Research On Android-based Monocular SLAM System

The localization of terminals means to obtain self-pose inside surroundings by information from built-in sensors. Simultaneous Localization and Mapping(SLAM) technology is an extension to build map for surroundings while localization, which can be taken as a basis for environment perception and interaction. Because of the visual sensor accessing large amount of information, wide application and many other advantages, visual SLAM has become a hot issue of researches in home and abroad.Android platform is a widely used smart operating system for smartphone. It provides a wide benefit for the HCI research on intelligent terminals. In this paper, after a detailed research for SLAM and computer-vison-related theory, we designed a visual SLAM system named MobileSLAM, which can be running on Android successfully.MobileSLAM algorithm modules of this article can be summarized as follows: feature extraction, camera pose calculation, Key-Frame selection,motion structure recovery, semi-dense map growing, loop closing and local map optimization.Firstly, we did a research on two kinds of conventional SLAM methods and compared its merits and drawbacks. Considering performance of Android, a semi-direct camera pose calculation method, along with a FAST feature detector, was adopted. This can on the one hand avoid heavy calculation of feature matching, on the other hand, it can take the advantage of the direct methods, to come out with a comparatively robust camera pose.Meanwhile, to ease the influence of inhomogeneous distribution features inside image, we designed a blocked FAST feature detect method to make feature detected better-distributed. What's more, to further correct camera pose and output a more robust result, we introduced IMU to fuse it with calculated camera pose by Extend Kalman Filter.Secondly, when in Key-Frame selection, to get a best match Key-Frame for current frame, a nearest neighbors Key-Frame pool was introduced. We utilized the brightness information and IMU data to computer the best Key-Frame, which can reduce the calculation and the redundancy of information.Then, the map got inside feature-based SLAM system is sparse which is hard to tell a real scene. To obtain a more dense map, a semi-dense map grow algorithm was came up with in this paper. We exploited existed sparse map as basis, relate current pixels with neighbored Key-Frames, to estimate and grow pixels inside gradient region. With a restrained epipolar line search and match, we get a collection of depth hypothesis, then a semi-dense map was generated by fusing all this depth hypothesis. Finally, the feasibility of the proposed monocular MobileSLAM algorithm is verified with experiment in both dataset and real scenes.