Real-time Face Detection

This project provides three face detection implementations using OpenCV:

• Version 0: Haar cascade classifiers (fast, lightweight, classical).
• Version 1: YuNET DNN detector (modern, accurate, robust).
• Version 2: Optimized YuNET DNN detector (with GPU support and multithreading).

Common Setup

1. Create a virtual environment (recommended):

python3 -m venv .venv
source .venv/bin/activate

2. Install dependencies:

pip install -r requirements.txt

---

Version 0: Haar Cascades (face_detection.py)

Run

From the Face-ID-Detection-System folder:

python3 face_detection.py

Options

• --camera: Camera device index (default: 0)
• --scale: Scale factor for detection (default: 1.1)
• --min-neighbors: Detection sensitivity; higher = fewer false positives (default: 5)
• --min-size: Minimum face size in pixels (default: 30)
• --width: Optional camera frame width (default: 640)
• --height: Optional camera frame height (default: 480)

Notes

• Uses OpenCV's built-in Haar cascade via cv2.data.haarcascades.
• Fast and lightweight; good for resource-constrained environments.
• Sensitive to pose, occlusion, and extreme lighting conditions.
• Allow camera access when prompted by your OS.

Example

python3 face_detection.py --scale 1.05 --min-neighbors 3 --min-size 50

---

Version 1: YuNET DNN (face_detection_version_1.py)

Requirements

YuNET requires OpenCV >= 4.5.0. The model is downloaded automatically on first run.

Run

From the Face-ID-Detection-System folder:

python3 face_detection_version_1.py

Options

• --camera: Camera device index (default: 0)
• --model: YuNET model filename (default: face_detection_yunet_2023mar.onnx)
• --conf-threshold: Confidence threshold for detections; higher = fewer false positives (default: 0.9)
• --nms-threshold: NMS threshold for duplicate detections (default: 0.3)
• --top-k: Keep top K detections (default: 5000)
• --width: Model input width (default: 160)
• --height: Model input height (default: 120)

Notes

• Modern DNN-based detector shipped with OpenCV.
• Outputs facial landmarks (2 eyes, nose, 2 mouth corners) overlaid as blue dots.
• Displays face count and FPS in the top-left corner.
• Better accuracy and robustness than Haar cascades.
• Can leverage GPU acceleration if OpenCV is built with CUDA support.

Example

python3 face_detection_version_1.py --conf-threshold 0.8 --nms-threshold 0.3

---

Version 2: Optimized YuNET DNN (face_detection_version_2.py)

Requirements

YuNET requires OpenCV >= 4.5.0. The model is downloaded automatically on first run. For GPU acceleration, OpenCV must be built with CUDA support.

Run

From the Face-ID-Detection-System folder:

python3 face_detection_version_2.py

Options

• --camera: Camera device index (default: 0)
• --model: YuNET model filename (default: face_detection_yunet_2023mar.onnx)
• --conf-threshold: Confidence threshold for detections; higher = fewer false positives (default: 0.9)
• --nms-threshold: NMS threshold for duplicate detections (default: 0.3)
• --top-k: Keep top K detections (default: 5000)
• --width: Model input width (default: 160)
• --height: Model input height (default: 120)
• --use-gpu: Use GPU acceleration if available

Notes

• Optimized version of YuNET with asynchronous detection using multithreading for smoother real-time performance.
• Captures at higher resolution (640x480) for better input quality, then resizes for detection.
• Supports GPU acceleration via CUDA for significant speedups on compatible hardware.
• Outputs facial landmarks (2 eyes, nose, 2 mouth corners) overlaid as blue dots.
• Displays face count and FPS in the top-left corner.
• Better accuracy and robustness than Haar cascades, with improved efficiency over Version 1.

Example

python3 face_detection_version_2.py --use-gpu --conf-threshold 0.8

---

Comparison

Feature	Haar Cascades (v0)	YuNET (v1)	Optimized YuNET (v2)
Speed	Very fast	Fast	Very fast (with GPU/threading)
Accuracy	Moderate	High	High
Pose robustness	Low	High	High
Landmarks	No	Yes (5 points)	Yes (5 points)
Model size	None (built-in)	~6 MB ONNX	~6 MB ONNX
GPU support	No	Yes (with CUDA)	Yes (with CUDA)
Multithreading	No	No	Yes
OpenCV version	3.0+	4.5.0+	4.5.0+
Overlays	FPS	Faces + FPS	Faces + FPS

---

General Notes

• Allow camera access when prompted by your OS.
• Press q or ESC to exit any detector.
• For best results, use adequate lighting and frontal/near-frontal face poses.
• Version 0 shows real-time FPS overlaid on the video.
• Versions 1 and 2 show face count and FPS overlaid on the video.