Predicting meltdown situation in Autism and ADHD in real-time through camera using deep learning algorithm.

Abstract

Neurodevelopmental issues such as Autism spectrum disorder (ASD) and Attention deficit hyperactive disorders (ADHD) are quite prevalent in small children detecting and differentiating them at a very early stage is necessary for the future of the affected children and their parents or care giver. This may require 24 hours of surveillance in level 3 cases in which the affected may experience meltdown situation. It is well known by clinical psychologist that sudden meltdowns are common in autistic children, which makes the situation difficult for the parents or care givers and is also a physical threat to the affected children and people around them as they most likely injure themselves. Research has shown that children diagnosed with autism spectrum disorder display specific behaviors that allow us to predict their violent outbursts. Our aim is to develop a CNN-based system that can identify these kinds of behaviors using real time camera.

In our study, we are trying to make a model that can perform Human Activity Recognition (HAR) in real time. Based on the available training data we have trained our model on a few common pre meltdown actions or gestures creating two classes of dataset. but in future we may take huge number of video frame of different types of gestures (using HMBD51 datasets) to train the algorithm so that it can practically identify the situation in real time and alarm the caregiver before they enter the meltdown situation, this will save the patients from self-inflicted injuries and panic attacks not just in the above mentioned two cases but many other brain disorders. The present model has achieved a training accuracy of 100% ,a satisfactory FPS(Frame processed per Second) and the validation accuracy is slightly increasing in each epoch.