Research of neural network algorithms for recognizing railway infrastructure objects in video images

The article describes the development of two neural network algorithms for recognizing objects of the railway infrastructure in video images. Both algorithms are aimed at improving railway traffic safety. One algorithm detects foreign objects on railway tracks and objects relating to the railway infrastructure. The other algorithm implements the semantic segmentation of main and auxiliary railway tracks, as well as trains within the visible range of the locomotive. The algorithms are implemented based on convolutional neural networks (CNN) YOLO and U-Net. The CNN is trained and tested using the image database of the Research Institute of Information, Automation and Communications in Railway Transport. The experimental studies conducted are aimed at increasing the efficiency of algorithms for object detection and segmentation through the use of data augmentation methods and additional preprocessing, as well as selecting an architecture and optimal network hyperparameters. The detection algorithm works in real time, achieving an average accuracy of 64% for 11 object classes according to the mAP metric. The operating speed of the semantic segmentation algorithm is 5 frames/s, the average accuracy for three classes of objects according to the IoU metric is 92%.