Automatic incident detection based on fundamental diagrams of traffic flow.

摘要

Freeway incident detection has been an active topic for both the research and practice for decades. As a part of the Advanced Traffic Management System (ATMS) under the Intelligent Transportation Systems (ITS) framework, Automatic Incident Detection (AID) algorithms have been serving as major incident detection tools for many years. They detect incidents based on traffic flow data collected by traffic surveillance systems. And they are fast, sensitive tools for the initial incident detection with only small software engineering cost added to existing surveillance systems. However, the effectiveness AID algorithms still cannot meet the requirements of Traffic Management Centers (TMCs). Advances in other competing technologies such as cell-phone call based detection, video based detection and freeway service patrol have reduced the dependence of TMCs on AID algorithms dramatically. Responses from AID researchers in the last decade focus on applying advanced learning methods from computer science and statistics. Some success has been reached. But in reality, the overfitting issue, the difficulties of understanding, implementing, calibrating and maintaining those algorithms still prevent them from being widely deployed at TMCs.;In this dissertation, a novel set of AID algorithms based on Fundamental Diagrams of Traffic Flow (FDs) is proposed. The research presented in this dissertation starts from a comprehensive review of existing AID algorithms and FDs. Then a novel methodology is developed to transfer the traditional traffic flow variables into more effective new incident detection features. There are two major parts in this methodology, coordinate transformation to generate incident detection features and the corresponding incident detection logics. The proposed incident detection algorithms (FD AIDs) are evaluated by comparing them with other existing AID algorithms popular in both the practice and research. The evaluation results show good potentials of FD AIDs to extend the detecting capability and increase the accuracy of existing AID algorithms.