Semantic Segmentation on Martian Terrain for Navigation Using Transformers

NASA has been at the forefront of extraterrestrial exploration, which has been the primary focus of human science research, for decades. Machine learning has been a major factor in enabling autonomy, increasing the cost-effectiveness and decreasing the duration of space missions. We focus on the topography of Mars for navigation of rovers.
AI4Mars Dataset includes terrain images from Curiosity, Opportunity and Spirit(MER) rovers. This project model recognises landscape features - Sand, Soil, Rock, Rover, Horizon and predicts the same.

In this research, a novel approach has been presented for multi-label semantic segmentation of the martian terrain for navigation using variants of a transformer model - SegFormer, which has resulted in a commendable accuracy of 90.86% and mIoU of 83.55% on the AI4Mars dataset.

Raw Dataset
Preprocessed Annotated Dataset

  Fig.1 - EDA - K-Means Clustering, Convex Hull

  

Fig.2 - Rover Detection

  

Fig.3 - Masking

SegFormer B2 outperformed B0 and B1. Per Class metrics are mentioned below:

Model Output

Purple - Rock, Yellow - Soil, Brown - Sand

☝️ Most of the files are large, hence cannot be viewed on github. To view .ipynb notebooks, clone this repository.
☝️ List of helpful links and articles in references.txt

Contributors

• Ankitha C
• Gagan Goutham
• Hita Juneja