People's Anthem

A project using facial recognition and a raspberry pi's camera to identify a person and play their favorite music.

The project provides an easy to setup configuration, with Docker -- taking care of all dependencies, including pre-compiled torch wheels optimized for the raspberry pi.

Requirements

• A raspberry pi (rpi)
• Docker installed on your OS (how to install)
• A raspberry pi camera
• A speaker plugged-in to the rpi via the audio jack
• A source of music (mp3 or a Spotify account) (More info here)

Install

To get started, you will need docker and an internet connection. The installation might take up to 150 minutes on a rpi4.

git clone [email protected]:alexantoinefortin/peoples-anthem.git
cd peoples-anthem
make build-peoples-anthem

Usage

In this section, we discuss how to:

1. Build a dataset of faces
2. Train a facial recognition algorithm
3. Use that model to recognize people and to play their Spotify playlist when they are recognized

1. Building a dataset of faces

1.1 Choosing a location for your rpi

The first step is to place your rpi and rpi-camera at a place where it can recognize people.

1.2 Running the code for recording the images of faces

make run-record-faces

The images will be saved, in this repository's directory, in: ./data/.

Let the container run for as long as necesseray. Try collecting images at different time of the day and in different lighting conditions, etc.

To kill the container use: docker kill peoples_anthem.

1.3 Creating a train/test set

Reorganize the images saved in ./data/ (see 1.2) according to the following directory structure:

./data/
├── train/
│   ├── alice/
│   ├── bob/
│   ├── misc/
├── test/
    ├── alice/
    ├── bob/
    ├── misc/

This means that you have to manually identify who is in these pictures and to put each pictures in the correctly named directory.

Note: To help the algorithm in identifying out-of-scope faces, the category misc, as seen above, should be added. This category should include images of:

• unrecognizable faces
• faces of people you do not want to recognize
• non-face items

2. Training a face recognition algorithm

Once step 1. is completed, run the model's training code:

make train-model

By default, the trained model will be saved in ./models/model.v1.pklz.

The code will print the train/test accuracy.

As a guideline, during the development of this project, the model was trained to recognize 4 different persons + 1 misc category. The model obtained an accuracy_train=0.85 and an accuracy_test=0.83.

Try adjusting the model's hyperparameters in ./code/train_face_recognition.py

3. Playing people's spotify playlist after recognizing them

Once step 2. is completed, we need to setup a config file so that the code can play the spotify playlist of your choice when it recognizes someone.

For guidance on how to do so, see: setting-up-spotify

Once this is completed, make sure that a speaker is plugged in the rpi's audio jack and run the code:

make run-peoples-anthem

To kill the container use: docker kill peoples_anthem.

Tips and gotchas

• A non-root user must have an active session for the raspberry pi to be able to play sounds through the speakers.

To ensure this, one can start a screen session (using /bin/bash screen) and leave that session open (CTRL+A D to disconnect from the session and to leave it open). By doing this, your non-root user will have an active session and the speaker will always be able to play music.

• Set the N_TRACKS parameter in ./code/peoples-anthem.py to change the number of track played after recognizing someone.

License

GPLv3