IBS Institute for Basic Science
Search

Deep learning based activity recognition of daily livings with wearable sensors for stroke survivors and non-disabled controls

Prof. Sujin Kim

December 16(Mon) - December 16(Mon), 2019

14:00 - 15:00

# 86314

CNIR Seminar


Date:  2:00 pm Tuesday, December 16th

 
Place: #86314

Speaker:  Sujin Kim Ph.D, PT

Assistant professor

Department of Physical Therapy

Jeonju University

 

 

 

Title: "Deep learning based activity recognition of daily livings with wearable sensors for stroke survivors and non-disabled controls"

 

Abstract: Individuals with stroke often show decreased use of the more-affected arms even if they have capability to use them. Recently, to assess real-world use of the more-affected arm, wearable sensors such as accelerometers or mobile devices have been widely used, and the success of these wearable devices depends on accuracy of recognizing daily activities of individuals with stroke. Previous research on human activity recognition has suffered from inaccuracy and inability to generalize a trained model due to various factors, such as insufficient data and not fine-controlled design of activities that often includes only gross movements, such as walking, jumping, or sitting. Here in this study, we attached five inertia measurement units on subjects’ upper extremity and trunk, and collected data from thirty young non-disabled subjects and ten individuals with stroke while they performed eighty daily activities of living including washing dishes, folding tower, and applying make-up lotions, etc. Each activity was segmented and labeled manually, and then the labeled data was trained in deep-learning based models for activity classification. We applied techniques to improve cross-subject generalization including data augmentation and transfer learning. Our model showed that applying deep learning methods improves accuracy of activity recognition of daily living for both non-disabled subjects and stroke survivors. Our work can be applied to generate a daily and/or weekly activity profile of patients, analyze patterns of habitual use of the upper extremity in post-stroke, and provide online visual or tactile feedback if they do not use their more-affected arms and hands for various daily activities.

 

 

Host: 김 성 신 박사