This paper focuses on the attitude estimation problem of redundant inertial measurement unit (IMU). The dynamic characteristics of the estimated system are modeled as a hybrid model of matrix Lie group and bias vector by using invariant extended Kalman filter structure, and a local IMU attitude estimator based on online covariance correction is designed. On this basis, the Lie group uncertainty is mapped to the vector space using logarithmic mapping, and then the distributed fusion problem depending on local attitude estimator parameters is constructed, thus designing a robust covariance intersection fusion criterion based on the Lie group. In particular, in order to solve the problem of unknown cross-covariance between local estimators, a prior tolerance is introduced to constrain the unknown cross-covariance, thereby providing a low-conservative covariance upper bound to improve fusion estimation performance, forming a two-layer structure method for redundant IMU attitude estimation. Finally, the effectiveness of the proposed algorithm is verified on the lower limb exoskeleton rehabilitation robot platform.