Children with autism spectrum disorders (ASD) benefit from specialized services throughout their lifespan. Eligibility for these services is determined by behavioral assessments and clinical observation. In an effort to make the assessment process faster, more reliable, and more accessible, there have been decades of research on the identification of markers that provide insight into ASD. For example, gaze following, eye contact, and gaze aversion are all cues relevant to the assessment of autism. Observations through behavioral assessments, especially in clinical research settings, are often videotaped for purposes such as repeat viewing, record keeping, training, and assessing the longitudinal impact of intervention. However, these videos are often also the basis for explorations of nuances and behavioral signals that may lead to researchers to even stronger and more reliable marker signals to the benefit of the clinical decision-making process. More recently, advances in computational methods such as computer vision techniques have begun to provide new avenues into the discovery and identification of novel behavioral markers associated with ASD. The tuning and training of these techniques relies on the availability of annotated video data for discovery. However, though assessment video datasets are uniquely valuable for these research purposes, it is difficult to share them because of privacy considerations. As a result, the current practice is to tightly control access to video sessions. This study proposes the development of privacy mechanisms for video observation sessions, specifically focusing on mechanisms that protect the facial identity of the child under observation while retaining gaze-based information critical for diagnostic assessments. Our approach is to implement known privacy mechanisms from the third-person video surveillance literature as well as novel mechanisms based on algorithms developed for non-photorealistic rendering and deep neural network based facial transfer that are applicable to this domain. We will evaluate these mechanisms on the extent of anonymization provided and the impact they have on behavioral assessment. The output of this research will be a characterization of known and novel privacy mechanisms with respect to utility for anonymization and behavioral observation. It is our goal that this study create technology to facilitate privacy enabled video data sharing, thus accelerating computational and behavioral research for ASD. Private videos can also be used to create training materials for behavioral analysts, thus increasing the breadth of content they can review in their training, allowing for faster training, enabling remote/online training, and creating opportunities for remote consultation in clinical practice.
This study creates methods to protect facial identity of children in videos of recorded behavioral observation sessions. Testing these methods on videos from 50 children with autism spectrum disorder (ASD), we will characterize well-established and novel video privacy mechanisms along two dimensions: (1) the extent to which they protect facial identity; and (2) the extent to which they retain the information needed for behavioral observations in the context of ASD assessment.
