Investigation of a surrogate measure-based safety index for predicting injury crashes at signalized intersections

The paper develops a machine learning-based safety index for classifying traffic conflicts that can be used to estimate the frequency of signalized intersection crashes, with a focus on the more severe ones that result in fatal and severe injury. The number of conflicts in different severity levels categorized by the safety index is used as an explanatory variable for developing statistical models for pro-actively estimating crashes.

Video-derived conflicts in different severity levels between left-turning vehicles and opposing through vehicles, a well-recognized severe injury crash typology at signalized intersections, were identified by jointly integrating the indicators of frequency and severity, using an autoencoder neural network integration method to develop anomaly scores. Regression models were then developed to relate crashes at the same intersections to the classified conflicts based on the value of their safety indexes. Cumulative Residual plots were investigated. Finally, equations defining the boundary between consecutive anomaly score levels were developed to facilitate application in practice.

Regression models for total and fatal plus severe (FSI) crashes utilizing classified extreme conflicts based on anomaly scores were found to outperform the models using total conflicts. The improvement is more pronounced for FSI crashes. The results also suggest that the machine learning integration method can efficiently classify conflicts accurately according to crash severity levels since the higher anomaly score is associated with a higher crash severity level (i.e., FSI).

The proposed framework represents a methodological advancement in traffic conflict-based estimation of crashes using a machine learning model to classify conflicts by their anomaly scores. For jurisdictions without the resources to develop such a model to classify conflicts for their own datasets, the simple equations defining the boundary between consecutive anomaly score levels could be used as an approximation.