Page 753 - AI for Good Innovate for Impact
P. 753
AI for Good Innovate for Impact
Use Case 5: AI-Powered Behavior Recognition for Developmental
Disability 4.9: Accessibility
Organization: SK telecom
Country: South Korea
Contact Person: Choong Hwan Choi, ch.choi@ sk .com
1 Use Case Summary Table
Item Details
Category Accessibility
Problem Addressed Manual behavior tracking in care centers is labor-intensive, inconsistent,
and delays intervention for individuals with developmental disabilities.
Caregivers are overburdened and lack data to support precise behavior
intervention planning.
Key Aspects of Detects challenging behaviors in real-time using Vision Artificial Intelli-
Solution gence (AI). The solution runs on edge devices with lightweight model
and personalizes recognition using face and clothing. Finally it provides
statistics and reports for intervention planning.
Technology Vision AI, Behavior Recognition, Edge AI, Knowledge Distillation, Pose
Keywords Estimation,Re-identification (Re-ID), Metadata Analytics
Data Availability Private (collected from care centers, actor-based simulations, and gener-
ated by Generative AI (gen-AI); labeled by professionals and social
enterprise).
Metadata (Type of Video frames, pose keypoints, time/location metadata, user annotations,
Data) behavior labels.
Model Training and Ensemble model using image, keypoint, and video streams.
Fine-Tuning Trained with 10M+ frames.
Techniques: attention, multi-stream fusion, knowledge distillation (using
SK Telecom’s VL2Lite framework).[1].
Testbeds or Pilot Deployed in 10+ care centers in South Korea (Seoul, etc.) [2].
Deployments
717
