State Lattice-based Motion Planning for Autonomous On-Road Driving

Since DARPA Urban Challenge 2007 (DUC), the development of autonomous vehicles has attracted increasing attention from both academic institutes and the automotive industry. It is believed that autonomous vehicles sophisticated and reliable enough would redefine mobility. The motion planner and sensor simulation presented in this thesis are intended to contribute to this prospect. The task of a motion planner for autonomous on-road vehicles is to generate a trajectory of motions for the vehicle to follow. The proposed motion planner employs a state lattice to construct a large variety of candidate trajectories and selects the best constraint-abiding one based on a set of cost criteria. The parallel computer architecture of CUDA is exploited to construct and evaluate the trajectories efficiently. The spatial planning horizon of the proposed planner can contain multiple segments with different widths. This feature helps the planner to adapt to various road layouts easily and to generate more consistent plans. During the construction of the state lattice, acceleration profiles are associated with the path segments to generate trajectory segments. Acceleration cubic polynomials and constant accelerations are applied in the proposed planner. The adopted association scheme makes it possible to span one acceleration profile over several trajectory segments, which, together with the applied smooth acceleration profiles, helps to enhance the feasibility of the trajectories. In the construction of cost maps of obstacles for the evaluation of the trajectories, the obstacles need to be dilated to compensate for the vehicle shape. A novel approach is proposed to analyse the sufficiency of the dilation strategy implemented in this work from the perspective of excluding all the trajectories that are practically not traversable. The scanning sensors are also simulated to increase the realistic level of the simulation experiments. Programmable shaders in the rendering pipeline of OpenGL are manipulated to record sensor-related data. Such implementation takes advantage of the parallel computer architecture of the GPU and thus enhances the computational efficiency of the generation process of the simulated sensor data. A novel macro-micro approach is proposed which can increase the accuracy of the simulated sensor data. Finally, the proposed planner is evaluated in a variety of simulated traffic scenarios. Given proper guidances from a behaviour planning layer, the proposed planner can generate reasonable plans in most scenarios.