Reconhecimento de atividades humanas usando medidas estatísticas dos sensores inerciais dos smartphones

Abstract

Human Activity Recognition (HAR) using different types of sensors has emerged as a revolutionary technology for monitoring people's lifestyles. With the popularization of smartphones, this task has been possible through embedded sensors (e.g. the accelerometer and gyroscope of inertial sensors) that allow recognize different types of activities using machine learning techniques. The data collected by the sensors are treated as time series and must go through several stages until the activity recognition is completed. Thus, the data must be processed, segmented and transformed into a set of features that represent the activities performed by the users. These features are traditionally expressed from mathematical measures such as mean, variance and standard deviation extracted from the sensor signals. This way of recognizing activities composes the most conventional way, has been successful and has made great strides in this field in recent years. However, using this approach requires that calculations be manually defined with the support of a domain expert. As new activities arise, the generated classification models lose performance, requiring the specialist to generate new measures to represent the new activities well. Thus, the adoption of this approach has affected the generalization capacity of recognition models over time. For this reason, more recent research has focused its efforts on solutions that learn the feature patterns and perform the feature extraction process automatically. Examples of such algorithms are deep neural networks and symbolic representation algorithms. In particular, symbolic representation algorithms automatically extract discrete data characteristics. In this context, this work presents the HAR-SR (Human Activity Recognition based on Symbolic Representation) method, which corresponds to a new approach to solve problems involving classification of human activities. The HAR-SR method performs automatic extraction of characteristics in the discrete domain using statistical quantifiers as new characteristics that represent activities. These quantifiers associate values to the time series according to their deterministic or stochastic nature. In addition, the HAR-SR uses data fusion strategies to combine the signals from the sensors and perform a dimensionality reduction in the data. The results of this research show that HAR-SR performs similarly to state-of-the-art researches in activity recognition. The differential of the proposed method, besides the good generalization capacity, is related to the computational cost, which is smaller in the process of feature extraction, since it uses a smaller number of characteristics to generate the classification model. Evaluation results using three real databases (SHOAIB, UCI, WISDM) in five scenarios show that it is possible to classify activities with 93% accuracy, and on average for all scenarios, presents 81% accuracy using leave-one-subject-out cross-validation.