Color Image Segmentation Based on Bayesian Theorem for Mobile Robot Navigation

Image segmentation is a fundamental process in many image, video, and computer vision applications. Object extraction and object recognition are typical applications that use segmentation as a low level image processing. Most of the existing color image segmentation approaches, define a region based on color similarity. This assumption often makes it difficult for many algorithms to separate the objects of interest which consist of highlights, shadows and shading which causes inhomogeneous colors of the objects’ surface. Bayesian classification and decision making are based on probability theory and choosing the most probable or the lowest risk. A useful property of the statistical classifier like Bayesian is that, it is optimal in the sense that it minimizes the expected mis classification rate. However when the number of features increased, Bayesian classifier is quite expensive both in terms of computational time and memory. This thesis proposes a Bayesian color segmentation method which is robust and simple for real time color segmentation even in presence of environmental light effect. In this study a decision boundary equation, which is acquired from class conditional probability density function (PDF) of colors, based on Bayes decision theory has been used for desired color segmentation. The estimation of unknown PDF is a common problem and in this study Gaussian kernel function which is most widely used nonparametric density estimation method has been used for PDF calculation. Comparisons were made between the proposed method to the k-nearest neighbor (KNN) and support vector machine (SVM), methods for image segmentation. Experimental results show that the proposed algorithm works better than other two methods in terms of classifier accuracy with result of more than 99 percent successful segmentation of desired color in varying illumination. In order to show the real time ability and robustness of proposed method for color segmentation, experimental results conducted on vision based mobile robot for navigation. First the robot was trained by some training sample of desired target color in environment. The decision boundary which acquired in the teaching phase has been used for real time color segmentation as the robot move in the environment. Spatial information of desired color in segmented image has been used for calculating the robot heading angle which is used by mobile robot controller for navigation. However, all of the existing color image segmentation approaches are strongly application dependent. This study shows that proposed algorithm successfully cope with the varying illumination which causes uneven colors of the objects’ surface. The experimental results show the proposed algorithm is simple and robust, for real time application on vision based mobile robot for navigation, in spite of presence of other shapes and colors in the environment

Preview PDF FK_2009_22a.pdf Download (573kB)

Actions (login required)

View Item