Deep neural network-based predictive modeling of road accidents

This work proposes to use deep neural networks (DNN) model for prediction of road accidents. DNN consists of two or more hidden layers with large number of nodes. Accident data of non-urban sections of eight highways were collected from official records, and dataset consists of a total of 2680 accidents. The data of 16 explanatory variables related to road geometry, traffic and road environment were collected from official records as well as through field studies. Out of a total of 222 data points of accident frequency, 148 were used for training and remaining 74 to test the models. To compare the performance of DNN-based modeling approach, gene expression programming (GEP) and random effect negative binomial (RENB) models were used. A correlation coefficient value of 0.945 (root mean square error = 5.908) was achieved by DNN in comparison with 0.914 (RMSE = 7.474) by GEP, and 0.891 (RMSE = 8.862) by RENB with the test dataset, indicating an improved performance by DNN in prediction of road accidents. In comparison with DNN, though lower value of correlation coefficient was achieved by GEP model, it quantified the effects of various variables on accident frequency and provided a ranked list of variables based upon their importance.