Skeleton-based visualization of poor body movements in a child's gross-motor assessment using convolutional auto-encoder

Abstract

This paper deals with human activity recognition (AR), which is the basic technology for understanding human behavior and movement in the field of sensing applications for human support systems. Focusing on children's gross motor (GM) skills as an AR target, a new visualization method to point out children's poor body movements is presented. The visualization is achieved as anomaly detection by an autoencoder (AE) trained with good GM movements with a complete rating score, and poor limb motion in GM is detected as anomal points by the GM-AE which is combined with authors' previous GM-AR and AE. In preparation for the GM-AE, the previous dataset and data augmentation have been improved, and new GM-AR was completely realized with an identification accuracy of 99.3 % to 148 actual assessment patterns out of 200 theoretical assessment combinations of the GM assessment tool TGMD-3. Using appropriate preparations proven by the GM-AR stage, we investigated some deep learning conditions related to GM-AE. Finally, it was confirmed that the presented visualization method can emphasize the points that match the evaluation items of TGMD-3.