Robot Localization In Fusion With Vision And Inertial Navigation

Automatic localization, which is base of robot navigation, is the core issue of robot research. Global localization methods like Global Position System have high accuracy but limited application. And visual system has advantages in accurate self-localization, widely application. However, accumulated error would be its problem. Therefore, we promote visual method by fusion with inertial navigation system.In this paper, we research localization algorithm by fusion with visual and IMU. Due to error from attitude estimation by IMU is accumulated in the dead reckoning, inertial localization comes out drift. We present inertial localization algorithm by fusion with wheel odometer. It avoids linear acceleration integration and projects output of wheel odometer into navigation coordinate by attitude estimated by IMU. Therefore, accuracy of inertial localization raised.Due to attitude (yaw, pitch, roll) estimation coupled in present visual odometry(VO), rotation error in one direction is projected onto the other two directions, and will be accumulated during large scale estimation, making direction of estimation diverge. A real-time adaptive EKF model is presented, with decoupled attitude estimation in VO based on IMU using gravity as vertical reference to correct accumulated error. According to state of motion, a fuzzy logic is used to adjust filter parameters, so that adaptive filter estimation was realized. This algorithm effectively makes VO more accurate and robust.Finally, either VO or IMU has limitation in global information, so that they can easily affect by accumulated error. However, we present simultaneous localization and mapping(SLAM) based on extend kalman filter. With IMU data for motion prediction, we prepare matched SIFT feature for data association to correct the prediction and simultaneously map with landmarks. The landmarks are used to decrease localization uncertainty in the filter.Meanwhile, during data analysis, we compared our localization results with the ground truth measured by total station, and our results showed the good efficiency and stability. It demonstrates the algorithm the high accuracy and practicality.