A monocular visual inertial odometry based on structural features

In response to the problem of poor robustness of traditional visual odometers that only rely on point features, a monocular visual inertial odometry that utilize point features, line features and structural features are deduced. The structural feature is an orthonormal basis that satisfy the Manhattan world hypothesis. This feature contains the parallel and orthogonal information of the structural object. The introduction of structural feature provide an overall constrain for visual inertial odometry and can reduce the accumulative error of rotation effectively. Meanwhile, to ensure the real-time performance of the system after the introduction of structural features, a robust and efficient algorithm is proposed to extract structural feature, which is based on the matching relationship of line features and the direction of gravity vector. The experimental results demonstrate that, compared with the existing open source systems, the proposed system can effectively reduce the accumulated rotation error while maintaining the real-time performance.