Autonomous Driving on Mars: From Dataset to Models - A Deep Learning Application on Martian Imagery

The use of deep learning methods has become of paramount importance for autonomous vehicles on Earth. These systems, integral to modern navigation technologies, leverage large deep learning models that are trained on vast datasets of images. These models are capable of segmenting images, providing highly accurate information to path planning and decision-making algorithms. However, these advanced systems have not yet been fully utilized for autonomous driving in planetary rovers, which could significantly benefit from deep learning models compared to traditional machine vision systems. The primary objective of this work is to explore the limitations and constraints that rovers face in the context of autonomous navigation. This exploration will lead us to investigate the realm of semantic segmentation systems, specifically developed for the task of segmenting images of Martian surfaces. We will discuss and analyze two of the most prominent publicly available datasets in this field, which consist of real Martian terrain images. Furthermore, we will propose an improved version of these datasets to enhance the learning process of deep learning models in this domain. Our study will also delve into various deep learning models, examining supervised, semi-supervised, and unsupervised learning approaches. This comprehensive overview aims to establish a solid foundation for future research in this area. To this end, we will also introduce several novel models designed to take advantage of depth information, enhancing automatic semantic segmentation. In addition, we have explored the potential of applying a semi-supervised method, originally created for anomaly detection, to determine its feasibility in the context of Martian surface analysis. The entirety of this research, including datasets and code, has been made publicly available on GitHub, ensuring that our findings and contributions can serve as a valuable resource for the research community and facilitate further advancements in the field