Headway Anomaly Detection for Lane-Based Queuing Status Classification Using Detector Presence Data: Comparative Study

Abstract

The identification of queuing status in traffic streams is a pivotal task with implications for various applications, such as level of service estimation and signal timing optimization. This study presents a comparative approach for binary classification of vehicles into queuing and non-queuing statuses using machine learning techniques based on vehicle headway anomalies. The proposed models leverage features from vehicle movement data at the stopline as input and queuing status as output. Four machine learning methods, the support vector machine, random forest, logistics regression, and Gaussian mixture model, are used and compared in the proposed headway anomaly detection framework. To validate the models, a simulation environment is constructed in VISSIM 4.3 and calibrated using real-world data. Results indicate that the random forest exhibits superior classification performance, showcasing its effectiveness as an ensemble approach. Notably, the resilience of these models is tested against the missing detection rate, with the random forest showing robust performance across different missing detection rate levels. Furthermore, feature importance analysis within the random forest model reveals “acceleration” and “headway” as significant predictors for classifying queuing status. The results advocate for the efficacy of the random forest model as a method for queuing status detection, indicating its utility for traffic analysis and the optimization of transportation network operations.