Abstract
<title>Abstract</title> <p>Accurate extrinsic calibration of onboard stereo cameras is fundamental to reliable 3D perception in autonomous driving. In real-world deployment, three interrelated challenges systematically degrade calibration accuracy: resolution degradation compromises feature localization precision; depth-dependent resolution loss renders distant correspondences unreliable; and dynamic objects violate the static-scene assumption underlying essential matrix decomposition. To jointly address these challenges, we propose a Hierarchical Progressive Self-Calibration (HPSC) framework that tackles all three within a unified three-stage pipeline. Stage 1 jointly enhances the binocular image pair via the AMCASSR stereo super-resolution network, recovering fine-grained texture while preserving cross-view geometric consistency. Stage 2 applies depth-stratified correspondence pruning, discarding far-range features whose localization uncertainty exceeds a tolerable bound. Stage 3 performs hierarchical dynamic suppression through coarse semantic segmentation followed by disparity-cluster-based geometric refinement, eliminating residual dynamic points that semantic methods miss. Experiments on real-world driving sequences demonstrate that the complete HPSC pipeline substantially reduces epipolar geometric error, outperforming each of the evaluated single-stage variants and the single-image super-resolution counterpart.</p>