ASD Behavioral Analysis from Video

In this project we developed machine learning pipelines that detect and interpret autism-related behaviors directly from video of young children. The goal is to turn hours of unstructured, naturalistic recordings into quantitative behavioral signals that clinicians can use for early screening — at a scale that manual coding cannot reach.

The system combines motion and pose features with temporal action-recognition networks (PyTorch / TensorFlow, OpenCV), trained on clinically annotated data. It recognizes fine-grained actions and social-communication cues that are associated with autism spectrum disorder, and contributed to peer-reviewed work on machine-learning-based ASD detection from videos (IEEE HEALTH-COM, 2020).