This dissertation combines data-driven applications as well as microscopic traffic simulation experiments to develop and validate artificial intelligence powered algorithms for traffic signal location detection and control. Both aspects are motivated by the need to evaluate the potential of person-based traffic signal control policies at scale to incentivize ride-sharing and people’s mode shift to higher occupancy vehicles (HOVs). In today’s world, this is imperative as single occupancy vehicles are the most attractive transportation alternative for many commuters, which leads to increased traffic congestion, air pollution, and loss of productivity. The topic is addressed in two main phases.
Two large-scale regulator identification systems leveraging streaming raw naturalistic GPS data are engineered to derive an accurate representation of traffic regulators in digital maps. The two models, Intersense and Linksense, are based on the eXtreme Gradient Boosting (XGBoost) algorithm, combining infrastructure and vehicle telemetry data features to analyze movement patterns vehicles follow when approaching intersections.
Intersense aggregates features at the intersection level. For the three most common regulator types (traffic light, stop sign and right of way), the achieved accuracy surpasses 96% without imposition of thresholds for minimum trajectory count per intersection or application of minimum confidence level to the predictions. The system is transferable across diverse cities with performance evaluated in San Francisco and San Jose and adapts to GPS data sources of uneven penetration and sampling rates.
Linksense employs domain knowledge inspired rules to rethink the problem at the road segment resolution, being capable of identifying intersections with mixed control. Despite being trained on an imperfect ground truth dataset, Linksense achieves to correct outdated representations in existing digital maps. Validation of such instances is provided through the Google Maps Street View API which tracks historical images of intersections. Linksense incorporates a map simplification algorithm to allow detection of the potential locations of regulators from complex map representations.
In the final phase, we present HumanLight, a novel decentralized adaptive traffic signal control algorithm designed to optimize people throughput at intersections. The proposed controller is founded on reinforcement learning with the reward function incorporating the transportation-inspired concept of pressure at the person-level. HumanLight rewards with more green times vehicles carrying more people to incentivize people’s mode shift to HOV alternatives. The proposed algorithm showcases significant headroom and scalability in different network configurations considering multimodal vehicle splits at various scenarios of HOV adoption. HumanLight introduces the concept of active vehicles, loosely defined as vehicles in proximity to the intersection within the action interval window, to more robustly represent traffic information in the state and reward embeddings achieving superior performance compared to state-of-the-art controllers.
Improvements on person delays and queues reach over 40% and 45% respectively for the moderate and high HOV adoption scenarios considering a 26% and 41% shift away from single occupancy vehicles respectively. HumanLight allows policymakers and traffic engineers to regulate the aggressiveness in the prioritization of the HOV fleet. Via a modification in the state embedding, the travel time benefits for the different vehicles types are controlled, allowing optimal tuning with input from behavioral studies on the elasticity of mode demand. Other tunable parameters, such as the discount factor, determining the importance of future rewards, are evaluated in depth. Performance as well as the generated phase profile are considered to inform system operators regarding the number of expected phase changes and phase duration patterns, critical aspects to be accounted for in acyclic signal controllers as they affect pedestrian waiting times.
Overall, this dissertation paves the way for person-based traffic signal control policies to be evaluated and applied at city-level. HumanLight’s scalable, decentralized design can reshape the resolution of traffic management to be more human-centric and empower policies that incentivize ride-sharing and public transit systems.
