Human Activity Recognition from Accelerometer with Convolutional and Recurrent Neural Networks

Abstract

Smartphones are present in most people’s daily lives. Sensors embedded in these devices open the possibility of monitoring users’ activities. The classification of the intricate data patterns collected through these sensors is a challenging task when considering hand-crafted features and pattern recognition algorithms. In this work, to face this challenge, we propose a convolutional neural network architecture along with two methods for transforming sensor data stream into images, and two recurrent neural networks, a long short time memory network and a gated recurrent unit network. The proposed model was evaluated using the UniMiB SHAR dataset. This dataset was acquired with accelerometers of mobile devices. The best macro average accuracy for classification of 17 types of activities, with 5-fold-cross-validation method, 95.49% was obtained with a gated recurrent unit network. The best macro average accuracy for classification with leave-one-subject-out method, 71.36%, was obtained with a convolutional neural network. These results are better than others previously published in the literature with the same dataset.