Generative AI for Trajectory Prediction

Motivation and Objectives

Generative AI is increasingly being applied across various fields, with image generation being a prominent area of focus. An intriguing application of this paradigm is its use in trajectory prediction, leveraging the advantages of image-based representations provided by Bird’s Eye View (BEV) scenarios. These representations are becoming increasingly prevalent as abstractions of sensor outputs. The goal of this project is to take an image representing the previous one-second trajectory as input (serving as a conditioning image) and generate the predicted trajectory for the next seconds as an image. The design and experiments are still in progress, and some preliminary results are presented here. This project is carried out in collaboration with colleagues from DLR.