This thesis describes a motion planning system for automated vehicles. The planning method is universally applicable in on-road scenarios and does not depend on a high-level maneuver selection automaton for driving strategy guidance. The majority of additional driving modes will therefore be processed in one particular processing unit for future high and full automation systems. The thesis pursues three research questions that have given rise to all the solutions presented. What are the capabilities of a universal method? How can perception quality benefit from motion planning? And how can robustness against motion uncertainty be ensured?
A planning framework using graphics processing units (GPUs) for task parallelization is presented. A method is introduced that solely uses a small set of rules and heuristics to generate driving strategies. It was possible to show that GPUs serve as an excellent enabler for real-time applications of trajectory planning methods. For further assessment, planning benchmark criteria have been derived and applied.
Like humans, computer-controlled vehicles have to be fully aware of their surroundings. Therefore, a contribution that maximize scene knowledge through smart vehicle positioning is evaluated. It can be shown through experimental results that areas marked as relevant are more often covered by sensors by minor adjustments of the initial optimal planning solution.
A post-processing method for stochastic trajectory validation supports the search for longer-term trajectories which take ego-motion uncertainty into account. This achievement is a gain in the trajectory’s temporal validity.
Autonomous vehicles are virtually regarded as the panaceas for the future of road transport due to numerous promising benefits. Thus, they have attracted wide attention from both academia and industry. Although associated technologies have been investigated and developed for decades, several obstacles still need to be overcome. One of the major obstacles is the total cost of the needed sensors. Therefore, it would be a long-term and effective solution to use less expensive sensors that can provide the same or even better performance.
The main focus of this dissertation is to design and implement a feature-based localization system. It aims to substitute the most expensive part of the present autonomous test vehicle, Applanix POS LV 510 which is four times more expensive than the vehicle. To do that, feature maps were first constructed through the log files of the test drive in Berlin. Then an algorithm is proposed to localize the test vehicle within the pre-built maps by using the Velodyne LIDAR and the Electronic Stability Program (ESP) associated sensors. The Velodyne LIDAR is employed to extract pole-like features from the previously mapped environments. Gyroscope and wheel speed sensors from the ESP are utilized to carry out the relative localization. The estimation does not need any GPS information after initialization. The performance of the proposed localization algorithm was evaluated through two datasets and the results indicate that it is comparable to the Applanix system. The real on-road tests also verified its effectiveness and robustness in terms of accuracy and precision. It is even more precise than the Applanix system as it shows higher repeatability.
The main innovations and contributions of this thesis can be summarized in three aspects. First, an innovative two-point localization scheme is proposed. It can greatly mitigate the influence of the wrong feature matching during the data association stage, thus it can get more accurate estimations. Second, an Ackermann constraint based trajectory smoothing method is proposed, which can smooth the trajectories especially during U-turns. Finally, the idea of using the online data to create feature maps is also evaluated and tested on the real roads. It is less accurate than the method of using the log files to create feature maps, but it can create city scale feature maps in a more efficient and convenient way.