A sequential bayesian framework for mapping and localization in dynamic 3D environments

Abstract

In this thesis, we present novel sequential Bayesian filtering approaches that enable environment modeling, environment perception and object-level recognition for autonomous robots using sequential visual measurements. For environment modeling, we propose a novel Bayesian filtering approach for key-frame-based visual SLAM with the motion-estimation-based process model. First, to ensure that the process and the measurement noise are independent (they are actually dependent in the case of a single sensor), we explicitly divide observations (i.e., image features) into two categories, common features consistently observed in the consecutive key-frame images and new features newly detected in the current key-frame image. Then two sets of image features are used for process and measurement models, separately. In addition, we formulate a novel Bayesian filtering framework for key-frame-based SLAM in order to solve the scalability problem and to improve the filter consistency. We demonstrate the performance of the proposed method in terms of the consistency of the global map and the accuracy of the estimated path. For environment perception, we present an efficient scene recognition scheme to localize a robot from the previously constructed map. For global localization, establishing accurate correspondences between current observations and the given map is required. Thus we present an efficient technique for determining feature correspondences from a lot of database images under scale and illumination changes. For this purpose, we introduce a scale optimization method to enhance the matching performance with the combination of the FAST detector and integral image-based SIFT descriptors that are computationally efficient. Then we present a particle-filter-based localization framework with decoupled visual measurements for process and measurement models. During robot navigation, the robots are required to automatically classify the moving objects from image sequences...