TY  - EJOU
AU  - Zhang, Lili 
AU  - Cheng, Jinming 
AU  - Wen, Yingyou 

TI  - Mamba-Transformer: A LiDAR Semantic Segmentation Algorithm for Adverse Weather
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Adverse weather conditions substantially degrade the performance of light detection and ranging (LiDAR) semantic segmentation because weather-induced interference often intensifies point-cloud noise, weakens the representation of long-range objects, and causes the loss of local structural details. These degradations make it difficult for segmentation models to maintain reliable perception across regions with different levels of uncertainty and complexity. To address these challenges, this paper proposes a reliability-guided Mamba–Transformer hybrid framework for adverse-weather LiDAR semantic segmentation. The proposed method first evaluates the reliability and difficulty of adverse-weather point clouds and constructs region-level priors to guide subsequent feature modeling and semantic inference. Based on these priors, the Mamba architecture is employed to capture global contextual dependencies, enabling effective modeling of long-range semantic relationships in degraded point clouds. Meanwhile, a Transformer module is introduced to perform local refinement on critical and difficult regions, thereby enhancing the discrimination of complex structures and improving segmentation robustness. Finally, an adaptive gating mechanism fuses global and local multi-source features so that complementary information from different modeling branches can be integrated for final semantic prediction. Experimental results on the SemanticSTF dataset demonstrate that the proposed method achieves strong overall performance in adverse-weather LiDAR semantic segmentation and shows advantages in overall segmentation accuracy and the recognition of complex regions.
KW  - LiDAR semantic segmentation; adverse weather; reliability guidance; Mamba; Transformer; RMT-Seg

DO  - 10.32604/cmc.2026.086285