Novel algorithms for longitudinal and lateral control for application in autonomous vehicles

Abstract

The studies undertaken in this project address some important problems associated with autonomous passenger cars. The word "autonomous" is used in a very broad sense and indicate a vehicle that can drive by itself through a road sharing with other vehicles. The studies also fall under the scope of Automated Highway System (AHS) that have been advocated by many academics as well as industrialists to curb the ever-increasing traffic problems worldwide. In the thesis, we focus on the lateral (steering) and the longitudinal (speed and distance) control problems of the autonomous vehicle that are essential in AHS implementation. The objective of lateral control is to maintain the vehicle in the center of the road at all possible speed and in the presence of disturbances. Based on the analysis of bicycle model, a genetic fuzzy proportional-plus-derivative (PD) controller is designed to guarantee stability and provide driver comfort. The fuzzy PD controller with feedforward compensation is initially designed based on the characteristic of bicycle model and summaries of human experience, then optimized off-line via genetic algorithms. Not only the scaling factors, but also the membership function of input variables and consequent parts of rules are optimized. Stability analysis studied through Lur's system and Lyapunov's direct method is also included. In this thesis, we also address the problem of longitudinal vehicle control. Its objective is to achieve automatic vehicle following in the longitudinal direction by following the speed response of the leading vehicle and at the same time keeping safe intervehcile spacing. A Fuzzified Radial Basis Function Networks (FRBFN) controller is designed to combine the advantage of fuzzy logics and neural network, which is known for approximation ability and fast training. The FRBFN is pre-structured to simulate the function and elements of a traditional fuzzy logic controller, which make the parameters of the neural networks have clear physical meaning. Then the output layer's parameters are on-line learned via gradient algorithm. Training data for off-line training and vehicle longitudinal dynamic model are not required for this method. Simulation results are provided to demonstrate effectiveness of the controller. The studies include simulation results as well as extensive experimental trials implemented on an in-house designed and built small-scaled prototype vehicle. This test bed is essentially an extensively modified Radio Controller (RC) car, which is driving and steering by the front wheels. Using the existing hardware (chassis, tires, steering actuation, and motor, etc), infrared and ultrasonic sensors are installed to measure distance, and encoder sensor to measure vehicle speed. An industrial computer combined with A/D card is mounted on the vehicle as the main control system. Vehicle lateral bicycle model and actuator model are obtained via system identification method through intensive experimental data. The small-scaled vehicle dynamic model exhibits the same properties with standard vehicle model. Thus, controller designed and implemented on this small-scaled vehicle should provide comparable results with that of an actual vehicle. The proposed control algorithms are implemented on the small-scaled vehicle. Experiment results show effectiveness and robustness of the control algorithms through simulating actually traffic conditions.