Scan Similarity-based Pose Graph Construction Method for Graph SLAM

Abstract

In order to perform missions such as exploring unknown target environment or rescuing people at human-inaccessible areas autonomously, robots are required to know their surrounding environment information. This environment recognition process is widely known as simultaneous localization and mapping (SLAM) which the robots build a map of surrounding environment and use this map to compute their pose simultaneously. The SLAM problem has been known one of the fundamental techniques for autonomous robot operation and studied over two decades in robotic society. It is still being actively researched and there is much room for improvement. In this dissertation, a novel pose graph construction method for graph SLAM is presented by defining the laser scan descriptor which reflects the geometrical property of the scan data and proposing the similarity computation method between two scans. Scan matching algorithms are used mainly to register two scan data in the front-end of the graph-based SLAM process. These scan matching algorithms have following two characteristics. The first one is that the success or failure of the registration process depends on the similarity of the geometric information of the environmental structure or the overlapping areas between the two scan data. The second one is that the registration error occurs due to the discontinuity of the laser sensor information and the measurement noise even if the two scans are successfully registered. Considering the first characteristic, the laser scan descriptor which can describe the geometrical information of the surrounding environments is proposed and the similarity computation method between two scan data is suggested based on the proposed scan descriptor. Generally, the errors are accumulated during the pose graph construction process because of the second characteristic. To alleviate this error accumulation phenomenon, dynamic keyframe selection method which selects the reference scan data every time step is proposed. Additionally, loop closure detection method is suggested by exploiting the proposed laser scan descriptor and the scan similarity computation method. Through the simulations, the odometry estimation is performed with the artificially generated sensor data and the ground truth data from the virtual environment simulating real world. The performance is evaluated by comparing with the conventional frame-to-frame and frame-to-keyframe methods. Additionally, the result of loop closure detection is shown by exploiting the proposed laser scan descriptor which can be used to compute the similarity between two scan data in the virtual environment. In order to show the objectivity of the proposed odometry estimation method, the experiments were conducted with the benchmark dataset. Additionally, to validate the applicability of the proposed method, the experiments were conducted by using the real world dataset with various robot trajectories. The performance is compared with the previous frame-to-frame and frame-to-keyframe methods referenced to provided benchmark ground truth or the near-ground truth poses which are obtained through surveying. Finally, the results of the graph-based SLAM are shown with the proposed graph construction method by applying the optimization method in case of loop closure in the real world environment.