**2. Literature Review**

### *2.1. AVs Introduction Impacts*

It is important to estimate AV's impacts on tra ffic because it is inevitable that AVs will be mixed with human driving vehicles (HVs) in road tra ffic. With AV's Market Penetration Rate (MPR) growing, the positive e ffects can become bigger in congested conditions [16]. Most studies have been conducted quantitatively using a microscopic simulation software, VISSIM: A Leksandra Deluka Tibljaš et al. [17] designed a rotary interchange and evaluated safety in a mixed tra ffic condition. Yan Wang et al. [18] confirmed that AVs introduction not only kept Level of Service (LOS) higher, but also improved safety at signaled intersections. However, when MPR reached over 50%, a negative impact began to appear with a growing tra ffic delay [19]. Lee at al. [20] also identified that the maneuvering of AVs should be properly controlled by various tra ffic and road conditions because the driving behavior of HVs is a ffected by the aggressiveness of AVs. Some studies have mentioned that AVs contribute to decrease tra ffic accidents (collision) and delays [21]. Kolarova et al. [22] conducted an online survey for analyzing the potential changes in the Value of Travel Time Savings (VTTS). It is shown for commuting trips that AVs reduce 41% of VTTS on average compared to HVs. For leisure or shopping trips, no significant changes in the VTTS were found. Tscharaktschiew & Evangelinos [23] investigated the impact of the transition in automated driving capabilities (driving mode choice) on road congestion pricing and vice versa, accounting for the interdependencies between tra ffic flow, the chosen level of autonomous driving, e ffective road capacity and marginal travel cost. The result suggested that when inconveniences related to autonomous driving are su fficiently high, the imposition of congestion tolls may lead to a situation where drivers abandon autonomous technologies entirely and opt instead for fully manual driving, not the generally expected positive e ffects.

### *2.2. AVs Cognition Survey*

Along with AVs impact studies, several AVs-related cognition surveys have been conducted. Most cognition studies used surveys, confirming that preference towards AVs was higher for men than for women [24] and higher for the younger generation than for the older [25]. According to Nordho ff et al. [26], most people think that impacts of AVs appear to be positive, but responded that AVs safety benefit needs to be experimentally verified [27]. Moreover, Im et al. [28] analyzed web articles and comments about AVs using text mining techniques, showing that the number of reports that include AVs as a keyword increased, and there are more negative views than positive views. Specifically, the articles and comments with negative views discussed AVs ethics, tra ffic accidents, and the problem of sudden unintended acceleration.

### *2.3. AVs Tra*ffi*c Accidents and Derived AVs Ethics*

We reviewed AVs traffic accidents researches and AVs ethics issues for building the design of preventive AVs. Hong et al. [29] & Yang et al. [30] classified types of AVs accident as follows: the negligence of the drivers, accidents due to mechanical defects, malfunction of S/W, and accidents caused by information error, hacking, weather, etc. In the case of the steadily rising trolley dilemma problem [31], doubts were brought up as to whether the trolley dilemma could apply to AVs. They suggested that deriving accident algorithms to respond to it could be misleading [32]. Bae & Lee [33] approached decision-making criteria for the protection priority in accident situations. They proposed two solutions: to enforce them by law and to leave them to the AVs Artificial Intelligence (AI) system itself. In regards to setting them on AVs AI system, Gogoll & Müller [7] discussed di fferences between Mandatory Ethics Setting (MES), considering society as a whole and Personal Ethics Setting (PES) considering individual interests. For the accident in an ideal condition, Goodall [34] proposed ethical collision algorithms, and Fleetwood [35] discussed Germany's AVs ethics guidelines and main ethical topics.

### *2.4. Deep Learning Application in Transportation Area*

There exist many transportation prediction studies using deep learning technologies. These can be bisected into two categories: prediction of either tra ffic flow or tra ffic accidents. First, studies of forecasting tra ffic flow compare proposed model performance to traditional classical algorithms [36–39]. Particularly, the Long-Short Term Memory (LSTM) model is compared with Statistically Adjusted Engineering (SAE), Radial Basis Function (RBF), Support Vector Machine (SVM), and Auto Regressive Integrated Moving Average (ARIMA) model, and DNNs is compared with Random Forest, a kind of machine learning model. Tra ffic accident prediction studies were carried out using Social Network Service data (Twitter) using Deep Belief Network (DBN) & LSTM [40]. For real-time accident detection, Chen et al. [41] analyzed the accident impact using GPS-based vehicle data. They used Stack Denoise Autoencoder (SDA), which is more e ffective to detect accident risk than the traditional models.

Furthermore, as rapid progress is made nowadays in AVs technology, backed by advances in the areas of deep learning and AI, various studies about AVs using AI have been implemented. Especially as AVs requires an accurate perception of surrounding environments to operate reliably, most studies are related to Convolutional Neural Networks (CNNs). That is why object detection is a fundamental function of AVs systems, including camera sensor (2D), Lidar (3D), radar, GPS, etc. Among them, 3D object detection for AVs studies have been carried out recently, and a new methodology (combining or extension) has been proposed. For instance, Li, P., et al. [42] proposed the Stereo R-CNN for 3D object detection as Faster Regions with CNN (Faster R-CNN) for stereo inputs extending to detect objects simultaneously with images on the left and right. The experiments on the challenging KITTI dataset show that their method outperforms the state-of-the-art stereo-based method (Stereo R-CNN) around 30% AP on both 3D detection and 3D localization tasks. Also, Chen, S., et al. [43] developed a CNN–LSTM based on prior knowledge and temporal information for AVs driving. The proposed algorithm was found to be approximately 85% accurate in mimicking human drivers. Zeng, W., et al. [44] proposed the Deep Structured Self-Driving Networks (DSDNet), which performs object detection, motion prediction, and motion planning with a single neural network. The algorithm showed that it has outperformed the state-of-the-art method (DSDNet) on several challenging datasets in general. Existing AVs datasets are limited in the scale and variation of the environments they capture, even though generalization within and between operating regions is crucial to the overall viability of the technology. In response, Sun, P., et al. [45] conducted the study based on the vast amount of real data it owned as a leading group in the study of AVs. This study is based on actual AVs data, and presents a large-scale multimodal camera-LiDAR dataset that is significantly larger, higher quality, and more geographically diverse than any other existing similar dataset. In addition, Djuric, N., et al. [46] introduced a deep learning-based approach that takes into account a current world state and produces raster images of each actor's vicinity for presenting an effective solution to a critical part of the AVs problem. The method first rasterizes actor contexts, followed by training CNNs to use the resulting raster images to predict the actor's short-term trajectory and the corresponding uncertainty. Also, they tested the framework (system), which was deployed to a fleet of AVs.
