Machine learning-based decision-making for autonomous vehicles at roundabouts

The central idea behind developing autonomous vehicle (AV) and advanced driver assistance systems (ADASs) is to learn human-like decision-making, in order to enhance driving safety when controlling vehicles. In these high-level control tasks, joining roundabouts safely and in a timely manner is a challenging problem, even for human drivers. A number of factors need to be considered for dealing with this dynamic and complicated environment. The locations, speeds and intentions of vehicles at a roundabout, or when waiting to join traffic, could influence control decisions made by either a human driver or an AV/ADAS. This thesis proposes three joining roundabouts decision-making approaches to tackle the problem. The first, a novel approach proposes grid-based image-processing approach (GBIPA) with cameras at roundabouts, to help machine learning algorithms learn the different traffic criteria required to join a roundabout safely. The second approach offers a multi-grid-based camera (MGC) system using multiple devices in order to improve learning performance from GBIPA. The MGC approach uses a fine grid to determine the speed and direction of approaching vehicles, whilst a larger grid assesses these vehicles’ positions. Issues with this solution include the AV’s position and orientation varying significantly when stopped at a roundabout, and recorded and real-time driver behaviours are not always identical. Besides, in order to optimize AV's decision time and accuracy, the third approach proposes a novel imitation learning-based decision-making framework (ILBDM) to take observations from a monocular camera mounted on a vehicle as an input device and then use deep learning networks to make the decision regarding when it is best to join a roundabout. The domain expert-guided learning framework could not only improve decision-making, but also speed up the convergence of deep learning networks. In total, 460 videos recorded the actions of human drivers approaching and joining various types of roundabouts, in order to help assess the three proposed approaches’ learning processes. This thesis contributes in three ways. First, the proposed methods enable AVs to make the right decision when reaching a roundabout. Second, the thesis provides a new roundabout-entering dataset for AV research. Large amounts of high-quality data were collected in this regard, representing roughly 50 roundabouts in 460 video sequences recorded at different times and in various weather conditions. As the data reflect real traffic conditions, the model learned from this information can be easily applied in real-world applications. Third, the proposed three approaches are assessed by using different learning architectures, in order to help achieve high levels of accuracy and short decision times before an AV enters a roundabout.