Residual Learning and LSTM Networks for Wearable Human Activity Recognition Problem

Abstract

Human activity recognition (HAR) has become an important research area in pervasive computing because of its extensive applications in solving real-life, human-centric problems. Due to the spatial complexity and temporal divergence of human behavior data, conventional machine learning methods can't handle this problem effectively. As one of the deep learning algorithms to solve HAR problem, Long Short-Term Memory (LSTM) is an excellent learning methods in handling time series problems, which can sufficiently use the historical data. But for traditional LSTM algorithms, when we add more depths to the network, the gradient vanishing arises. To this end, a deep long short-term memory (LSTM) network with residual connection is proposed. Generally, the proposed network shows improvements on both the temporal (using LSTM) and the spatial (residual connections stacked deeply) dimensions. Experiments with the Opportunity Dataset show that, comparing with baseline LSTM, our algorithm can recognize the human activities with an F1 score of 0.908, increased by 1.4%.