Sky and Ground Segmentation in the Navigation Visions of the Planetary Rovers

Area of Science:

• Computer Vision
• Robotics
• Remote Sensing

Background:

• Sky and ground segmentation are critical for autonomous systems like rovers.
• Existing methods often require extensive manual annotation, limiting practical application.

Purpose of the Study:

• To develop an efficient sky and ground segmentation framework for rover navigation.
• To leverage weak supervision and transfer learning for improved performance with limited data.

Main Methods:

• Proposed a novel neural network architecture: Network in U-shaped Network (NI-U-Net).
• Introduced a conservative annotation method to minimize manual intervention.
• Utilized transfer learning and weak supervision techniques.

Main Results:

• Achieved state-of-the-art results on the Skyfinder dataset, with high accuracy (99.232%) and IoU (98.223%).
• The NI-U-Net model operates at 40 frames per second, ensuring real-time processing.
• Demonstrated superior performance with limited manual annotation.

Conclusions:

• The proposed framework effectively bridges the gap between lab research and real-world rover applications.
• Provides essential semantic information for robust rover navigation.
• The NI-U-Net and conservative annotation method offer a practical solution for sky and ground segmentation.