Universal Background Subtraction Based on Arithmetic Distribution Neural Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Background subtraction is crucial for video analysis.
• Existing methods struggle with complex temporal pixel distributions.
• A novel approach is needed for robust background modeling.

Purpose of the Study:

• To propose a universal background subtraction framework using Arithmetic Distribution Neural Network (ADNN).
• To leverage arithmetic distribution operations for learning temporal pixel distributions.
• To enhance accuracy through an improved Bayesian refinement model.

Main Methods:

• Developed an ADNN model incorporating product and sum distribution layers.
• Utilized histograms as probability density functions in network layers.
• Integrated a GPU-accelerated Bayesian refinement model using neighboring information.

Main Results:

• The ADNN framework demonstrated superior performance on standard benchmarks.
• Achieved promising results with a simple architecture compared to CNNs.
• Outperformed state-of-the-art traditional and deep learning background subtraction methods.

Conclusions:

• The proposed ADNN framework offers an effective solution for universal background subtraction.
• This work pioneers the use of arithmetic distribution operations in network layers for background subtraction.
• The method provides a simple yet powerful architecture for learning pixel distributions.