URTDepth: A Benchmark Dataset for Depth Completion in Urban Rail Transit Scenarios

Depth completion datasets are crucial for training and evaluating depth completion networks. However, long-distance datasets are scarce for urban rail transit perception tasks. Therefore, we introduce URTDepth, a novel benchmark dataset for depth completion collected in real-world urban rail transit environments. The dataset comprises 12,000 synchronized samples, each containing high-resolution RGB images, sparse LiDAR depth maps, dense ground-truth depth maps (ranging from 0 to 500 m), and precise calibration data. The data processing pipeline includes millisecond-level synchronized acquisition of point clouds and images, 3D long-range high-precision mapping and localization of railway tunnels using point cloud-based SLAM, and line-of-sight detection to filter occluded background points. We conduct comparative experiments using multiple state-of-the-art depth completion methods on this benchmark, demonstrating the dataset’s effectiveness for evaluating performance in challenging long-range, low-light urban rail transit scenarios.