Skip to content
/ MMM-Face-Reco-DNN Public
forked from nischi/MMM-Face-Reco-DNN

Face recognition with opencv and deep neural network

License

MIT license
0 stars 47 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

jimbydude/MMM-Face-Reco-DNN

BranchesTags
 
 

Repository files navigation

Face Recognition with Open CV and Deep Neural Network

This module detects and recognizes your face for MagicMirror². This module is mainly inspired by MMM-Facial-Recognition-OCV3 and a tutorial from Adrian Rosebrock. It uses the new DNN (Deep Neural Network) provided by OpenCV 4.1 and is much more accurate than the old Haar Cascade method. The old method is still available in this module because for a Raspberry Pi Zero the new way will be too heavy.

With this module you can show and hide all other modules depending on which person is recognized. For example, if you are in front of the mirror, the mirror will show you your agenda, but if a stranger is in front, the agenda will not be shown.

Development Status

The module is finished as far as I can tell. I have tested the module on my development machine as well as on the MagicMirror and it works fine with the PiCamera.

Screenshot

This module works in the background, and so no screenshots are available.

Dependencies

I used a Raspberry Pi Spy Cam which i built into my mirror. It also works fine with the regular Raspberry Pi camera module.

Installation

OpenCV

You will find a lot of good tutorials to install OpenCV 4 for your computer / Raspberry Pi. I am not going to write another one here. The module has only been tested with OpenCV 4.1. If you use another version, please let me know whether it works so I can make a list of all compatible versions.

If you install OpenCV in a virtual environment, please do not forget to set the correct pythonPath in the settings.

OpenCV 4.1.2 quick installation (wthout compliation) for Raspbian Buster

If you are planning to run OpenCV on Raspbian Buster, you can follow these steps to install pre-compilled OpenCV v.4.2.1 package. This will save time for OpenCV compilation.

wget https://github.com/cyysky/OpenCV-4.1.2-for-Raspbian/raw/master/opencv_4.1.2-1_armhf.deb
sudo dpkg -i opencv_4.1.2-1_armhf.deb
sudo apt-get -f install
sudo dpkg -i opencv_4.1.2-1_armhf.deb

pip

If you have both python v2 and python v3 installed then you may need to be specific in the use of either pip or pip3. If you are targetting python v3 then you will need to substitute pip3 instead of pip in the example commmands below. If you have only a single version of python installed then you may use pip3.

dlib

You can install dlib easily using pip with the following command.

pip3 install dlib

face_recognition

You can install the face_recognition library over pip too.

pip3 install face_recognition

If you have issues installing, this tutorial may be helpful.

imutils

You can install imutils over pip.

pip3 install imutils

numpy

You can install numpy over pip.

pip3 install numpy

Compatibility

I have tested the following dependency versions. Let me know if the module works with another version so I can extend this list.

Dependency Versions
OpenCV 4.1
dlib 19.17.0
face_recognition 1.2.3
imutils 0.5.3

Install the Module

Load this module directly from GitHub using the following commands.

cd ~/MagicMirror/modules/
git clone https://github.com/nischi/MMM-Face-Reco-DNN.git
cd MMM-Face-Reco-DNN
npm install

Usage

Here is a summary of the most important points to use this module. If you are interested how it works, have a look at Adrian Rosebrocks' tutorial.

Face Dataset

First of all you need to create your face dataset. The module is preconfigured such that you store your dataset in the folder dataset in the root of the module. It is possible to use another folder if you want to.

Put each person in a seperate subdirectory like this:

- dataset
  - thierry
    - img01.jpg
    - img02.jpg
    - img03.jpg
  - adam
    - img01.jpg
    - img02.jpg
    - img03.jpg
  - james
    - img01.jpg
    - img02.jpg
    - img03.jpg
  - john
    - img01.jpg
    - img02.jpg
    - img03.jpg

It's enough if you have around 10 pictures for each person. Try to use pictures in different settings and lighting if possible as it may help to build a better model of the person's face.

Facial Recognition Embeddings

After you set up your dataset, we need to go to the embeddings for the recognitions. I prepared a script for that. It took a while to run on my Macbook and will take much longer on a Raspberry Pi, so if possible, use it on a computer with more power and then transfer the resulting file.

You will find a script called encode.py in folder tools of the module. The following arguments are default:

python3 encode.py -i ../dataset/ -e encodings.pickle -d hog

# Arguments can be omitted to use the defaults above
python3 encode.py
Argument Description
-i / --dataset Path to dataset directory containing face images
Default Value: ../dataset/
-e / --encodings Path to serialized database of facial encodings
Default Value: encodings.pickle
-d / --detection-method Which face detection model to use. "hog" is less accurate but faster on CPUs. "cnn" is a more accurate deep-learning model which is GPU/CUDA accelerated (if available).
Default Value: hog

After that you are ready to configure the module and use it on your MagicMirror.

Building encodings.pickle in a Container

Another option for building the encodings.pickle file is to use a Docker container. This way all dependencies and libraries are isolated from the host OS, you will only need make and docker.

cd tools;
make  # generate an encodings.pickle with hog, or alternatively

make encoding=cnn  # overide the default to use cnn encoding

Module Usage

To setup the module in MagicMirror², add the following section to the config.js file in the MagicMirror/config directory.

{
    module: 'MMM-Face-Reco-DNN',
    config: {
      // Logout 15 seconds after user was not detected any more
      // If they are detected within this period, the delay will start again
      logoutDelay: 15000,
      // How often the recognition starts in milliseconds
      // With a Raspberry Pi 3+ it works well every 2 seconds
      checkInterval: 2000,
      // Module set used for when there is no face detected ie no one is in front of the camera
      noFaceClass: 'noface',
      // Module set used for when there is an unknown/unrecognised face detected
      unknownClass: 'unknown',
      // Module set used for when there is a known/recognised face detected
      knownClass: 'known',
      // Module set used for strangers and if no user is detected
      defaultClass: 'default',
      // Set of modules which should be shown for any user ie when there is any face detected
      everyoneClass: 'everyone',
      // Set of modules that are always shown - show if there is a face or no face detected
      alwaysClass: 'always',
      // XML to recognize with haarcascade
      cascade: 'modules/MMM-Face-Reco-DNN/tools/haarcascade_frontalface_default.xml',
      // Pre-encoded pickle with the faces
      encodings: 'modules/MMM-Face-Reco-DNN/tools/encodings.pickle',
      // Use Raspberry Pi camera or another type
      // 1 = RasPi camera, 0 = other camera
      usePiCamera: 1,
      // If using another type of camera, you can choose
      // i.e. 0 = /dev/video0 or 'http://link.to/live'
      source: 0,
      // Rotate camera
      rotateCamera: 0,
      // Method of facial recognition
      // dnn = deep neural network, haar = haarcascade
      method: 'dnn',
      // Which face detection model to use
      // "hog" is less accurate but faster on CPUs
      // "cnn" is a more accurate deep-learning model which is GPU/CUDA accelerated
      detectionMethod: 'hog',
      // How long in milliseconds modules take to hide and show
      animationSpeed: 0,
      // Path to Python to run the face recognition
      // null or '' means default path
      pythonPath: null,
      // Should a welcome message be shown using the MagicMirror alerts module?
      welcomeMessage: true,
      // Capture new pictures of recognized people, if unknown we save it in folder "unknown"
      // So you can extend your dataset and retrain it afterwards for better recognitions
      extendDataset: false,
      // If extendDataset is true, you need to set the full path of the dataset
      dataset: 'modules/MMM-Face-Reco-DNN/dataset/',
      // How much distance between faces to consider it a match. Lower is more strict.
      tolerance: 0.6,
      // allow multiple concurrent user logins, 0=no, any other number is the maximum number of concurrent logins
      multiUser: 0,
    }
}

Notifications

The module sends notifications if a user is logged in or logged out. In addition you can request the list of logged-in users to check if somebody is in front of the mirror. You can then use it for your own module, such as MMM-MotionControl.

Notification Direction Description
USERS_LOGIN out Sent if new users are detected in front of the mirror
USERS_LOGOUT out Sent if users leave the camera's view
LOGGED_IN_USERS out All logged in (in front of mirror) users
GET_LOGGED_IN_USERS in Request all logged in users

Configuration

Option Description
logoutDelay Log out this long after user was not detected any more. If they are detected within this period, the delay will start again.
Default Value: 15000
checkInterval How often the recognition starts in milliseconds. With a Raspberry Pi 3+ it works well every 2 seconds.
Default Value: 2000
noFaceClass Module set used for when there is no face detected ie no one is in front of the camera.
Default Value: noface
unknownClass Module set used for when there is an unknown/unrecognised face detected.
Default Value: unknown
knownClass Module set used for when there is a known/recognised face detected
Default Value: known
defaultClass Module set used for strangers or if no user is detected.
Default Value: default
alwaysClass Set of modules that are always shown - show if there is a face or no face detected.
Default Value: always
everyoneClass Set of modules which should be shown for every recognised user.
Default Value: everyone
cascade XML to recognize with haarcascade.
Default Value: modules/MMM-Face-Reco-DNN/tools/haarcascade_frontalface_default.xml.
encodings Pre-encoded pickle with the faces.
Default Value: modules/MMM-Face-Reco-DNN/tools/encodings.pickle
usePiCamera Use Raspberry Pi camera or another type. (1 = RasPi camera, 0 = other camera)
Default Value: 1
method Method of facial recognition. (dnn = deep neural network, haar = haarcascade)
Default Value: dnn
detectionMethod Which face detection model to use. "hog" is less accurate but faster on CPUs. "cnn" is a more accurate deep-learning model which is GPU/CUDA accelerated (if available).
Default Value: hog
animationSpeed How long in milliseconds modules take to hide and show.
Default Value: 0
pythonPath Path to Python to run the face recognition.
Default Value: null
welcomeMessage Should a welcome message be shown using the MagicMirror alerts module?
Default Value: true
extendDataset Capture new pictures of recognized people, if unknown we save it in folder "unknown". So you can extend your dataset and retrain it afterwards for better recognitions.
Default Value: false
dataset If extendDataset is true, you need to set the full path of the dataset as well.
Default Value: modules/MMM-Face-Reco-DNN/dataset/
multiUser Allow multiple concurrent user logins, 0=no, 1=yes
Default Value: 0

Facial Recognition States

There are 3 states that of facial recogntion that this module runs in. The states are not directly configurable but rather you use the classes to define which modules are activated in the various states.

State Description
noface This is the state when facial recognition has found no faces
unknown This is the state when facial recognition has found a face that it does NOT recognise
known This is the state when facial recognition has found a face that it does recognise

Classes

In order for this module to do anything useful you have to assign classes to your modules. Do not set any classes on this module itself, it has no output and will not work.

There are a series of classes that drive the showing/hiding of modules based on the state of facial recognition. The following table gives a view of which classes drive which states.

When a state changes, the classes from the old state are hidden and the classes from the new state are shown. Anytime that a specific class appears in both states, nothing happens to the modules tagged with that specific class.

For example:

  • starting from no faces detected, when an unknown face is detected, any modules tagged with the noface class will disappear and any modules tagged with the unknown class will appear.
  • starting from no faces detected, when an unknown face is detected, any modules tagged with the default class will remain unchanged, because default provides both the Noface and the Unknown states.
  • starting from no faces detected, when an known face is detected, any modules tagged with the default class will be hidden and modules tagged with known or "specific user's name" ie to specify modules for a certain user, use their name as the class name

The state to class mappings are:

States that the class is active in
Class Noface Unknown Known
noface Y
unknown Y
known Y
default Y Y
everyone Y Y
always Y Y Y
"specific user's name" Y 
{
    module: 'example_module',
    position: 'top_left',
    // Set your classes here seperated by a space
    // Always shown
    classes: 'always'
},
{
    module: 'example_module_2',
    position: 'top_left',
    // Only shown for Thierry and James
    classes: 'thierry james'
},
{
    module: 'example_module_3',
    position: 'top_right',
    // Only shown for James
    classes: 'james'
}
{
    module: 'example_module_4',
    position: 'top_right',
    // Only shown for known (recognised users)
    classes: 'known'
}

Known Issues

  • Support for multiple concurrently logged in users is not yet complete.

Credits

About

Face recognition with opencv and deep neural network

Resources

Readme

License

MIT license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

8 tags

Packages

No packages published

Languages

  • JavaScript 62.3%
  • Python 35.7%
  • Dockerfile 1.2%
  • Makefile 0.8%