Currently, the improvement of the quality of transport and logistics services provided is directly related to the introduction of new and modernization of existing technologies of informatization and digitalization. One of the most urgent tasks solved by the introduction of digital technologies into existing technological processes is to improve the safety of train traffic. The analysis of domestic and foreign works devoted to the development of train safety improvement systems has shown that one of the methods of solving the task is the development and implementation of vision systems for detecting infrastructure objects and obstacles in the course of train movement. This is especially true when train speeds increase when it is difficult for the driver to correctly assess the current situation and make an operational decision. This paper describes the implementation of a vision system for unmanned trains. Within its framework, a new approach to the training of a highly specialized mask neural network was implemented. The main task of this system is to recognize obstacles and human figures against the background of the railway infrastructure determine their location relative to the tracks and assess this situation from the point of view of traffic safety. To obtain a higher-quality mask, the approach of simultaneous use of images of standard CVS cameras and cameras with the higher resolution was used. This method is able toimprove the quality of recognition, especially at large distances, when the object of interest is not noticeable in the complex environment surrounding it. The work performed has shown good results in identifying objects on railway tracks. The creation of a prototype of such a system and equipping it with traction rolling stock will allow for the timely detection of obstacles and people on the train path, which contributes to improving the level of train safety.
