Registering Unmanned Aerial Vehicle Videos in the Long Term

Area of Science:

• Computer Vision
• Robotics
• Image Processing

Background:

• Unstable viewpoints from unmanned aerial vehicles (UAVs) hinder automatic video processing.
• Traditional methods like registration and stabilization suffer from jitter and drifting errors.
• Existing motion estimators have limitations, especially for airborne video data.

Purpose of the Study:

• To develop an improved video registration technique for UAVs.
• To address the limitations of 2D-rigid transforms in airborne video analysis.
• To create a real-time, jitter-free video registration solution for UAVs.

Main Methods:

• Extending prior work on motion estimator modeling for UAV video.
• Adapting the model to utilize perspective transforms for enhanced accuracy.
• Developing a lightweight implementation for real-time processing.
• Evaluating the approach with long-duration traffic surveillance videos.

Main Results:

• The proposed hybrid solution effectively mitigates jitter and drifting errors.
• Perspective transforms provide significantly higher accuracy for UAV videos compared to 2D-rigid transforms.
• The lightweight implementation enables real-time automatic registration of stationary UAV videos.
• The approach demonstrates potential when integrated with background subtraction tasks.

Conclusions:

• The refined theoretical framework and practical adaptation to perspective transforms offer a robust solution for UAV video registration.
• The developed method achieves jitter-free registration in real-time.
• This technique is particularly beneficial for applications like traffic surveillance using UAVs.