Human activity recognitionnew approaches based on machine learning and deep learning

Abstract

Currently, the scientific community is giving significant attention to the field of human activity recognition (HAR), which has gained remarkable prominence as a topic of discussion. Since the irruption of smartphones and wearable devices in daily life, the costs and ease of conducting studies in this field have improved significantly. Moreover, its applicability in various research areas such as medicine, fitness, or home automation makes this topic even more attractive for researchers in the field. However, despite the remarkable advances made in the last decade, it is not possible to transfer that acquired knowledge to a real-life environment. That is because most of the related work has been carried out under laboratory conditions. In other words, with pretty specific indications, placing the measuring devices and performing the actions in an explicit way that does not represent at all the variability present in the real world. For those reasons, this Thesis has focused on orienting the research in this field towards a real-life environment. To that end, a dedicated dataset has been constructed to carry out the main research, based on the personal smartphone sensors of 19 different individuals. The main difference between that dataset and those already existing in the scientific community is that those individuals have been given as much freedom as possible to use their smartphones during data collection. Thus, even when performing the same action conceptually, the resulting data may vary, as each individual may use the smartphone differently, as is the case in everyday life. Hence, once the data was obtained, an in-depth study was carried out, in search of the best machine learning and deep learning models to classify the data, according to the actions studied. The results confirm the possibility of transferring the acquired knowledge to a real-life environment. In terms of their performance, it is worth mentioning tree-based models like Random Forest and other deep learning models such as Convolutional Neural Networks (CNN) or recurrent neural networks based on the Long Short-Term Memory (LSTM) technique, among the various methods used.