Spatial-Temporal Modeling for Railway Point Cloud: An Efficient 3D Framework with Mamba and A New Benchmark

• 1、School of Computer Science and Technology, Beijing Jiaotong University, Beijing, 100044

Abstract：3D point cloud semantic segmentation plays a significant role in scene understanding and has been widely applied in various downstream tasks, such as autonomous driving and augmented reality. In this paper, we focus on point cloud segmentation on railway scenarios that remains underexplored in previous studies. First, we meticulously construct a railway point cloud dataset, Railway360, which is collected from a main railway line in China, covering 4.1 kilometers and approximately 800 million points divided into 10 categories. Thanks to the usage of a single-line high-frequency LiDAR and the specific vertical scanning pattern, Railway360 provides not only the position coordinates, intensities, and category labels, but also the timestamps of the point cloud collection. This allows us to obtain a continuously complete and uniformly sampled point cloud dataset, enabling more accurate temporal analysis of the scene, which can largely benefit subsequent tasks such as scene segmentation and track equipment measurement. Second, we propose a novel Mamba-based framework, named RailSeg-Mamba, for railway 3D segmentation, where spatial and temporal features are fully integrated by leveraging the significant capabilities of modeling the sequence features of the selective state space models. To our knowledge, this is the first framework that takes the temporal features of point cloud into account. We extensively evaluate the performance of the state-of-the-art methods and provide a new benchmark on our dataset. The results demonstrate that modeling point clouds as sequences provides a promising solution for railway segmentation.

Keywords： Computer Application Technology Deep Learning 3D Point Cloud