With computer vision we’re able to localize assets and determine their condition. In this talk we will dive into the details of convolutional neural networks and we’ll show how they help ProRail maintain 7000 km of railway track.
ProRail manages and maintains over 7000 km of railway track. In order to prevent and solve delays, accurate knowledge about the location and state of the various assets is required. Special camera-equipped trains cover the entire railway track twice a year, providing over 50 million images of the railway. Using computer vision, these images can be used to both localize assets and determine their condition. This technology is currently being put into production at ProRail, where the focus is first on the localization of important assets and next on quantifying their state.
In this talk, we will dive into how computer vision is applied at ProRail. First, we will discuss the convolutional neural network that we use in order to detect our main asset of interest: the insulation joint. We will elaborate on the specific choices we made and the tools we used during the implementation. However, a simple cnn might not cut it for us when trying to distinguish between 28 slightly different assets. In the second part of our talk we’ll touch on more advanced techniques such as transfer learning, and compare them on the use case.
No prior knowledge is required for this talk.