Selection of Artificial Intelligence Algorithms in Computer Vision Tasks for Detection and Localization of Objects on Railways

this article addresses the problem of selecting and substantiating neural network architectures for computer vision systems in railway transport. In the context of digital transformation within the industry, the development of reliable intelligent decision support systems capable of real-time operation has become a critical priority. This study presents a comparative analysis of the effectiveness of modern convolutional neural networks and emerging transformer-based architectures for the semantic segmentation of key railway infrastructure objects. The research is based on a specialized dataset comprising 8,203 images captured by onboard locomotive cameras under diverse weather and lighting conditions. Five artificial intelligence models were evaluated: U-Net, U-Net++, DeepLabV3+, MAnet, and SegFormer. Performance assessment was conducted using criteria including the Mean Intersection over Union (mIoU) metric, inference speed, and analysis of Type I (false alarm) and Type II (missed detection) error probabilities. Experimental results demonstrate that the U-Net++ architecture with a DenseNet-121 encoder delivers the most balanced performance, minimizing critical missed detections while maintaining high localization accuracy. The findings and proposed evaluation criteria provide a structured framework for the substantiated selection of machine vision algorithms in the design of autonomous control systems and intelligent decision support systems for operators of track maintenance machinery.