Localization is an important part in the field of mobile robot, also it is precondition and guarantee for mobile robot’s navigation. In mobile robot’s localization, visual information are widely adopted. But visual information are potentially interfered by different noise exiting around the surroundings. So robustness of localization algorithm may be depressed because of these noises. To solve the “perception alias” problem and uncertainty in the large scale environment, this thesis proposed mobile robot’s localization based on monocular visual sensor . HSI-SIFT interest point feature is an robust algorithm for feature description under complex environment and variable illumination, which can achieve successful feature matching against scale change, partial occlusion, cluttered background, and scene dynamic. HSI-SIFT local visual features are extracted from multiple images obtained in the tranining stage. PCA is applied to reduce the quantity, size and complexity, and then, a modified robust RPCL algorithm is implemented to learn the natural cluster of these reduced features. The Location Prototype Feature Models of each topological nodes are constructed according to individual locations. As is well known, topological and metric mapping both have its own advantages and disadvantages. A hybrid hierarchical environment modeling is proposed which combines topological modeling and feature-based metric modeling method to extend to large scale environments, thus reducing the computation complexity and improving the precision of environment. The main idea is to connect local metric models by means of multi-scale hierarchical topological models. THSI allows a compact environment model which does not require global metric consistency and permits both precision and robustness. Global localization is the most important behaviors for autonomous navigation of robot, for which robustness and precision a necessity. Localization based on our proposed hybrid hierarchical environment modeling can achieve robust global localization and precise positioning. Door-plate numbers are selected as visual landmarks for topological localization. An adaptive colorful segmentation approach based on PCA and scale-space filtering is proposed. We can also use HSI-SIFT local feature in query image to recognize current locations via voting schemes. Furthermore, the probabilistic method based on a recursive Bayesian Filtering can be adopted to solve the problem of perceptual aliasing and image variability. A method which can obtain the metric pose estimation by monocular vision is proposed which borrows the concept of binocular vision.
修改评论