Feature Augmentation using Tubular Image Filtration for Autonomous On-board Classification of Mars Dust Devil Tracks

High resolution Mars orbital images, such as the HiRISE Imager on the Mars Reconnaissance Orbiter (MRO), are unable to downlink every image they take due to bandwidth constraints. To ensure efficient data collection, scientists decide which regions to image a priori, leaving many regions of Mars unobserved by high-resolution cameras. Regions of Mars that are not of immediate interest to scientists may be overlooked due to this hand-selection. We introduce a novel approach to high-resolution orbital imaging that leverages low-power machine learning algorithms to identify regions of interest to scientists using pixel-based image classification. Instead of identifying which regions of Mars to image, scientists would be able to identify which specific surface features to downlink. The onboard software would then perform classification on the surface as it images and only downlink images that contain the desired features.

To demonstrate this, regions of dense dust devil trac­­­ks were selected as a unique surface feature to classify. We use TextureCam, a low-power random forest-based classifier, for pixel classification. In order to increase the accuracy of pixel-based classification of dust devil regions on the Mars surface, we implemented a feature augmentation strategy to highlight regions of high-density dust devil tracks by outlining tubular objects. This filter based machine vision approach enables the use of less resource-intensive machine learning algorithms. We demonstrate a novel approach to image filtration using a modified Frangi extraction algorithm originally used to detect hairs of various thicknesses. This novel filtration system combined with TextureCam provides an accurate classifier with compute and power requirements suitable for onboard orbital image classification.