Date of Award

12-2017

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Automotive Engineering

Committee Member

Dr. Beshah Ayalew, Committee Chair

Committee Member

Dr. Andrej Ivanco

Committee Member

Dr. Yunyi Jia

Committee Member

Dr. Simona Onori

Abstract

Autonomous vehicle technology is a promising development for improving the safety, efficiency and environmental impact of on-road transportation systems. However, the task of guiding an autonomous vehicle by rapidly and systematically accommodating the plethora of changing constraints, e.g. of avoiding multiple stationary and moving obstacles, obeying traffic rules, signals and so on as well as the uncertain state observation due to sensor imperfections, remains a major challenge. This dissertation attempts to address this challenge via designing a robust and efficient predictive motion planning framework that can generate the appropriate vehicle maneuvers (selecting and tracking specific lanes, and related speed references) as well as the constituent motion trajectories while considering the differential vehicle kinematics of the controlled vehicle and other constraints of operating in public traffic. The main framework combines a finite state machine (FSM)-based maneuver decision module with a model predictive control (MPC)-based trajectory planner. Based on the prediction of the traffic environment, reference speeds are assigned to each lane in accordance with the detection of objects during measurement update. The lane selection decisions themselves are then incorporated within the MPC optimization. The on-line maneuver/motion planning effort for autonomous vehicles in public traffic is a non-convex problem due to the multiple collision avoidance constraints with overlapping areas, lane boundaries, and nonlinear vehicle-road dynamics constraints. This dissertation proposes and derives some remedies for these challenges within the planning framework to improve the feasibility and optimality of the solution. Specifically, it introduces vehicle grouping notions and derives conservative and smooth algebraic models to describe the overlapped space of several individual infeasible spaces and help prevent the optimization from falling into undesired local minima. Furthermore, in certain situations, a forced objective selection strategy is needed and adopted to help the optimization jump out of local minima. Furthermore, the dissertation considers stochastic uncertainties prevalent in dynamic and complex traffic and incorporate them with in the predictive planning and control framework. To this end, Bayesian filters are implemented to estimate the uncertainties in object motions and then propagate them into the prediction horizon. Then, a pair-wise probabilistic collision condition is defined for objects with non-negligible geometrical shape/sizes and computationally efficient and conservative forms are derived to efficiently and analytically approximate the involved multi-variate integrals. The probabilistic collision evaluation is then applied within a vehicle grouping algorithms to cluster the object vehicles with closeness in positions and speeds and eventually within the stochastic predictive maneuver planner framework to tighten the chanced-constraints given a deterministic confidence margin. It is argued that these steps make the planning problem tractable for real-time implementation on autonomously controlled vehicles.

Share

COinS