%0 Thesis %1 bastos2015reconhecimento %A Bastos, Igor Leonardo Oliveira %C Salvador, Bahia, Brazil %D 2015 %E Ângelo, Michele Fúlvia %E Loula, Angelo Conrado %E Apolinário Júnior, Antonio Lopes %E Gonzaga, Adilson %I Universidade Federal da Bahia (UFBA) %K artificial-neural-network brazilian-sign-language bsl gesture-recognition histogram-of-oriented-gradients libras real zernike-invariant-moments %T Reconhecimento de sinais da Libras utilizando descritores de forma e redes neurais artificiais %U https://repositorio.ufba.br/ri/handle/ri/19374 %X Gestures are nonverbal bodily actions used for the expression of some meaning. These include movements of hand, face, arms, fingers, and others, and they are addressed by work that aim at recognizing them to promote human interactions with computer systems. Because of the wide applicability of the gesture recognition, it has been noticed that these works are becoming increasingly common, and the techniques and methodologies applied are getting even more elaborated, providing better and better results. The choice of techniques applied to the recognition of gestures varies according to the strategy employed in each work and which aspects are used for this recognition. For instance, there are studies based on use of statistical models. Other studies are based on the acquisition of geometrical characteristics of hands and body parts. On the other hand, studies, including the present one, uses descriptors and classifiers, responsible by extracting features of images that are relevant for recognition and classifying the gestures based on this information. In this context, this work aims to develop, implement and present an approach to recognize gestures, being based in a literature review about the main techniques and methodologies used for this purpose. The Brazilian Sign Language (Libras) was choosed as practical field. In order to extract the image information, we composed a feature vector resultant from the application of the descriptors Histogram Oriented Gradients (HOG) and Invariant Zernike Moments (MIZ), which gather information about shapes and edges in the images. For the recognition, the classifier Multilayer Perceptron was used, which is arranged in a architecture where the classification process is divided into two stages. Due to the absence of public Libras datasets, we created, with the help of experts and deaf students, a dataset containing 9600 images related to 40 Libras signs. In this way, the approach started with the creation of the image dataset and had the classification of the signs as final step. At the end, tests were conducted and yielded 96.77% recognition rate, revealing a high success rate. This result was validated considering potential threats to the approach, such as tests considering a non-present person at the training set of the classifier and tests applying the approach in a public gesture dataset.

@mastersthesis{bastos2015reconhecimento, abstract = {Gestures are nonverbal bodily actions used for the expression of some meaning. These include movements of hand, face, arms, fingers, and others, and they are addressed by work that aim at recognizing them to promote human interactions with computer systems. Because of the wide applicability of the gesture recognition, it has been noticed that these works are becoming increasingly common, and the techniques and methodologies applied are getting even more elaborated, providing better and better results. The choice of techniques applied to the recognition of gestures varies according to the strategy employed in each work and which aspects are used for this recognition. For instance, there are studies based on use of statistical models. Other studies are based on the acquisition of geometrical characteristics of hands and body parts. On the other hand, studies, including the present one, uses descriptors and classifiers, responsible by extracting features of images that are relevant for recognition and classifying the gestures based on this information. In this context, this work aims to develop, implement and present an approach to recognize gestures, being based in a literature review about the main techniques and methodologies used for this purpose. The Brazilian Sign Language (Libras) was choosed as practical field. In order to extract the image information, we composed a feature vector resultant from the application of the descriptors Histogram Oriented Gradients (HOG) and Invariant Zernike Moments (MIZ), which gather information about shapes and edges in the images. For the recognition, the classifier Multilayer Perceptron was used, which is arranged in a architecture where the classification process is divided into two stages. Due to the absence of public Libras datasets, we created, with the help of experts and deaf students, a dataset containing 9600 images related to 40 Libras signs. In this way, the approach started with the creation of the image dataset and had the classification of the signs as final step. At the end, tests were conducted and yielded 96.77% recognition rate, revealing a high success rate. This result was validated considering potential threats to the approach, such as tests considering a non-present person at the training set of the classifier and tests applying the approach in a public gesture dataset.}, added-at = {2019-09-13T16:47:15.000+0200}, address = {Salvador, Bahia, Brazil}, author = {Bastos, Igor Leonardo Oliveira}, biburl = {https://www.bibsonomy.org/bibtex/2f5bb15ab17f2b24609c2e6a5ac4e493d/jpmor}, editor = {Ângelo, Michele Fúlvia and Loula, Angelo Conrado and Apolinário Júnior, Antonio Lopes and Gonzaga, Adilson}, id = {http://repositorio.ufba.br/ri/handle/ri/19374}, interhash = {17247e4dfa96111f6457b3a7e2b77bce}, intrahash = {f5bb15ab17f2b24609c2e6a5ac4e493d}, keywords = {artificial-neural-network brazilian-sign-language bsl gesture-recognition histogram-of-oriented-gradients libras real zernike-invariant-moments}, publisher = {Universidade Federal da Bahia (UFBA)}, school = {Universidade Federal da Bahia (UFBA)}, timestamp = {2020-10-07T13:36:50.000+0200}, title = {Reconhecimento de sinais da Libras utilizando descritores de forma e redes neurais artificiais}, type = {Master Thesis}, url = {https://repositorio.ufba.br/ri/handle/ri/19374}, year = 2015 }