Quantified Evaluation on Automatic Freeway Bottleneck Identification Algorithms Considering Data Quality Issues of Loop Detectors

摘要

Computer algorithms used to identify recurrent bottlenecks have been studied since the wide deployment of loopdetecting systems in the US. Such algorithms automatically analyze the archived loop detector data, and identifypotential recurrent bottleneck and their characteristics such as activation rate, location, time of day, for furtherinvestigation. In a highway congestion mitigation project, such automatic algorithms can save a lot of resources andlabor hours for initial screening of bottlenecks on a large freeway network, when compared with traditional drivingand traffic video inspection methods. Representative bottleneck identification algorithms include Chen's, Ban's andJin's algorithms. From practitioners' point of view, it is necessary to understand the strengths and conditions of thesealgorithms against real-world loop detector data with quality issues. However, the lack of effective evaluationmethod makes such analysis difficult. In this paper, a new evaluation method is proposed based on a novel set ofindexes defined on the spatial-temporal diagram. With these indexes, Chen's, Ban's and Jin's algorithms arecalibrated and evaluated using field data for two freeway corridors (US 12/14 and I-894) in the State of Wisconsin,USA. Ground truth data for this study come from the manual inspection of 287,055 traffic video snapshots. The 5-min loop detector data are used as inputs to all three algorithms. The evaluation results show different characteristicsof tested algorithms against noisy loop data. Chen's method cannot handle detector data noises effectively, whileBan's method and Jin's method have some capabilities to reduce the impact of some data quality issues of loopdetectors. In addition, a simple speed cleaning method is proposed for Chen's and Ban's algorithms and itseffectiveness is also evaluated in this study. The paper is concluded by summarizing the characteristics of eachalgorithm, limitations of this evaluation study and future work.