Author(s): P Rajesh, M Karthik, P Deepthi Sai Sree, M Praveen, M Geethika Navya and P Harshitha

Abstract: The recognition of sign language among hitherto unheard-off people is also an issue that is not so easily tackled since the styles of signing, patterns of motions and the forms of hands are varied among different users. The performance of many current recognition systems has only been shown to be high when tested on signers that were part of the training set thus limiting the applicability of these systems to practice. The current paper puts forward a signer-neutral recognition system that can be assessed with the help of the AUTSL dataset comprising 226 isolated sign types delivered by several subjects. The system does not handle raw video frame, but instead, it uses gestures in sequences of hand and upper-body landmarks that are extracted on a frame-by-frame basis. Such landmark sequences are done using a normalization that minimizes appearance-dependent variability and modeled with a Transformer encoder with the capacity to reflect long range temporal relations. An attention-based temporal aggregation mechanism is employed to emphasize informative motion segments while suppressing redundant frames. Tests with a rigorous signer-disjoin evaluation program show validation and test accuracies of 85.45 and 83.52 suggesting high test precision in the case of signers who are not seen.

DOI: https://www.doi.org/10.33545/26633582.2026.v8.i2a.273

Pages: 24-32 | Views: 198 | Downloads: 47

Download Full Article: Click Here