Awesome

Air-Guitar-CS249R

Welcome to Air Guitar CS249R! This Repo contains the code, resources and supporting materials for the Air Guitar final project for the class CS249R: Machine Learning!

At a high level, this project is concerned with two different machine learning models running on the Teensy 4.1: one project classifies gesture IMU data into two classes: strum or no strum, while another model is concerned with classifying data from flex sensors attached to four fingers into 13 different chords.

The repo is organized as follows:

The base folder contains supporting information and resources, including the schematic for the circuit used.

The audio files folder includes the audio files used for the project. Specfically, the chord_recordings subdirectory holds all the audio files played back upon playing the corresponding chord. The files are named with the name of the associated chord and all of them are .wav format.

The data directory holds all the data collected for this project used for training and validation. The flex subdirectory holds the data from flex sensors, while the imu subdirectory holds the data from the IMU. All data are stored in .csv format as time series data. All data belonging to one class are stored in one .csv file with the name of the associated class, and separate data points are separated by empty rows.

The data_collection directory holds all the scripts used to collect data for both models. There are two subdirectories, one for each of the models. Both of them include a python script as well as a Platform.IO project. The python script includes further instructions in the top as comments.

The deployment directory contains the final and compiled source code of the project. It is a Platform.IO project, and includes the header files for the models used. Most of the source code is included in the src subdirectory. The folder also contains the deployment.py script, which reads the serial data from the Teensy running the deployment project, and plays the desired audio file. More detailed instuctions are at the top of the python file.

The experimention folder contains multiple Arduino/Platform.IO projects that we used in our testing to interface with the Teensy board.

The notebooks folders contains all the Jupyter Notebooks that we used to train our models. The folder contains two notebooks for each sensor type, one for training a fully-connected model and another for training a convolutional model.

Lastly, the results directory includes the results of all of our model training attempts for both sensors, including confusion matrices, as well as the TFLite files and the model header files. The two models that we ende up using for the deployment project were the flex-cnn-onlyonefclayer for the Flex sensor model and the strum-cnn-onlyonefclayer for the IMU model.

Chords Gestures

This project supports 13 chords. They are the following: A, A Minor, B, B Minor, C, C Minor, D, D Minor, E, E Minor, F, G, G Minor

Since it is hard to use the actual hand positions of these chords, both physically and for data collection and classification, we have come up with our gestures to represent these chords. Each gesture is composed of a combination of 4 fingers: index, middle, ring, and pinky, in that order. A 1 signifies that the finger is held straight, while a 0 signifies that a finger is bent towards the palm of the hand. Using this encoding, the following are the gestures for each of the supported Chords:

a: 0001

am: 1001

b: 1000

bm: 1111

c: 0110

cm: 1011

d: 0010

dm: 0100

e: 0000

em: 1100

f: 1101

g: 1101

gm: 1010

Hardware Design and Setup

The hardware aspect of the project is a core component of the overall system. The circuit diagram of the components and their respective connections are shown below.

This project uses sensors attached to gloves to collect the data and run the application. On one glove the IMU is attached. This is to detect the strumming. The other glove has 4 flex sensors attached, one for each finger except the thumb, and this is used to identify the chords. The gloves have wires running from their pins and is connected to the main board, which contains the MCU and the additional components.