Robust color-based vision for mobile robots
MetadataShow full item record
An intelligent agent operating in the real world needs to be fully aware of the surrounding environment to make the best decision possible at any given point of time. There are many forms of input devices for a robot that gather real-time information of the surroundings, such as video cameras, laser/sonar range finders, and GPS to name a few. In this thesis, a vision system for a mobile robot navigating through different illumination conditions is investigated. Many state-of-the-art object recognition algorithms employ methods running on grayscale images, because using color is difficult for several reasons: (a) The object-of-interest's true colors may not be recorded by the camera hardware due to illumination artifacts, and (b) colors are often too ambiguous to be a robust visual descriptor of an object. In this dissertation, we address these two challenges and present new color-based vision algorithms for mobile robots that are robust and efficient. The first part of this dissertation focuses on the problem of color constancy for mobile robots under different lighting conditions. Specifically, We use a generate-and-test methodology to evaluate which simulated global illumination condition leads to the generated view that most closely matches what the robot actually sees. We assume the diagonal color model when generating views of the object of interest under previously unseen conditions. In the second part of the dissertation, we present a vision framework for mobile robots that enables observation of illumination artifacts in a scene and reasoning about the lighting conditions to achieve robust color-based object tracking. Before implementing this framework, we first devise a novel vision-based localization correction algorithm with graphics hardware support, and present how to find possibly shaded regions in the recorded scene by using techniques from 3D computer graphics. We then demonstrate how to integrate a color-based object tracker from the first part of this dissertation with our vision framework. Even with the contributions from the first two parts of the dissertation, there remains some degree of uncertainty in robot's assessment of an object's true color. The final part of this dissertation introduces a novel algorithm to overcome this uncertainty in color-based object tracking. We show how an agent can achieve robust color-based object tracking by combining multiple different visual characteristics of an object for more accurate robot vision in the real world.