Fischer, Johannes (Karlsruhe Institute of Technology), Stiller, Christoph (Karlsruhe Institute of Technology)

Data-Driven Online Estimation of Driver Model Parameters for Vehicle Trajectory Prediction

Scheduled for presentation during the Poster Session "Operation and navigation of automated vehicles" (FrAT16), Friday, September 27, 2024, 10:30−12:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 14, 2024

Abstract

Fast and accurate trajectory prediction is crucial for the development and validation of automated driving systems. Using driver models for prediction is a promising approach to achieve this since they can produce realistic driving behavior at low computational cost. To produce high-quality predictions, the driver model parameters need to be adapted to the current traffic situation and observed driving behavior online. Our work combines data-driven methods with driver models to obtain realistic short-term trajectory predictions. We propose to train machine learning models to predict the driver model parameters that best capture the observed behavior of other vehicles. We use attention-based architectures to process sequential input data and predict the driver model parameters as a weighted sum of prototypes, thus ensuring that the predicted driving model parameters are realistic. Compared to particle filter-based state-of-the-art methods, our approach profits from the rich representational capabilities of learned models and the high online runtime efficiency of driver models. We show that our approach outperforms state-of-the-art methods for online driver model parameter estimation on a real-world traffic dataset.