Obstacle Detection at Level Crossings under Adverse Weather Conditions -- A Survey

TL;DR

This survey addresses obstacle detection at railway level crossings under adverse weather by comparing inductive loops, cameras, radar, and LiDAR, and by evaluating how weather degrades each modality. It analyzes per-sensor mitigation strategies (e.g., dehazing, deraining, rain models, and full-waveform LiDAR) and systematically reviews multi-sensor fusion paradigms—data-level, feature-level, and decision-level—to enhance reliability and fault tolerance. The paper highlights novel model-based and learning-based methods, datasets like RADIATE, and the promise of adaptive fusion and weather-resilient data pipelines for real-time deployment. The findings underscore the importance of robust, fail-safe fusion architectures and comprehensive datasets to support safe railway operations in a wide range of environmental conditions.

Abstract

Level crossing accidents remain a significant safety concern in modern railway systems, particularly under adverse weather conditions that degrade sensor performance. This review surveys state-of-the-art sensor technologies and fusion strategies for obstacle detection at railway level crossings, with a focus on robustness, detection accuracy, and environmental resilience. Individual sensors such as inductive loops, cameras, radar, and LiDAR offer complementary strengths but involve trade-offs, including material dependence, reduced visibility, and limited resolution in harsh environments. We analyze each modality's working principles, weather-induced vulnerabilities, and mitigation strategies, including signal enhancement and machine-learning-based denoising. We further review multi-sensor fusion approaches, categorized as data-level, feature-level, and decision-level architectures, that integrate complementary information to improve reliability and fault tolerance. The survey concludes with future research directions, including adaptive fusion algorithms, real-time processing pipelines, and weather-resilient datasets to support the deployment of intelligent, fail-safe detection systems for railway safety.

Figures (6)

• Figure 1: Examples of camera deployment in railway systems: (a) fixed surveillance camera at a level crossing, (b) on-board sensors mounted on a train denzler2022multi.
• Figure 2: Example LiDAR point clouds collected at a level crossing using a Velodyne VLP-16 sensor. Three types of obstacles are shown: (a) a pedestrian, (b) a cyclist, and (c) a car crossing the tracks.
• Figure 3: Phase difference pdf under different propagation scenarios.
• Figure 4: Illustration of the array measurement model with the rain distortion.
• Figure 5: RMSE comparison between calibrated root-MUSIC-based DoA estimation and conventional root-MUSIC without calibration.