Hybrid-Feature-Based Visual Odometry For Robots

Accurate pose estimation is an important prerequisite for robots to realize autonomous navigation.As robotic applications become more widespread,the requirements on the accuracy and robustness of pose estimation are increasing.Aiming at the man-made environment satisfying the Manhattan world hypothesis,in this thesis,we utilize the complementarity of different visual features and the structural regularity of the scene,using points and vanishing points to realize the visual odometry(VO)based on hybrid features.Firstly,we improves the RANSAC-based vanishing point estimation algorithm by nonlinear optimization.Accurate estimation of vanishing points estimate is the key for subsequent use of this information.After analyzing the classical RANSAC-based vanishing point estimation algorithm,the consistency measurement used in the lines' classification are also used to construct the nonlinear least squares problem,and the Gauss-Newton iterative method is used for optimization.For the case of prior information constraint,the problem is redefined as the least squares problem with constraints.The Lagrangian Multiplier method is used to optimize the solution to obtain the optimal estimation results.Secondly,by introducing global constraints using the vanishing point,a monocular visual odometry aided by vanishing points is proposed.The vanishing point is used as global information to represent the transformation relationship between the world coordinate system and the camera coordinate system.Taking the monocular VO pose estimation result based on the point feature as the initial value,we transform the world vanishing point obtained during the initialization to the current camera coordinate system,use the line corresponding to the three orthogonal vanishing directions in the current image to construct the projection residual model,and furthermore construct the quaternion-based nonlinear least squares problem by the projected residuals.The nonlinear optimization method is used to optimize the attitude estimation and finally optimize the position estimation.Finally,by combining the re-projection residual of the point feature with the projection residual corresponding to the vanishing point,the hybrid residual is constructed under the maximum likelihood probability estimation model.After being transformed into a nonlinear least squares problem based on Lie algebra pose,the graph optimization method is used during the pose estimation,jointly optimized based on the hybrid features,and finally the monocular visual odometry based on the tight coupling of the hybrid features is realized.In order to verify the effectiveness of the improved vanishing point estimation algorithm,we perform experiments on public dataset and the self-designed simulation dataset,and obtain the more accurate vanishing point estimation results.To verify the effectiveness of the monocular visual odometry aided by vanishing points and based on hybrid feature tight coupling,we compare them with ORB-SLAM based on the EuRoc MAV dataset and the experience in the actual scene,and validate the accuracy and effectiveness of the proposed two VO algorithms.