From prototype to deployment: a systematic review of real-time YOLO-based deep learning methods for railway track defect detection

Railway infrastructure safety relies on the timely detection of track defects, the failure of which can result in severe operational disruptions and accidents. Recent advances in deep learning and computer vision, particularly YOLO-based object detection, have produced encouraging solutions for in-service railway inspection. Yet substantial inconsistencies persist in how real-time performance is defined, evaluated, and translated into deployable systems. This study presents a systematic literature review of YOLO-based railway track defect detection methods published between 2020 and 2026, conducted in accordance with the PRISMA 2020 guidelines. From 131 records retrieved from Web of Science and Scopus, 44 studies were retained after rigorous screening. The review is structured around seven research questions covering computer vision tasks, real-time detection capability, model architectures, system limitations, computational platforms, dataset and metric standardization, and deployment feasibility. Within the reviewed corpus, object detection was the most frequent task (72.7% as a standalone task), with YOLO-based architectures serving as the proposed primary method in roughly 89% of studies—a composition that, by the review’s YOLO-focused design, characterizes this subset rather than the prevalence of YOLO across railway inspection at large. Although the entire corpus is framed around real-time operation, an explicit speed-and-hardware rule classified only 63.6% of studies as offering high real-time feasibility, and reported throughput varied considerably (from below 1 to about 279 FPS) across heterogeneous hardware, reflecting the absence of standardized evaluation criteria. Although most studies motivate their designs by onboard or edge deployment, only 7 (15.9%) measure inference on edge/embedded hardware or on board a vehicle, and only 2 run their systems on track. Despite frequently reported accuracies exceeding 90% mAP, no study demonstrates fully operational, system-level deployment: coded against a TRL-based rubric, all 44 studies remain at levels L1–L3, and 84.1% do not progress beyond offline evaluation (L1–L2), revealing a marked gap between experimental performance and operational readiness. The review further identifies recurring challenges in dataset diversity, class imbalance, real-world validation, and metric reporting. Four contributions are made: (i) a PRISMA-compliant synthesis of recent YOLO-based literature, (ii) a systematic account of how the term “real-time” is operationally underspecified, (iii) an evidence-based quantification of the deployment maturity gap, and (iv) recommendations for methodological standardization. Bridging the divide between algorithmic performance and operational deployment remains the central challenge for next-generation intelligent railway inspection systems.