Analysis of models for image processing when implementing an object detection system on a railway track

Abstract

With the development of the railway industry, informatization of society and the automation of many technological processes, it becomes possible to create an automatic control system, diagnostics and safety of locomotive movement. One of the most important systems of this complex is the system for detecting objects on railway tracks, ruptures of the railway bed and its turns. Such a system can be developed in the form of a camera installed on a locomotive and information processing systems on board each rolling stock, or a global system for remote processing of information from several locomotives. Regardless of the implementation of the system, there is a need to create a block for detecting objects on images coming from cameras. The implementation of this block is possible using interacting full-convolutional and convolutional neural networks and training on a dataset covering various situations occurring on the railway tracks.