Conflict-Based Real-Time Road Safety Analysis: Sensitivity to Data Collection Duration and its Implications for Model Resilience

Abstract

Conflict-based approaches to real-time road safety analysis can provide several benefits over traditional crash-based models. In particular, as traffic conflicts are much more frequent than crashes, models can be trained with significantly shorter collection periods. Since existing literature has not investigated the sensitivity of real-time conflict prediction models (RTConfPM) to data collection duration, here we aim to fill this gap and discuss the implications for model resilience. A real-world highway case study was analyzed. Methodologically, various traffic variables aggregated into 5 min intervals were selected as predictors, synthetic minority oversampling technique (SMOTE) was applied to deal with unbalanced classification issues, and support vector machine (SVM) was chosen as classifier. The dichotomous response variable separated safe and unsafe intervals into two classes; the latter were defined considering a minimum number of rear-end conflicts within the interval, which were identified using a surrogate measure of safety (SMS), that is, time-to-collision. Several RTConfPMs were trained and tested, considering different data collection durations and different criteria to define the unsafe situation class. The results, which were shown to be robust with respect to the machine learning classifier used, indicate that the models were able to provide reliable predictions with just three to five days of data, and that the increase in performance with collection periods longer than 10 to 15 days was negligible. These findings can be generalized by considering the number of unsafe situations corresponding to the data collection period of each tested model; they highlight the relevance of RTConfPM as a more flexible and resilient alternative to the crash-based approach.