Learning image-based localization

Image-based localization plays a vital role in many tasks of robotics and computer vision, such as global localization, recovery from tracking failure, and loop closure detection. Recent methods based on regression forests for camera relocalization directly predict 3D world locations for 2D image locations to guide camera pose optimization. During training, each tree greedily splits the samples to minimize the spatial variance. This thesis develops techniques to improve the performance camera pose estimation based on regression forests method and extends its application domains. First, random features and sparse features are combined so that the new method only requires an RGB image in the testing. After that, a label-free sample-balanced objective is developed to encourage equal numbers of samples in the left and right sub-trees, and a novel backtracking scheme is developed to remedy the incorrect 2D-3D correspondence in the leaf nodes caused by greedy splitting. Furthermore, the methods based on regression forests are extended to use local features in both training and test stages for outdoor applications, eliminating their dependence on depth images. Finally, a new camera relocalization method is developed using both points and lines. Experimental results on publicly available indoor and outdoor datasets demonstrate the efficacy of the developed approaches, showing superior or on-par accuracy with several state-of-the-art baselines. Moreover, an integrated software and hardware system is presented for mo- bile robot autonomous navigation in uneven and unstructured indoor environments. This modular and reusable software framework incorporates capabilities of perception and autonomous navigation. The system is evaluated are in both simulation and real-world experiments, demonstrating the efficacy and efficiency of the developed system.Applied Science, Faculty ofMechanical Engineering, Department ofGraduat