Vision-based self-localization in non-stationary environments by using support vector machines

Self-localization is one fundamental problem in robotics, and important for various tasks. Most previous methods for self-localization is based on comparison between an environment-map and observed features (or landmarks). These approaches often fail in a dynamic and large environment with noisy sensors. To solve this problem, we propose a vision-based method for learning-based self-localization by using support vector machine (SVM). We designed an effective filter to extract features robust against sensor uncertainty as well as object movement. Also, we propose to use a set of SVMs to minimize misrecognition rate. In experiments with a real robot and an omni-directional vision sensor, effectiveness of the proposed method will be demonstrated.