Article Preview
Top1. Introduction
With artificial intelligence technologies, deep learning technology, which has recently been receiving great attention, has been researched based on artificial neural networks developed in the 1940s, but after it was stopped due to a problem with optimization, error backpropagation in the 1980s. The backpropagation algorithm and the research started again. Self-driving cars are cars that can drive to their destination by self-aware of the surrounding environment, the judgment of danger, and explosive growth is predicted in the future. High-level artificial intelligence technology is required to distinguish people who are vulnerable to walking at risk of accidents. Not only IT companies such as Google and NVidia but also global automotive companies need to study deep learning technology of artificial intelligence as a core technology for autonomous driving. In this study, a deep learning algorithm of artificial intelligence was applied to build the autonomous driving system of automobiles, (Wang et al., n.d.)which is being activated recently, and attempted a new approach to the autonomous driving system by simulating it in the simulation environment, CARLA. The deep learning algorithm, which is a core part of this study, aims to efficiently access all of the above important points by using a CNN that shows excellent results in image (Chang et al., 2016) recognition among the existing deep learning algorithms and a model that is effective for outputting behavioral results with continuous values.
Imitation Learning, one of the learning methods of deep learning models applied for self-driving cars, uses data collected by experts. And, through neural network learning, the deep learning model can imitate the expert's behavioral policy. However, imitation learning is difficult to converge in a regression problem that not only requires vast amounts of data to learn complex behavioral policies but also predicts continuous values such as steering and acceleration that determine the behavior of a vehicle. In addition, researches on autonomous vehicles are currently being conducted for the purpose of autonomous driving in general urban terrain, so studies on autonomous vehicles in special environments are insufficient(Byung-yoon, 2016). As research on autonomous vehicles using deep learning is actively conducted in various industries and research facilities, various autonomous vehicle simulators exist. Various simulators such as AirSim, Carla, and LG SVL have been released, and these simulators provide not only vehicles based on physics engines for autonomous vehicles, but also various sensors required for information acquisition such as LiDAR, GPS, and cameras. In this study, among various (Dosovitskiy & Ros, n.d.)simulators, the AirSim simulator is used to conduct research. We used CARLA to study the performance of three approaches to autonomous driving. We present CARLA (Car Learning to Act), an open simulator for urban driving. CARLA was developed from the ground up to support the training, prototyping, and validation of autonomous driving models, including perception and control. CARLA is an open platform. Exceptionally, the content of urban environments provided with CARLA is also free. The content was created from scratch by a dedicated team of digital artists employed for this purpose. It includes urban layouts, a multitude of vehicle models, buildings, pedestrians, road signs, etc. Simulation The platform supports flexible configuration of sensor networks and provides signals that can be used to train driving strategies, such as GPS coordinates, speed, acceleration and detailed collision data, and other offenses.
Deep learning algorithms can be seen as a sophisticated and mathematically complex evolution of machine learning algorithms. The field has received a lot of attention lately and for good reason: recent developments have led to results that were not previously believed possible. Deep learning describes algorithms that analyze data with a logical structure similar to how a human would draw conclusions. Please note that this can happen through both (Krizhevsky et al., 2012)supervised and unsupervised learning. To achieve this, deep learning applications use a layered structure of algorithms called an artificial neural network (ANN). The design of such an RNA is inspired by the biological neural network of the human brain, leading to a learning process much more efficient than standard machine learning models.