Human Activity Recognition on smartphone using Convolutional Neural Networks

Human activity recognition (HAR) is a key research area in Human Computer Interaction (HCI) and plays an important role in people’s daily life, through computational technologies. HAR can be intended as a typical pattern recognition problem or more specifically as a classification problem with the aim to identify a variety of daily activities (ADLs) performing by an individual at a given moment, to maintain healthy lifestyle, help patient rehabilitation and to detect and diagnose automatically precocious illnesses, such as the Parkinson's disease. HAR can be divided into two approaches: vision-based HAR and sensor-based HAR. In this thesis sensor based HAR is used, that allows to collect data extracted from two type of wearable sensors: accelerometer and gyroscope. Sensors data are acquired by three public available datasets, that are selected among different public databases containing kinematic data of human subjects and are the most common used in human activity research fields based on wearable sensors. The proposed model to make prediction and identify classes for new data is the Convolutional Neural Network (CNN), that is a Deep Learning Network that learns directly from data, eliminating the need to manually extract features. The model is built using the Keras deep learning library and developed on five different architectures having different layers combinations. The hyper-parameters tuning is computed on each architecture of each database to identify the best model configuration. The optimized model is fit on the training set, evaluated on the validation set and used to make prediction on the test set. Moreover, the CNN model is fit on the training set of one database and used to make prediction on other database. Different performances metrics are computed in order to compare results from the various architectures of each dataset. The performances evaluated on the test set of each database are high, with accuracies of about 95%. On the other hand, the performances evaluated to make prediction on other datasets are smaller with accuracies ranging from about 50% to 80%. The performances evaluated by the model confirms the possibility to make prediction on test set of each database correctly and suggest how the model is not reliable to make prediction on other databases.