Theoretical and Practical Aspects of 2D Shape Assessment using Selective Fixations and Artificial Potential Functions
Hulya Yalcin


Abstract


In this thesis, we present a mathematical framework for selective fixation generation and its usage through artificial potential functions. Recent research suggests that computational advantages may be realized for vision problems when cameras are capable of fixating on areas of interest. Studies on the formalization of the task of selecting a sequence of fixation points and saccade directions that efficiently examine the area being searched have been limited. In our earlier work, we have developed a simple heuristic algorithm based on simple computations on the periphery region of current fixation point to determine a new visual target. The MS research presented here aims to construct the underlying theory of selective fixation control using artificial potential functions. We implemented the approach on BUVIS, an inspection system endowed with visual attention capability. Experiments demonstrating the robustness to lighting conditions and real-timeness of the system are presented for industrial objects with varying illumination conditions.


Introduction

In machine vision, there is a growing trend towards goal-oriented systems, where effectiveness in the world is taken as a structuring constraint in designing systems. With this philosophy, the need to manage resources more efficiently has become a crucial feature for vision systems operating in real-time dynamic environments. A system using real-world visual input for decision-making must ignore the irrelevant visual stimuli, attend to the salient image points and place reliable priorities on tasks and resources. Such work must confront the issues of deciding what to perceive "next". Traditionally, research in machine vision has concentrated on thorough analysis of acquired images. This contrasts with human perception. For human visual behaviour, selectively gathering information about the environment is characterized by the ability to fixate on a point of interest and the ability to select new fixation locations. Humans can shift the attention by concentrating on a part of the field of view. Motivated by human visual system, recent studies in machine vision have tried to mimic this behaviour. These studies depend on detection of visual targets and interrogation around those targets instead of processing whole image. The process of identifying and selecting new visual targets is referred to as selective fixation control. Remarkable efficiency on computational grounds has been observed in machine vision systems motivated by fixation control. In this manner, visual resources are allocated to process only a small part of the whole scene. In our earlier work, we have devised a visual inspection system - based on a simple heuristic algorithm - which selectively fixates on the interesting parts of an incoming image and uses the attentional sequence thus gathered in a task-dependent manner - specifically the task of tracing an object's outline. In this thesis, we present a unified mathematical framework based on artificial potential functions - which formalizes this algorithm. It turns out that this formalism tantamounts to a feedback based approach - that naturally leads to the automatic generation of camera actuator commands that cause the camera to find a sequence of fixation points and use the attentional sequence thus generated a in top-down driven task - such as the tracing of salient parts of the objects. In contrast to open loop, process-all systems, the provable correctness of such a clo! sed loop system can be investigate.

                                                                [ More about this in html... ]

                                                                My thesis [pdf]

Resume | Research | Main Page
Carnegie Mellon University, Robotics Institute
5000 Forbes Av., Pittsburgh, PA, 15213

hulya@ri.cmu.edu