Appearance-based mapping using minimalistic sensor models

摘要: This paper addresses the problem of localization and map construction by a mobile robot in an indoor environment. Instead trying to build high-fidelity geometric maps, we focus on constructing topological maps as they are less sensitive poor odometry estimates position errors. We propose modification standard SLAM algorithm which assumption that robots can obtain metric distance/bearing information landmarks is relaxed. Instead, registers distinctive sensor "signature", based its current location, used match positions. In our formulation this non-linear estimation problem, infer implicit measurements from image recognition algorithm. method for incrementally building uses panoramic camera images at various locations along path features it tracks update map. The very general does not require environment have uniquely features. Two algorithms implemented address problem. Iterated form Extended Kalman Filter (IEKF) batch-processed linearized ML estimator compared under odometric noise models.