Study On Real-time Single Camera Based SLAM

This thesis aims to realize the simultaneous localization and mapping of robotin3D model with the monocular camera, based on the extended Kalman filteringalgorithm. The approach can not only accurately obtain the robot pose, but also builta map of the environment concurrently. There are many existing works carried outon SLAM, e.g., full state vector filter, real-time SLAM at low frame, unified inversedepth parameter method, external physical characteristic based approach, etc. Themethods make the algorithm of EKF be widely used in SLAM; however, thecomplexity of calculation and the accuracy of the features are not yet welladdressed.The thesis has developed an SLAM approach where a single camera is the onlysensor of robot, i.e., monocular SLAM. Because the lack of depth dimension for3Dmodel, the unified inverse depth parameter method is used in the modeling of3Dsimultaneous localization and mapping. The state of the SLAM problem is describedby the pose of camera (i.e., robot), direction angle, elevation angle and unifiedinverse depth. The observation model and motion model have been carefullyanalyzed and built based on the EKF algorithm. With the built models, theinnovation between the predicted feature observation and matched actualmeasurement can make the state variance converge, and finally the SLAM isimplemented.The thesis applies several methods to address the data association problemwhich results in some serious issues in the practical applications, such as,mismatching issue, low matching accuracy and add one point repeatedly. Throughthe evaluation of the different methods in feature matching performance, thecharacteristic description and feature image block are integrated. In addition, thematched features will be classified by the algorithm of one-point RANSAC intoinliers and outliers. Finally, after measurement updating, features managementmodel will extract and delete the features with invalid information and instability,and add these instability features into blacklist to improve the efficiency of programimplementation and strengthen the robustness of the system.