Skip to Main Content
This study sought to identify constraints that might lead to a concise system of recognizing fingerspelling hand shapes. Previous studies of grasping suggested that hand shape is controlled using combinations of a small number of neuromuscular synergies, but fingerspelling shapes appear to be more highly individuated and, therefore, might require a larger number of degrees of freedom. Static hand postures of the American Sign Language manual alphabet were recorded by measuring 17 joint angles. Principal components (PCs) analysis was compared to the use of subsets of individual variables (i.e., joint angles) for reduction in degrees of freedom. The first four PCs were similar across subjects. Classification using weightings from these four components was 86.6% accurate, while classification using four individual variables was 88.5% accurate (thumb abduction, as well as flexion at the index and middle finger proximal interphalangeal joints and the ring finger metacarpalphalangeal joint). When chosen for each subject, particular four-variable subsets yielded correct rates above 95%. This superior performance of variable subsets over PC weighting vectors suggests that the reduction in degrees of freedom is due to biomechanical and neuromuscular constraints rather than synergistic control. Thus, in future application to dynamic fingerspelling, reasonable recognition accuracy might be achieved with a significant reduction in both computational and measured degrees of freedom.