Multi-sensor Fusion Using Fuzzy Inference System for a Visual-IMU-Wheel Odometry

This article presents an adaptive fusion method for a multisensor odometry using a fuzzy inference system and applies it to a visual-inertial measurement unit (IMU)-wheel odometry. This article first presents the mechanism of the multisensor odometry with an error state Kalman filter and points out that the assumption of the constant covariance in the prediction and measurement models are invalid due to dynamical interior and exterior abnormities. The previous work of fuzzy logic Kalman filter can only deal with uncertain uncertainty in the measurement model and thus not suitable for this double stage variant uncertainty problem. This article provides a solution using adaptive network fuzzy inference systems. The abnormal working conditions of the two models can be detected online separately, and the fusion gains are tuned accordingly. The training method of the networks is also provided to further improve the fuzzy inference system when the ground truth of the trajectories is available. Experimental results show that the proposed adaptive fusion method can significantly improve the robustness and accuracy of the fused trajectory estimation, even under severe working conditions.