Aplicación de factorización de trayectoria espacial en reconocimiento automatico de señales para uso en sistema de control de dispositivos domóticos.

Abstract

This project presents the application of a novel Machine Vision (MA) technique, known as Trajectory Space Factorization (TSF), which reconstructs the 3D structure of objects in video, from their image sequences (frames) in 2D. Moreover, the idealization and execution of a Domotic Device Control System (CODDO), which translates a defined set of gesture signs used in the American Sign Language, into instructions that handle intelligent terminal devices by means of Amazon Alexa Smart Home Skill. The signs or signals are automatically recognized by retrieving the 3D structure of the hands. Subsequently, their 3D representation is mapped to a single dimension, this being the input to a translated sign classifier model. The chosen classification model is the result of a comparison between a Support Vector Machine (SVM) model and a Recurrent Neural Network (RNN) model, of which the latter is the most accurate. The final part of the document contains the results of the TSF application, and implementation of the CODDO system for the proposed case study, as well as conclusions and recommendations on the development of this project.