Comparative Analysis of GRU and LSTM based Models for Pose Estimation in Pakistan Sign Language Recognition

Abstract

This study explores Sign Language Recognition (SLR) within the context of Pakistan Sign Language (PSL), aiming to bridge communication gaps between signers and non-signers. Sign languages employ handshapes, body gestures, and facial expressions to facilitate communication, addressing the worldwide linguistic needs of deaf communities. While significant efforts have been devoted to global SLR and Sign Language Translation (SLT) systems, limited attention has been paid to PSL. To address this gap, we propose a novel approach for dynamic word-level SLR, incorporating manual and non-manual features. The proposed method utilizes pose estimation RNN-based architectures (GRU and LSTM) on both our proprietary pronoun-based video dataset and the PkSLMNM dataset. By extracting key points from 3D coordinates within individuals, we propose several optimization functions for original and augmented datasets. We then compare the sequential classification potential of GRUs and LSTMs. Our findings reveal that GRU outperforms LSTM, achieving a 4% improvement in real-time classification accuracy on both augmented and original datasets, with an overall accuracy of 98.61%.