Date

2022-2-08

Author

Yılmaz, Tuğçe Alara

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

266
views

317
downloads

Cite This

People perform a wide variety of activities every day. Systems that can automatically distinguish these activities, i.e. human activity recognition models, have improved markedly, especially in the last decade. Deep learning is demonstrating increasingly promising outcomes in overcoming the problem of human activity detection as technology improves at a rapid pace. However, validating activity recognition in real-world situations is critical for practical solutions that work in natural contexts. Establishing systems that could achieve automatic activity recognition with real-life settings such as the devices that people use every day naturally and the environment they live in, might require lots of computational complexity. A lightweight neural network model is adopted for this purpose, one that could run swiftly even in the background process without taking up a lot of space when embedded into smartphones. Four inertial sensory data are represented in color coded image form and fused with three channelled audio data image representation to perform recognition task. The resulting fusion images allow rapid recognition performance because the size of the each image is so small. This thesis also underlines that audio sensor data, which require considerably bigger window sizes for identification on their own, improve automated recognition performance when used in conjunction with inertial sensors, even when divided into small window sizes to interact with other sensors simultaneously. In addition, this thesis provides a strategy that helps the computer better discern real-world behaviors by introducing activities, contexts, and placements in a hierarchical manner to perform accurate activity recognition. By merging auditory images with inertial color coded images, representing multiple activity pairs with hierarchical groups according to activity-context-placement, and employing a lightweight model, high accurate recognition performance score competitive to state-of-art, nearly 91\% success rate is achieved. We believe that this research could be classified as a "quality of experience" because it presents a lightweight model that could be used to predict behavior of individual by data collected from devices such as smartphone and smartwatch that everyone uses each day naturally.

Subject Keywords

Human activity recognition, Deep learning, Inertial sensor, Audio sensor

URI

https://hdl.handle.net/11511/96320

Collections

Graduate School of Natural and Applied Sciences, Thesis