An online Bayesian classifier for object identification

Many autonomous mobile robots use a camera as a primary sensor for object recognition in the environment. The problem is that classifying an object in a camera image can be difficult for a robot controller. One possible solution is to use a Bayesian classifier with online learning to help the robot identify objects in an unstructured, realistic environment. This paper describes the work that has been done to develop an online Bayesian classifer for use with a low-cost color camera on a mobile robot. The theory behind the classifier is briefly described, followed by the experimental results of a Bayesian classifier using off-line learning of RGB values for identifying the colors of m&m candies by a sorting robot. The extension of this classifier to incorporate on-line learning is then described, followed by a proposed approach to incorporate the classifier on a mobile robot with a larger field of view than the sorting robot.