arduino-tiny-ml

The repository contains a gesture recognition system implemented on Arduino nano 33 ble sense board using a neural network based on gyroscope and accelerometer inputs. It uses:

• Google Colab for training a TensorFlow Lite neural network and deploy the arduino header file containing the model.
• Arduino IDE for collecting the data in .csv files for the training and the inference.

Training data collection

The arduino project input_capture contains the source code that collects features from the gyroscope and accelerometer in widows of 128 samples:

• Mean value
• Standard deviation
• Root mean square
• Minimum value
• Maximum value
• Power Spectral Density

These measurements are computed directly in the arduino microcontroller and saved in .csv files for seprate gestures, ready for the actual neural network training in google colab. In our implementation, we considered four types of movements:

• Rest (nothing)
• Shake (a left-right movement)
• Up-Down
• Circle

The measurements are stored in the training_data folder.

Note

During training and inference, the Arduino board is handled with the usb port facing upwards and the chips of the board facing right.

Neural Network training

The colab notebook contains the steps for the creation and validation of the model automatically using its confusion matrix.

The results that we obtained indicates a low amount of false positives/false negatives and good accuracy levels.

Total accuracy: 0.975

Per class specs: 
-> Rest
	 accuracy 1.0
	 true positive rate 1.0
	 true negative rate 1.0
	 positive predictive value 1.0
	 f1 score 1.0 
-> Left-right
	 accuracy 0.975
	 true positive rate 1.0
	 true negative rate 1.0
	 positive predictive value 0.9
	 f1 score 0.9473684210526315 
-> Down-Up
	 accuracy 1.0
	 true positive rate 1.0
	 true negative rate 1.0
	 positive predictive value 1.0
	 f1 score 1.0 
-> Circle
	 accuracy 0.975
	 true positive rate 0.92
	 true negative rate 0.9649122807017544
	 positive predictive value 1.0
	 f1 score 0.9583333333333334 

Inference

In order to evaluate the model in real time we provided an arduino project training_data with the source code for the inference.