Identifying Children's "Stranger Danger" Behaviors

One of the most common mental health problems among children are anxiety disorders, with many children maintaining their anxiety well into adulthood. Anxiety impacts children’s ability to make friends, develop proper social skills, and participate in class which leads to poor educational outcomes. Therefore, it is important that researchers find ways to identify which children are at most risk for developing anxiety as early as possible to get them the treatments necessary. One of the most common predictors of anxiety is behavioral inhibition, which is an early personality trait that is characterized by fear and withdrawal to new situations and people. Being able to identify these behaviors helps us identify which kids should receive extra support to reduce their chances of developing anxiety later in life. Traditionally, researchers and therapists will watch how each child reacts to new situations and identify a variety of behaviors that are classified as risk factors for anxiety. While this type of behavioral coding is the most common and accurate way to assess risk behaviors among children, this is very time consuming to watch hours of videos and requires many hours of specialized training to make sure all behaviors are identified in the same manner across different coders. Recent advancements in machine learning have created the ability to analyze human movement and behavior using computer vision, however few researchers in the field of psychology have attempted using computer vision to digitize their traditional behavioral coding schemes. This continuation project builds on work from Spring 2025, where the student team successfully developed a web-based tool for analyzing video footage of children in "Stranger Danger" scenarios. The system labeled children’s proximity to others in the room and estimated fear and freezing behaviors using pose estimation and facial recognition models. The goal of the Fall 2025 project is to further improve the robustness, scalability, and accuracy of the existing system. Key areas of focus include reducing frame detection failures, increasing prediction precision, adding behavioral and audio event detection capabilities, and expanding these features to a “scary robot” scenario directly following the "stranger danger” scenario. Proposed extensions include: • Enhanced person detection in relation to a static location (parent chair, robot) to improve accuracy of proximity • Behavior recognition possibly by utilizing pos detection: approach, withdrawal, gaze shifts, speaking, crying • Audio-event synchronization: detect when the stranger speaks, child responds, toy is offered/taken, when experiment is started More detailed labels for the data are in the experiment and please look at the previous work under the background reading section to