Independent Research Presentation

Title: Testing Real-time LSTM-based Smart Fall Detection

Speaker: Awatif Yasmin


Advisor: Dr. Anne Hee Hiong Ngu


When: Thursday, December 1 @ 11:00 a.m.


Where: https://txstate.zoom.us/j/7910806695 

 

Abstract:

Fall detection is a technique that can help elderly people by detecting their movements, especially falls. SmartFall Detection App is a watch-based application where the user wears a smartwatch on his/her left wrist which can detect falls. If the user did fall and can also signal for help if the user needs that. The main issue with this application is the model's performance gap between the version used in the real world and the offline trained version. The offline version of the model gave almost perfect accuracy in fall detection, but the real-world version which is the lite version of the offline version has a high false positive rate. In this independent study, I will outline the various investigations and testing techniques devised to fix the real-time model. During the investigation, we noticed every user’s data is different and the accuracy of the model can be increased if we use the same type of training and testing data for each user. So, we started working on the personalization process. The main concept is to train the model with a standard fall dataset and load it into a smartphone. The phone and the watch are connected by Bluetooth. The user will wear the watch and do their usual daily activities. When they take the watch off and deactivate the system, the system will send the user data to the Couchbase server. The server will train an incremental model using  that user data. The next time the user activates the system a new personalized model will load which is trained by that user’s data. We will give a demonstration of the personalization process.  

 

Deadline: Dec. 28, 2022, midnight