[Paper Review] Impact of Traffic Lights on Trajectory Forecasting of Human-driven Vehicles Near Signalized Intersections

Geunseob Oh, Huei Peng|arXiv (Cornell University)|Jun 2, 2019

Autonomous Vehicle Technology and Safety24 references4 citations

TL;DR

This paper proposes a novel trajectory forecasting framework that leverages vehicle-to-infrastructure (V2I) communications to access future traffic light (TL) states, enabling more accurate prediction of human-driven vehicle behavior near signalized intersections. By modeling driver responses to TL phases (green, yellow, red) and timing via deterministic and probabilistic Human Policy Models, the approach reduces trajectory forecasting errors by up to 150× for long-horizon predictions (e.g., 15s GYR scenario), especially in complex scenarios like the yellow-light dilemma zone.

ABSTRACT

Forecasting trajectories of human-driven vehicles is a crucial problem in autonomous driving. Trajectory forecasting in the urban area is particularly hard due to complex interactions with cars and pedestrians, and traffic lights (TLs). Unlike the former that has been widely studied, the impact of TLs on the trajectory prediction has been rarely discussed. In this work, we first identify the less studied, perhaps overlooked impact of TLs. Second, we present a novel resolution that is mindful of the impact, inspired by the fact that human drives differently depending on signal phase (green, yellow, red) and timing (elapsed time). Central to the proposed approach is Human Policy Models which model how drivers react to various states of TLs by mapping a sequence of states of vehicles and TLs to a subsequent action (acceleration) of the vehicle. We then combine the Human Policy Models with a known transition function (system dynamics) to conduct a sequential prediction; thus our approach is viewed as Behavior Cloning. One novelty of our approach is the use of vehicle-to-infrastructure communications to obtain the future states of TLs. We demonstrate the impact of TL and the proposed approach using an ablation study for longitudinal trajectory forecasting tasks on real-world driving data recorded near a signalized intersection. Finally, we propose probabilistic (generative) Human Policy Models which provide probabilistic contexts and capture competing policies, e.g., pass or stop in the yellow-light dilemma zone.

Motivation & Objective

• To identify the overlooked impact of traffic lights on trajectory forecasting accuracy in urban environments.
• To develop a behavior-cloning-based framework that models how human drivers respond to dynamic TL states (phase and timing).
• To enable long-term trajectory forecasting with improved accuracy by incorporating future TL state information via V2I communications.
• To address multi-modal behaviors such as the yellow-light dilemma (stop vs. go) using probabilistic policy modeling.
• To demonstrate the effectiveness of the approach through ablation studies on real-world driving data near signalized intersections.

Proposed method

• The method uses Vehicle-to-Infrastructure (V2I) communication to obtain future traffic light phase and timing information, which is integrated into the forecasting model.
• It introduces deterministic Human Policy Models that map sequences of vehicle and TL states to acceleration actions, modeling driver behavior as a function of TL state and elapsed time.
• A probabilistic extension of the Human Policy Model is proposed to capture competing policies (e.g., stop or proceed in yellow phase), enabling multi-modal trajectory forecasts.
• The framework combines the Human Policy Models with a known system dynamics transition function for sequential trajectory prediction, enabling both deterministic and generative forecasting.
• Monte Carlo sampling is used to estimate prediction uncertainties and generate diverse, realistic trajectory samples from the probabilistic model.
• The approach is evaluated using real-world data from a signalized intersection, with ablation studies comparing models with and without TL state awareness.

Experimental results

Research questions

• RQ1 How does the uncertainty in future traffic light states affect the accuracy of trajectory forecasting for human-driven vehicles near intersections?
• RQ2 To what extent can future traffic light phase and timing information improve long-term trajectory forecasting performance?
• RQ3 Can a behavior-cloning-based model effectively capture state-dependent driver actions at traffic signals, including transitions between green, yellow, and red phases?
• RQ4 How well can a probabilistic model represent competing driver policies, such as stopping or proceeding during the yellow-light dilemma zone?
• RQ5 What is the impact of incorporating V2I-derived future TL states on forecasting error across diverse intersection scenarios?

Key findings

• The proposed method reduces trajectory forecasting error (MAE, TWAE, ADN) by up to 150 times for long-horizon predictions (e.g., 15s GYR scenario) compared to models without TL state awareness.
• For 5-second prediction horizons, the method achieves less than 10 milliseconds inference time per most-probable trajectory on a standard laptop, demonstrating computational efficiency.
• The probabilistic model successfully captures the dual behavior in the yellow-light dilemma zone, forecasting both 'stop' and 'proceed' trajectories with appropriate probabilities, unlike deterministic models that fail in such cases.
• The ablation study confirms that access to future TL states significantly improves forecasting accuracy across all tested scenarios (G, R, GY, YR, RG, GYR), with the largest gains in long-term forecasts.
• Outliers in predictions are primarily attributed to edge cases (e.g., sudden braking) and competing policies, which the probabilistic model effectively captures through multi-modal outputs.
• The framework is generative and enables Monte Carlo sampling for uncertainty estimation, offering a practical tool for robust autonomous driving planning.