Human Activity Recognition Based on Multichannel Convolutional Neural Network With Data Augmentation

Abstract

In view of the excellent portability and privacy protection of wearable sensor devices, human activity recognition (HAR) of wearable devices has increased applications in human-computer interaction, health care, etc. Therefore, it is necessary to recognize various human activities accurately and efficiently. In this paper, we propose a multi-channel convolutional neural network with data augmentation for HAR, denoted AMC-CNN. First, the sliding windows in time series are used to construct the feature window, and then the feature window is augmented by data transformation and data addition. Through the horizontal connection of the augmented feature windows, the reconstructed feature samples are obtained. Second, a relatively lightweight multi-channel convolutional neural network is designed. In order to mine deep correlations between sensor data more efficiently, we use reasonable convolution kernel settings to compute multi-channel convolutions and extract parallel multi-scale features, and then train the model to recognize human activities. Finally, experiments are carried out on WISDM and MHEALTH datasets to verify the effectiveness of the data augmentation. Moreover, the other three models are adopted for comparative experiments. The evaluation indexes show that the method proposed in this paper has a better recognition effect, which verifies that the AMC-CNN model is effective in both single-sensor and multi-sensor HAR.