOLOv8: https://github.com/ultralytics/ultralytics
• SORT by Alex Bewley: https://github.com/abewley/sort

---

Have ideas or feedback? Open an issue using the templates or start a discussion. Contributions are welcome!