Use this URL to cite or link to this record in EThOS:
Title: Machine vision for shape and object recognition
Author: D'Souza, Collin
ISNI:       0000 0001 3401 2630
Awarding Body: Nottingham Trent University
Current Institution: Nottingham Trent University
Date of Award: 2000
Availability of Full Text:
Access from EThOS:
Access from Institution:
Robots today can perform assembly and material handling jobs with speed and precision, yet compared to human workers, robots are hampered by their deficiency of sensory perception. In most instances in the industry, the robots have to be reprogrammed if the positions of the assembly components change. This is a tedious task. The current work has been motivated by the concept of using machine vision in the initial stages to recognise and locate components in order to aid a robot in performing assembly autonomously. Machine vision is a useful robotic sensor since it mimics the human sense of vision and allows for non-contact measurement of the environment. A 3-D object gives rise to an infinite variety of 2D images or views, because of the infinite number of possible poses relative to the viewer. In order for a vision system to be effective in assisting a robot to approach an object autonomously, two things must be known -"what" object is seen and "where" it is located ie. the object has to be recognised and its co-ordinates must be known. The objects used in this investigation are of the polyhedral type, which resemble mechanical objects. In many instances, where machine vision has been used, only 2-D silhouettes of the objects have been made use of for recognition. This work considers recognition of 3-D objects from any orientation, considering both inplane rotation and rotation in depth. Only 2-D information is used to infer 3-D information. The use of artificial neural networks (ANNs) has been made for learning and recall. In several approaches utilising ANNs, some transforms are used for extracting invariant features. Opposed to this, this investigation explores the area of extracting salient features and relating them to recognise objects. The system developed utilises two CCD cameras in a stereoscopic set-up for obtaining 3-D information about the object. A hierarchical system has been developed in software for object recognition. Training of each object is done by presenting characteristic views of each polyhedral object. Ideally, the vision system on the robot arm should be moved around the object to obtain the characteristic views. In the simulations however, the objects are rotated and displaced to mimic robot movement. Each image of an object is processed and features such as corners and edges are extracted. The relationships between the features are determined to identify the types of surfaces. The relationships between the surfaces is then encoded and input into the artificial neural network. Incremental learning of several views of the object is done using Fuzzy ARTMAP for all objects. The system has been improved later by adding a pattern rotation layer and modifying the ANN to minimise training of the system to a few characteristic views. When a single novel image of an object is presented, the correct object can be recognised. Since stereo vision is used, the location of the object with respect to the cameras is also determined. The system was tested for recognising four objects viz. a cube, pyramid, triangular prism and pentagonal prism having five facet types. Typical recognition times were 18 seconds on a computer with a 166 Mhz processor. The vision system has the potential to be implemented on a robot arm in the future in the 'eye-in-hand' configuration. The robot arm can be moved precisely to obtain the two images or a miniature stereo setup can be mounted close to the gripper. This system could thus be used to identify, locate and approach mechanical objects autonomously.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Stereo; Neural networks