A Comparative Study of the Multi-state Constraint and the Multi-view Geometry Constraint Kalman Filter for Robot Ego-Motion Estimation

Accurate ego-motion estimation is critical for an autonomous robot. For this task, image-aided inertial navigation is a good solution. In this paper, the applicability and effectiveness of two algorithms for camera/IMU integration are compared under different ego-motion estimation context: the Multi-State Constraint Kalman Filter (MSCKF) and the Multi-View Geometry Constraint Kalman Filter (MVGCKF) based on Tri-focal Tensor Geometry. Both algorithms estimate the motion of a robot with an IMU (Inertial Measurement Unit), while observations of salient features from a monocular camera are used to correct the inertial navigation error in a Kalman filter. In contrast to traditional vision-aided inertial navigation algorithms, such as EKF-SLAM or the sliding window filter, the measurement models of the both two approaches do not include 3D feature information in filter states, in this way the computing complexity are reduced. We present experimental characterizations of the MSCKF and MVGCKF on KITTI dataset. Theoretical analysis and experimental results show that MSCKF is better for the low speed scenarios such as robot and pedestrian navigation, while MVGCKF is a better choice for the high dynamic and high speed environment such as unmanned land vehicles which need high update rate for navigational parameters. MVGCKF diverges rapidly at stationary and low speed state although its computation burden is less than that of MSCKF.