A Deep Learning Approach for Improving Activity Detection on Portable Smart Devices
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Engineering and Technology
Date of Award
Spring 5-2021
Abstract
Activity trackers have become very popular in the recent decade. Human activity recognition (HAR) has been receiving more attention for years because of its easy-to-use features in numerous applications to identify activities. Activity tracker users want to measure their activity for health, fitness, or similar purposes. Trackers are loaded with sensors and researchers are aimed at increasing software functionality of these devices. Furthering simple tasks already presented in many trackers, such as number of steps and number of calories burned, we developed pattern recognition software to be used with the activity trackers. We advanced a small-size artificial intelligent agent to detect user activity. Our classification model can be deployed on activity tracker sensors-based devices, such as phones, smart watches, and smart activity trackers. Our primary aim is to recognize four human activities: walking, walking upstairs, walking downstairs, and standing. We used two deep learning methods. They are Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) artificial recurrent neural network to learn users’ activity characteristics and predict activity. For the dataset, we utilized one of the most used open-source datasets of Smartphone –Based Recognition of Human Activities and Postural Transitions, which was collected from 30 individuals who performed postural transitions and basic activities. These were recorded using the smartphone’s sensors. We used TensorFlow as the primary software library with Keras API. The significant finding of this research is the increased detection accuracy by exceeding previously published results. The leave-one-subject-out experiments yielded more than 97% classification accuracies in the recognition of walking, walking upstairs, walking downstairs, and standing. A feasible window size on data and deep learning architecture are determined through the experiments. Our research provides many benefits to people and healthcare workers who would like to accurately measure the four fundamental activities we targeted. Healthcare providers are the primary beneficiary for newly developed software as they may prescribe these activities to patients.
Advisor
Mutlu Mete
Subject Categories
Computer Engineering | Engineering
Recommended Citation
Demirhan, Huseyin, "A Deep Learning Approach for Improving Activity Detection on Portable Smart Devices" (2021). Electronic Theses & Dissertations. 390.
https://digitalcommons.tamuc.edu/etd/390
