Abstract Manager

The use of safety performance functions (SPFs) to predict crashes is growing in Canada, with many agencies adopting analytical techniques embodied in Part C of the Highway Safety Manual. However, even though SPFs for pedestrians are theoretically possible, it has so far been rare for agencies to develop or calibrate this type of SPFs, with the result that these advanced techniques are not being applied for the most vulnerable modes.

A key reason why developing or calibrating SPFs for pedestrians has not taken off in Canada is that the low frequency of crashes at individual intersections creates a dataset where predictive modelling based on a traffic and pedestrian volume approach has been difficult.

Miovision undertook a research project with Toronto Metropolitan University where the objective was to create pedestrian SPFs that use near-miss data as inputs to predict pedestrian collisions instead of using volumes as the inputs. The near-miss data is collected from intelligent transportation systems (ITS).

Near-miss data was collected at 44 intersections across 5 Canadian cities, resulting in 1000 near-miss events, or an average of more than 23 near-misses per intersection.  This is a very rich dataset compared to traffic volume and crash data alone.

A novel machine learning approach (anomaly scoring) was used to automatically classify near-misses as severe and not severe based on their kinetic attributes. Then we attempted to make SPFs that predict the number of actual pedestrian crashes based on the near-miss counts of various severities. The end result was that the count of the most extreme conflicts was best at predicting the actual number of pedestrian crashes, achieving an r-squared value of 0.69.

This study opens the door to better use of pedestrian SPFs in Canada which leverage new data sources from ITS, and it highlights the critical importance of sorting near-miss data using a kinetic energy approach into severe events which have the power to predict crashes and non-severe events which are not relevant for predicting crashes.