Abstract
Simultaneous Localization And Mapping (SLAM) technology is the key technology in the
field of autonomous robot, autonomous driving, autonomous UAV and AR. The most frequently
used SLAM sensors include lidar, monocular camera, stereo camera, RGB-D camera and so on.
It is regared as a hot and difficult issue because of the advantage of the low cost, portability, rich
information, easily integration with other sensor information and so on in the field of visual
SLAM. In addition, the RGB-D camera is more stable indoors than outdoors and can build a
dense map through the tcechnology of SLAM. Therefore, indoor SLAM is also worth
investigating. In practical applications, for example, the fly of UAV is generally localized by the
GPS, but the localization accuracy is low and it would lead to invalidation in the scene of the
indoors , tunnnel, cave, vally etc.To overcome the faults above, the monocular SLAM algorithm
has been researched in this thesis, and it was applied to the video sequences collected by the DJI
UAV which offers theoretical bases for the autonomous flight of the UAVs.
A perfect SLAM system includes front-end design, back-end optimization, map building
andloopdetection. In this thesis, basic principles of SLAM system are deeply analyzed and
mathematical principles are deduced and the program has been designed from these four aspects.
1. In the front-end design, the essential matrix E and the homography matrix H had been
calculated exploiting the epipolar geometry for the monocular SLAM, and these two matrices
had been decomposed to get the correct pose after verification, then 3D map points were solved
out by triangulation exploiting this pose, the subsequent poses and map point would be estimated
through Perspective-n-Point(PnP); While the pose and map point estimation of the RGB-D
camera front-end was directly calculated through PnP algorithm; However, in the front-end
design based on the direct method, the pose and the map point are iteratively solved by
minimizing the Photometric Error. In this thesis, feature extraction and matching, pose estimation
and triangulation and the other principles has been analyzed through the monocular UAV pictures,
then the front-end system has been constructed based on RGB-D dataset, and it works and can be
improvec and expanded in the future.
2. In the back-end optimization, Levenberg-Marquardt algorithm and other algorithms has
been used for iterative optimization by exploiting Ceres library and G2O library, so the camera
pose estimation and the map point has been optimized, we found that Ceres library worked faster
in the dataset we had choosed.
3. In the part of map building, the sparse map, dense map and occupancy grid map were
built with the ORB-SLAM2, LSD-SLAM and RGB-D dataset respectively, functions of
localization, navigation, and obstacle avoidance could be realized according to different map
types, and the occupancy grid map has the advantage of small size memory occupation.
4. In the looping detection, emerged pictures were detected by unsupervised learning
algorithm based on machine learning, and the dictionary based on ORB feature training was
constructed, then bag of words model was constructed and the picture similarity was calculated
by TF-IDF weights construction algorithm and the correct loop was detected.
Key Words：Monocular SLAM；RGB-D SLAM；Nonlinear Optimization；Mapping；
3-D Reconstruction
评论0