Mathematical Modelling of Geometric Transformations for Mizo Sign Language Gesture Recognition

Abstract

The work introduces the first ever computational dataset for Mizo Sign Language, a sign language from India which has limited resources and lacked a digital presence. 25 distinct hand gestures were captured using a Kinect V2 sensor, and the system captured them in different lighting conditions. To enhance the variability of the gesture classes, the original 3 GB dataset of images was preprocessed using a mathematical normalization technique which increased it to 10 GB. Since each gesture in the dataset is stored as a high-dimensional tensor, posebased techniques and attention models based on Transformer architectures could be used. This goal of this paper is suited for advancing gesture recognition computationally based on mathematical and machine learning methods which especially need spatial and temporal considerations. This advances fields of computer vision, linguistics, and mathematics.