Multi-stage localization given topological map for autonomous robots

Vision-based place recognition is of a particular importance for autonomous systems that aim to navigate intelligently in a human-inhabited environment. Given a topological map of an indoor environment, the autonomous system shall localize itself invariantly with different illumination and imaging conditions. To address these challenges, we propose to use global-local feature extraction and classification in multiple stages. Scale Invariant Feature Transform (SIFT) is used as a local feature detector and descriptor which has been proven to be a robust local invariant feature descriptor. Fourier Transform, Hue Saturation Value (HSV), and Hough Transform are used as global features. The Support Vector Machines (SVM) is used to localize the system by classifying the global features. However the K-nearest neighbors matching technique (K-NN) is used to support SVM's classification in ambiguous decisions by classifying the local features.