Abstract
Movement-based and physical activity programs are central tools in autism intervention, so recognizing the activities a child performs during therapy is valuable for objective progress tracking. Manual monitoring of these sessions is time-consuming and subjective, and raw video raises privacy concerns because it shows identifiable children. We address autism therapy activity recognition from privacy-preserving 2D skeletons, and we focus on the practical difficulty that several therapy activities differ only in subtle motion details. As a backbone we adopt ProtoGCN, a graph convolutional network that represents each action as a combination of learnable motion prototypes. This contrastive backbone organizes all classes at once, however, so it does not enforce a margin between the few pairs that remain entangled after training. We therefore introduce a Refine-Confusable (RC) module, a training-only regularizer that pushes apart the empirically most-confused class pairs using a hinge-margin loss over momentum-updated class centroids. The module changes neither the backbone nor the inference cost. On the MMASD dataset restricted to the ten-class 2D-skeleton configuration, the RC module improves the base model under both splitting regimes, with a gain of 1.21 points of pooled accuracy on the stricter subject-independent split. Under the protocol-matched holdout it reaches 96.30% accuracy with 0.959 macro-F1, which surpasses recent 2D-skeleton baselines while it keeps a lightweight and privacy-preserving modality. A per-fold analysis shows that this gain is consistent across folds rather than tied to a single split, and t-SNE and prototype visualizations show that the learned representation is discriminative and interpretable.

This publication has 10 references indexed in Scilit: