Resilience in Smart Transportation: A Multi-layer Ground Segmentation for Robust Autonomous Vehicle Localization in Complex Scenes

Autonomous vehicles (AVs) are becoming an essential segment of smart transportation (ST), but their effectiveness depends on how reliably they can percept and localize themselves within complex urban environments. A key challenge they encounter remains ensuring precise localization in zones with multi-level roads, such as bridges, tunnels, and steep ramps, where traditional ground-segmentation techniques frequently fail. These failures in detecting actual ground points can interfere with path planning and compromise overall system safety. In response, we propose a novel multi-stage ground segmentation framework that significantly enhances the robustness of AV perception. This work develops Iterative Ground Plane Fitting, the first approach capable of robustly and accurately segmenting distinct ground planes across multiple layers simultaneously. This further gets complemented by a slope-aware block-wise segmentation component leveraging inertial data for the correct classification of inclined surfaces, thereby ensuring reliable performance on diverse geotechnical terrains. The framework is further stabilized with a two-stage polar partitioning strategy and a spatio-temporal consistency refinement module, which guarantee real-time performance and stability against noise and occlusions. On the SemanticKITTI benchmark, our method achieves 97.25% Precision, 98.29% Recall, and 97.76% F1-score (macro-averaged over the evaluated sequences), outperforming strong baselines including Patchwork++ (≈96.92% F1) and GndNet (≈95.95% F1). Moreover, the method runs in real time (17.13 ms/frame), while maintaining robust performance in sloped and multi-deck scenes. Given the foundational solution for reliable perception within the most challenging urban areas, this work directly contributes to the safety and efficiency of autonomous systems which support the development of resilient ST.