Human Activity Recognition Based on Wavelet-CNN Architecture

Abstract

Human activity recognition (HAR) has become an important research field in pervasive computing and has attracted extensive attention from researchers at home and abroad. However, traditional recognition methods rely heavily on artificial feature extraction, which greatly affects the generalization ability of the model. Therefore, this paper designs a deep learning model based on wavelet transform and convolution neural networks. Firstly, the waveform data of multi-channel sensor is decomposed into low-frequency and high-frequency components by wavelet transform (WT) after window sliding segmentation, which are combined as the input data of network model. Then, convolution neural networks with different convolution kernels are used to extract multidimensional features efficiently, and the max-pooling layers are used to filter the interference noise caused by human unconscious jitter. Finally, through the output classification of full connection layer, the accurate recognition of human activity state is realized. In order to verify the effectiveness of the designed model, we evaluate the performance of the model from the convergence speed, loss and accuracy of the model, and compare it with the more advanced recognition model on the public dataset of OPPORTUNITY. Finally, the proposed architecture Wavelet-CNN achieves 91.65% F-measure and has higher activity recognition ability.