RE 2 Mamba: a Mamba-based multimodal framework for road elevation estimation

Abstract High-precision estimation of road elevation is essential for ensuring ride comfort and driving safety. However, the complexity of real-world road scenes and the influence of external factors on sensor measurements hinder existing methods from achieving both high accuracy and computational efficiency. To tackle this challenge, this paper presents an efficient multimodal Bird’s-eye view (BEV) framework for road elevation estimation, termed RE 2 Mamba, which integrates LiDAR geometric data with RGB appearance information. Rather than engaging in costly global reasoning on dense BEV grids, the proposed method performs global multimodal context modelling within a compact representation using a Mamba state-space model, generating high-resolution BEV outputs through geometrically consistent alignment and BEV backfilling. To further improve cross-modal complementarity and long-range consistency, this paper introduce dual-path feature propagation for LiDAR and image tokens, while broadcasting global context to each BEV cell. The network predicts a discretised elevation distribution and decodes continuous elevations by taking expectations over bin centres. On the RSRD dataset, with a BEV resolution of 0.03 m, RE 2 Mamba achieves a mean absolute error of 0.57 cm and a root mean square error of 0.73 cm, with an end-to-end latency of 10 ms per frame. Ablation studies further substantiate the effectiveness of token-domain global modelling, as well as BEV backfilling and broadcasting. These results illustrate a favourable accuracy-latency trade-off and highlight the potential of the proposed framework for resource-constrained real-time applications.