Detecting arm flapping in children with Autism Spectrum Disorder using human pose estimation and skeletal representation algorithms

Stereotypical behaviour such as arm flapping is among the prominent early signs for young children with Autism Spectrum Disorder (ASD). Diagnosis of arm flapping requires clinicians to use the standard Repetitive Behaviour Scale-Revised (RBSR) which is a structured questionnaire with the caregivers to detect the arm flapping behavioural patterns or cues. This method involves clinicians in multiple long sessions, risking a delay in diagnosis and usually an expensive process. Moreover, trained clinicians may not be available in some areas. The focus of this work is to propose a development of a computational framework to automate the diagnosis process of arm flapping. Here, we show how the human action recognition (HAR) techniques, namely, the pose estimation and the skeletal representation are utilized simultaneously to segment parts of the human body (head, neck, elbows and shoulders) into stickman model. We show how the stickman model allows us to estimate arm asymmetry (during arm flapping) which indicates possible sign of autism. The framework developed has been tested against data taken from a public database and has shown a high accuracy in detecting the repetitive behavioural pattern among young children. The results show that our method can provide efficient results in clinical assessment.