Authors: Gorade Pranita Bhimaji, Bangar Abhishek Popat, Bhosale Sachin Bajirao
Abstract: Railway level crossings remain among the highest-risk elements of the railway system because their safety depends on the behaviour of drivers, pedestrians and cyclists, while conventional protection based on train detection, barriers, warning lights and inductive loops cannot confirm that the crossing area is clear before a train arrives. This paper presents a structured narrative review of 22 verified sources published between 2015 and 2026, including two labelled preprints, organised into seven themes: crossing risk and road-user behaviour; camera based deep learning; trespass and near-miss video analytics; LiDAR, laser and radar sensing; sensor fusion, adverse weather and datasets; machine learning (ML) for crash-risk and injury severity prediction; and warning and edge deployment. Convolutional neural networks and YOLO-type detectors achieve high reported precision for occupancy and road-user detection, LiDAR and laser systems detect small obstacles on low-power edge hardware, camera–radar fusion with redundant voting is the most frequently recommended architecture, and ML models outperform count regression for crash prediction. However, evidence is dominated by singlesite and fair-weather studies, public multi-sensor crossing datasets are scarce, and missed-detection rates, latency and safety integrity are rarely reported. The paper contributes a thematic synthesis, a comparison of fourteen representative studies, a proposed conceptual seven-layer architecture and a prioritised research agenda.
