Human Activity Recognition Based on Convolutional Neural Network

Abstract

It is increasingly essential to monitor clinical signs and physical activities of elderly, looking for early warning signs or to recognize abnormal situations, such as a fall. In recent years, the usage of wearable sensors has increased significantly. Data from wearable devices can be used to recognize human movement patterns while performing various activities. Accelerometers have been widely used in human activity recognition systems, however, instead of traditional techniques used for feature extraction, the scientific community is currently developing classifiers based on deep learning techniques, seeking better performance and lower computational cost. Convolutional neural networks (CNN) are the main deep learning technique used in this context. These networks adjust filter coefficients that are applied to small regions of the data, extracting local patterns and their variations. This paper presents a human activity recognition system based on convolutional neural networks to classify six activities—walking, running, walking upstairs, walking downstairs, standing and sitting—from accelerometer data. Results demonstrate the ability of the proposed CNN-based model to obtain a state-of-art performance, with accuracy of 94.89% and precision of 95.78% for the best configuration.