Road shape classification based map matching algorithm for robust localization system

Abstract

Autonomous cars require highly reliable perception that models the surrounding environment to perform level 3-5 autonomous driving. For this reason, approaches that use High Definition (HD) maps as recognition information are becoming the key research in order to overcome limitations of the recognition system through sensors. HD maps can provide point cloud data as well as feature information such as RSMs (road surface markers), traffic signs, and traffic lights. To exploit HD maps, a precise localization system is required.

The most widely used approach for localization is a fusion of the GPS and IMU-based on the Bayesian filter. However, the precision of this approach can be degraded by limitations of sensors such as drift errors from IMU, uncertainty from low-cost GPS, and multi-path errors from GPS in urban areas. To solve this problem, the measurement from matching between the HD map and the perception can be used to correct the estimated pose.

In this thesis, lane marks are used for map matching because of their high observation frequency and detectability [1]. They can be represented to polyline segments by HD maps. For matching these segments from a map and detected lanes obtained by sensors, optimization-based registration algorithms such as the Iterative Closest Point (ICP) are utilized. However, when applying them to specific shapes such as the straight line and circular arc segments in structured driving environments, unexpected map matching errors can occur due to insufficient constraints in the optimization process. These issues are known as under-constrained problems. In order to deal with them, this paper proposes the road shape classification-based map matching algorithm. The proposed algorithm classifies input lane segments from the perception system and HD map considering their geometry. According to shapes of classified lane segments, two methods are used to handle under-constrained problems. In this paper, the geometry-based map matching and covariance estimation method is utilized to align under-constrained segments. In the case of matching constrained segments, Iterative Closest Point (ICP) and its covariance estimation method is used. The result of the proposed algorithm is used to correct a pose estimation system based on the Extended Kalman Filter (EKF).

In the simulation environment, the proposed method was quantitatively evaluated in the test scenario which contains various road shapes. It was also qualitatively validated by using experimental data from an autonomous car platform. Comparing with the ICP-based localization system, the proposed algorithm improved the fault matching rate significantly, and satisfied with centimeter-level precision of the root mean square error (RMSE) for lateral distance errors.