Road Shape Classification-Based Matching Between Lane Detection and HD Map for Robust Localization of Autonomous Cars

Many map matching-based localization algorithms estimate the autonomous car's pose using a registration between lanes measured by a camera and the High-Definition (HD) map. However, the registration methods based on numerical optimization, such as the Iterative Closest Point (ICP) and Normal Distributions Transform (NDT), can cause underdetermined problems due to no unique solution when matching under-constrained lane shapes such as straight lines and circular arcs. This paper proposes a robust localization algorithm with centimeter-level accuracy by applying different matching techniques depending on the road shape. We proposed a road shape classification-based map matching algorithm to overcome the under-constrained problems, which have no unique solution due to insufficient constraints. The proposed algorithm classifies lane segments into line, arc, and clothoid curves considering their curvature characteristics. After that, we find correction information through a map matching and covariance estimation method using lane pairs with the same type of their shape. For the under-constrained shapes, the geometry-based map-matching algorithm and covariance estimation method are applied to avoid underdetermined results. Finally, the measurement calculated from the correction information and predicted pose of the ego-vehicle is exploited for the measurement update of the Extended Kalman Filter (EKF). The proposed method was quantitatively evaluated in the simulation environment, which contains various road shapes, and qualitatively validated for experimental data from an autonomous driving platform. The proposed algorithm shows more robust matching capability, efficient computation time, and higher accuracy than the localization system based on ICP-based lane matching.