Monitoring of Traffic Anomalies using Microscopic Traffic Variables in Vehicular Transportation Networks
Author(s)
Thajchayapong, Suttipong
Type
Thesis
Abstract
This thesis proposes methodologies to monitor traffic anomalies using microscopic
traffic variables measured by equipped vehicles sharing information
with one another and/or localized road-side infrastructure. The proposed
methodologies can identify not only traffic anomalies that lead to traffic
incidents, but also small transient deviations that are usually difficult to
detect.
Firstly, the thesis addresses the issue of anomaly detection where novel
supervised and unsupervised algorithms are proposed. The unsupervised
algorithm uses the change in variability of microscopic traffic variables to
detect traffic anomalies, which is also shown to outperform previous algorithms
monitoring ideally placed loop detectors. The supervised algorithm
can identify anomalies under different traffic regimes with 100% detection
rate and low false alarm rate when applied to real-world data, which presents
a signi cant improvement over the unsupervised algorithm. It is also shown
that the proposed algorithms can detect anomalies even when the microscopic
traffic variables are aggregated and missing.
Secondly, three classification algorithms are proposed, which can be integrated
with the previously proposed detection algorithms. The first algorithm
identifies a lane-blocking, which is a well-known type of anomaly that
often leads to traffic incidents, and is shown to outperform existing algorithms. The second algorithm identifies real-world cases of transient anomalies
as well as incident precursors by assessing spatial-temporal changes of
microscopic traffic variables. The third algorithm addresses the problem of
misclassifications under different traffic regimes by employing a certainty-based
decision function, and it is shown to successfully classify all anomaly
cases in the real-world data set.
Finally, the study is extended to the inference of traffic anomalies at a
location where traffic variables could not be measured directly. The key
contributions of the proposed algorithm are the ability to infer both normal
and anomalous traffic conditions at a target location by assessing only
microscopic traffic variables from adjacent locations, and the ability to estimate
lane-level traffic
flow, time occupancy and inter-arrival time. Based
on real-world data, it is shown that the proposed algorithm outperforms a
Kalman filter-based approach.
traffic variables measured by equipped vehicles sharing information
with one another and/or localized road-side infrastructure. The proposed
methodologies can identify not only traffic anomalies that lead to traffic
incidents, but also small transient deviations that are usually difficult to
detect.
Firstly, the thesis addresses the issue of anomaly detection where novel
supervised and unsupervised algorithms are proposed. The unsupervised
algorithm uses the change in variability of microscopic traffic variables to
detect traffic anomalies, which is also shown to outperform previous algorithms
monitoring ideally placed loop detectors. The supervised algorithm
can identify anomalies under different traffic regimes with 100% detection
rate and low false alarm rate when applied to real-world data, which presents
a signi cant improvement over the unsupervised algorithm. It is also shown
that the proposed algorithms can detect anomalies even when the microscopic
traffic variables are aggregated and missing.
Secondly, three classification algorithms are proposed, which can be integrated
with the previously proposed detection algorithms. The first algorithm
identifies a lane-blocking, which is a well-known type of anomaly that
often leads to traffic incidents, and is shown to outperform existing algorithms. The second algorithm identifies real-world cases of transient anomalies
as well as incident precursors by assessing spatial-temporal changes of
microscopic traffic variables. The third algorithm addresses the problem of
misclassifications under different traffic regimes by employing a certainty-based
decision function, and it is shown to successfully classify all anomaly
cases in the real-world data set.
Finally, the study is extended to the inference of traffic anomalies at a
location where traffic variables could not be measured directly. The key
contributions of the proposed algorithm are the ability to infer both normal
and anomalous traffic conditions at a target location by assessing only
microscopic traffic variables from adjacent locations, and the ability to estimate
lane-level traffic
flow, time occupancy and inter-arrival time. Based
on real-world data, it is shown that the proposed algorithm outperforms a
Kalman filter-based approach.
Date Issued
2010-12
Date Awarded
2011-01
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Barria, Javier
Creator
Thajchayapong, Suttipong
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)