Light detection and ranging (LiDAR) is extensively utilized for self-pose estimation within the domain of simultaneous localization and mapping (SLAM) for autonomous driving. Nevertheless, SLAM-based methods face the challenge of error accumulation over time in long-term and large-scale missions, rendering them unsuitable as standalone navigation approaches. Matching localization based on prior map can achieve drift-free state estimation, especially by providing global constraints in global navigation satellite system (GNSS)-denied environments. However, fluctuations and partial absence of prior map can significantly impact localization performance. Therefore, to achieve accurate and robust positioning in diverse scenarios, this article proposes an integrated approach that utilizes the voxel mapping technique while also considers adaptive map updates to fully leverage their respective strengths and achieve complementarity. First, we compute the frequency of map update based on the current level of environmental matching. Then, the point cloud of LiDAR odometry is utilized for updating the prior map, balancing both the efficiency and accuracy. Additionally, we derive uncertainty models for LiDAR odometry and map-based localization, considering the observation noise and average residual, thus enabling effective covariance determination. The fusion is achieved through iterative error state Kalman filter (IESKF) for state updating. Validation was conducted on long-term and changing public datasets as well as on a customized device. Our algorithm exhibited superior localization accuracy compared to other state-of-the-art (SOTA) methods. Corresponding ablation studies further confirmed the effectiveness of submodules.
