Crowd Detection Camera

Shows your webcam feed and alerts you when too many people are detected.

---

Requirements

• Windows PC
• Python 3.10 or newer (download here)
• A webcam (built-in laptop camera works fine)

---

Setup (one-time only)

Open Command Prompt and run:

pip install opencv-python

---

How to Run

1. Open Command Prompt
2. Navigate to the folder where you saved this project, e.g.:

   cd path\to\crowd-detection
   

3. Start the program:

   python app.py
   

4. A window opens showing your webcam feed.
   • Green boxes appear around detected people
   • The count shows in the top-left corner
5. Press Q to stop.

---

Change the Alert Threshold

Open config.json with Notepad and change the number:

{
  "threshold": 5
}

This means: alert when more than 5 people are detected.

Save the file and restart the program for the change to take effect.

---

What the Alerts Mean

What you see	What it means
Green boxes + "People: N"	Normal — N people detected
Red border + "OVERCROWDED" + beep	Too many people (count exceeds threshold)

---

Troubleshooting

Problem	Fix
ERROR: Cannot open webcam	Another app is using the camera. Close it and try again.
No green boxes appear	Move further back — HOG needs to see full or partial body
Program won't start	Make sure you ran pip install opencv-python

---

Advanced Configuration

All settings live in config.json. Open with Notepad to edit. Restart after saving.

{
  "threshold": 5,
  "min_confidence": 0.3,
  "smoothing_window": 8,
  "detect_every_n_frames": 3,
  "hog": {
    "win_stride": [8, 8],
    "padding": [8, 8],
    "scale": 1.03
  }
}

Field	Default	What it does
threshold	5	Alert triggers when people count exceeds this number
min_confidence	0.3	Filters weak detections (trees, signs). 0.0 = keep all, 1.0 = only very confident. Lower → more detections. Higher → fewer false positives.
smoothing_window	8	Frames to average count over. Higher = more stable, slower to react.
detect_every_n_frames	3	Run detection every N frames. 1 = every frame (slowest). 3 = ~3x faster, boxes update every 150ms.
hog.win_stride	[8, 8]	HOG sliding window step in pixels. Smaller = more thorough, slower.
hog.padding	[8, 8]	Context pixels around each detection window. Larger catches partially-visible people.
hog.scale	1.03	Image pyramid scale step. Smaller (e.g. 1.03) catches more size variation, slower.