Local and Global Correspondence Establishing Techniques For Simultaneous Localization and Mapping

Abstract

Understanding 3D scene is an necessary task for human beings. Using the information from the 3D scene geometry, the human can operate various actions, such as moving without collision, navigating the 3D space, carrying objects, and physically interacting with other agents. Those are considered as very simple tasks for human beings, but it is a very difficult task for robots to acquire the capability of 3D perception. With this goal, we try to figure out what components construct the 3D perception. Generally, there are two main steps: localization and mapping. In the localization step, we can find where the robot agent is located in the 3D space. The location of the robot can be used as a clue to improve the subsequent mapping process. In the mapping step, we can construct the environment around the robot agent. The well-constructed environment map could help the robot agent to localize its position in the map. Since both the localization and mapping processes should operate at the same time, we call the algorithmic structure as the simultaneous localization and mapping (SLAM). Simply speaking, SLAM is a computational problem of reconstructing the 3D scene and tracking the robot trajectory at the same time. In this dissertation, we focus on how to improve the localization and mapping performance by correctly establishing the local and global correspondences. First, we present a new 3D scene reconstruction framework with the improved camera tracking capability to reduce the drift problem. There are two types of constraints in this work: colorimetric and geometric constraints. For the colorimetric constraint, we impose the more weights on the reliable feature correspondences obtained from color image frames. For the geometric constraint, we compute the consistent surface normal vector for the noisy point cloud data. Experimental results show that the proposed framework reduces the absolute trajectory error representing the amount of the drift and shows a more consistent trajectory in comparison to the conventional framework.

Second, we present a loop closure detection method using the bag-of-visual-word method with a local patch descriptor obtained from the learning based approach. We have trained a neural network model with a place-oriented dataset and extract the descriptors for the local patches from the trained neural network model. Besides, we have constructed the ground-truth label for the evaluation. Our experiment shows promising results, compared to the state-of-the-art loop closure detection method. The presented methods in this dissertation show sufficient and reliable results. Hence, such techniques are expected to be beneficial for various applications of simultaneous localization and mapping.