Real-time freeway sideswipe crash prediction by support vector machine

This study presents the applications of a pattern classifier named support vector machine (SVM) in predicting freeway sideswipe crash potential. Historical loop detector data for sideswipe crashes and corresponding non-crash cases were collected from Interstate-894 in the Milwaukee, Wisconsin, USA. Two sets of significant explanatory features were aggregated from the collected detector data to capture the prevailing traffic state and variances between adjacent lanes. Then, three SVMs with different nonlinear kernel function were formulated with the significant features as inputs. To comparatively evaluate the performance of SVM models against other commonly applied crash potential predictors, the multi-layer perceptron (MLP) artificial neural network models were also developed to predict sideswipe crash potential. The results showed that SVM models offers similar overall accuracy as the premier MLP model, but SVMs achieved better sideswipe crash identification at higher false alarm rates. The research also investigated the potential of using the SVM model for evaluating the impacts of traffic factors on sideswipe crash. Sensitivity analysis conducted on the trained SVM models successfully identified the variables' impact on sideswipe crash. These results affirmed the superior performance of SVM technique in crash potential prediction analysis.