Reconhecimento de gestos da Língua Brasileira de Sinais através de máquinas de vetores de suporte e campos aleatórios condicionais ocultos

Abstract

This work investigates the use of Support Vector Machines and Hidden Conditional Random Fields in the recognition of signs from the Brazilian Sign Language (Língua Brasileira de Sinais, Libras). Employing basic concepts from Vapnik’s Statistical Learning Theory, the gesture recognition problem is cast as a supervised learning problem defined over images and image streams, avoiding the inherent ill-conditioning present in many density estimation problems through the use of discriminative classification models. From linguistic studies on the structural formation of the Libras sign, a two-layer recognition architecture has been created to operate over features extracted from depth images captured through a depth sensor. This work utilizes quantitative approaches for performance assessment, performing comparisons through contingency tables and statistical hypothesis tests; revealing statistically significant results favoring the aforementioned choice of classification models. Results have shown how the multiclass SVMs organized in Directed Acyclic Graphs provided a needed balance between efficiency and accuracy in the classification of the sub-lexical structures of the Libras, whereas Hidden Conditional Random Fields boosted the system’s recognition rates without for this sacrificing its generalization for unobserved instances.