A structured review of railway obstacle detection as an engineering informatics problem: Sensing, representation, decision support, and traceability

Railway obstacle detection (ROD) is often presented as a perception or benchmark problem, but deployable ROD also depends on explicit representation, decision support, traceability, and information infrastructure. Unlike road-scene autonomy, railway operation involves route-constrained motion, long braking distances, clearance-envelope logic, heterogeneous sensing degradation, and a low tolerance for missed hazards. These characteristics make detector accuracy alone insufficient. This paper presents a structured review with three linked contributions. First, it reframes ROD as an engineering informatics problem. Second, it organizes the literature through a five-level evidence-maturity framework (M0–M4) that separates benchmarkable, operationally contextualized, deployable, and assurance-oriented claims. Third, it synthesizes the technical chain from sensing, datasets, representations, algorithms, and multi-modal fusion to deployment assurance and post-event traceability. The review compares nine railway datasets and resources and examines visual, 3D, fusion, tracking, uncertainty, and anomaly-detection methods. It further analyzes how explicit scene representations, runtime monitoring, degraded-mode handling, cooperative information flows, and lifecycle records support railway decision making. Across the literature, public benchmark evidence is increasing, but clearance-envelope semantics, uncertainty reporting, information infrastructure, auditable records, cross-context scalability, and generality beyond paper-specific settings remain limited. Progress in ROD should therefore be assessed by how well sensing, representations, algorithm outputs, and deployment evidence are connected to railway safety decisions and traceable engineering assurance.