Visual Bearing-Only Simultaneous Localization and Mapping with Improved Feature Matching

In this this paper, we present a solution to the simultaneous localization and mapping (SLAM) problem for a robot equipped with a single perspective camera. We track extracted features over multiple frames to estimate the depth information. To represent the joint posterior about the trajectory of the robot and a map of the environment, we apply a Rao-Blackwellized particle filter. We present a novel method to match features using a cost function that takes into account differences between the feature descriptor vectors as well as spatial information. To find an optimal matching between observed features, we apply a global optimization algorithm. Experimental results obtained with a real robot show that our approach is robust and tolerant to noise in the odometry information of the robot. Furthermore, we present experiments that demonstrate the superior performance of our feature matching technique compared to other approaches.