**6. Conclusions**

Automated Vehicles are under the spotlight as an alternative to diminishing traffic accidents. However, AVs accidents that occurred during several test drives led to worry about AVs safety. When AVs are commercialized and mixed with human driving vehicles, AVs-related accidents can occur even under ideal conditions. In order to solve this problem, studies on AVs ethics have been conducted, but it was di fficult to obtain an adequate answer. So, Germany and the United States issued AVs ethical guidelines emphasizing the prevention of tra ffic accidents. This study analyzed the tra ffic accident data based on DNNs to propose preventive measures for AVs accidents.

For the analysis of tra ffic accidents, we preprocessed TAAS, standard node-link data, and TOPIS data. In the DNNs analysis, the input layer consists of external factors and additional factors, and the output layer was set to the degree of injuries for both o ffenders & victims. Also, to construct the optimal hidden layers, we controlled the learning data range, the epoch, and the number of nodes in the hidden layers. For the analysis, we conducted learning by adding input factors incrementally. The results show that in the o ffender case, 81% of prediction accuracy was achieved when only external factors were considered, while the accuracy increased to 85% when factors such as violation of law and vehicle types are added to the analysis. In the victim case, the prediction accuracy remained at 61% with external factors only, but the accuracy increased to 67% when additional factors were taken into consideration. The main reason determining about 20% prediction accuracy deviation between o ffender and victim comes from the level of data accuracy of the key injury determinant between o ffender and victim. The o ffender's degree of injury is largely determined by the type of vehicle, which did not require the data collection. On the other hand, the speed which a ffected the degree of victim injury most were from the merged data of TOPIS and TAAS with many missing data. The analysis revealed that factors such as vehicle type, time, and day were found important in determining the degree of injury. In addition, a random forest importance analysis identified that the injury determinants of the o ffender were ordered vehicle type, speed, time, and day, while those of victims were speed, time, vehicle type, and day.

As discussed, vehicle type and speed were identified as the main factors of determining injury, respectively. Accordingly, in the future, we plan to conduct researches that can more accurately prevent accidents by recognizing vehicle types using the CNNs (YOLO, etc.) technique, and that can predict the degree of injury according to speed or suggesting an appropriate vehicle safety distance. Also, future study will apply a developed methodology (Meta AI etc.) using updated tra ffic accident data. However, since data from HVs were utilized to develop preventive measures for AVs accidents, there is a limit to directly apply the results to AVs. Nevertheless, since the analysis focused on external factors (such as weather, road conditions, road types, etc.) that are di fficult to control even AVs, it is expected that they will be used as basic data for analyzing the impact of external factors related to tra ffic accidents related to AVs in the future. Future study is needed to analyze tra ffic accidents by using data on the causes (e.g., weather, body defects, etc.) of AVs accidents as input values.

**Author Contributions:** Conceptualization, K.H.; Data curation, J.S.; Formal analysis, M.K.; Investigation, M.K.; Methodology, M.K.; Project administration, K.H.; Software, M.K. and J.S.; Visualization, M.K.; Writing—original draft, M.K.; Writing—review & editing, M.K. and K.H. All authors have read and agreed to the published version of the manuscript.

**Funding:** This paper was supported by a gran<sup>t</sup> (code 20TLRP-B148970-03) from Transportation and Logistics R&D Program (TLRP) funded by Ministry of Land, Infrastructure and Transport of Korean government.

**Acknowledgments:** This paper is a revised and complemented content of "A Study on DNN (Deep Neural Network)-based Tra ffic Accident Context Analysis for The Design of Preventive Automated Driving System" Minhee Kang's Master Degree.

**Conflicts of Interest:** The authors declare no conflict of interest.
