Research On Graph-based Monocular Vision SLAM Problem

Simultaneous localization and mapping(SLAM), which is defined as the problem of building a map while at the same time localizing the robot within that map, has been intensively studied in the robotics community in the past. With the development of sensors and computer vision research community, cameras, which are accurate, compact, well-understood and most importantly cheap and ubiquitous today, have gradually been at the center of robot SLAM comparing to other sensors such as laser range-finders and sonar. In this study we focus on visual SLAM, especially the monocular visual SLAM, which got lower cost but more complicated solution for the lack of absolute depth information in contrast to binocular visual SLAM and RGB-D SLAM. Monocular SLAM approaches could be divided into two categories, filter-based SLAM and graph-based SLAM. The filtered-based SLAM such as EKF SLAM has a higher computation cost and is not suitable for real-time application, so this study concentrates mainly on research of monocular visual SLAM based on graph optimization.This thesis studies on the ORB-SLAM method which parallelizes the SLAM tasks of tracking, local mapping and loop closing and has achieved accurate results in real-time. Though of the attractive idea that ORB-SLAM tries to make use of the same ORB feature for all SLAM tasks, some flaws do exist. This thesis tries to find some solutions to fix the flaws. Considering ORB is invariant to rotation and scale only in a certain range, a large number of experiments on performance of different kinds of feature detection algorithms for the monocular SLAM tasks in this thesis have been made to find an optimized feature detection algorithm respectively. Given that the quantity and distribution of features have an effect on tracking results and pose estimation of the camera, a feature extraction algorithm is proposed to control the quantities and distribution of features. ORB-SLAM triangulates the map points using the Linear-LS algorithm, which has no geometric meaning and may cause inaccuracy, thus we improve the triangulation algorithm to Iterative–LS, which is very simple to program and requires only a trivial adaptation to the linear methods. Because the feature descriptor this thesis uses is FREAK, so the improved SLAM algorithm is called FREAK-SLAM.An extensive experimental validation of FREAK-SLAM has been performed in indoor and outdoor environment. The result is FREAK-SLAM tracks more stably with designed feature numbers and feature distribution. More importantly, the localization accuracy has been improved in contrast to ORB-SLAM.