University Honors

Human pose estimation tools are designed to automatically detect facial and body key points and can be used to track humans in the environment. Typically, these are applied on adult data, and their application with young humans, such as infants, has not been adequately explored. These tools could be proven very useful in applications and research with infant populations that focus on movement assessment and diagnosis. The goal of this work was to examine the validity of a pose estimation tool, OpenPose, to track infant arm movement during reaching. A dataset involving videos from 36 infants (7 neurodivergent infants) who are less than 12 months old was used. The total number of reaching actions analyzed was 435. The wrist joint of each infant was manually tracked to obtain the 2D coordinates of its position throughout the reaching action. OpenPose was run in the same reaching segments to obtain the coordinates of the same keypoint. The ability of OpenPose to accurately track wrist joint motion was assessed through the direct comparison between the two methods. Certain parameters, such as different camera angles were also considered in the comparison. Bland-Altman and correlation analysis was performed in the coordinates obtained via the two methods and across the different camera angles. Our findings support a satisfactory performance of OpenPose to track infant reaching motion across the different angles, when compared to manual tracking. This work provides valuable insights into the use of OpenPose for motion tracking specifically in infants, an important yet under-represented population in this area of research.