Title:	Trajectory Prediction Model Dataset Generalization
DNr:	Berzelius-2025-310
Project Type:	LiU Berzelius
Principal Investigator:	Erik Frisk <erik.frisk@liu.se>
Affiliation:	Linköpings universitet
Duration:	2025-10-01 – 2026-04-01
Classification:	20208
Homepage:	https://liu.se/medarbetare/erifr93
Keywords:

Abstract

This project investigates the transferability of trajectory prediction models across diverse real-world datasets in autonomous driving.Leveraging deep latent scene embeddings, the research seeks to quantify dataset similarity and inform the selection of effective pretraining sources, thereby enhancing domain adaptation and generalization. The approach trains a unified embedding architecture on heterogeneous datasets, evaluates similarity using metrics such as KL divergence, and validates transferability through cross-dataset trajectory forecasting experiments.