Optimal Features and Classes for Estimating Mobile Robot Orientation Based on Support Vector Machine

In order for a mobile robot to perform its assigned tasks, it often requires a representation of its environment such as knowledge of how to navigate in its environment, and a method for determining its position in the environment. A major problem in computer vision and machine learning is to achieve a good feature as it can largely determine the performance of a vision system. A good feature should be informative, invariant to noise or a given set of transformations, and fast to compute. Also, in certain settings sparsity of the feature response, either across images or within a single image, is desired. Our objective of this paper is to obtain optimal features as well as determining the optimal class of angle in order to estimate mobile robot orientation single or unified images from two camera orientations. We introduce feature selection process before classifying features based on support vector machine classifier. We achieve better accuracy rate by only reducing its feature number from 30 features down to only 17 features on unified images. Furthermore, we also find that only 5 classes of robot angles are sufficient to estimate robot orientation correctly.