Abstract
Common screening tasks for developmental-behavioral disabilities require human judgement to decide pass/fail on checklists, which possibly causes subjective biases. On the other hand, professional requirements for an assessment build a barrier for the accessibility to such screening tests. Therefore, we applied a combination of computer vision techniques to automatically perform cognition assessment on toddlers. To tackle insufficient data, multi-person scene, and unexpected movements of toddlers, YOLOv5, Mediapipe, LOFTR, and depth prediction model trained from Mannequin Challenge dataset are utilized to accurately focus our detection model on assigned areas to generate better results. We believe that similar concepts could be further extended to other sub-fields in childhood developmental-behavioral screening and improve clinical practice.