This study was based on the black box images of traffic accidents on highways, cluster analysis and prediction model comparisons were carried out. As analysis data, vehicle driving behavior and road surface conditions that can grasp road and traffic conditions just before the accident were used as explanatory variables. Considering that traffic accident data is affected by many factors, cluster analysis reflecting data heterogeneity is used. Each cluster classified by cluster analysis was divided based on the ratio of the severity level of the accident, and then an accident prediction evaluation was performed. As a result of applying the Logit model, the accident prediction model showed excellent predictive ability when classifying groups by cluster analysis and predicting them rather than analyzing the entire data. It is judged that it is more effective to predict accidents by reflecting the characteristics of accidents by group and the severity of accidents. In addition, it was found that a collision accident during stopping such as a secondary accident and a side collision accident during lane change act as important driving behavior variables.