Neural radiance fields assisted by image features for UAV scene reconstruction.

Area of Science:

• Computer Vision
• Photogrammetry
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV) applications require robust 3D scene reconstruction.
• Existing Neural Radiance Field (NeRF) methods struggle with large-scale, sparse, and complex UAV-captured scenes, leading to blurred details and incomplete reconstructions.
• Challenges include lack of edge information and objects appearing in few images.

Purpose of the Study:

• To develop an improved NeRF-based 3D scene reconstruction method for UAV imagery.
• To address limitations of current NeRF techniques in handling large-scale, complex aerial scenes.
• To enhance the accuracy and clarity of reconstructed UAV-captured environments.

Main Methods:

• Proposed a hybrid image encoder combining Convolutional Neural Networks (CNNs) and Transformers for feature extraction.
• Extended NeRF architecture with a branch to estimate uncertainty for transient regions, suppressing dynamic content.
• Refined the loss function to improve training optimization and synthesis quality.
• Utilized a custom UAV aerial imagery dataset for experiments.

Main Results:

• The proposed method effectively reconstructs and renders complex UAV-captured scenes with improved detail.
• Demonstrated enhanced accuracy in handling large-scale and sparsely viewed environments compared to baseline NeRF.
• Successfully suppressed dynamic elements to focus on static structure reconstruction.

Conclusions:

• The hybrid image encoder and uncertainty estimation significantly improve NeRF-based 3D scene reconstruction for UAVs.
• The refined approach offers a more robust solution for challenging aerial imagery datasets.
• This work advances the capabilities of vision-based 3D reconstruction in remote sensing and target detection applications.