Sparse Optical Flow Implementation Using a Neural Network for Low-Resolution Thermal Aerial Imaging.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Deep learning optical flow algorithms offer high accuracy but are computationally expensive.
• Current algorithms are often too resource-intensive for small Unmanned Aerial Vehicles (UAVs) due to size, weight, and power constraints.
• Dense optical flow fields contain redundant information for navigation tasks.

Purpose of the Study:

• To reduce the computational load of optical flow neural networks for UAVs.
• To enable real-time optical flow computation on resource-constrained aerial platforms.
• To maintain accuracy in optical flow estimation using thermal imagery.

Main Methods:

• Applied a feature extractor based on the Shi-Tomasi technique to thermal aerial images.
• Extracted relevant features to compute optical flow, reducing redundant data.
• Trained the RAFT-s model with both full images and the proposed feature-extracted input.

Main Results:

• Achieved a substantial increase in computational speed compared to using full images.
• Maintained the accuracy of the optical flow estimation, particularly in high thermal contrast scenarios.
• Demonstrated the feasibility of running advanced optical flow models on UAVs.

Conclusions:

• Feature extraction is an effective method to optimize deep learning optical flow for UAVs.
• The proposed approach balances computational efficiency and accuracy for real-time navigation.
• This technique facilitates the deployment of advanced computer vision capabilities on small aerial robots.