LIBRAS-VC: uma plataforma web para reconhecimento e tradução em Língua Brasileira de Sinais por visão computacional

Abstract

Communication between deaf and hearing people in Brazil is frequently hindered by the lack of technological tools capable of automatically translating Brazilian Sign Language (LIBRAS). This paper presents LIBRAS-VC, a web platform developed following the Design Science Research (DSR) methodology, aimed at recognizing and translating LIBRAS signs in real time through computer vision. The system integrates the MediaPipe library to extract body landmarks from hands, face, and upper trunk, composing a 156- dimensional feature vector per frame. For static sign classification, the Random Forest algorithm is employed; for dynamic signs, Dynamic Time Warping (DTW) is used. A complementary linguistic post-processing module, based on Large Language Models (LLM), converts sequences of recognized tokens into natural Portuguese sentences. The system architecture consists of a FastAPI backend and a React frontend, featuring 21 REST endpoints and the MVVM pattern in the interface. The constructed dataset contains 426 samples distributed across 17 classes of the LIBRAS alphabet and special signs. Results demonstrate the feasibility of the proposed computational approach for promoting digital accessibility and social inclusion of the deaf community.