Driver intention analysis and anomaly detection for human-like automated driving vehicles

Automated driving technology is promising a safer transport future due to its potential to avoid human errors that are currently responsible for most road traffic accidents. While the technologies gradually evolve towards full automation, human driven vehicles and vehicles of different levels of automation will continue to share the road in the foreseeable future. Ensuring the safety of the automation functions as well as improving public acceptance of such technologies are getting the most attention from industry and academia.

The aim of this project is to develop reliable machine learning models to help understand the intentions of fellow human driven vehicles (e.g. the social etiquette to give way) and identify abnormal driving behaviours, with which automated driving cars can act accordingly and operate safely, efficiently and in a way that is accepted by the general public. To achieve such understanding, it is unlikely that hard-code rule-based logics in the automation systems will tackle all the real world scenarios, not to mention the edge cases that have not been observed in the training data. Machine learning and associated machine learning models may be the only effective way to emulate human like understanding of intentions and identify abnormal driving behaviours. To train such models, data collection would be expensive using dedicated test vehicles and edge cases are rare to come by. Therefore, it is proposed to access crowd-sourced data from video based social media platforms such as Youtube, where dashcam footage of various dangerous driving behaviours are voluntarily uploaded and shared in large quantity for financial incentives. On the other hand, several opensource datasets, such as KITTI, can serve as examples of normal driving behaviours. From the video data, various computer vision algorithms will be efficiently implemented in an integrated fashion to extract enough features for building a digital twin of normal driving as well as edge cases in the simulation environment so that subsequent simulation using randomised parameters will generate sufficient data for model training and validation.

This is a collaborative project linking both UKRI CDT in Accountable, Responsible and Transparent AI (ART-AI) and EPSRC CDT for Advanced Automotive Propulsion Systems (AAPS). The successful applicant will enjoy the technical and academic support from an interdisciplinary supervisory team composed of experts from both CDTs. Internally, a human-driver-in-the-loop test facility, a scaled autonomous vehicle test platform as well as a ground truth data acquisition system which are built in parallel to the PhD project will enhance the research proposed; while externally there are ongoing industrial and governmental collaborations that provide opportunities for the student to implement their research in real world applications such as notifying the emergency service with timely information of a traffic accident. They will also be encouraged to publish academic papers and to attend international conferences in machine learning and the field of intelligent transportation.

Applicants should hold, or expect to receive, a first or upper-second class honours degree in computer science or information engineering, or a closely related discipline. A master level qualification, good machine learning or deep learning academic profile, and good coding skills would be advantageous. Prior knowledge in autonomous driving is desirable, but not required.

Formal applications should include a research proposal and be made via the University of Bath’s online application form. Enquiries about the application process should be sent to art-ai-applications@bath.ac.uk.

Start date: 2 October 2023.