Efficient Structure from Motion for Large-Size Videos from an Open Outdoor UAV Dataset.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Modern unmanned aerial vehicles (UAVs) generate high-resolution video data, challenging existing structure from motion (SfM) and simultaneous localization and mapping (SLAM) algorithms designed for smaller, lower-resolution scenes.
• Current SfM/SLAM methods face computational cost issues when processing large-scale, high-resolution UAV videos.

Purpose of the Study:

• To develop an efficient and accurate video-based SfM method tailored for high-resolution, large-size UAV videos.
• To address the computational challenges posed by processing extensive UAV video data for 3D reconstruction and localization.

Main Methods:

• Utilizes a visual SLAM (VSLAM) system for efficient keyframe and keypoint extraction from downsampled videos.
• Implements a novel two-step keypoint adjustment technique to refine existing keypoints at the original video scale.
• Employs rotation-averaging constrained global bundle adjustment (BA) for pose and structure refinement.

Main Results:

• Achieves an average efficiency improvement of 100% on a newly collected large-size dataset and 45% on the EuRoc dataset compared to other methods.
• Demonstrates superior localization accuracy against state-of-the-art SLAM and SfM techniques.
• Introduces a large-scale (3840 × 2160) outdoor video dataset with millimeter-accuracy ground control points.

Conclusions:

• The proposed SfM method offers significant efficiency and accuracy gains for processing high-resolution UAV video data.
• The new dataset provides valuable resources for advancing research in large-scale SLAM and SfM.
• This work contributes to more robust and scalable 3D reconstruction and localization from aerial imagery.