Knowledge Discovery in Vehicle Identification Sensor Networks

Abstract

Intelligent Transport Systems (ITS) are driving innovation in road transport by integrating advances in communication and information systems with traditional engineering practices. Core to ITS development is the collection and analysis of traffic data. One class of traffic sensors, known as Automatic Vehicle Identification, is characterised by the ability to identify vehicles using unique identifiers. Through vehicle re-identification, such sensors can provide reliable estimates of travel time and inform route choices, both at the individual and aggregate levels, across all levels of the road hierarchy. In particular, Automatic Number Plate Recognition (ANPR) video cameras require just a visible number plate instead of specialised devices for vehicle detection. The benefits of ANPR technology for traffic monitoring have led to its adoption in cities across the world, forming complex sensor networks with increased potential to power ITS solutions. Despite successful application in traffic forecasting, two technical barriers prevent a more widespread and diverse adoption of ANPR networks: • The lack of technical guidance on pre-processing ANPR data. We address this by developing a data pipeline which documents the various data sources and processing steps required to produce traffic data ready for analysis. In addition, we benchmark the pipeline against a real ANPR network, located in the North East of England. • The methodological gap in representing and extracting popular travel routes (corridors) from observed data. We develop a mathematical framework for corridor identification, which highlights route importance in connecting and distributing regional road traffic. The second part of this thesis focuses on two new ITS applications of ANPR networks. They demonstrate how traffic authorities can collect evidence of corridor performance and safety issues in order to prioritise transport improvements: • Bottleneck detection and impact assessment is a critical traffic monitoring activity largely confined to highways. By developing a detection algorithm for ANPR monitored corridors, bottleneck detection is scaled to an entire urban network. Bottlenecks are categorised by frequency of occurrence, and their impact quantified against other sources of congestion, indicating that recurring bottlenecks account for over 75% of urban traffic congestion. Our method is the first to use ANPR sensors to automatically identify traffic bottlenecks and quantify their impact across an urban road network. • Frequent overtaking and lane-changing behaviour can have negative impacts on traffic flow. We investigate the link between overtaking rate and traffic conditions as a proxy to understanding and quantifying corridor safety levels. Our findings

suggest that overtaking rate increases with vehicle concentration and inversely with speed, albeit with a scaling relationship that greatly depends on road characteristics. Our method is the first to be able to quantify the scaling effect of vehicle overtakings for a variety of roads and traffic conditions. Successful traffic management increasingly relies on continuous data collection and analysis. Using our pipeline for data processing and new methodologies of data analysis, stakeholders can extract added value from their ANPR sensing infrastructure and better position themselves to fully realise the vision of intelligent traffic management systems