Reconocimiento de objetos deformables en condiciones variables usando modelos estadísticos de forma y apariencia

This thesis contains mathematical and computational tools to solve the problem of recognize deformable objects in uncontrolled environments, in variable lighting conditions and object with partial occlusion. With those tools is possible to propose and implement algorithms that allow to improve exist...

Descripción completa

Detalles Bibliográficos
Autor: PAVEL HERRERA DOMINGUEZ
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2010
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/515
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/515
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Reconocimiento de objetos/Object recognition
info:eu-repo/classification/Detección de objetos/Object detection
info:eu-repo/classification/Visión por computadora/Computer vision
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
Descripción
Sumario:This thesis contains mathematical and computational tools to solve the problem of recognize deformable objects in uncontrolled environments, in variable lighting conditions and object with partial occlusion. With those tools is possible to propose and implement algorithms that allow to improve existing methods, these improvements are the main contributions of this thesis. First, we propose to partition the object in parts formed with the elements of the object representation, as a result we have a graph or a set of clusters. Thus, with the structured representation of the object we propose several algorithms to recognize the object, two for the partial occlusion case, and one algorithm for the variable lighting environments. For occlusion, the rst method consists in arrange the clusters in a hierarchy; later the algorithm uses the hierarchy to merge the parts and retrieve partial information of the object. The second method uses the graph to set a neighborhood system to build a Markov random eld; after, this eld is employed to infer the model parameters that are assumed to be hidden by the occlusion. Additionally, the proposed algorithm for recognizing the object in variable lighting conditions uses the graph to perform a search over the elements assumed to be in the image in order to get the rest of the elements not found because of the lighting conditions. Experiments were performed in order to validate the proposed algorithms. For the occlusion case the experiments were done over faces and car-side classes from Caltech database. With the objective of investigate the behavior of the algorithms, images from the database were modied, adding synthetic occlusion. The experiment for the algorithm that considers variable lighting environments was performed over images taken from Yale B database. The results for all the performed experiments show improvements up to 20% points in the detection rates, comparing them with the results obtained by using SumMaxMaps, the original algorithm for the used representation.