An incorporation of deep temporal convolutional networks with hidden markov models post-processing for sensor-based human activity recognition
Linh Trinh, Bach Ha · 2022
The use of sensors for human activity recognition (HAR) is one of the most active research fields. Several machine learning techniques for classifying human actions have been proposed in HAR. However, because they rely so heavily on the quality of handcrafted features, these techniques demand extensive feature engineering. Recent approaches to deep learning have attempted to provide comprehensive training. In this paper, we present a deep temporal convolutional neural network with Hidden Markov Chain for HAR post-processing. Our proposed method comprises an enhancement to ED-TCN [13], an efficient algorithm to generate very large, noisy data for training model with weight initialization, and a post-processing technique for smoothing the prediction of model. Experiments indicate that our proposed model outperforms other state-of-the-art models on the PAMAP2 and WISDM datasets, suggesting that our model is effective at recognizing human activity using sensors. Our method’s implementation is available on github: https://github.com/khaclinh/EFTCN-HMM.