Smart Home Monitoring with Motion Sensors

MIT 2.156 Artificial Intelligence and Machine Learning for Engineering Design

September 2024 - December 2024

Cambridge, Massachusetts

Machine Learning, Human Activity Recognition, Time Series Windowing, Unbalanced and Sparse Datasets, Data Annotation and Analysis

Technology designed to aid in eldercare can reduce the demand on nurses who help individuals with activities of daily living (ADLs), while also providing autonomy to seniors, who are reported to generally prefer living at home independently. Over a decade worth of research in machine learning (ML) has been done in this field, yet ML models for in-home human activity recognition (HAR) have been unable to translate from academia to commercial products. The goal of this project was to determine the limiting factors in current academic literature that have caused a gap between research and real-world implementation. From there, the second goal was to demonstrate that removing these hypotheses does, in fact, have a significant impact on a model’s ability to classify activities. Finally, I aimed to improve upon these models without relying on the limiting assumptions.

This was an individual research project in which I was responsible for identifying the research question, implementing machine learning models, and analyzing my findings.

I identified two primary assumptions in literature that limit the generalizability of machine learning models for human activity recognition in homes. I improved upon two current models by removing these assumptions while maintaining good accuracy in classification tasks. Finally, I presented my work at the end of semester poster session to 50+ other students. The project recieved strong, positive feedback from my professor and the TAs for the course.

This project helped me understand the forefront of academic literature for human activity recognition in smart home monitoring systems. This was pivotal to my graduate reseasearch since I am passionate about pursing this field for my masters and PhD thesises. Ultimately, I came to understand that the data available to researchers is very limiting in what machine learning models can extract while also maintaining user privacy. This motivated a partnership with Cherish, a start up company focused on detecting humans in their homes using radar data for health and security.