Please use this identifier to cite or link to this item:
Title: Recognition and localization of objects in relative scale for robotic applications
Authors: Md. Saiful Islam.
Keywords: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
Issue Date: 2007
Source: Md. Saiful Islam. (2007). Recognition and localization of objects in relative scale for robotic applications. Doctoral thesis, Nanyang Technological University, Singapore.
Abstract: Three-dimensional object recognition and localization are the central problems for many applications of computer vision. Especially, for many robotic applications, it is necessary to recognize and localize the objects of interest in cluttered environment with high efficiency. Examples of such applications include object manipulation, visual inspection, surveillance system, and landmark localization for visual navigation. An object in different image planes could have undergone different kinds of geometric and photometric transformations. Moreover, the background could be cluttered with other objects and the objects of interest could be partially occluded. The recognition and localization methods should be robust to all such transformations and conditions. In this work, a method of object recognize and localize in relative scale is proposed. The models of a set of objects of interest are represented by local shapes in a reference scale. Invariant moments of a small image patch around an interest point are used as a local feature, which is invariant to geometric and photometric transformations as well as robust to cluttering and partial occlusions. Local shape is represented by these local features extracted from several reference images of an object captured from different viewpoints. All these extracted local features are inserted into a hash table. Spatial relations between local features are also represented as shape graphs. In the recognition phase, features are extracted from a scene image in the estimated relative scale of the object. This is done after the compensation for affine deformation caused by viewpoint change. Then the potential matches for a feature of a scene image are obtained by indexing the hash table. Generalized Hough transformation (GHT) method is used for clustering matched features and hypothesizing about the identity and orientation of the object in the scene. The clusters are verified by shape graphs with the help of the least squares method. Features of the object of interest in stereo images matched with same reference features are used to solve correspondence problem in order to estimate the location of the identified object in a three-dimensional environment. In this work, we have proposed a novel method of object recognition and localization by minimizing the number of scales: model representation of the objects is carried out in a reference scale and recognition is carried out in a relative scale. In fact, we have developed a framework for recognition and localization of objects in relative scale. As both the representation and recognition are single scale processes, they are computationally efficient unlike a multi-scale method. Repeatability rate of detected interest points by the proposed relative scale method is higher than that of the existing methods. We have also introduced a new local invariant descriptor to facilitate feature matching in relative scale. Another key contribution of this work is a PCA-based hashing technique that makes the feature matching efficient. Furthermore, we have proposed a method to solve stereo correspondence problem using the knowledge about local features belonging to an object of interest which makes the localization easier. The proposed relative scale method has been implemented and experiments have been carried out on objects from well known object-image-libraries as well as a set of objects imaged in our laboratory. By the recognition method we have obtained up to 100% recognition rate. The recognition process takes only a few seconds by our primary MATLAB implementation in Windows environment. It may be possible to obtain real-time response by an optimized and careful implementation using faster hardware and software. Some potential applications of the relative scale method are also demonstrated. The method proposes an alternative for object recognition and localization and should be very useful for many practical applications.
Rights: Nanyang Technological University
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:SCSE Theses

Files in This Item:
File Description SizeFormat 
  Restricted Access
Main report12.43 MBAdobe PDFView/Open

Page view(s) 5

checked on Oct 20, 2020

Download(s) 5

checked on Oct 20, 2020

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.