🚦 Real-Time Traffic Signal Decision System

Live public traffic camera → YOLO11n vehicle detection → density-based signal control

Hridam Biswas · KIIT University

---

The Problem

Traffic signals in most cities run on fixed timers — a car sits at a red light on an empty road for 2 minutes while no one crosses the other way. This system reads a live traffic camera, counts how much road is covered by vehicles, and makes a real-time signal decision. No fixed timers. No wasted time.

---

Pipeline

 Public YouTube Traffic Camera (yt-dlp)
             │
             ▼
  ┌──────────────────────┐
  │   Frame Capture      │  OpenCV VideoCapture → raw BGR frame
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │   YOLO11n Detection  │  Detects: car · truck · bus · motorcycle
  │   (COCO classes)     │  Outputs: bounding boxes + class IDs
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────────────────────────┐
  │   Density Calculation                    │
  │                                          │
  │   density = Σ(bbox areas) / frame_area   │
  │             ───────────────────────────  │
  │                      × 100 %             │
  └──────────┬───────────────────────────────┘
             │
             ▼
  ┌──────────────────────┐
  │  Rolling Average     │  Last N frames smoothed → removes flicker
  │  Smoothing (N=8)     │
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────────────────────────┐
  │   Signal Decision                        │
  │                                          │
  │   density < 10%  →  🟢  GREEN            │
  │   density < 30%  →  🟡  YELLOW           │
  │   density ≥ 30%  →  🔴  RED              │
  └──────────┬───────────────────────────────┘
             │
             ▼
  ┌──────────────────────┐
  │   Live Display       │  Bounding boxes · signal badge
  │   (OpenCV window)    │  density bar · vehicle count · FPS
  └──────────────────────┘

---

Signal Logic

Smoothed Density	Signal	Meaning
< 10 %	🟢 GREEN	Road is clear — let traffic through immediately
10 % – 30 %	🟡 YELLOW	Moderate traffic — prepare to stop or go
≥ 30 %	🔴 RED	Heavy congestion — hold all crossing traffic

Density is rolling-averaged over the last 8 frames to prevent the signal flickering on every detection jitter.

---

Live Window Layout

┌────────────────────────────────────────────────────┐
│  Vehicles: 7                          ┌──────────┐ │
│  FPS: 24.3                            │  GREEN   │ │
│                                       └──────────┘ │
│                                                    │
│          [ live camera feed + bounding boxes ]     │
│                                                    │
│  Density: 8.4%  ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░  │
└────────────────────────────────────────────────────┘

Element	Description
Bounding boxes	Orange rectangles around every detected vehicle
Signal badge	Top-right color block — GREEN / YELLOW / RED
Density bar	Bottom-left progress bar; color matches signal state
Vehicle count	Total vehicles detected in current frame
FPS	Live inference + display frame rate

---

Quick Start

git clone https://github.com/Hridambiswas/traffic.git
cd traffic
pip install -r requirements.txt

# Random public camera from config
python main.py

# Specific YouTube live stream
python main.py --source "https://www.youtube.com/watch?v=ByED80IKdIU"

# Stricter detection + wider smoothing window
python main.py --conf 0.5 --window 12

# Lower resolution stream (faster)
python main.py --height 360

Press Q to quit.

---

CLI Reference

Flag	Default	Description
--source	random from config.py	YouTube live URL or direct stream URL
--conf	0.40	YOLO confidence threshold (higher → fewer false positives)
--window	8	Rolling average window size in frames
--height	480	Preferred stream resolution height

---

Configuration

All tunable parameters live in config.py:

VEHICLE_CLASSES   = {2, 3, 5, 7}   # car, motorcycle, bus, truck (COCO IDs)

GREEN_THRESHOLD   = 10.0            # % density below which signal is GREEN
YELLOW_THRESHOLD  = 30.0            # % density below which signal is YELLOW

SMOOTH_WINDOW     = 8               # rolling average frame count
CONF_THRESHOLD    = 0.40            # YOLO detection confidence

DISPLAY_WIDTH     = 960             # output window width
DISPLAY_HEIGHT    = 540             # output window height

Parameter	Lower value	Higher value
GREEN_THRESHOLD	More likely to stay yellow/red	Easier to get green signal
YELLOW_THRESHOLD	Red trips sooner	More tolerance before red
SMOOTH_WINDOW	Reacts faster, more flicker	Stable signal, slower response
CONF_THRESHOLD	Detects more vehicles	Only high-confidence detections

---

Camera Sources

Public live traffic cameras are configured in config.py:

Camera	Location	URL
Cam 1	NYC Times Square	youtube.com/watch?v=ByED80IKdIU
Cam 2	Chicago Traffic	youtube.com/watch?v=AdUw5RdyZxI
Cam 3	LA Highway	youtube.com/watch?v=1EiC9bvVGnk

Add any public YouTube live stream to CAMERA_SOURCES in config.py. The system randomly picks one on start and auto-reconnects if the stream drops.

---

YOLO Vehicle Classes

COCO Class ID	Label	Included
2	car	✅
3	motorcycle	✅
5	bus	✅
7	truck	✅
0	person	❌
9	traffic light	❌

Only vehicle classes contribute to density. Pedestrians and static objects are ignored.

---

Project Structure

traffic/
│
├── main.py          ← entry point · CLI args · stream reconnect loop
├── config.py        ← all thresholds, paths, camera URLs
│
├── stream.py        ← yt-dlp handler · random source with fallback
├── detector.py      ← YOLO11n detection · bbox-area density calc
├── signal.py        ← rolling-average smoothed signal decision
├── display.py       ← OpenCV overlay renderer
│
└── requirements.txt

---

Dependencies

Package	Version	Purpose
ultralytics	≥ 8.4.0	YOLO11n model + inference
opencv-python	≥ 4.11.0	Frame capture, rendering, live window
numpy	≥ 1.26.0	Array ops for density computation
yt-dlp	≥ 2026.1.0	Public YouTube stream URL extraction

---

Built by Hridam Biswas · KIIT University