The rail industry is moving towards full automation. This means that the development of a suitable environment perception systems capable of efficiently performing safety-related functions such as object detection is going to be necessary. Ensuring that these systems are robust enough to work under the different variations of the operational design domain is not an easy task, especially if you only rely on real data and their limitations when acquiring and labelling new samples.
In this video, Daniel Ochoa de Eribe, from CAF Signalling, explains how, within the boundaries of VALU3S, they managed to generate a representative dataset of synthetic images in a semi-automatic manner and evaluate their side sign and light signal detection system with it. In order to do that, they utilised a video game called Train Simulator, the labelling tool DarkLabel and two tools developed with their partner Ikerlan: DaGe4V and VaTRA.