A Computing Approach Using Probabilistic Neural Networks for Instantaneous Appraisal of Rear-End Crash Risk

Computing and information technology has significantly increased the capabilities to collect, store, and analyze freeway traffic surveillance data. The most common forms of such data are collected using the underground loop detectors. In the recent past the potential of using these data for identification of crash-prone conditions has been explored. In the present work, application of probabilistic neural networks (PNN) is explored to identify conditions prone to rear-end crashes on the freeway. PNN is a neural network implementation of the well-documented Bayesian classifier. In this research the rear-end crashes observed on the Interstate-4 corridor in Orlando FL are divided into two groups based on the average traffic speeds observed around the crash location prior to the crash occurrence. Using decision tree-based classification it was observed that although these two groups of crashes have comparable frequencies, traffic conditions belonging to one of the groups (characterized by a low-speed traffic regime) are comparatively rare on the freeways. Hence, if those conditions are encountered on the freeway in real time, then conditions may be considered prone to rear-end crashes. As conditions belonging to the other group of rear-end crashes (characterized by a medium-to-high speed regime) are more commonly observed on the freeway, PNN-based classification models are developed for this group. The rear-end crashes along with a sample of randomly selected noncrash cases were used to calibrate the classifiers. The output layer of the PNN models was modified to provide a measure of crash risk, instead of the binary classification based on an arbitrary threshold. A desirable threshold on this output may be established to separate crash-prone conditions from “normal” freeway traffic.