Research On Monocular SLAM Of The Indoor Robot

With the progress of scientific technology,requirements on intelligence are much higher,and how to realize the highly intelligent of mobile robots become the hotspot of current research.Simultaneous localization and mapping(SLAM)of mobile robots are one of main research in the study.They are also extremely important for realizing the high intelligence of mobile robots.The basic problem SLAM has solved is that the mobile robot with a variety of sensors in the unknown environment can establish the environment model and estimate its own motion.Compared with other sensors,the camera is easy to install and light and can receive large amount of information,so the vision-based SLAM research is more and more attractive for SLAM researchers.In addition,monocular SLAM has attracted more and more researchers' attention due to its low cost and easy calculation.Because monocular SLAM has large amount of data need to be computed and be stored,its development has been limited by its computational and storage capabilities,which limits developments of the Monocular SLAM.Until 2007,the first pure monocular SLAM system was put forward.There are still many problems to be overcome in its theory and practice.This thesis mainly are based on the front-end part of the monocular SLAM.This thesis analyzes the front-end of visual SLAM of the classic algorithms and their disadvantages.Combined with knowledge of other computer vision,design a SLAM front-end algorithm that is more suitable for light changes and scenes with more moving objects This thesis mainly studies and improves the keypoints of detection and matching,and pose estimation of the front-end part of the feature-based SLAM algorithm,which is mainly divided into three parts:Firstly,the classic SURF detection algorithm is introduced.Because of its insensitivity to corners,the algorithm is proposed that SURF based on grids is added with the corner detection.The algorithm divides image and detect the corners in the parts of image that has less keypoints,which expand the distribution of keypoints in the scene.Secondly,various matching algorithms are analyzed and FREAK which is more stable numbers in the various scenes of lightness changing algorithm is chosen.It is proposed to the linearly interpolation of FREAK,that simplify the sampling model and linearly interpolate the descriptors so as to obtain more matching keypoints in the scenes of lightness changing and resolution changing.Thirdly,the RANSAC algorithm,the 1-point RANSAC algorithm and the 1-point RANSAC algorithm for EKF are introduced to and state their shortcomings.Then,the 1-point RANSAC for EKF algorithm with good robustness in a dynamic scene.It is improved by distinguishing the static and dynamic regions for each image,where the keypoints matched are located.With the position and pose estimation of the moving objects in the scene,it can more accurately calculate the pose of camera in scene with moving objects.The results show that this algorithm has better robustness to indoor scenes(scenes with many changes of lightness and moving objects).