| Ego-motion estimation plays an important role in a mobile robot,since it provides funda-mental information for a robot to realize tasks such as obstacle avoidance,trajectory planning,autonomous navigation.Hence,ego-motion estimation is one of the major research topics in the robotics community.This thesis is aimed at proposing two visual-based methods to estimate the ego-motion of a quadrotor without external global positioning assistance.By using the Hamilton quaternion representation,we propose an EKF-based monocular visual-inertial odometry.We use a distorted camera model in our method to better fit the real imaging model.Besides,features clustered in the texture region are redundant in constructing the geometric constraints of feature points and camera poses,so we follow the approach by[22]to select feature points that are evenly distributed in each image.In[4],the algorithm updates when features are lost or observed by all cameras in the sliding window.We use a depth filter in[23]to estimate the inverse depth of features during movements.When the inverse depth of feature points converge,the reprojection error can be established to participate in the filter updating,which reduces the algorithm error.Finally,we establish a multi-state constrained measurement equation to achieve a monocular visual-inertial odometry.Compared with the indirect method,the direct method does not need to calculate the descriptor of feature points and feature matching.On the other hand,stereo visual odometry does not require the same complicated initialization as monocular visual odometry,and can abtain the actual scale information of ego-motion estimation directly.Therefore,we design a sparse direct method stereo visual odometry in chapter 4.The inaccurate depth estimation of feature points leads to the failure of front-end tracking,so we combine the indirect method to improve this situation,making the algorithm more robust.Finally,we test the performance of our algorithm via open source datasets. |
PDF Full Text Request