Adaptive Sensor Fault Detection and Isolation using Unscented Kalman Filter for Vehicle Positioning

Abstract

There is an increasing demand for sub-meter vehicle localization for advanced safety and autonomous systems. Fault detection and isolation (FDI) for sensor systems, such as camera, LIDAR, GNSS, and V2X has been a challenge because their performances are significantly affected by weather, geographical changes, and even spoofing. In this paper, a sensor FDI using Student's t-distribution based adaptive unscented Kalman filter is presented. The proposed filter evaluate each sensor by Hotelling's T2test utilizing the predicted sensor output and its covariance. This method can assess the correlation between data that is generated within the same sensor, for accurate fault detection. In addition, measurement noise is adaptively updated by identifying both the covariance and the degree of freedom of the outlier robust Student's t-distribution. The robustness and accuracy of the localization and measurement noise estimation is confirmed through simulation and an experiment on a highway scenario. Furthermore, the result also shows that the precise FDI can be achieved without any prior information regarding sensor measurement noise. The proposed algorithm enhances the reliability of future position based systems such as autonomous control or V2V safety brake.