**5. Result**

### *5.1. Extracting Tra*ffi*c Accidents Context Fatures*

Learning based on the optimization model was carried out to identify factors affecting the degree of offender injury. As shown in Table 7 and Figure 4, the -1 set, which only uses external factors, showed 81.16% prediction accuracy. It means that more than 80% of the injury to the offender can be predicted using external factors only. In addition, 84% prediction accuracy was achieved on the degree of injuries if additional factors of the vehicle were applied to the -1 set, and 84.38% accuracy was confirmed when the vehicle type and the violation of the law were taken into consideration. As with offender factors, 61.05% accuracy was confirmed for the victim's injury when using external factors only. In addition, 66.46% of accuracy was observed in predicting victim's injuries when vehicle type was added to -1 Set and 64.43% accuracy when adding time and day factors along with the vehicle type.

As a result of the analysis of the degree of injury of the offender, high accuracy is confirmed when the vehicle type and the violation of the law are added to the speed, the external factor. In addition, when the time, day, and vehicle types are set as additional factors in addition to external factors, high accuracy is achieved in the analysis of the victim's injury. The time series and vehicle type are considered to be the main factors in determining the victim's injury (See Table 7 & Figure 3).

Following the learning through the optimization model, we identified specific features that determine the degree of traffic accident injury. Based on the analysis through DNN learning, a random forest feature importance analysis is performed, which is a machine learning technique, to confirm factors more accurately. As a result, the main factors of injuring offenders were identified in the order of "Vehicle Type, Speed, Time, and Day of the week", and those of the victims were in the order of "Speed, Time, Victim Vehicle Type, and Day of the Week". It was found that the factors that determine the

degree of injury to the offender and the victim are similar. However, for the offender, the vehicle type is found more critical, while for the victim, the link speed is identified as more important (See Table 8 & Figure 4).


**Table 7.** Offender/Victim Injuries Prediction Accuracy using optimal DNNs model.

(**a**) Offender injuries accuracy

**Figure 3.** (**a**) Offender Injuries Accuracy by Optimal DNNs; (**b**) Victim Injuries Accuracy by Optimal DNNs: The red dash line rectangular stands for the highest accuracy at that factors.

**Table 8.** Result of Importance Analysis for Traffic Accident Factors.


(**a**) Offender injuries Feature Importance 

(**b**) Victim injuries Feature Importance 

**Figure 4.** (**a**) Injuries Feature Importance; (**b**) Victim Injuries Feature Importance
