Algorithms for Recognizing Railway Infrastructure Objects in Images

Two algorithms for recognizing railway infrastructure objects in images are presented. The first algorithm is detecting potentially dangerous railway infrastructure objects and elements that determine the movement of trains. The second algorithm implies segmenting the main and auxiliary railway tracks, as well as trains in the direction of motion of the locomotive. The algorithms are implemented using YOLO and U-Net convolutional neural networks (CNN). The RZD_dataset proposed by the Research and Design Institute of Design Information, Automation, and Communication on Railway Transport was used to train and test the CNN. To increase the precision of object detection, we expanded the source image base, selected the best topology of the algorithm and augmented the illumination in the HSV space. The algorithms work in real time. The average precision for 11 classes of the detection algorithm based on the mAP metric is 63.8%. The average precision for three classes of the segmentation algorithm based on the IoU metric is 84.5%.