Deep Semi-supervised Network in Activity Recognition Using Wearable Sensor Data

Abstract

Activity recognition in wearable devices is mainly classified by extracting data features from triaxial accelerometers and gyroscopes However during the process of data collection the traditional labeling method is typically based on an observation of video-based information and manually corrected which is very time-consuming and costly At the same time human-labeled results are prone to a large amount of noise which might affect the performance and efficiency of the feature engineering Therefore this thesis uses a semi-supervised probabilistic generative model combined with a deep neural network to classify the sensor data into actions This method can still make the classification model a reliable classifier under conditions with limited labeled data In the feature extraction stage the Variational Autoencoder (VAE) in the deep neural network generates the latent features of the original data after dimension reduction and uses the latent features to assist the classification of the deep neural network model For unlabeled data the latent features and the unknown action label are regarded as implicit variables in the two probabilistic models; for the labeled data the prior probability distribution of the latent feature and the action label are regarded as two sets of probabilistic models Then uses the Variational Inference to approximate the posterior probability in Bayesian statistics and gradually increases the target's Evidence of Lower Bound (ELBO) to reach Expectation-Maximization In this study three groups of public datasets were assessed The experimental results show that the accuracy of this study is competitive to that of other models using datasets with 20% of the label samples or more

Date of Award	2019
Original language	English
Supervisor	Ren-Shiou Liu (Supervisor)