Deep Learning-Based Automated Vehicle Steering

Abstract

Autonomous Vehicle applications are full of open challenges. Despite the advanced technologies, the lack of robust systems still exists due to the high complexity of the surrounded environments. The automated steering is one of the most complex autonomous driving system's application. Model predictive control is the most common control strategy used to implement the automated steering tasks due to its ability to solve an online quadratic optimization problem in the real-time, in addition to its efficiency in handling the constraints of the system's environments. MPC controller is used to drive the vehicle autonomously along the centerline of the road based on two main factors, the lateral deviation and relative yaw angle. Deep learning technology has been widely used in recent years because of the promising performance achieved in different applications and tasks. In this context, we suggested that the implementation of the Deep Neural Network (DNN) will provide a great improvement and it can be more computationally efficient than solving an online quadratic problem (QP), that will naturally lead to reduce the time, the complexity, and the computational loads of implementations. The main aims of this paper are to design a deep learning-based approach for automated vehicle steering based on the behaviour of the traditional MPC controller. In addition, to study the efficiency of the full replacement of the MPC controller by the suggested DNN model. The study is based on performing a comparison between the implementations of both controllers (MPC and DNN model) in terms of the performance and the execution time. The performance indicator is the ability of the controller to drive the decision variables (lateral deviation and yaw angle) to be close to zero in order to drive the vehicle autonomously along the desired path.