In order robots to manipulate objects of interest autonomously, the 3D information about the objects of interest is required. In order to obtain the 3D object information, various range sensors have been traditionally being used in the computer vision society. Among the available sensor types, stereo vision system can be used to retrieve the 3D object information. The hardware and software developments and the new computer vision systems that are being developed prompt lot of researchers to use stereo camera systems for this purpose. Humans have great learning capability; therefore humans can easily identify and handle new objects from the environment. Such capability to computers is strongly limited with respect to time, intelligence, and computational power. There are infinite kinds of objects available in the nature; though some may differ small for humans, any change in object property can significantly influence the performance of computer vision system. Therefore often computer vision researchers try to develop algorithms and systems addressing only specific kind of objects; they aim that their developed systems are safe, functional, reliable, and robust.
Object localization using stereo vision requires stereo feature points to be uniquely identified. Objects without any surface texture hardly provide unique stereo features. Therefore localization of objects without any surface texture is a difficult task. The content of the thesis presents the methods that are meaningful for the identification, localization, and 3D reconstruction of domestic objects in a service robotic environment using stereo vision. The considered objects have no surface texture information. A novel and simple approach for tracking the object pose using two object points is presented in the thesis. As the stereo feature extraction is a critical part for the localization of the considered objects various methods have been investigated to find the stereo features. The investigated methods aim at minimal 3D object reconstruction errors. Constrained and unconstrained approaches using single and stereo cameras have been suggested for the considered objects’ localization. Special attention has been given to analyze the cause of 3D object reconstruction errors. Selection of stereo features for the localization of objects without surface texture information also has been investigated. In the object localization process, robust color object segmentation using attention and closed loop methods have been used. Shape based object identification using Hu-Moments has been considered.
The effectiveness of the proposed and considered methods is demonstrated through the experimental results of domestic objects localization in a service robotic application.