Real-Time Face Gate with OpenCV + OpenVINO ( USB Camera, Rpi4)

Leave a Comment / Tech / By

This post documents a reproducible mini-project: a real-time “face gate” prototype built with OpenCV for camera capture and OpenVINO for inference.

Source code:<CashStolen/rpi-myriad-face-reg>

1)pipeline overview

The core pipeline is:

  1. Capture a frame (BGR) with OpenCV.
  2. Run face detection to get one or more bounding boxes.
  3. For each face ROI:
  • Crop the face region (optionally with padding).
  • Run emotion recognition on the face ROI.
  • Select a corresponding emoji PNG (e.g., happy/sad/angry/neutral/surprise).
  • Overlay the emoji sticker above the face (or centered on the forehead).

This design keeps the code modular and makes it easy to swap models or tune thresholds.

2)environment

  • OS: Debian 13 Rpi4
  • Python: 3.10 (my project uses a dedicated environment)
  • Key dependencies:
  • opencv-python 4.10.0.844
  • openvino 2022.3.2
  • numpy 1.24.4

Install dependencies:(the requirement.txt lies in my git project)

Bash
pip install -r requirements.txt

3)models

This project uses two OpenVINO IR models (.xml + .bin):

  • face-detection-adas-0001
  • emotions-recognition-retail-0003

The expected path after your download:

Bash
models/intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml
models/intel/face-detection-adas-0001/FP16/face-detection-adas-0001.bin

models/intel/emotions-recognition-retail-0003/FP16/emotions-recognition-retail-0003.xml
models/intel/emotions-recognition-retail-0003/FP16/emotions-recognition-retail-0003.bin

4)quick start

First of all, get the index of your camera

Then, after finishing the necessary connection and settings of your Rpi4 or 5 and camera device, it’s time to run our project.

Example command:(your can also write in a line)

Bash
python <ENTRY_SCRIPT>.py \
  --cam 1 \
  --device MYRIAD \
  --fd_model "models/intel/face-detection-adas-0001/FP16/face-detection-adas-0001.xml" \
  --emo_model "models/intel/emotions-recognition-retail-0003/FP16/emotions-recognition-retail-0003.xml" \
  --emojis_dir "emojis" \
  --conf 0.6

--cam: camera index (USB camera often 1)

--device: MYRIAD (default) or another OpenVINO target(maybe CPU?)

--conf: face confidence threshold (higher = fewer false positives)

5)how does it work?

a.How the Emoji Overlay Works

A clean overlay uses alpha blending:

  1. Load emoji PNG with alpha channel (RGBA).
  2. Resize emoji to match face size (or a fixed ratio of face width).
  3. Compute the overlay position (usually above the face box):
    • emoji_w ≈ face_w * 1.0 (or slightly larger)
    • emoji_h scaled with aspect ratio
    • x = face_center_x - emoji_w/2
    • y = face_top_y - emoji_h * 0.9 (place it above the head)
  4. Alpha blend the emoji onto the BGR frame region.

Recommended placement styles:

  • Above head (like “campus gate terminal” look)
  • Centered on forehead (more comedic, but can cover facial features)

b.Accuracy and Threshold Tuning

If the emoji feels “wrong” too often:

  • Increase face detection threshold, e.g. --conf 0.7
  • Crop with padding (too tight crops can hurt emotion recognition)
  • Ensure the emotion model receives the correct input format:
    • correct resize (model input H×W)
    • correct channel order (usually BGR)
    • correct normalization (depends on model)

Also note:

  • Emotion recognition is sensitive to lighting, pose, occlusion, and resolution.
  • For small faces in the frame, results will be unstable (this is normal).

c.Performance Tips (Lower Latency)

Our FPS is a bit low(about 10)using MYRIAD and lower(only 5)with CPU on RPi4.

If you see lag:

  • Reduce camera resolution (e.g., 640×360)
  • Run emotion inference every N frames (e.g., every 2–4 frames)
  • Limit processing to the largest face (optional)
  • Avoid expensive per-frame image conversions

A practical configuration is:

  • Face detection every frame or every 2 frames
  • Emotion recognition every 2–4 frames per face

ooo. Troubleshooting

The window opens but feels delayed

  • Lower camera resolution / FPS
  • Reduce inference frequency
  • Ensure you are not buffering too many frames

Face boxes exist but no emoji appears

  • Check emojis_dir path
  • Verify PNG files have alpha channel (RGBA)
  • Confirm overlay coordinates are clamped within image bounds

Emojis appear but are tiny or off-screen

  • Increase scale ratio (use face width as reference)
  • Clamp (x, y) so the emoji stays within the frame

Summary

This project demonstrates a complete real-time CV demo using OpenVINO:

  • robust face detection
  • emotion classification per face
  • a fun and readable UI via emoji overlay

In the mean time, I have to emphasize that I cannot upload any photos related to our project for the reason that they are too private for my friends and I. And I believe that all of you can understand my project easily!^_^

And just feel free to exchange the project into your better one!<CashStolen/rpi-myriad-face-reg>

Leave a Comment

Your email address will not be published. Required fields are marked *