Self-Supervised Visual Representation Learning with Masked Autoencoders for Traversability Estimation in Outdoor Environments
9.03 MB
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Lataukset8
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
Autonomous robots operating in unstructured outdoor environments require reliable perception systems to assess traversability and navigate safely. However, gathering large amounts of labeled data to train these perception models is time-consuming and expensive. Self-supervised learning has recently emerged as a promising approach for leveraging large amounts of unlabeled data to learn useful visual representations.
In this thesis, we use Masked Autoencoders (MAE) for self-supervised visual pretraining to support traversability prediction. A domain-adaptive approach is applied, where a pretrained MAE model is further fine-tuned on unlabeled outdoor images to learn representations suitable for robotic navigation. The trained encoder is then adapted for downstream traversability tasks using both patch-level classification and pixel-level segmentation models. We also explore multi-scale feature aggregation and the fusion of LiDAR-derived depth data to improve the understanding of terrain geometry.
All experiments are carried out using the GOOSE dataset, which contains a variety of outdoor environments, including forests, grasslands, uneven terrain, and roads. The models are evaluated using standard metrics such as F1-score, precision, and recall. The results show that self-supervised pretraining improves performance when labeled data is limited and helps the model efficiently learn useful terrain features. At the same time, the experiments highlight a limitation of vision transformer-based encoders: patch-based tokenization can
make it more difficult to preserve fine spatial details.
Overall, the experiments demonstrate the potential of self-supervised visual representation learning for traversability estimation in robotics and provide insights into important design considerations when using transformer-based architectures for perception tasks.