Data-Driven Monitoring of Operations and Safety at Signalized Intersections using Multi-Source Traffic Data

Abstract

Monitoring traffic operations and safety at signalized intersections is critical for transportation agencies to optimize signal timing, reduce congestion, and enhance roadway safety. Signalized intersections in the U.S. are typically equipped with advance and stop bar detectors, which vary in configuration, including single-channel and lane-by-lane detection setups. While queue length is a well-established measure for monitoring traffic signal performance and intersection operational efficiency, its estimation becomes challenging with single-channel detector configurations, which do not distinguish lane-specific vehicle arrivals. Furthermore, with the commonly deployed detection infrastructure at signalized intersections, tracking vehicles across the approach area remains difficult, limiting the availability of driver behavior data essential for monitoring both operations and safety. Additionally, accurate modeling of the dilemma zone while accounting for the complex interactions between driver behavior and vehicle dynamics is critical for safety monitoring and assessment at signalized intersections. To fill these challenges and gaps, this dissertation developed a three-component framework leveraging multi-source traffic data, such as high-resolution events and crowdsourced trajectories, to address these challenges with data-driven monitoring of operations and safety at signalized intersections. The first component estimated cycle-based maximum queue length using high-resolution event data from single-channel advance video detectors. The second component introduced a machine learning-based optimization framework for vehicle reidentification using non-visual detection data from loop detectors. The third component modeled the Type I dilemma zone using crowdsourced trajectory data and evaluated the accuracy of existing dilemma zone quantification methods. Queue length is one of the most important metrics required for the performance monitoring of signalized intersections. However, the current methodology of estimating queue length in the literature suffers from several drawbacks, including unstable estimation and the requirement of multiple data sources. Moreover, manual parameter calibration is required for single-channel advance detection, a common signal control detection configuration in many U.S. cities. To bridge these gaps, the first component of this dissertation proposed a cycle-based maximum queue length estimation method based on a) the empirical observation of breakpoints in the time gap between successive actuation and b) the identification of queue status for all detector actuation in a cycle. Maximum queue length for cycles with long queues was estimated based on the saturation flow rate and the trajectory of the last vehicle in the queue. The proposed methodology was implemented on two study intersections in Tucson, Arizona. Results showed that queue length can be estimated using the proposed method with mean absolute percentage errors of 14.77% and 15.1% and mean absolute errors of 25 ft and 42.5 ft. The results showed significant improvement in queue length estimation from single-channel detection data compared to similar methods in the current literature. The proposed method can help transportation agencies accurately estimate queue length at intersections with single-channel advance detection without the need for manual field data collection and without installing lane-by-lane detection. The advance and stop-bar detectors deployed at signalized intersections detect vehicles at discrete locations without linking or reidentifying them over the approach area. Accurate tracking and reidentification of vehicles between these detectors could provide valuable driver behavior data, especially during the safety-critical yellow onset periods. However, reidentifying vehicles using non-visual detection data is challenging and not well-explored, with existing analytical models relying on a priori-calibrated parameters. To this end, the second component proposed a machine learning (ML)-based reidentification framework for accurately tracking vehicles over the advance and stop bar loop detectors. The framework comprised two major components: advanced ML and deep learning (DL) models for accurately predicting the travel time between detectors and a novel optimization model that utilized these predicted travel time and actuation events for reidentifying vehicles. Tests carried out on a major intersection approach in Phoenix, Arizona showed that the optimization framework based on Neural Oblivious Decision Ensemble (NODE) reidentified vehicles even at congested conditions with 94.5% precision and 92.1% recall, outperforming state-of-the-art analytical, conventional ML, and comparable DL models. The low false alarm rate and high recall of this reidentification framework enable obtaining driver behavior data at the yellow onset to monitor traffic for analyzing stop/go behavior, dilemma zone entry/exit, red light running, and crossing conflicts at signalized intersections. The stop/go dilemma drivers face at the yellow onset is highly correlated with the potential risks of rear-end collisions and red-light running-related right-angle crashes at signalized intersections. This dilemma has been physically characterized using the Type I and Type II definitions. Unlike the Type II definition with several limitations, the Type I counterpart incorporates the dynamics of driver-vehicle attributes to quantify the dilemma zone accurately but requires high-quality vehicle trajectory data. Such trajectory data in existing studies are extracted from field-setup video cameras or radar, undergoing manual trajectory reduction and labor-intensive data processing challenges. Moreover, accurate modeling of the Type I dilemma zone dynamics and accuracy evaluation with the Type II methods remain major research gaps in the existing literature. The final component of this dissertation addressed these gaps and challenges by accurately quantifying the Type I dilemma zone using a large sample of crowdsourced vehicle trajectory data. Quantile regression was implemented to capture the dynamics of individual driver-vehicle attributes directly into the minimum stopping and the maximum clearing distances. Results across 15 intersection approaches consistently showed that the Type I dilemma zone is created if vehicles approach at a very high speed. Accuracy evaluation yielded low root mean squared errors of 14.8 ft and 25.1 ft in estimating the start and end of zone boundary, demonstrating the proposed method’s superiority over other dilemma zone quantification methods. Besides boundary comparison, driver behavior at the approach area was analyzed to understand potential rear-end and right-angle collision risks. This dissertation component advances the understanding of dilemma zone boundary dynamics among transportation researchers and practitioners and provides a sound empirical basis to support the development of efficient dilemma zone protection and signal timing strategies to improve intersection safety.