A Data-Centric Approach to 3D Semantic Segmentation of Railway Scenes

TL;DR

The paper addresses robust long-range LiDAR semantic segmentation for autonomous trains in open railway environments, where data scarcity at greater distances hampers safety-critical predictions. It introduces two data-centric augmentations—Track sparsification and Pedestrian Instance Pasting—designed to balance point density and diversify distant pedestrians, and evaluates them on the railway-specific OSDaR23 dataset using a state-of-the-art 3D segmentation network. It also reports the first 3D segmentation benchmark on OSDaR23 and proposes mean range IoU as a metric to assess distant-range performance. The results show meaningful improvements in far-range IoU without sacrificing near-range accuracy, demonstrating the value of data-centric approaches in railway perception and suggesting extensions to temporal data and multi-sensor fusion for further gains.

Abstract

LiDAR-based semantic segmentation is critical for autonomous trains, requiring accurate predictions across varying distances. This paper introduces two targeted data augmentation methods designed to improve segmentation performance on the railway-specific OSDaR23 dataset. The person instance pasting method enhances segmentation of pedestrians at distant ranges by injecting realistic variations into the dataset. The track sparsification method redistributes point density in LiDAR scans, improving track segmentation at far distances with minimal impact on close-range accuracy. Both methods are evaluated using a state-of-the-art 3D semantic segmentation network, demonstrating significant improvements in distant-range performance while maintaining robustness in close-range predictions. We establish the first 3D semantic segmentation benchmark for OSDaR23, demonstrating the potential of data-centric approaches to address railway-specific challenges in autonomous train perception.

Figures (8)

• Figure 1: Example of a segmented pointcloud from the OSDaR23 dataset tagiewOSDaR23OpenSensor2023
• Figure 2: Schematic representation of three main deep learning-based methods for semantic segmentation of point cloud data. Adapted from xuRPVNetDeepEfficient2021.
• Figure 3: Recall for the class track across the validation set. High recall is observed close to the sensor, with performance decreasing beyond 60 m.
• Figure 4: Effect of the tracks sparsification transformation on scene 3_ fire_ site_ 3.1, frame 58 from the OSDaR23 dataset.
• Figure 5: Visualisation of the person instances pasting transformation. Best viewed zoomed in.