Abstract Manager

Safety evaluation, especially for innovative treatments and those targeted at certain crash types, is challenging to accomplish with conventional crash-based analyses. Despite the growing interest in using traffic conflicts as a surrogate measure of safety, validation studies relating conflicts measures to crash data are few. This research work aims to provide a methodology that uses surrogate measures of safety obtained from video analytics to evaluate safety at signalized intersections and validate the results. The data for this study were generated by Transoft video analytics software, for the sites in the region of Durham, Ontario, along with a 5-year period crash data for the same sites obtained from the Regional Municipality of Durham. In the first phase, this study explores how traffic conflicts and crashes are compared in ranking the sites for the purpose of network screening, as the first stage of road safety management process. In doing so, two approaches of using a conflict indicator threshold and defining a conflict risk score, which considers the conflicting speed, are examined. In the next phase, using negative binomial approach, crash prediction models are developed for 5 different collision types at intersections; rear-end (including sideswipe), merging, left-turn opposed, right-angle, and pedestrian-related collisions. A cross-validation approach is then accomplished to examine the validity of the developed models on test data samples. When comparing the conflict-based and crash-based site rankings, the results indicate that the two conflict-based approaches (i.e., using threshold and risk score) are introducing the sites with higher risks of collisions very closely to the sites crash history. Specifically, the average ranking precision for the conflict-based approaches is 92%. In terms of crash prediction models, the goodness of fit measures show strong crash-conflict correlation.  It was also found considering different scenarios of traffic conflicts play an important role in improving the accuracy of predictions. The cross validation further confirms the results. The predicted values for unseen data samples (validation datasets) very well follow the observed crash counts. The methodological approach introduced here is viable for quickly conducting network screening, in particular, where there is an implemented treatment which knowledge on safety effects is sparse or non-existing. On the matter of crash prediction, although the findings are very promising, further validation studies are needed to generalize the models to other regions and different types of sites.