A classification method of the static and dynamic states of continuous fingerspelling

This study proposes a method for classifying static and dynamic states of continuous fingerspelling. We perform two types of classification: 3-class classification distinguishing static, dynamic, and transition states, and 2-class classification separating transition intervals from non-transition intervals (fingerspelling segments). Using MediaPipe to detect hand skeleton coordinates and a Transformer model, it incorporates the SVM recognition results. Adding SVM results improved classification recognition rates from 35.9 % to 51.3 % for 3-class classification and from 82.1 % to 84.6 % for 2-class classification.