Page 140 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Use Case 18: Smart Rehabilitation: AI-Driven Exercise Assessment

and Live Feedback for Enhanced Recovery

Country: India

Organization: Anna University, Chennai

Contact Person(s):

Dr. Dhananjay Kumar, dhananjay@ annauniv .edu
Mr. B. Naren Karthikeyan, oraclenaren2004@ gmail .com
Mr. A John Prabu, johnprabu0702@ gmail .com

1 Use Case Summary Table

Item Details
Category Healthcare

Problem Addressed Patients with musculoskeletal disorders lack real-time corrective
feedback and progress tracking when performing rehabilitation
exercises at home, which can hinder recovery and lead to improper
form or ineffective therapy [1].

Key Aspects of Solution • Real-time visual and audio cues to correct exercise through
examining the extracted pose. Processing is to be done on a
Raspberry Pi.
• Assessment of exercise to be done through a custom
Spatio-Temporal Graph Neural Network (STGCN) on the cloud

Technology Keywords STGCN, Raspberry Pi, Mediapipe, cross-platform Neural Network
PACKage (XNNPACK)

Data Availability Public (UI-PRMD Dataset: [7])
Metadata (Type of Data) Video data

Model Training and • Live feedback is comparing joint angles and shifts in distance
Fine-Tuning between the reference pose and the patient’s pose.
• STGCN is to be trained using a triplet loss where the anchor is a
correctly performed exercise, the positive sample is a correctly
performed exercise other than the anchor, and the negative
sample is an incorrect exercise.
• The final expected outcome is a latent space where correctly
performed exercises are together and incorrect ones are further
away from them.

104