**1. Introduction**

Amid an active discussion of the Fourth Industrial Revolution, Automated Vehicles (AVs) are expected to play an important role in leading the Fourth Industrial Revolution. AVs are defined as vehicles capable of navigating, controlling, and avoiding risk partly or totally without human assistance [1]. According to the Society of Automotive Engineering [2], AVs can be categorized into six levels, ranging from none auto-system (SAE level 0) to full auto-system (SAE level 5). Human intervention is minimized from SAE level 3 and driverless driving is possible at level 5. With such features as driving safety improvement, increase in convenience and mobility [3,4], AVs are highly evaluated as key future mobility of reducing traffic accidents. The benefits mentioned above will be accomplished when AVs fully take root. However, some researches have indicated that the public still expresses a low level of acceptance for AVs [5–7]. It is mainly attributed to AVs traffic accidents arisen during the test driving by Google, Uber, etc. In particular, a fatal pedestrian accident involving Uber has been at the forefront of ethical controversy over AVs. Neither of the types of traditional ethics (deontology, utilitarianism) fit in well to provide a proper answer to this accident, nor the trolley dilemma excuse is unsuitable [8–10]. In response, the AVs guidelines, including provisions of preventive design and safety, were issued in Germany and the United States [11,12]. Specifically, German AV ethics guidelines state that "Automated and connected technology should prevent accidents wherever this is practically possible" in its fifth clause. In addition, various studies emphasized that trust in AVs is the most important determinant to accept AVs for their mobility, and that the trust is decided by perceived safety risk, compatibility, and system quality [13–15].

To respond to the prevention of potential AVs related accidents, this study proposes Preventive Automated Driving System (PADS) of using Deep Neural Networks (DNNs)-based tra ffic accident context analysis. The study conducts experiments to identify key features a ffecting tra ffic accidents caused by unpredictable conditions such as black ice, sink-hall, centerline crossing, and so on.

The paper is organized as follows: Section 2 reviews studies on the e ffects of AVs and various tra ffic accident cases, and examines deep learning applications used in transportation research. Section 3 introduces methods to collect and process tra ffic accident data. In Section 4, we introduce how to build an optimal DNNs algorithm for forecasting the severity of accident injuries and extract factors causing accidents. Section 5 validates the important factors extracted in Section 4 by using a random forest-based machine-learning algorithm. Finally, Section 6 concludes the paper with a summary of empirical findings and derives future researches and implications related to preventive AVs.
