AGGRESSIVE ACTION IDENTIFICATION IN AUTISM SPECTRUM DISORDER USING VIDEO ANALYSIS

Abstract

Autism Spectrum Disorder ASD is a clinical condition associated with brain development. It affects how a person perceives the world, Pathological symptoms of ASD varies from mild to severe. Repetitive action or pattern of action is often associated with ASD. One of the main concerns of autistic children is the aggressive behavior that can be directed towards them, which may lead to self injuries. Constant monitoring is required in case of autism with aggressive behavior to prevent self injuries. The proposed work aims to recognize three classes of violent action namely head banging against wall or any object, arm flapping continuously and spinning uncontrollably from video data. The proposed work uses a 3DCNN model with Skeleton Joint features for recognizing the said actions. The accuracy of proposed model is 83.56% in dataset validation. The cross data accuracy is about 65%. The proposed work also aims at analyzing the difficulties in live video analysis and recognition of actions from live stream video. The self-stimulatory behavior dataset – SSBD is used is this work. KEYWORDS— 3DCNN, Deep Learning, Action Recognition, Autism Spectrum Disorder