Analysis of Fisheye Stereo Camera based Visual SLAM for indoor environment
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2022-05-16
Department
Major/Subject
Autonomous Systems
Mcode
ELEC3055
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
67+21
Series
Abstract
Robots operating in indoor and outdoor environments need accurate positioning on the map. SLAM (Simultaneous Localization And Mapping) is an essential technique to estimate the position and reconstruct the structure of the environment. Path planning, motion control, obstacle avoidance and other major modules depend heavily on the pose estimation of the robot. Sensors like GPS, LIDAR, camera, IMU are used with sensor fusion algorithms to estimate the robot pose. However, in indoor environments, GPS signals are blocked or reflected by walls, making them inefficient to use. LIDAR gives more accurate data in indoor environments, but they are more expensive than a camera in terms of price, power and computational capability. Vision sensors used in SLAM is a vital technique in computer vision that utilizes the camera characteristics and pixel information to obtain the position and orientation of the camera. In this thesis, a fisheye stereo vision camera is used to evaluate the performance of three novel visual SLAM algorithms to two different trajectories enforcing offline execution. To benchmark these algorithms, LIDAR-based NDT-SLAM data is collected along with the raw frame data captured from cameras. The datasets consist of fisheye stereo images with a resolution of 848x800 captured at 30fps. This thesis is focused on evaluating three indirect vSLAM methods, RTABMAP, OpenVSLAM and ORBSLAM3, which extracts the features from the image and computes the estimates based on the feature tracking. It is observed that OpenVSLAM-ORBSLAM performed better with APE statistics than the other when compared with the reference trajectory. A maneuver-wise analysis concluded that OpenVSLAM is best operable in straight and left maneuvers while ORBSLAM3 has a slight edge over the other two in the right maneuver. ORB feature extraction methods prove computationally faster and more efficient in map storage as ORBSLAM algorithms generate sparse maps compared with the RTABMAP's dense map.Description
Supervisor
Fontanelli, DanieleThesis advisor
Zhou, QuanKeywords
visual SLAM, ORBSLAM, RTABMAP, feature extraction