MonoSeg: An Infrared UAV Perspective Vehicle Instance Segmentation Model with Strong Adaptability and Integrity.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Unmanned aerial vehicle (UAV)-based infrared vehicle detection faces challenges from viewpoint variations, platform instability, low contrast, and thermal crossover effects.
• Edge computing platforms require efficient algorithms to balance real-time processing with detection performance.

Purpose of the Study:

• To present MonoSeg, a novel instance segmentation framework specifically designed for UAV perspective infrared vehicle detection.
• To improve detection accuracy and computational efficiency in challenging infrared imaging conditions.

Main Methods:

• Developed the Ghost Feature Bottle Cross (GFBC) module to enhance backbone feature extraction efficiency and reduce computational overhead.
• Introduced the Scale Feature Recombination (SFR) module for adaptive multi-scale feature fusion in the Neck stage.
• Implemented a Comprehensive Loss function for precise instance boundary delineation.

Main Results:

• MonoSeg achieved state-of-the-art performance on benchmark datasets, evidenced by high Box mAP and Mask mAP scores.
• The framework demonstrated substantially lower computational requirements compared to existing methods.
• Successfully addressed challenges posed by dynamic viewpoints and infrared imaging limitations.

Conclusions:

• MonoSeg offers a significant advancement in UAV-based infrared vehicle detection, providing high accuracy and efficiency.
• The proposed modules and loss function effectively enhance feature extraction, fusion, and boundary delineation.
• MonoSeg presents a viable solution for real-time vehicle detection in complex environments using edge computing platforms.