Object Segmentation Using Machine Vision in Automated Control Systems of Railway Rolling Stock

Abstract: The paper presents the results of a study into the application of YOLOv11 convolutional neural network model configurations to object segmentation tasks. It is hypothesized that YOLOv11 will be utilized in the domain of machine vision within the context of an automated railway rolling stock control system. The primary focus of this study was to assess the performance and quality of image processing implemented by various model configurations. A comparative analysis of 25 YOLOv11 configurations was conducted, varying in architecture and input image resolution within the range of 640 × 640 and 1920 × 1920 pixels. The creation of a specialized visual dataset was instrumental in facilitating the training of the models.This dataset comprised 20,000 annotated images of railway infrastructure, systematically distributed across 40 object classes. The performance and segmentation accuracy of all trained models was evaluated using the mAP metric (0.5–0.95). The results obtained are informative when choosing the configuration of theYOLOv11 model with the most suitable parameters for use in on-board machine vision systems of automated control systems, depending on the requirements for the accuracy of object segmentation and the power of available computing resources.