Unsupervised Monocular Visual Odometry for Fast-Moving Scenes Based on Optical Flow Network with Feature Point...

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate visual odometry is crucial for robot navigation and pose estimation.
• Fast-moving scenes present challenges like image blur and disparity, degrading feature tracking stability.
• Existing methods struggle with robustness in dynamic environments.

Purpose of the Study:

• To develop an unsupervised monocular visual odometry framework.
• To enhance feature tracking robustness and accuracy, particularly in fast-moving scenarios.
• To fuse information from optical flow networks and traditional point feature extractors.

Main Methods:

• Proposed an unsupervised framework combining optical flow and traditional point feature extraction.
• Implemented a training process using FlannMatch for outlier filtering and a flow network with forward-backward consistency.
• Introduced the AvgFlow estimation module to select optimal matched point pairs based on scene motion.

Main Results:

• The trained optical flow network demonstrated superior robustness compared to SURF in complex, fast-motion scenarios.
• The proposed fusion approach and AvgFlow module effectively improved feature matching stability.
• Experiments on the KITTI Odometry dataset confirmed the effectiveness of the trajectory estimation, especially in challenging dynamic scenes.

Conclusions:

• The novel visual odometry framework provides robust and accurate pose estimation.
• The fusion of optical flow and traditional features significantly overcomes limitations in fast-moving environments.
• The approach offers a promising solution for real-world applications requiring reliable visual odometry.