The aim of this thesis is to implement a human activity recognition system using the Sensortile.box development board, which is used as a wearable device by the user. The system consists of firmware that captures data on the development board and a neural network model that classifies the captured data in real time. Acceleration and rotational velocity are measured. Using the data acquired during various activities by different subjects, several neural network models were trained: the Multi-layer Perceptron Neural Network (MLP), the Convolutional Neural Network (CNN) and the Long Short-Term Memory Neural Network (LSTM). These models were initially trained and tested using the Tensorflow library on a personal computer. An evaluation of the performance and effectiveness of the selected models in human activity recognition was conducted. The test showed that the MLP neural network model achieves the best results according to the performance metric used (F1 score). When implementing the trained neural network models on the Sensortile.box development board, it was found that due to their size, additional adjustments were needed, primarily quantization of parameters. After these adjustments, a slight drop in the performance of all neural network models was observed. The MLP neural network model was the only one that met the real-time classification requirement, achieving good performance despite the limitations of the embedded system on the Sensortile.box development board.
|
