Rock Hunting With Martian Machine Vision
This work addresses the challenge of low-power, customized training for computer vision on Mars rovers, which is incremental as it builds on existing deep learning methods for space applications.
The study tackled the problem of onboard object recognition for Mars rovers by developing deep learning methods for classifying and detecting Martian rocks, achieving over 97% accuracy for binary rock vs. rover classification and enabling geo-located bounding boxes with rock counting, but with a trade-off in low-power implementations showing 37% accuracy at 1 frame per second.
The Mars Perseverance rover applies computer vision for navigation and hazard avoidance. The challenge to do onboard object recognition highlights the need for low-power, customized training, often including low-contrast backgrounds. We investigate deep learning methods for the classification and detection of Martian rocks. We report greater than 97% accuracy for binary classifications (rock vs. rover). We fine-tune a detector to render geo-located bounding boxes while counting rocks. For these models to run on microcontrollers, we shrink and quantize the neural networks' weights and demonstrate a low-power rock hunter with faster frame rates (1 frame per second) but lower accuracy (37%).