Nonlinear Bayesian Filtering for Road Geometry Estimation Using Multi-Source Observations

Abstract

A nonlinear Bayesian filtering framework is proposed for robust road geometry estimation in Advanced Driver Assistance Systems (ADAS). The method is designed for challenging driving conditions in which conventional road feature detection is unreliable, such as snow-covered roads and degraded or occluded lane markings. The road ahead is represented as a sequence of connected clothoid segments, and recursive state estimation is performed using an Extended Kalman Filter (EKF) for prediction and a Cubature Kalman Filter (CKF) for measurement updates. The framework combines information from multiple onboard sensors and digital map data. The considered measurement sources include lane markings, road edges, barriers, map data, and surrounding vehicle trajectories. Vehicle trails are incorporated through a clothoid-fitting procedure, and statistical gating based on the Normalized Innovation Squared (NIS) is used to reject inconsistent observations. In addition, an adaptive segmentation strategy based on map data information is proposed to improve the alignment between the road representation and the underlying road geometry. The proposed framework is evaluated using recorded vehicle data from highway and snow-covered rural driving scenarios. The results show that the inclusion of map data and surrounding vehicle observations improves estimation accuracy, particularly at longer look-ahead distances. The results further show that the framework maintains a stable estimate of the road centerline even when primary road feature detections are weak or unavailable. Overall, the proposed method demonstrates robust road geometry estimation performance across varying driving environments and highlights the value of combining onboard sensing with map data information.