SegMam: Multi-Scale Deformable Cross-Mamba Attention With Pseudo-LiDAR From Multi-Camera Depth for BEV Segmentation
IEEE Access · 2026 · 14 · 93284-93301
DOI: 10.1109/access.2026.3705050 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Junhan Kim und 4 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Basic Science Research Program, National Research Foundation of Korea, Ministry of Education, Science and Technology, Ministry of Trade, Industry and Energy (MOTIE) and the Korea Electronics Technology Institute (KETI), in 2026.
Abstract (Originalsprache)
Bird’s Eye View (BEV) semantic segmentation serves as a critical perception technology for spatially coherent understanding of complex urban environments from multi-camera inputs in autonomous driving and intelligent transportation systems. While recent Transformer-based methods have demonstrated excellent results, they face challenges in handling high-resolution BEV representations, particularly for mid- to long-range perception and robust operation under adverse weather conditions. To address these challenges, this paper proposes a novel framework utilizing multi-scale deformable Cross-Mamba attention for multi-camera BEV semantic segmentation. Our approach replaces conventional cross-attention mechanisms with selective state space models, enabling efficient spatial integration during inference while effectively preserving extended contextual dependencies, thereby supporting object and drivable area segmentation. To improve geometric distance estimation without relying on LiDAR or Radar sensors, monocular depth maps extracted from RGB imagery are converted into Pseudo-LiDAR point clouds that serve as the value stream of the Cross-Mamba module, where they are aggregated with multi-scale RGB BEV queries lifted from hierarchical visual features. This dual-stream design facilitates direct interaction between three-dimensional geometric information and multi-scale visual features, yielding consistent improvements under nighttime and rainy conditions and for mid-range vehicle perception, while remaining competitive in inference latency with state-of-the-art camera-only baselines. To foster future research, we provide the monocular depth parameter files, along with our code and trained model checkpoints at https://github.com/KimJunHan/SegMam.git
Förderung
- Basic Science Research Program, National Research Foundation of Korea
- Ministry of Education, Science and Technology
- Ministry of Trade, Industry and Energy (MOTIE) and the Korea Electronics Technology Institute (KETI), in 2026
Zitieren
APA
Kim, J., Yu, H., Ahn, K., Kim, S., & Kang, Y. (2026). SegMam: Multi-Scale Deformable Cross-Mamba Attention With Pseudo-LiDAR From Multi-Camera Depth for BEV Segmentation. IEEE Access, 14, 93284-93301. https://doi.org/10.1109/access.2026.3705050
Vancouver
Kim J, Yu H, Ahn K, Kim S, Kang Y. SegMam: Multi-Scale Deformable Cross-Mamba Attention With Pseudo-LiDAR From Multi-Camera Depth for BEV Segmentation. IEEE Access. 2026;14:93284-93301. doi:10.1109/access.2026.3705050
BibTeX
@article{Kim2026,
author = {Kim, Junhan and Yu, Hankyeol and Ahn, Kyungjae and Kim, Suyeon and Kang, Yeonsik},
title = {SegMam: Multi-Scale Deformable Cross-Mamba Attention With Pseudo-LiDAR From Multi-Camera Depth for BEV Segmentation},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {93284--93301},
doi = {10.1109/access.2026.3705050}
}
EUVIMED