Multistage Learning-Based Fusion Framework for Autonomous Vehicle Localization in GNSS-Denied Environments

Learning-based multi-sensor fusion offers a promising solution for the autonomous vehicle localization in GNSS-denied environments. However, existing methods primarily focus on the real-time pose estimation while neglecting the suppression of accumulated errors in historical poses. To address this problem, we propose a multi-stage learning-based fusion framework that integrates pose transformation estimation, local pose correction, and global pose optimization. In Stage1, we deeply fuse geometric structure data from LiDAR, color texture data from camera, and motion state data from IMU to accurately estimate the pose transformation. Continuous transformations are accumulated to generate a preliminary pose sequence. Entering Stage2, we constantly introduce absolute pose data from UWB to rapidly correct Stage1 poses in the local region. Accumulated errors in Stage1 poses are evaluated in this region by comparing poses before and after the UWB correction. Reaching Stage3, we fully leverage the spatio-temporal correlation of accumulated errors to globally optimize Stage1 poses. Measured accumulated errors in the UWB-covered region are employed as the benchmark to recursively predict accumulated errors in the non-covered region, generating a more accurate pose sequence. We conduct qualitative and quantitative experiments on KITTI dataset and our dataset to compare our fusion framework with state-of-the-art methods. Experimental results show the competitiveness of our fusion framework in improving the localization performance in GNSS-denied environments.