DAN-SuperPoint: Self-Supervised Feature Point Detection Algorithm with Dual Attention Network.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Traditional feature point detection methods struggle in low-texture environments.
• Visual Odometry (VO) systems require robust feature extraction for accurate navigation.

Purpose of the Study:

• To develop a self-supervised feature extraction network for the VO front-end.
• To improve feature representation and detection accuracy in challenging visual conditions.

Main Methods:

• Utilized a feature pyramid structure for multi-scale feature fusion.
• Incorporated position and channel attention modules to enhance feature dependency.
• Employed confidence and tolerance loss terms to improve prediction accuracy and convergence speed.

Main Results:

• The proposed network demonstrated satisfactory performance on the Hpatches and KITTI datasets.
• The method effectively enhances feature representation for visual odometry tasks.
• Achieved reliable feature extraction in low-texture scenarios.

Conclusions:

• The developed self-supervised network offers a reliable solution for feature extraction in visual odometry.
• The attention mechanisms and novel loss functions contribute to improved performance.
• The network shows promise for real-world applications requiring robust visual navigation.