Conditional Unscented Autoencoders for Trajectory Prediction

• DOI: 10.48550/arxiv.2310.19944
• Publication Date: 2023-01-01
• Authors: Faris Janjoš; Marcel Hallgarten; Anthony Knittel; Maxim Dolgov; Andreas Zell; J. Marius Zöllner
• Keywords: Leverage (statistics); Code (set theory)

The CVAE is one of the most widely-used models in trajectory prediction for AD. It captures interplay between a driving context and its ground-truth future into probabilistic latent space uses it to produce predictions. In this paper, we challenge key components CVAE. We leverage recent advances VAE, foundation CVAE, which show that simple change sampling procedure can greatly benefit performance. find unscented sampling, draws samples from any learned distribution deterministic manner, naturally be better suited than potentially dangerous random sampling. go further offer additional improvements including more structured Gaussian mixture space, as well novel, expressive way do inference with CVAEs. wide applicability our by evaluating them on INTERACTION dataset, outperforming state art, at task image modeling CelebA baseline vanilla Code available https://github.com/boschresearch/cuae-prediction.