A new algorithm for estimating 3D structure and robot motion using visual tracking and IMU/compass

Monocular SLAM (Simultaneous Localization and Mapping) is popular in SLAM researches of the past few years. Filtering approaches and bundle adjustment style optimization are main popular strategies, with a lot of applications. This paper proposes a novel adaptive estimation based SLAM algorithm with application to a lake surface robot. Orientation and linear velocities of the robot, and accurate SURF feature tracking work as prerequisites of the algorithm. The algorithm is theoretically proved and experimentally validated in the paper. Key frames are selected in the SLAM process and stored with metric information of features points, to generate the environment map. Robot localization and sparse point based map could be estimated online at 50Hz with assistance of GPU. Moreover, dense point based map could be recovered offline for visualization. Finally, the corresponding experiments are carried out to validate performance of the new monocular visual SLAM algorithm.