Case study
An experimental traffic-control system that watches an intersection, counts what is actually there, and adjusts signal timing instead of running a fixed cycle. Vehicle detection comes from YOLOv8, the road network is a SUMO microscopic simulation driven through TraCI, and a Streamlit dashboard shows the live state. A separate track trains a PPO reinforcement-learning agent.
Fixed-time signals cannot respond to real demand. The objective was a system that observes an intersection, estimates how much traffic is waiting, and makes signal decisions from that estimate - then presents the whole loop in a live dashboard so the behaviour is visible step by step.
The repository holds two implementations. The Pro version wires YOLO detection, the SUMO/TraCI runtime, and a Streamlit UI together. The Basic version is a focused reinforcement-learning evaluation built on a custom environment and Stable-Baselines3 PPO.
The currently verified runtime does not rely on live quantum hardware. The repository contains Qiskit-based circuit definitions for research, but during simulation the runtime uses classical NumPy logic to compute traffic pressure. CARLA is an optional, separate experimental track and is not required for the main workflow.
Detection, simulation, and control are separate concerns joined by a small state file so the dashboard can run as an independent process.
YOLOv8 Nano processes camera frames and counts cars, buses, trucks, and motorcycles, mapped onto a 3x3 zone grid to estimate lane density.
ImplementedSUMO runs the microscopic traffic simulation and TraCI reads per-traffic-light queue lengths each step, producing the state the controller consumes.
ImplementedThe runtime computes a pressure score across junction queues with a NumPy softmax and sets the signal phases accordingly. The repo notes this is classical heuristic logic, not hardware quantum execution.
ImplementedEach step writes status to a JSON file used for inter-process communication between the simulation and the UI.
ImplementedStreamlit reads the state file and renders live analytics plus the camera feed with detections drawn on top.
ImplementedIn the Basic version, a custom Gymnasium environment (18-value observation, discrete phase actions) is used to train a PPO policy that rewards lower cumulative waiting time and higher throughput.
Researchyolov8n.pt)live_status.json, state.json), intentionally untrackedThe simulation writes JSON that the dashboard reads. This keeps the two processes independent and easy to restart, but the repository notes that file communication may add latency at scale and would need an API or service layer for a city-wide deployment.
SUMO is scriptable, reproducible, and confirmed working end to end. CARLA offers richer 3D visuals but is heavier and is kept as an optional, separate experimental track rather than a dependency.
The runtime uses a NumPy pressure calculation rather than an RL or quantum policy. That keeps the running system predictable and fast; the RL and quantum work is evaluated alongside it, not shipped into the live loop.
The code resolves its own directory and falls back sensibly when SUMO_HOME is unset, so the same scripts run across machines without hardcoded absolute paths.
cv2.VideoCapture(0)) rather than network video.Install SUMO 1.26.0 and make sure it is on PATH or set SUMO_HOME, then install the Python packages with pip install streamlit pandas ultralytics traci opencv-python. The Pro version needs two terminals - one for the simulation, one for the dashboard:
cd 01_Pro_Version/environment
python main.pycd 01_Pro_Version/ui
python -m streamlit run app.pyThe Basic reinforcement-learning evaluation is self-contained:
cd 02_Basic_Version
pip install -r requirements.txt
python main.pyIt loads the trained PPO policy and runs a short smoke test in the custom environment, printing rewards and traffic states.
This is the most systems-heavy build here; it shares its applied-vision and simulation character with the microplastic and QBioForge projects.