Multisensor information fusion has been extensively used in the fields of navigation and localization. Inertial and visual sensors are combined for vehicle navigation in unfamiliar environments, leveraging their complementary strengths. However, during data collection, the complex and dynamic motion environment introduces measurement noise, which reduces positioning accuracy. Traditionally, measurement noise is assumed to be uniformly distributed white Gaussian noise with constant mean and covariance. In practice, however, the measurement noise varies considerably, severely impacting positioning accuracy. To address this issue and enhance the positioning accuracy of visual-inertial navigation systems, this study proposes a monocular visual-inertial odometer based on the adaptive variational Bayes algorithm. This approach accounts for unknown measurement noise by modeling the probability density function (pdf) of the system's measurement noise matrix using the inverse Wishart distribution with a summed mean. This method modifies the Gaussian characteristics of the measurement noise to more accurately represent the real noise. To further enhance system accuracy and avoid measurement bias associated with using only a monocular camera, the high error measurements are mitigated using an innovation chi-square test. This approach reduces the number of iterative approximations while improving accuracy. The proposed algorithm was validated using data from public datasets in various environments. The results demonstrated that the proposed algorithm achieves higher accuracy and better positioning performance compared to the visual-inertial multistate-constrained Kalman filter (MSCKF) fusion algorithm.
原文链接
分享
引用
