A Monocular Visual-Inertial Odometry Based on Hybrid Residuals

In recent years, visual-inertial odometry for state estimation has achieved a huge success in terms of robustness and accuracy. However, current visual-inertial odometry algorithms still struggle with high computational cost, which severely hinders its application on computational resource limited platforms. In this paper, a novel tightly coupled visual-inertial odometry algorithm that can maximize utilize the information provided by 3D map points and IMU pre-integration to reduce the algorithm's computational cost is proposed. In the front-end, an accelerated optical flow tracking algorithm using the depth known 3D map points and IMU pre-integration between two consecutive frames is proposed to reduce the tracking time. In the back-end, a joint optimization model based on hybrid residuals is proposed, which consists the geometric, the photometric and the IMU residuals. It's worth noting that partial feature points with good tracking quality from all the feature points that the algorithm is tracking on are selected to construct photometric residuals. Therefore, those selected feature points can provide double constraints for the motion. The one feature point provide double motion constraints strategy allows our odometry algorithm use fewer feature points under the condition of same number of motion constraints which really reduced the amount of feature points also the computational resource needed by our algorithm to achieve a comparable pose estimation accuracy. Experiment results on public available dataset demonstrate the effectiveness of our proposed approach.