Abstract
This research focuses on examining the interacting relationship between vehicle crashes and road engineering factors. The back-propagation artificial neural network (BPN) method with feed-forward structure is used to develop the model. Each intersection is decomposed into several approach sets to establish the proposed microscopic relationship with vehicle crashes. A total of 1,225 records at 69 primary/secondary intersections in Tainan, Taiwan, were used to calibrate the bi-level models of 11 crash types. First is a distinguish model to tell if an approach set will have crash; the second is a prediction model to forecast crash numbers by type for each approach set. Error rates for all distinguish models are under 15% with an average of 6.1%. Error rates for all prediction models are under 25% with an average of 6.7%. An overall RMS is 0.1617. Practical field test also proves model's validity.
