A multi-sensor-based mobile robot localization framework

This paper presents a novel multi-sensor-based robot localization framework inspired by human coarse-to-fine recognition mechanism to realize fast and robust localization in the process of robot navigation. This localization framework consists of two parts: coarse place recognition and accurate location estimation. The coarse place recognition is realized using an onboard camera, whereas an image retrieval system is employed. The coarse localization system utilizes feature matching between the observed image and the re-ranked retrieval images to infer the possible locations of the robot. To obtain the accurate pose of robot, a modified particle filter which is mainly based on the laser radar data is implemented. To integrate the two localization stages, the current states of the particles are monitored. Once the information entropy changed greatly or the number of the effective particles is in a low level, a set of pointed particles is generated based on these estimated locations and then injected into the modified particle filter timely to ensure that enough particles are surviving around the correct pose of robot. Experiments are conducted extensively in an office environment and the results exhibits great improvement on speed and stability of mobile robot localization compared to conventional localization methods.