Efficient Non-Consecutive Feature Tracking for Robust Structure-From-Motion.

Area of Science:

• Computer Vision
• Photogrammetry
• Robotics

Background:

• Structure-from-motion (SfM) relies heavily on accurate feature tracking.
• Disjointed feature tracks due to occlusions or noise degrade SfM performance, especially in large-scale scenes requiring multiple sequences.
• Existing methods struggle with interrupted tracks across different subsequences or videos.

Purpose of the Study:

• To develop an efficient non-consecutive feature tracking framework for SfM.
• To address the challenge of matching interrupted feature tracks in large-scale and multi-sequence environments.
• To improve the robustness and accuracy of SfM for complex datasets.

Main Methods:

• Proposed a framework for solving feature "dropout" issues caused by noise or distortion.
• Developed methods for rapid recognition and joining of common features across different subsequences.
• Introduced a segment-based coarse-to-fine SfM algorithm for handling large datasets.

Main Results:

• The proposed framework effectively matches interrupted tracks across different subsequences and videos.
• The system demonstrates robust performance on challenging video data with large-scale scenes.
• The segment-based SfM algorithm efficiently processes large datasets.

Conclusions:

• The non-consecutive feature tracking framework significantly enhances SfM robustness.
• The approach provides an effective solution for large-scale scene reconstruction using multiple image sequences.
• The system offers improved performance for SfM applications in challenging visual conditions.