+44 (0)1509 263171
Please use this identifier to cite or link to this item:
|Title: ||Developing an advanced collision risk model for autonomous vehicles|
|Authors: ||Katrakazas, Christos|
|Issue Date: ||2017|
|Publisher: ||© Christos Katrakazas|
|Abstract: ||Aiming at improving road safety, car manufacturers and researchers are verging upon autonomous vehicles. In recent years, collision prediction methods of autonomous vehicles have begun incorporating contextual information such as information about the traffic environment and the relative motion of other traffic participants but still fail to anticipate traffic scenarios of high complexity. During the past two decades, the problem of real-time collision prediction has also been investigated by traffic engineers. In the traffic engineering approach, a collision occurrence can potentially be predicted in real-time based on available data on traffic dynamics such as the average speed and flow of vehicles on a road segment. This thesis attempts to integrate vehicle-level collision prediction approaches for autonomous vehicles with network-level collision prediction, as studied by traffic engineers.
An interaction-aware motion model (i.e. a model which describes the motion of each vehicle and the interactions between vehicles) based on Dynamic Bayesian Networks (DBNs) is extended in order to accommodate both network-level collision prediction and vehicle-level information. The corresponding datasets contain a) collision and traffic data from the M1 and M62 motorways on the Strategic Road Network of England during 2012 and 2013 and two expressways in Greece, b) highly disaggregated simulated traffic and conflict data from M62 and c) vehicle-level data acquired using the radar sensor of an instrumented vehicle.
The prevailing traffic conditions just before reported collisions as well as traffic conditions during normal operations act as inputs to the network-level classifiers in order to estimate the probability of a collision happening in real-time. Network-level collision prediction is performed by six machine learning classifiers, i.e. k-Nearest Neighbours (kNN), Support Vector Machines (SVMs), Relevance Vector Machines (RVMs), Random Forests (RFs), Gaussian Processes (GPs) and Neural Networks (NNs). Moreover, as normal traffic conditions are usually overrepresented in traditional real-time collision prediction studies all the network-level collision prediction classifiers are treated with imbalanced learning techniques to assure proper identification of both hazardous and safe traffic.
The network-level classification results imply that imbalanced learning crucially increases the power of all network-level classifiers. Undersampling cases representing safe traffic conditions is found to work better with traffic data aggregated in 5-minute or 15-minute intervals. On the other hand, oversampling dangerous traffic conditions along with undersampling safe cases performs better in highly disaggregated data (i.e. in 30-second or 1-minute intervals).
By integrating network- and vehicle-level information in the interaction-aware DBN, it has been found that when traffic conditions are classified as hazardous, then the identification of dangerous traffic participants is notably enhanced. Even when traffic data aggregated at 30-second intervals are utilised, the identification of vehicles posing an imminent threat to the ego-vehicle is reinforced by 9-14%. However, when traffic conditions are deemed as normal, the interaction-aware model demonstrated that network-level information does not boost the detection of dangerously driving vehicles.|
|Description: ||A Doctoral Thesis. Submitted in partial fulfilment of the requirements for the award of Doctor of Philosophy of Loughborough University.|
|Appears in Collections:||PhD Theses (Design School)|
Files associated with this item:
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.