Long Term Localization And Mapping For Mobile Robots

With the development in the filed of mobile robotics, the autonomy of the robots has been significantly promoted, leading to larger working area, lower deployment prerequisites and richer handled tasks, which is important for further application of robots. In the technique side, the first problem for mobile robots is the localization. Currently, a usual scheme is to build the map first and then run the localization algorithm. The map building is achieved by the simultaneous localization and mapping (SLAM). However, this scheme is not appropriate for long-term operating robots, such as warehouse robots and inspection robots, due to the out-of-dated map. Focusing on this problem, this thesis proposed a multi-session SLAM system which supports long term operation. Specifically, its main idea is to divide the SLAM process into multiple sessions. During each session, the environment is assumed to be static. The dynamic can occur across sessions. Upon this idea, the multi-session SLAM is developed to enable the robots adaptive to the environment changes and build in-date maps, thus leading to long term localization with acceptable complexity. The system consists of three parts:(1) The information perspective. By exploring the Kullback-Leibler divergence, the quantita-tive measure of the difference between the two sets of poses generating the maps is derived, which is utilized for the development of pose pruning algorithm. After pruning the pose with redundant observations, the sparse factors generation is proposed for reserving geometric information in the pruned poses. With these two methods, the size of the graph is related to the size of the map instead of the length of the trajectory.(2) The observation perspective. The pose estimation is stated as a probabilistic model inde-pendent of the sensor types. This model is found to be able to represent many conventional pose estimation algorithms. Derived from the model, a framework for multi-sensor fusion is proposed for better estimation. Turning to the dynamic environment, the dynamics is inserted into the model as a state, transforming the detection of dynamic environment to a probabilistic inference problem with higher efficiency and accuracy.(3) The architecture perspective. The single session SLAM is divided into multiple sessions. The pose pruning and dynamic environment detection are connected as the graph complexity con-trol module. By deploying the module across sessions, the SLAM can be applied to the scenario of long-term operation with slight loss of accuracy.During the experiments, the system is tested on many datasets including lots of variations such as sensor type, environment and geometric region. The results all validate the efficiency and effectiveness of the proposed system and the localization error is found to be non-accumulative. Finally, these works are believed to provide the primary theoretic and practical efforts to support the SLAM of long-term operating robots.