Skeleton-Based Activity Recognition for Children with Autism Using Graph Convolutional Networks

Area of Science:

• Computer Vision
• Machine Learning
• Autism Intervention

Background:

• Movement-based programs are key in autism intervention.
• Manual progress tracking is subjective and time-consuming.
• Video data raises privacy concerns due to identifiable individuals.

Purpose of the Study:

• To develop a privacy-preserving method for recognizing autism therapeutic activities using 2D skeletons.
• To address the challenge of distinguishing activities with subtle motion differences.
• To improve the accuracy of activity recognition models.

Main Methods:

• Utilized ProtoGCN, a graph convolutional network, as the backbone for action representation.
• Introduced a Refine-Confusable (RC) module, a training-only regularizer, to enhance class separation.
• Applied a hinge-margin loss over momentum-updated class centroids to penalize confused class pairs.

Main Results:

• The RC module improved the base model's performance across various evaluation splits (random, session-independent, subject-independent).
• Highest gains were observed in the subject-independent evaluation, reaching 96.30% accuracy and 0.959 macro-F1.
• The method surpassed recent 2D-skeleton baselines while maintaining a lightweight and privacy-preserving approach.

Conclusions:

• The proposed RC module effectively enhances autism therapeutic activity recognition from 2D skeletons.
• The privacy-preserving approach offers a valuable tool for objective progress tracking in autism intervention.
• Learned representations are discriminative and interpretable, showing potential for clinical application.