Manutenção ferroviária. Inspeção de wayside. Visão computacional. Detecção de objetos. Segmentação semântica. Otimização de modelos.
Nombre: AIURY SANT'ANNA JURESWSKI
Fecha de publicación: 07/04/2026
Junta de examinadores:
Nombre |
Papel |
|---|---|
| DANIEL CRUZ CAVALIERI | Examinador Externo |
| FILIPE WALL MUTZ | Examinador Interno |
| RAQUEL FRIZERA VASSALLO | Presidente |
Sumario: Rail transport plays a crucial role in the global economy, standing out for its efficiency, safety, and high capacity. In Brazil, with a vast railway network of approximately 31,000 kilometers, mining companies make extensive use of this mode, which makes railway maintenance a constant challenge. This task, which requires meticulous planning, specialized labor, and considerable resources, is further complicated by the vast stretches of railway lines, often located in remote areas and exposed to adverse weather conditions that accelerate infrastructure deterioration.
In this context, continuous monitoring of railway infrastructure is essential to ensure the system’s safety and efficiency, allowing for early detection of failures and anomalies, and enabling preventive actions. The presented work proposed an automatic method for inspecting waysides using computer vision, by implementing object detection and semantic segmentation techniques. In the first phase, the elements of interest were labeled for training, validation, and testing of the models. In the second phase, the object detection modules were individually trained and tested, with the YoloV5 Nano model standing out by achieving 0,983 precision and 0,869 mAP50-95, and the YoloV5 Extra Large model showing good results in spring measurement, with a MAE of 3,325 pixels and R2 of 0,976.
In segmentation, the combination of ResNet-50 and DeepLabV3 achieved an IoU of 0,9064, while MobileNet V2 with FPN performed best in band measurement, with a MAE of 2,480 pixels and R2 of 0,930.
Additionally, the work implemented several optimizations in the proposed pipeline, such as pruning, quantization, dataloader, and warmup. The combination of the Nano model in Torchscript with MobileNet V2 and FPN in Torch achieved an impressive 75.72 frames per second in GPU and 33.63 frames per second in CPU, both surpassing the real-time execution threshold. Meanwhile, the Extra Large model achieved 26.48 frames per second in GPU, slightly below the ideal for real-time execution.
With the proposed pipeline and the applied optimization methodologies, the work demon strated the feasibility of performing inspection and measurement of railway assets in a short time using only one camera for data capture, achieving real-time execution on both GPU and CPU. This contributes significantly to the efficient and precise maintenance of railways.
