Adaptive EKF SLAM with Semantic Landmark Typing for Robust Navigation of Highly Automated Vehicles

Classic implementations of simultaneous localization and mapping algorithms based on the extended Kalman filter (EKF SLAM) demonstrate limited effectiveness in weakly structured environments, which are often encountered by highly automated vehicles (HAVs). The lack of detection of additional features and the processing of all landmarks according to a single template, without considering their nature and reliability, leads to the accumulation of errors and a loss of robustness. This paper proposes an approach that aims to solve this problem by deeply integrating semantic information into the filter’s probabilistic model. The method is based on classifying observed landmarks into types (“discrete obstacle,” “extended road edge”) and then applying mathematical models that take into account the individual statistical properties of each type of landmark (e.g., different detection accuracies). In particular, individual measurement noise covariance matrices and other thresholds for data association are used. The effectiveness of the proposed approach is demonstrated in the Webots simulation environment using the ROS 2 software platform. The results show that in poorly structured environments without discrete obstacles, the proposed method maintains high localization accuracy. This is achieved through the reliable use of extended structural elements, such as road edges, as landmarks, which are ignored or processed inconsistently by classical approaches.