Implementation and evaluation of a prediction algorithm for an autonomous vehicle

TL;DR

This paper develops and evaluates a high-frequency prediction algorithm for autonomous vehicles, comparing kinematic and dynamic bicycle models, with the dynamic model showing superior accuracy especially at higher speeds.

Contribution

The paper introduces a novel measurement procedure for cornering stiffness and integrates a dynamic bicycle model into an extended Kalman filter for improved vehicle trajectory prediction.

Findings

Dynamic model outperforms kinematic model by 82.6% in accuracy.

Positional deviation of only 1.25 cm per meter achieved.

Novel optical measurement method for cornering stiffness.

Abstract

This paper presents a prediction algorithm that estimates the vehicle trajectory every five milliseconds for an autonomous vehicle. A kinematic and a dynamic bicycle model are compared, with the dynamic model exhibiting superior accuracy at higher speeds. Vehicle parameters such as mass, center of gravity, moment of inertia, and cornering stiffness are determined experimentally. For cornering stiffness, a novel measurement procedure using optical position tracking is introduced. The model is incorporated into an extended Kalman filter and implemented in a ROS node in C++. The algorithm achieves a positional deviation of only 1.25 cm per meter over the entire test drive and is up to 82.6% more precise than the kinematic model.