Reduced Image Classes in Modified U-Net for Mars Rover Navigation

Abstract

Rover navigation currently relies on algorithms to automatically determine their path. This is because the distance from Earth to Mars means that real time communication is impossible. Current navigation algorithms have difficulties in identifying terrain, causing problems such as becoming stuck in soft terrain. Furthermore, the available computation power and memory are limited on a rover. This paper presents both a modified U-Net model to identify parts of the terrain and combining multiple classes to have less output classes. The proposed method was to combine classes like soil and bedrock into more generalized classes, like traversable and untraversable, to reduce memory usage and needed computational power. Combining the classes shows that a model can be trained faster, and in some cases even improve. Testing this method on low resolution images has shown improved results in testing. After training, a three-class model is able to yield a higher mIoU of 0.4583 on a test set compared to the full five-class model, which achieved 0.3451. This method is non-specific to U-Net and can be applied to many different models. Combining this method with other models and larger datasets during training could be an option of improving the accuracy of models running on less processing power, allowing for use on platforms such as Mars rovers.