Classification of Recurrence Plots’ Distance Matrices with a Convolutional Neural Network for Activity Recognition

Abstract

With the emergence of ubiquitous sensing technologies, it is now possible to continuously monitor users during their everyday activities in order to provide personalized feedback and interventions. For example, fitness trackers can count steps and motivate users to keep active. With their rich set of sensors, smartphones are also capable of monitoring user behavior such as physical activity and location. Smartphones’ inertial sensors have been used to recognize different types of activities. Usually, statistical sets of features from time and/or frequency domain are extracted from the raw signals to train machine learning models. The final performance of the system will depend on the set of features which needs to be defined by the researcher. In this work, we propose a method based on recurrence plots’ distance matrices and convolutional neural network (CNN) that does not require feature engineering. A recurrence plot is a visualization of the recurrent states of a dynamical system. For the activity recognition task, the raw acceleration signal is transformed into an image-like representation of recurrent states and a CNN is then trained with those images. The results show that this method is able to achieve better results than a feature based approach in terms of accuracy and recall.