| Localization is a foundation issue for mobile robots, which is critical to autonomous navigation. Although localization problem has been widely studied for many years, there are still a lot of problems to solve for stable applications. The one is how to localize a robot in indoor environments by low cost sensor, such as RGB-D camera. Another challenge is how to localize in dynamic environments which is continuously change and full of dynamic obstacles.In this thesis, a RGB-D camera based localization solution is researched and im-plemented on a service robot under indoor office environment. In the meanwhile, a localization method in dynamic changing environment is studied and improved for in-dustry railless intelligent AGV(Automated Guided Vehicle). The main contributions are as follows:1. A point-plane feature based RGB-D visual odometry algorithm is proposed. It combines FAST features and plane features to match pairwise RGB-D frames. Features are associated by a consistency constrain graph, while transformation is estimated by optimizing a joint error function. Experiments on dataset demon-strate that our approach gains both high accuracy and efficiency.2. A Monte Carlo localization(MCL) method for two wheeled robots with a RGB-D camera is proposed. It calibrates camera poses online by ground plane extraction to solve camera swinging problem and improves the particle filter algorithm in global localization by selectively filter as well as particle traversability map to handle the field of view limit of camera.3. A dual-map dynamic localization approach based on MCL algorithm is proposed for the large environment changing challenge. Once change is detected, the ap-proach begin doing simultaneous localization and mapping(SLAM) and use local map to track robot poses precisely. After the observation is consistent with global static map, it recovers to normal particle filter procedure. Experiments demon-strate that our approach is more stable and more robust than MCL algorithm. |
PDF Full Text Request