Abstracts Category : Other

Learning image-based localization

by Lili Meng

Image-based localization plays a vital role in many tasks of robotics and computervision, such as global localization, recovery from tracking failure, and loop closuredetection. Recent methods based on regression forests for camera relocalizationdirectly predict 3D world locations for 2D image locations to guide camera poseoptimization. During training, each tree greedily splits the samples to minimizethe spatial variance. This thesis develops techniques to improve the performancecamera pose estimation based on regression forests method and extends its application domains. First, random features and sparse features are combined so thatthe new method only requires an RGB image in the testing. After that, a label-freesample-balanced objective is developed to encourage equal numbers of samplesin the left and right sub-trees, and a novel backtracking scheme is developed toremedy the incorrect 2D-3D correspondence in the leaf nodes caused by greedysplitting. Furthermore, the methods based on regression forests are extended to uselocal features in both training and test stages for outdoor applications, eliminatingtheir dependence on depth images. Finally, a new camera relocalization method isdeveloped using both points and lines. Experimental results on publicly availableindoor 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 simulationand real-world experiments, demonstrating the efficacy and efficiency of the developed system.