Human activity recognition via accelerometer and gyro sensors

• Main Author: Tee, Jia Lin
• Format: Final Year Project / Dissertation / Thesis
• Published: 2023
• Subjects: QD Chemistry; T Technology (General); TD Environmental technology. Sanitary engineering
• Online Access: http://eprints.utar.edu.my/6046/1/fyp_CS_2023_TJL.pdf; http://eprints.utar.edu.my/6046/

In recent years, the scholarly community has shown great interest in Human Activity Recognition (HAR) as a result of its wide applications and growing significance spanning different domains. While much research has been conducted, focusing on the development of algorithms and techniques for HAR, less emphasis was placed on the improvement of HAR research’s efficiency in terms of sensor data collection, annotation, and storage, resulting in the use of incomplete, inefficient, and time-consuming data engineering systems. This project aims to address the issue of inefficient data engineering infrastructure being used in current HAR research by developing an efficient, comprehensive, and user-friendly data engineering system for data collection, annotation, and storage. To implement the data engineering system proposed, two mobile applications, SensorData and SensorDataLogger with user-friendly interfaces and intuitive functionalities are developed using Java programming language and Android Studio. The dataset created using the proposed data engineering system is then used to train unidirectional Long Short Term Memory (LSTM) model to evaluate the efficiency of proposed system in terms of accuracy and generalization capabilities. In other words, if the dataset created using the proposed system can achieve good accuracy and generalization during training and testing, it means that the proposed data engineering system is effective. To prevent overfitting, early stopping is used to monitor validation loss during training and dropout rate of 0.3 are applied. This project proves that the proposed data engineering system is efficient, which is able to achieve an accuracy of 96.57%. In conclusion, this project will be a significant contribution to the development of HAR in multiple aspects. Firstly, it advances the domain enhancing the data engineering system’s efficiency. Next, it improves the accuracy and reliability of HAR research by allowing the researchers to produce dataset of high-quality. Furthermore, it improves scalability and reproducibility by allowing researchers to expand projects to large scope or reproduce other research with least effort. Moreover, it reduces the barriers of entry for non-technical researchers to engage in HAR research. Lastly, the project paves the way for establishment of standardized dataset, with streamlined data collection, data annotation and data storage, and allow comparative research and benchmarking.