Road Barlow Twins: Redundancy Reduction for Motion Prediction

Royden Wagner, Marvin Klemp, Carlos Fernandez Lopez, Omer Sahin Tas

Keywords: Self-supervised, pre-training, motion prediction, self-driving

TL;DR: Self-supervised pre-training for motion prediction in self-driving applications

Abstract:

Anticipating the future trajectories of other traffic agents, i.e., motion prediction, is crucial for self-driving vehicles to operate safely in dynamic environments.
In this work, we introduce a novel self-supervised pre-training method for motion prediction.
Our method is based on Barlow Twins and applies the redundancy reduction principle to embeddings generated from HD maps.
Through our method, deep learning models learn augmentation-invariant features of HD maps.
We hypothesize that an understanding of the environment can be learned faster using these features.
We pre-train several large transformer models and subsequently fine-tune them on motion prediction.
Our experiments reveal that the proposed pre-training method can improve mADE and mFDE by 12\% and 15\% and outperform contrastive learning with PreTraM and SimCLR in a semi-supervised setting.
Code: https://github.com/KIT-MRT/road-barlow-twins