Deep learning assessment of child gross-motor

Abstract

The acquisition of gross motoer (GM) skills during childhood is very important for physical and psychological development. Various body function measurement tests have been designed to assess a child's GM performance, but the assessment process is a laborious manual task; hence, IT automation combined with activity recognition (AR) is highly desirable. This paper focuses on GM assessment deep-learning (DL) by expanding the previous fruitful result of GM classifiction, which utilized OpenPose to detect childrens' skeletons, a specific person tracking algorithm to recover the OpenPose's drawbacks, conversion of the skeleton's time-series data into motional time-series images, and its data augmentation technieque. A procedure for building a database containing assessment information is presented, and a new CNN-based deep learning network that performs both GM classification and evaluation simultaneously is proposed. Applying these methods to actual GM assessment including 13 types GM motions including 155 combinations of GM assesment scores, the new GM-AR could classify them with a very high accuracy of 99.6%.