An Empirical Evaluation of Deep Learning Techniques for Human Activity Recognition

• Main Author: Lu, Weijie (Author)
• Other Authors: Yongchareon, Sira (Contributor), Yu, Jian (Contributor)
• Format: Others
• Published: Auckland University of Technology, 2020-05-18T04:20:03Z.
• Subjects: Resident Activity Recognition; Deep Learning Models; Performance improvement; CASAS Datasets; Hybrid architectures; Thesis
• Online Access: Get fulltext

 The recent advancement and development of human-activity recognition technology have led to the gradual entrance of smart home induction systems into residents' lives, stimulating the demand for associated products and services. With these developments, human activity recognition based on deep learning models has earned an increasing share of attention. This research evaluates the ability of nine baseline deep-learning models to classify five CASAS datasets. The study aims to find the baseline deep learning model that best recognises resident activity and to establish methods that improve the performance of baseline deep-learning models. Specifically, we hypothesise that the bidirectional and hybrid architectures will improve the performance of classifying residential activity. To test this hypothesis, we incorporate the hybrid architecture into the convolutional neural network (CNN), and the bidirectional architecture into the long short-term memory and gated recurrent unit (GRU) classifiers. We then verify whether these extensions improve the performances of the baseline models. Finally, we alter the groupings and compare the performances of the baseline deep learning models by different evaluation metrics and the Friedman test. Among the nine deep-learning models tested, the BI-GRU model best recognised various human activities. Our hypothetical improvement method, the bidirectional architecture, significantly improved the model's performance.