Abstract
Camera and IMU are two complementary sensors for motion estimation in robotics perception. By combining the IMU information, visual-inertial odometry (VIO) can achieve more accurate and robust pose estimation than visual-only odometry. In most visual-inertial optimization frameworks, to reduce the increased computational complexity caused by relinearizing high-frequency IMU dynamics, uncertainty-weighted preintegrated IMU residuals are involved as pseudo-measurements between selected keyframes. During IMU preintegration, the orientation residual is constructed by the integration of corrected angular velocity. However, in classical IMU complementary orientation filtering, both the gyroscope and the accelerometer information are used for orientation estimation. In this work, we propose a loosely coupled roll and pitch angle estimation framework that combines tightly coupled fixed-lag visual-inertial odometry and IMU complementary filter. A more robust VIO rotation estimate is achieved by exploring the gravity direction in accelerometer measurements. The experimental result shows that the proposed method achieves competitive performance in academic benchmarks.