Weighted Kernel Filter Based Anti-Air Object Tracking for Thermal Infrared Systems.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Surveillance Technology

Background:

• Traditional visual object tracking methods are optimized for RGB data, limiting their use in infrared (IR) domains.
• Existing IR tracking datasets and methods are insufficient for specialized applications like anti-air surveillance.
• There is a need for robust IR-based trackers and dedicated datasets for anti-air surveillance systems.

Purpose of the Study:

• To develop a reliable IR-based visual object tracking framework for anti-air surveillance.
• To create a novel dataset of anti-air thermal infrared (TIR) images for training and evaluation.
• To address the limitations of RGB-optimized trackers in IR surveillance applications.

Main Methods:

• Collected anti-air TIR images from an electro-optical surveillance system to construct a new dataset.
• Proposed an end-to-end convolutional neural network framework utilizing a Siamese network for feature extraction.
• Implemented a detection-by-tracking approach with continuously updated kernel filters for robust target tracking.

Main Results:

• The proposed framework demonstrated robust learning of target structural information in the IR domain.
• Kernel filters effectively enabled robust tracking of anti-air targets like UAVs and drones.
• Experimental results on the new dataset showed outstanding performance with real-time processing at 40 FPS.

Conclusions:

• The developed IR tracking framework and dataset significantly advance anti-air surveillance capabilities.
• The method offers a robust and efficient solution for tracking small, fast-moving targets in challenging IR conditions.
• This work provides a valuable resource for future research in IR-based visual object tracking for defense applications.