Validation of safety risk prediction tools and traffic models for railway level crossings

There are approximately 6000 level crossings in Britain where the trains and road users cross at the same level. In the ten-year period from 2006 to 2016, there were 86 fatalities as a result of collisions between trains and road users at level crossings. Around the world there are a number of safety risk prediction tools in use by road and railway authorities which consider physical and operational features of a level crossing as inputs and produce a prediction of the safety risk for the crossing. There is little information regarding the method of calculation used in any of the tools and no evidence can be found of validation of the results produced by the tools. There is also a large degree of variety between the features that are considered by the tools; the only commonality that can be found is that every tool uses an underlying traffic model to account for how safety risk varies as road traffic volume increases at level crossings. The most common traffic model is traffic moment – which is the product of road and rail traffic in a day – although some other models are used notably the hypothesis developed by Stott (1987) and the model developed by Peabody and Dimmick.

Until recently it has not been practical to test the degree to which any of the traffic models correlate with observed collisions due to unavailability of the data. The GB railway infrastructure manager has made information available on the numbers of road users traversing each level crossing, together with the numbers of collisions that have occurred. As such it is now possible to perform more rigorous tests of the degree to which the outputs of traffic models correspond with collisions. Furthermore, in recent years there have been advances in computer technology that have introduced new techniques to obtain information from observation data; these techniques include machine learning methods that can be used to identify trends and, in many cases, extract meaningful information, from observed data. There is no information in the available literature that shows that either these data, nor these emerging computation techniques have been applied to the study of safety risk prediction tools, which provides a clear avenue for research that is explored in this work.

This work tests:
• whether it is reasonable to expect safety risk prediction tools to be able to produce reliable estimates of risk;
• the degree to which the risk predictions from current safety risk prediction tools correlate with observed rates of collision; and
• whether it is possible to use modern data analysis methods to determine a more accurate method of risk prediction.

The outcomes of this work make a number of contributions to the prior knowledge on level crossing safety, in particular:
• Whilst safety risk prediction tools are widely used around the world, no evidence can be found of the predictive accuracy of any of the tools.
• The various tools are all based on underlying traffic models although, again, there is no evidence of the accuracy of any of the models. Newly available data make it possible to test the models for level crossings in Britain.
• When tested, it was found that the most commonly used traffic model – traffic moment – provides a good theoretical model in idealised conditions but does not appear to correlate well with real-world conditions.
• In fact none of the traffic models that can be tested were found to correlate well with observed collisions. Remarkably a model based on observation of collisions in the 1930s is better at describing collision rates than a model specifically created in the 1980s to describe British level crossings.
• It was found that, whilst none of the traffic models correlates well with observed collisions, there does appear to be a power-law that describes collision rates. Importantly it appears that the rate of collisions per road user decreases as the number of road users increases at a level crossing. This finding is especially significant as it provides the first evidence to support the practice of level crossing closure as a means of improving safety.
• A study was undertaken using machine learning techniques to determine whether it was possible to correlate data on physical and operational features of level crossings with rates of collisions. It was found that, as with the previous studies, traffic volumes do correlate to a small degree, however no other correlation can be found in the data.

Whilst undertaking this work, additional contributions were made, specifically:
• a meaningful unit of level crossing safety was established, and
• a method for comparing observed collision rates against theoretical models that can be used for overdispersed data was identified.

As well as advancing the theoretical knowledge on level crossing safety, this work provides meaningful results that are useful to the day-to-day management of the railway.