Leveraging Computer Vision and Humanoid Robots to Detect Autism in Toddlers

University of Minnesota Ph.D. dissertation. December 2018. Major: Computer Science. Advisor: Maria Gini. 1 computer file (PDF); xvi, 150 pages.Autism Spectrum Disorder is a developmental disorder often characterized by limited social skills, repetitive behaviors, obsessions, and/or routines. Early intervention significantly improves long-term outcomes for toddlers identified in the second year of life and is the best approach for affecting lasting positive change for children with an ASD. Research shows that children with autism especially enjoy technology, including autonomous (or seemingly autonomous) robots. Tying these together, we hypothesize that observing play interactions between very young children (2 - 4 years old) and a humanoid robot can help us identify children with autism; this first requires us to generate a very large, thoroughly characterized dataset of typically developing children. We begin with an eye tracking experiment comparing four different robots and a young human peer; this shows us which type of robot may be of most interest to children in an in-person, real-life play scenario, and if that robot is as interesting as a peer. Using the robot found to be most interesting in the eye tracking experiment, we next detail a human-robot interaction experiment that engages 2 - 4 year old children in a series of social games with a small humanoid robot; we then analyze the social distances, or proxemics, of the child throughout the interaction. To generate the proxemics data, we use a highly automated person detector which utilizes two state-of-the-art convolutional neural networks; with the proxemics and other development assessment data, we compare and group participants and discuss the implications of those results. A subset of robot interaction participants also finished the eye tracking task, so we discuss the relationship between the human-robot interactions and eye tracking results. Lastly, to validate the generalizability of our automated tracker, we test the system on two other child development experiments, a multiple-participant in-group bias play scenario for 5 and 8 year old children, and an unsolvable box task for toddlers