This thesis describes the development of neural network models for Automatic Incident Detection (AID) on arterials, using simulated data derived from loop detectors and probe vehicles. While automatic incident detection models have been under development for the past two decades, research has focussed on development and implementation of freeway automatic incident detection models. Development of arterial automatic incident detection models is far more complex due to a number of factors including the lack of conservation of flow, limited detection infrastructure and interruption of flow by traffic lights. There is general agreement in the literature, however, that the success of arterial incident detection models relies to a large extent on supplementing loop detector data with other sources of traffic information and on the development of advanced data fusion systems that can combine the different inputs and produce a reliable indication of the occurrence of an incident. This study extends previous research efforts and builds on the findings supporting the use of neural networks for data fusion. The study also addresses the limitations of previous research by developing and testing an advanced neural network data fusion system based on both loop detector and probe vehicle data, providing a comparison of model performance for various probe vehicle penetration rates.
Due to the difficulties in collection of data from real road networks, this study relied on data generated from a calibrated and validated traffic simulation model for a commuting corridor in Brisbane. Incidents were simulated and data was collected from loop detectors and probe vehicles at two locations on the network. Two detector configurations were tested, one making use of detectors located downstream of the intersection (Configuration 1), effectively treating the link as a freeway link, and the other conforming to the standard configuration on the road network (Configuration 2).
A set of 108 incidents was generated for the two detector configurations, with varying model features including incident location on link, road and detector configuration, incident duration and severity and prevailing traffic flow conditions, to ensure the model developed was as general as possible. Probe vehicle penetration rates of 5, 10 and 20 percent were considered.
A number of neural network architectures were developed for both Detector Configurations.Various types of Multi-Layer Feedforward (MLF) and Modular neural networks were tested. It was found that a Jordan/Elman MLF network, a modified MLF network providing data from previous time steps, performed best for both Detector Configurations.
From the literature review, it was expected that model performance would improve with inclusion of Flow/Occupancy ratio data, which it did. However, it was surprising that addition of Flow data further improved the model performance. The best performance obtained for Detector Configuration 1 was a detection rate of 59% for a false alarm rate of 0.5%. The data used consisted of Flow, Occupancy and Flow/Occupancy ratio from loop detector data and probe vehicle data for a penetration rate of 20%. Better performance was obtained for incidents occurring in the slow lane and for incidents occurring under high flow conditions.
The best performance obtained for Detector Configuration 2 was a detection rate of 86% for a false alarm rate of 0.36%. The data again consisted of Flow, Occupancy and Flow/Occupancy ratio from loop detector data and probe vehicle data for a penetration rate of 20%. However, there was very little difference in model performance, with detection rates of 86% also achieved for penetration rates of 0% to 10%). False alarm rates remained below 0.5%. For Detector Configuration 2, better performance was obtained for incidents occurring in the slow lane and for incidents occurring under low flow conditions. Incident detection performance was also strongly affected by incident severity. The better performance obtained for the Detector Configuration 2 data is an excellent outcome as most arterial road networks are already capable of providing data from detectors set up in this configuration which means that AID models can be developed for existing infrastructure. Another favourable result for the Detector Configuration 2 data is that while probe vehicle data improved model performance, good performance was still achieved without probe travel time data. It is interesting to note, however, that when Detector Configuration 2 data was augmented with speed data derived from loop detectors, performance improved further, with a detection rate of 90% achieved for a false alarm rate of 0.5%). It is possible that speed data from probe vehicles could have a similar impact but this would need to be explored in future research.
This research demonstrated the feasibility of developing a neural network model for detection of incidents on arterials using loop and probe vehicle data. Where it is not possible to collect probe vehicle data, satisfactory performance can be obtained using loop detector data alone. Data collected from the standard detector configuration (Detector Configuration 2) was most effective. If speed data is available at a location, it is highly recommended. Some recommendations for improvement and further study are also made.