End-to-End Deep Learning-Based Human Activity Recognition Using Channel State Information

Abstract

The automatic recognition of human activities in the house using Channel State Information (CSI) has received wide attention due to the potential use of wide range of intelligent services. Two-dimensional Convolutional Neural Network (2D-CNN) is one of the most popular approaches for human activity recognition (HAR). This method first applies signal transform to convert a time-series CSI signal into a 2D image, and then uses the image to train a complex 2D-CNN model. In this paper we will present simple deep neural networks including multi-layer perceptron (MLP) and one-dimensional Convolutional Neural Network (1D-CNN) for HAR. Our proposed networks are fully end-to-end automatic learning from feature extraction/selection and classification, and do not require extra signal transform and denoising. Experimental results indicate that the proposed networks not only achieve much better
recognition performance but reduces the network complexity significantly, as compared to the existing methods.