Highly Accurate Visual Method of Mars Terrain Classification for Rovers Based on Novel Image Features.

Area of Science:

• Planetary Science
• Robotics
• Computer Vision

Background:

• Autonomous mobility for Mars exploration rovers requires accurate terrain classification to ensure safe traversal.
• Current visual terrain classification methods achieve less than 90% accuracy, posing risks of rover sinking or damage.
• Developing a high-accuracy vision-based system is essential for enhancing Mars rover operational safety and efficiency.

Purpose of the Study:

• To present a novel, high-accuracy vision-based method for classifying Mars terrain.
• To introduce new image features specifically designed for analyzing Martian terrain characteristics.
• To evaluate the performance of different machine learning classifiers for Mars terrain classification.

Main Methods:

• Proposed novel image features: multiscale gray gradient-grade, multiscale edges strength-grade, multiscale frequency-domain mean amplitude, multiscale spectrum symmetry, and multiscale spectrum amplitude-moment.
• Utilized three classifiers: K-nearest neighbor (KNN), support vector machine (SVM), and random forests (RF).
• Conducted experiments using the Mars image dataset MSLNet, collected by the Mars Science Laboratory (MSL) rover, with images at 256x256 resolution.

Main Results:

• The random forests (RF) classifier achieved the highest accuracy of 94.66% in classifying Mars terrain.
• The proposed novel image features demonstrated effectiveness in distinguishing between different Martian terrain types.
• Experimental results validated the superiority of the developed vision-based method over existing approaches.

Conclusions:

• The developed vision-based method significantly improves Mars terrain classification accuracy.
• The random forests classifier, combined with the proposed features, offers a robust solution for autonomous Mars rover navigation.
• This advancement contributes to safer and more efficient exploration of the Martian surface.