Aiming at the limitation of the traditional log-linear model widely applied to frequency prediction of traffic accidents, in which the elasticity is assumed as constant and the interrelation of the explosive variables is simplified, a new frequency prediction model of traffic accidents based on translog production function was proposed. Adopting a dynamic elasticity, the model predicts the accident frequency according to two major variables: traffic flow and the length of section. The Ontario accident statistics from Canada has verified its effectiveness and better goodness of fit compared with the traditional log-linear model.