Support Vector Machines Approach for Predicting Realtime Rearend Crash Risk on Freeways

Abstract:

The paper aims to study the relationship between rearend crash potential and lanelevel traffic data collected by single pair of loop detectors located on the G60 Freeway in Shanghai, China. The matched casecontrol method with support vector machines was applied to modelling the traffic data of different time slices before the crashes respectively, 5~10minutes, 10~15 minutes and 15~20 minutes before the crash. Results indicate that support vector machines classifiers based on the traffic data of 5~10 minutes before the crashes have the highest crash prediction accuracy of 84.85% and a false alarming rate 0.33%. The model proves to be valid to predict the realtime crash risk, which is helpful in freeway traffic management.

Key words:freeway; rearend crash; realtime prediction; crash risk; support vector machines