Robust Global Localization by Using Global Visual Features and Range Finders Data

Global localization is a challenging problem of using sensor data to estimate the pose of a robot in an environment when the starting pose is unknowm. The conventional probabilistic algorithms called Monte Carlo Positioning (MCL) is one of the most popular methods to solve this problem. MCL algorithms use a set of weighted particles to approximate the distribution probability of where the robot is located and it requires a wandering motion to converge to a single, high likelihood pose during global localization. Sometimes this wandering motion is not allowed in actual industrial applications. This paper presents a framework which incorporates image-based localization module into a conventional MCL algorithm. The core module in our proposed approach is called Double Re-localization Decision Process (DRDP) by performing two selection of relocation decisions before and after the pose update process with two different sensor sources. A compact global descriptor is used for fast image association and a scan matching using vanilla ICP (Iterative Closest Point) of Point-to-line metric is applied to obtain further pose of the proposal candidate. Several experiments are designed to verify the effectiveness of the our approach in indoor environment.