Even though some of the driver assistant systems have been commercialized to provide safety and convenience to the driver, they can be applied for autonomous driving in limited situations such as highways. In this paper, we propose a supervisor agent, which can enhance the driver assistant systems by using deep distributional reinforcement learning. The supervisor agent is trained using end-to-end approach which directly maps both a camera image and LIDAR data into action plan. Because the well-trained network of deep reinforcement learning can lead to unexpected actions, collision avoidance function is added to prevent dangerous situations. In addition, the highway driving case is a stochastic environment with inherent randomness and, thus, its training is performed through the distributional reinforcement learning algorithm, which is specialized for stochastic environment. The optimal action for autonomous driving is selected through the return value distribution. Finally, the proposed algorithm is verified through a highway driving simulator which is implemented by the Unity ML-agents.

Deep learning is a fast-growing field of research, in particular, for autonomous application. In this study, a deep learning network based on various sensor data is proposed for identifying the roads where the vehicle is driving. Long-Short Term Memory (LSTM) unit and ensemble learning are utilized for network design and a feature selection technique is applied such that unnecessary sensor data could be excluded without a loss of performance. Real vehicle experiments were carried out for the learning and verification of the proposed deep learning structure. The classification performance was verified through four different test roads. The proposed network shows the classification accuracy of 94.6% in the test data.

With the commercialization of various Driver Assistance Systems (DAS), those vehicles have some autonomous functions like Smart Cruise Control (SCC) and Lane Keeping System (LKS). It is believed that autonomous driving can be achieved by combining the DAS functions in the limited situations such as on highways. However, in order to coordinate the DAS functions for autonomous driving, a supervisor is needed to select an appropriate DAS function. In this paper, we propose a method for training a supervisor that selects proper DAS by deep reinforcement learning. The driving policy operates based on camera images and LIDAR data that are accessible in autonomous vehicles. Therefore, deep reinforcement learning network model is designed to analyze both camera image and LIDAR data. This system aims to drive in simulated traffic situation of highway without collision and with high speed. Unlike the systems which learn how to throttle, brake and steering directly, the proposed method can guarantee safe driving because the learned driving policy is based on the existing commercialized DAS functions. In order to verify the algorithms, a simulation tool is developed using Unity for highway environment with multiple vehicles and autonomous driving performance is compared with the proposed supervisor.

This paper presents an efficient hierarchical motion planning framework with a long planning horizon for autonomous driving in structured environments. A 3D motion planning with time information is a non-convex problem because there exists more than one local minimum point and various constraints such as roads and obstacles. Accordingly, to deal with high computational complexity and the problem of falling into a local minimum in an enormous solution space, a decoupled method is utilized, that consists of two steps: Long-term planning and short-term planning. First, the long-term planner provides reasonable far-sighted behavior through two processes. In the first process, a rough path that includes a driving strategy is generated in the 2D spatial space. Then, the jump point search algorithm is utilized with time information on the path to reduce the computational burden of A*, giving an acceptable quality of solution at the same time. In this step, a safe, comfortable, and dynamically feasible trajectory is generated. Next, the short-term planner optimizes a short-sighted trajectory using particle swarm optimization. In this method, a steering angle set is encoded as a particle, resulting in a safe, comfortable, and kinodynamically feasible trajectory. The proposed algorithm is implemented and evaluated in a range of vehicle-in-the-loop simulation scenarios, which include various virtual static and dynamic obstacles generated by Intelligent Driver Model. In the evaluation results, the proposed method reduced the computation time by up to 0.696 s with increasing the step cost by up to about 3%. The proposed algorithm is executed every 100 ms for a planning horizon of 10 seconds, and the average computation time is 31 ms with the worst-case computation time of 94 ms.

The safety of an automated vehicle requires accurate information of surrounding conditions, because a false sensor output can lead to a fatal accident during driving. Thus, monitoring of abnormalities in every sensor is important for robust perception of the environment. Since it is difficult to obtain anomalous data, it is hard to develop a robust detection algorithm using only a relatively small number of anomalies. In this paper, we propose a data augmentation method for oversampling minority anomalies in lane detection. Using a generative adversarial network that makes the generator learn to estimate the distribution of anomalous data, it generates synthesized minority anomalies. The generated anomalies are used to train an anomaly detection network while minimizing latency for use in real situations. During training, the generated anomalies, with various mixed quality, are sampled differently according to their quality. This helps the detection network to be optimized with better quality data. Experimental result shows that when using the proposed anomaly detection framework for monitoring lane abnormality, it improves the performance by 12% when compared to the vanilla recurrent neural network.

This paper proposes semi-autonomous control that minimizes intervention considering driver's steering and braking intentions. The biggest challenge of this problem is how to fairly judge driver's intentions that appear differently in the lateral and longitudinal directions and how to minimize controller intervention. A decoupled model predictive control (MPC) and optimal intervention decision methods are proposed considering driver incompatibility. Several MPCs are designed first considering the fact that the driver can avoid obstacles either by braking or moving to the left or right lanes. The control input to avoid the collision is calculated for each MPC such that its intervention can be minimized reflecting the driver's intention. After driver incompatibility is formalized, the optimal input is selected to minimize the incompatibility among the paths that can avoid accidents. The proposed algorithm is validated in simulations where collision can be avoided while minimizing the intervention.