Date of Award

12-2013

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Mechanical Engineering

Advisor

Dr. Ardalan Vahidi

Committee Member

Dr. Mohammed Daqaq

Committee Member

Dr. Thomas Kurfess

Committee Member

Dr. Joshua Summers

Committee Member

Dr. Paul Venhovens

Abstract

This dissertation expounds on algorithms that can deterministically or proba-bilistically predict the future Signal Phase and Timing (SPAT) of a traffic signal by relying on real-time information from numerous vehicles and traffic infrastructure, historical data, and the computational power of a back-end computing cluster. When made available on an open server, predictive information about traffic signals’ states can be extremely valuable in enabling new fuel efficiency and safety functionalities in connected vehicles: Predictive Cruise Control (PCC) can use the predicted timing plan to calculate globally optimal velocity trajectories that reduce idling time at red signals and therefore improve fuel efficiency and reduce emissions. Advanced engine management strategies can shut down the engine in anticipation of a long idling interval at red. Intersection collision avoidance is another functionality that can benefit from the prediction.

We start by exploring a globally optimal velocity planning algorithm through the use of Dynamic Programming (DP), and provide to it three levels of traffic signal information - none, real-time only, and full-future information. The no-information case represents the average driver today, and is expected to provide an energy efficiency minimum or baseline. The full-information case represents a driver with full and exact knowledge of the future red and green times of all the traffic signals along their route, and is expected to provide an energy efficiency maximum. We propose a probabilistic method that seeks to optimize fuel efficiency when only real-time only information is available with the goal of obtaining fuel efficiency as close to the full-future knowledge example as possible. We used Monte-Carlo simulations to evaluate whether the fuel efficiency gains found were merely the result of lucky case studies or whether they were statistically significant; we found in related case studies that up to 16% gains in fuel economy were possible. While these results were promising, the delivery of relevant and accurate future traffic signal phase and timing information remained an unsolved problem.

The next step we took was towards building The next step we took was towards building traffic signal prediction models. We took several prescient techniques from the data mining and machine learning fields, and adapted them to our purposes in the exploration of massive amounts of data recorded from traffic Management Centers (TMCs). This manuscript evaluates Transition Probability Modeling, Decision Tree, Multi-Linear Regression, and Neural Network machine learning methods for use in the prediction of traffic Signal Phase and Timing (SPaT) information. signal prediction models. We took several prescient techniques from the data mining and machine learning fields, and adapted them to our purposes in the exploration of massive amounts of data recorded from traffic Management Centers (TMCs). This manuscript evaluates Transition Probability Modeling, Decision Tree, Multi-Linear Regression, and Neural Network machine learning methods for use in the prediction of traffic Signal Phase and Timing (SPaT) information.

Finally, we evaluated the influence of providing SPaT data to vehicles. To that end, we investigated both smartphone and in-vehicle proof-of-concepts. An in-vehicle velocity recommendation application has been tested in two cities: San Jose, California and San Francisco, California. The two test locations used two different data sources: data directly from a TMC, and data crowdsourced from public transit bus routes, respectively. A total of 14 test drivers were used to evaluate the effectiveness of the algorithm. In San Jose, the algorithm was found to produce a 8.4% improvement in fuel economy. In San Francisco, traffic conditions were not conducive to testing as the driver was unable to significantly vary his speed to follow the recommendation algorithm, and a negligible difference in fuel economy was observed. However, it did provide an opportunity to evaluate the quality of data coming from the crowdsourced data algorithms. Predicted phase timing com-pared to camera-recorded ground truth data indicated an RMS difference (error) in prediction of approximately 4.1 seconds.

Share

COinS