Deep Learning-Based Human Action Recognition Framework to Assess Children on the Risk of Autism or Developmental Delays.

Automatic human action recognition of children with machine learning and deep learning methods using play-based videos can lead to developmental monitoring, early identification, and efficient management of children at risk of neurodevelopmental disorders (NDD) and Autism Spectrum Disorders (ASD). Advancements in deep learning make it feasible to develop human action recognition models with large datasets, enhance clinician capacity, and improve access, affordability, and quality of care. However, data collection is challenging due to ethical, legal, and limited datasets of children with NDD and the enormous amount of human tasks involved in video annotation. Therefore, we propose a new method to overcome these challenges by training several deep learning models using a publicly available action dataset comprising adults performing various actions. We demonstrate the effectiveness of our multiple models to recognize similar actions of children in a custom-collected video dataset of children with NDD, ASD, and Typical development. Our method assist child psychologists in intelligently detecting children at risk of NDD and measuring their progress from their videos captured in the natural environment.