Layout-Agnostic Change Point Detection for Enhanced Human Activity Recognition in Smart Homes
Author(s)
Stennett, Tyler
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Human Activity Recognition (HAR) in smart home environments relies heavily on accurate temporal segmentation, yet most existing methods use fixed, overlapping windows that often blur activity boundaries and degrade downstream classifier performance. This thesis introduces two layout-agnostic change point detection (CPD) frameworks, TAD and TCPC, that achieve state-of-the-art results across multiple smart home datasets. TAD (Textual Descriptions of Sensor Triggers-based Activity Distribution) Change Point Detection uses semantically rich sensor metadata and a BiLSTM HAR classifier to detect changes in activity distributions between non-overlapping windows, while TCPC (Textual Descriptions of Sensor Triggers-based Contrastive Predictive Coding) Change Point Detection eliminates the need for labeled data by using a contrastive model that identifies change points through variations in learned latent representations. Evaluated on four CASAS smart homes (Aruba, Milan, Kyoto, and Cairo), both methods consistently outperformed statistical and embedding-based baselines, improving the geometric mean of true positive and true negative rates for change point detection by 1.20% to 16.45%, and enhancing downstream classification with a BiLSTM model by 1.66% to 8.18% in Macro F1 score. Leveraging lightweight sentence-transformer embeddings and modest recurrent architectures, these approaches are suitable for real-time deployment and offer a practical, representation-driven alternative to traditional segmentation techniques, paving the way for future work involving transformers, richer datasets, and energy-efficient models.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis