Embedded machine learning for gesture recognition using Arduino Nano 33 BLE Sense. Includes source code for programming the board and notebooks for training the model.

This project is about developing a gesture recognition system using Arduino Nano 33 BLE Sense device to classify two arm gestures; a punch and a flex. First the motion data for these two actions are captured through the Nano board using its onboard accelerometer. Then the obtained data is used to train a Neural Network model using Tensorflow on Google Colab. The trained model is validated and exported into Tensorflow Lite format. This is uploaded to the Arduni Nano board to classify and predict the action of the arm. This project was carried out using this article as a reference.

Shown below are what each file and folder contains.

• IMU_Capture - Contains Arduino sketch for capturing motion data.
• IMU_Classifier - Contains Arduino sketch for predicting the motion using the trained model.
• serial_logger.py - Used to copy the output from the Arduino Serial Monitor to a CSV since manual copying cannot be done successfully.
• model_training.ipynb - Jupyter Notebook including the data pre-processing, building, training and validation of the model.
• data - This folder contains the CSV datafiles captured by the Nano 33 BLE Sense using the serial_logger.py Python script.
• model_files - Contains the Tensorflow lite model and its header file version.
• library_files - Contains the Tensorflow lite library (the correct version used for this project).

1. Open IMU_Capture.ino using Arduino IDE. Connect the Nano 33 BLE Sense board to the computer and set up the serial port correctly.
2. Download the MBed OS library for Nano boards using Arduino IDE Library Manager. Also add the Arduino_TensorFlowLite-2.4.0-ALPHA.ZIP file in the library_files folder by Sketch > Include Library > Add .Zip Library.
3. Compile the code, upload it to the board and verify data is being read using the Serial Plotter. (Pick up the board and do a punch or a flex motion with the board in your hand)
4. Reset the board (using reset switch) and run the serial_logger.py script using Python. Do 10 repititions of punch motion and press control+c to exit. A new .csv file will be created. Rename this with the name of the moiton.
5. Do the same for the flex motion.
6. Run the Jupyter Notebook (preferably on Google Colab) and upload the two .csv files. Train the model and obtain the model.h file. Copy the contents of this file.
7. Open the IMU_Classifier.ino sketch. Create a new tab named model.h and paste the copied contents to this file.
8. Upload the code to the board and open the serial monitor. Perform a motion and observe the predicted classification in the serial output.

• Jupyter Notebook contains a cell which gives an error. Fix it. Shape mismatch.