Robust visual inertial monocular using nonlinear optimization

Recently, visual inertial has became popular due to its excellent result. However, the excellent result severely depends on the accuracy of estimation of initial parameters. The existing method is not effective on estimating the initial parameters and lacks the function to perform the closed loop detection, which will cause the error accumulation and low accurate estimation to system's state. In the paper, to estimate high accurate initial parameters (scale, IMU biases, gravity direction and velocity), we propose an effective method to estimate these parameters by using nonlinear optimization method. Besides, we also address the issue of state accumulation drift with preintegral theory among selected keyframes and perform an closed loop detection. Experiments on EuRoc datasets show that our method helps get good initial result with scale factor error less than 0.01, the gravity magnitude converging to 9.8, accelerometer biases converging to 0 and gyroscope biases converging to the level of 10E-3. We also get results with error less than 1 degree in rotation and 0.08m in translation. Our method performs better effect comparing with other state-of-the-art visual inertial monocular SLAM methods.