STORAGE REDUCTION OF IOT DATA FOR HUMAN ACTIVITY RECOGNITION

Master of Science (Computer Science): Adarsh Muralidharan

04 December 2025
11:00 AM Atlantic

Hybrid Defence

CARN 410

Join the meeting now
Meeting ID: 242 362 190 331 6
Passcode: J76yP6SB

Thesis Committee:

Dr. Sazia Mahfuz, Supervisor
Dr. Jiju Poovvancheri, SMU, External Examiner
Dr. Amir Eaman, Internal Examiner
Dr. Darcy Benoit, Director of School
Dr. Zoe Panchen, Chair of the defence

Abstract

Human Activity Recognition (HAR) is a key area within machine learning and human centered artificial intelligence, enabling systems to interpret human motion for applications in healthcare, elderly assistance, and smart environments. However, large-scale deployment of HAR systems faces a critical bottleneck in the storage and transmission of high-dimensional time-series sensor data, which leads to significant computational and memory overhead.

This thesis proposes a deep learning-based framework for efficient storage reduction and representation of sensor data in HAR systems. The framework employs autoencoder architectures to compress raw accelerometer and gyroscope signals into compact latent embeddings that preserve the essential temporal and discriminative characteristics required for activity classification, along with the metadata. Multiple deep neural architectures were designed and evaluated, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, hybrid CNN-LSTM models, and Transformer-based encoders. Each was systematically assessed for its reconstruction capability and classification accuracy using the UCI HAR dataset, comprising sensor recordings from 30 participants performing six common daily activities.

The procedure was conducted, evaluating reconstruction using Mean Squared Error (MSE) and classification using accuracy, precision, recall, and F1-score. Experimental results demonstrate that the proposed CNN-LSTM autoencoder combined with a CNN classifier achieves a data reduction exceeding 99%, while maintaining high classification performance. These findings confirm that the proposed reduction framework reduce sensor storage requirements without compromising recognition accuracy, making it suitable for real-time and resource-constrained HAR deployments on wearable and edge devices.

About Adarsh…

I am pursuing a Masters degree in Computer Science at Acadia University. My interests lie in machine learning, software development, and building practical technology solutions that bridge research and real-world applications. I have a good background in software development and enjoy designing efficient, reliable systems for the web and cloud. I hope to continue developing innovative tools that make intelligent systems more accessible and impactful.