In order to solve the problem that intelligent train visual positioning system based on deep learning is difficult to test, this paper proposes a safety test method for intelligent train visual positioning. Firstly, based on the idea of Image-to-Image translation, we construct a generative adversarial network (GAN) to generate test cases. Then we implement the quantitative evaluation of the error detection ability of test cases based on deep mutation testing. Finally, according to the characteristics of urban rail operation organization, we propose a parallel test platform architecture of "virtual-reality, semi-reality, reality" to support the construction of the test case generation model and test execution. The method proposed in this paper provides a basis for ensuring the safety of intelligent visual train positioning, provides a new research idea for the safety application of intelligent visual perception technology in the autonomous running of trains, and plays an essential role in ensuring the safety of trains.

urban rail transit  /  intelligent visual train positioning  /  machine learning testing  /  test case generation  /  GAN  /  mutation testi