Fusion Attention Mechanism for Foreground Detection Based on Multiscale U-Net Architecture

Area of Science:

• Computer Vision
• Deep Learning
• Video Processing

Background:

• Foreground detection is crucial for video surveillance and computer vision.
• Traditional methods struggle with complex, dynamic real-world scenes.
• Accurate foreground target extraction remains a challenge.

Purpose of the Study:

• To propose an advanced foreground detection method using deep learning.
• To enhance foreground object identification in complex video environments.
• To improve upon the limitations of traditional unsupervised approaches.

Main Methods:

• Developed a foreground detection model based on a multiscale U-Net architecture.
• Integrated a fusion attention mechanism via skip connections within the U-Net.
• Utilized spatial information from single RGB images and fused spatiotemporal information from multiple images.

Main Results:

• Achieved an F-measure of 0.9785 using single RGB images (spatial information).
• Reached an F-measure of 0.9830 by fusing multiple images (spatiotemporal information).
• Demonstrated superior performance, particularly in the Low Framerate category, exceeding state-of-the-art methods.

Conclusions:

• The proposed multiscale U-Net with a fusion attention mechanism effectively detects foreground objects.
• The method shows significant improvements in accuracy and robustness for foreground detection.
• The approach offers a superior alternative to existing methods, especially in challenging scenarios.