Continuous Personalized Fall Detection and Data Collection

Falls are very real problems for elderly people. Falling is the leading cause of injury and death in older Americans according to the Centers for Disease Control and Prevention. Currently there is no software system for reliably detecting falls for elder people that is unobtrusive, affordable, and easily accessible. While the idea of using a wrist worn smart watch to collect accelerometer data to predict falls has been proposed before, the fall detection algorithms are all trained using simulated fall data from healthy young adults. Any system trained for fall detection that utilizes only a simulated dataset cannot be recommended for real-world application as it does not reflect the characteristics of fall from the target population. We propose a solution that will move fall detection into the real world by being capable of collecting real data from elder patients unobtrusively. This system utilizes an online database and a GPU server to collect labeled data from each user and retrain the model specific to that user dynamically. This allows the system to learn the patterns in a user’s linear acceleration data combined with a pre-existing synthetic dataset to determine if a user has fallen or not. That is, it will be capable of personalizing the model to each user to improve precision while maintaining the recall. Additionally, it is capable of reliably collecting all linear acceleration data (falls or activities of daily life) from a user via positive and negative feedback data. This will help to provide future research with real-world fall datasets to create even more reliable models. This system is built to scale using Couchbase, Tensorflow, and commodity smartwatch and smartphone devices.

Discipline thesis:degree_discipline

Computer Science

Grantor

Texas State University

Year dc:date.issued

2020

Author dc:creator

• Coyne, Shaun

Advisor dc:contributor.advisor

• Ngu, Anne H.H.

• fall detection
• RNN
• deep learning
• Couchbase
• data collection
• labeled dataset
• user feedback
• automation
• continuous training
• activity recognition

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10877/12103

OAI identifier oai:identifier

oai:digital.library.txst.edu:10877/12103