Mobile robot self-localization in unstructured environments based on observation localizability estimation with low-cost laser range-finder and RGB-D sensors

Abstract

When service robots work in human environments, unexpected and unknown moving people may deteriorate the convergence of robot localization or even cause failure localization if the environment is crowded. In this article, a multisensor observation localizability estimation method is proposed and implemented for supporting reliable robot localization in unstructured environments with low-cost sensors. The contribution of the approach is a strategy that combines noisy laser range-finder data and RGB-D data for estimating the dynamic localizability matrix in a probabilistic framework. By aligning two sensor frames, the unreliable part of the laser readings that hits unexpected moving people is fast extracted according to the output of a RGB-D-based human detector, so that the influence of unexpected moving people on laser observations can be explicitly factored out. The method is easy for implementation and is highly desirable to ensure robustness and real-time performance for long-term operation in populated environments. Comparative experiments are conducted and the results confirm the effectiveness and reliability of the proposed method in improving the localization accuracy and reliability in dynamic environments.