Support vector machine model for prediction of accidents on non-urban sections of highways

A support vector machine (SVM) model is proposed for accident prediction on non-urban sections of highways as an alternative to conventionally used fixed-effect or random-effect negative binomial (FENB/RENB) regression models. Road accident data over a period of 8 years on different sections of eight highways in India were collected from police records. In addition, data relating to road geometry, traffic and road environment related variables were collected through field studies. A total of 222 data points was gathered by dividing highways into sections with certain uniform geometric characteristics. Two modelling approaches were used (RENB and SVM models) to predict accident frequencies. The results showed encouraging performance of the SVM model in comparison with the RENB model in terms of both the correlation coefficient and root mean square error values. The SVM, originating from statistical learning theory, can better solve over-fitting and local minima problems and the results indicate that this approach can effectively be used as an alternative to the RENB approach if the sole aim is to predict crashes. The results clearly indicate that, to improve safety on Indian highways, minor access roads to highways and service roads need to be properly designed and controlled and the dispersion of speeds needs to be reduced.