Gate violation prediction at highway-rail grade crossings using tree-based ensemble techniques

Abstract

Highway-rail grade crossing (HRGC) safety is one of the priority areas in the United States transportation system that requires for greater research efforts not just limited to crash analysis, but also to gain a deeper understanding of surrogate safety measures such as driver behavior-based traffic violations at HRGCs. This paper uses vehicle profile data to identify the key variables and develop prediction models for gate violations and examine the relationship between model accuracy and the key input variables. A data set of 256 vehicle-train events was collected at two HRGC testbeds in Lincoln, Nebraska. Among them, 76 events are gate violations, and 180 events are non-violations. Two tree-based ensemble techniques, the bootstrap forest and the boosted tree, were applied to the data set. It was found that once a vehicle is within 190 feet from the HRGC stop line, the model was approximately 80 percent accurate in predicting a gate violation. It was also found that as vehicles came closer to the HRGC, the prediction error decreased. With the advent of vehicle profile data collection, tree-based ensemble techniques are ideal for safety studies as they can utilize the highly non-linear vehicle profiles and relate these to safety surrogate metrics.