This research explores the use of deep learning for autonomous Martian rover navigation, focusing on the challenges and limitations of existing segmenta tion methods in the context of planetary surfaces. Given the success of deep learning in Earth-based autonomous vehicles, we investigate its potential to re place traditional machine vision in planetary exploration. Our study includes an extensive evaluation of two primary Martian terrain datasets [1] and [2], along with an improved dataset version designed to address class imbalance and enhance model learning. Wecompare several deep learning approaches—including supervised and semi supervised models—for their ability to accurately classify Martian surface fea tures, such as rocks and sand. To optimize feature extraction, we tested vari ous loss functions, including Cross-Entropy and Dice loss, demonstrating that Dice loss offers improved performance in handling class imbalances and spa tial information for segmentation tasks. Additionally, we experimented with an autoencoder-based anomaly detection model to capture unique terrain fea tures; however, results indicated the model’s limitations in the complex Mar tian environment. The study provides insights into the role of loss function choice, dataset qual ity, and architecture modifications for developing more effective segmentation models in planetary settings. All datasets, code, and findings are shared pub licly on GitHub [3] to support future advancements in autonomous navigation for planetary rovers, with implications for enhancing exploration missions on Mars and beyond.

Autonomous Driving on Mars: Dataset and Models for Martian Terrain Segmentation

COCCO, ALESSIO
2023/2024

Abstract

This research explores the use of deep learning for autonomous Martian rover navigation, focusing on the challenges and limitations of existing segmenta tion methods in the context of planetary surfaces. Given the success of deep learning in Earth-based autonomous vehicles, we investigate its potential to re place traditional machine vision in planetary exploration. Our study includes an extensive evaluation of two primary Martian terrain datasets [1] and [2], along with an improved dataset version designed to address class imbalance and enhance model learning. Wecompare several deep learning approaches—including supervised and semi supervised models—for their ability to accurately classify Martian surface fea tures, such as rocks and sand. To optimize feature extraction, we tested vari ous loss functions, including Cross-Entropy and Dice loss, demonstrating that Dice loss offers improved performance in handling class imbalances and spa tial information for segmentation tasks. Additionally, we experimented with an autoencoder-based anomaly detection model to capture unique terrain fea tures; however, results indicated the model’s limitations in the complex Mar tian environment. The study provides insights into the role of loss function choice, dataset qual ity, and architecture modifications for developing more effective segmentation models in planetary settings. All datasets, code, and findings are shared pub licly on GitHub [3] to support future advancements in autonomous navigation for planetary rovers, with implications for enhancing exploration missions on Mars and beyond.
2023
Autonomous Driving on Mars: Dataset and Models for Martian Terrain Segmentation
Autonomous Driving
Segmentation
Deep Learning
Mars
File in questo prodotto:
File Dimensione Formato  
Cocco_Alessio.pdf

accesso riservato

Dimensione 12.85 MB
Formato Adobe PDF
12.85 MB Adobe PDF

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/78067