Date of Award
Doctor of Philosophy (PhD)
Computational Modeling & Simulation Engineering
Modeling and Simulation
Traffic incident analysis is a crucial task in traffic management centers (TMCs) that typically manage many highways with limited staff and resources. An effective automatic incident analysis approach that can report abnormal events timely and accurately will benefit TMCs in optimizing the use of limited incident response and management resources. During the past decades, significant efforts have been made by researchers towards the development of data-driven approaches for incident analysis. Nevertheless, many developed approaches have shown limited success in the field. This is largely attributed to the long detection time (i.e., waiting for overwhelmed upstream detection stations; meanwhile, downstream stations show light traffic volume) and the concerns about the costly false alarms (e.g., dispatching response teams to non-incident cases). With the advancements in advanced machine learning algorithms and emerging data sources, there are opportunities to leverage such algorithms and a variety of data to significantly improve incident analysis practices.
As such, this dissertation first aims to develop an incident detection framework based on advanced machine learning algorithms that can leverage large-scale sensor data to enhance the predictive performance. Artificial neural network (ANN) is selected as a representative artificial intelligence (AI) module to predict incident occurrence based on lane-based data or station-level average loop detector data with occupancy, speed, and flow information. The memory unit and relevant knowledge database are integrated to refine the prediction result of AI module and provide the framework the ability to evolve and learn from historical prediction records. Compared with the benchmark approach California algorithm (CA#7), the proposed framework demonstrates its augmented prediction performance in terms of the shorter time to detection (TTD), lower false alarm rate (FAR), and higher detection rate (DR).
Secondly, we notice the existence of inaccurate labeled incident occurrence time and its impact on the incident detection framework. Therefore, we propose to utilize the unsupervised learning approach, fuzzy c-means (FCM) clustering, to relabel incident occurrence times and to further examine its impact on three different incident detection approaches (i.e., CA#7, ANN, and support vector machine (SVM)). In order to better automatically relabel three types of inaccurate mapping between reported incident occurrence times and loop detector measurements, Bayesian information criterion (BIC) values and additional restriction rules are applied. The evaluation results based on simulated incident scenarios demonstrate that the proposed relabeling strategy helps improve the performance of three traffic incident detection (TID) approaches in terms of a higher DR and a lower FAR.
Finally, we propose a data-driven analysis framework for identifying secondary incidents (SI). The proposed approach intends to leverage the untapped potential of ubiquitous probe vehicle data for SI identification. The developed framework consists of three major components: detecting the impact area of a primary incident (PI), estimating the boundary for the impact area, and identifying SIs within the boundary. The proposed framework has been tested based on probe data collected from different simulation models. The results show that the impact area induced by a PI can be well represented by the estimated boundary, especially by the genetic algorithm (GA-) and ant colony optimization (ACO-)based methods.
"Enhanced Traffic Incident Analysis with Advanced Machine Learning Algorithms"
(2020). Doctor of Philosophy (PhD), Dissertation, Computational Modeling & Simulation Engineering, Old Dominion University, DOI: 10.25777/ptja-qv26
Available for download on Friday, December 23, 2022