A tiny Convolutional Neural Network (CNN) for Human Activity Recognition (HAR), based on a single wearable flexible sensor applied to the knee joint, is presented; it is capable to distinguish three activities taken from a public dataset (w-HAR): jump, walk, sit. HAR has gained increasing interest in recent years, particularly with the widespread adoption of Artificial Intelligence and Neural Networks. A defining characteristic of this research domain is its versatility, spanning various fields, including healthcare and sports. The signal processing introduced in this work involves a grey-scale image encoding followed by a tiny CNN that consists of only one Convolutional layer. Activation functions employ a leaky ReLU while a Softmax function was used as a nonlinear component. Notably, during training the images were grey-scale encoded through a normalization while for the validation and test the images were built only with a routing of the time series data. Results in terms of test accuracy (94.43%) suggest the potentiality of HAR based on flexible and stretchable sensors using innovative image processing deep learning techniques and tiny CNN that could lead to intelligent sensors integrated with embedded AI capabilities.
Human Activity Recognition using Flexible Sensor and Tiny Convolutional Neural Network
Longo G.;Fasolino A.;Liguori R.
;Di Benedetto L.;Rubino A.;Licciardo G. D.
2024-01-01
Abstract
A tiny Convolutional Neural Network (CNN) for Human Activity Recognition (HAR), based on a single wearable flexible sensor applied to the knee joint, is presented; it is capable to distinguish three activities taken from a public dataset (w-HAR): jump, walk, sit. HAR has gained increasing interest in recent years, particularly with the widespread adoption of Artificial Intelligence and Neural Networks. A defining characteristic of this research domain is its versatility, spanning various fields, including healthcare and sports. The signal processing introduced in this work involves a grey-scale image encoding followed by a tiny CNN that consists of only one Convolutional layer. Activation functions employ a leaky ReLU while a Softmax function was used as a nonlinear component. Notably, during training the images were grey-scale encoded through a normalization while for the validation and test the images were built only with a routing of the time series data. Results in terms of test accuracy (94.43%) suggest the potentiality of HAR based on flexible and stretchable sensors using innovative image processing deep learning techniques and tiny CNN that could lead to intelligent sensors integrated with embedded AI capabilities.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.