Learning Temporal Distribution and Spatial Correlation Toward Universal Moving Object Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Video Analysis

Background:

• Moving object segmentation aims to differentiate moving objects from static backgrounds in videos.
• Existing methods often struggle with universality across diverse natural scenes.
• Developing a general solution for varied environmental conditions remains a significant challenge.

Purpose of the Study:

• To propose a universal model for moving object segmentation applicable to diverse natural scenes.
• To address the limitations of scene-specific methods in video analysis.
• To enhance the robustness and accuracy of moving object detection.

Main Methods:

• Introduced the Learning Temporal Distribution and Spatial Correlation (LTS) method.
• Developed the Defect Iterative Distribution Learning (DIDL) network for scene-independent temporal distribution learning, incorporating an improved product distribution layer.
• Proposed the Stochastic Bayesian Refinement (SBR) Network to learn spatial correlation and refine segmentation masks.

Main Results:

• The LTS approach demonstrated scene-independent segmentation capabilities.
• The method achieved improved accuracy through spatial correlation learning.
• Experiments on multiple datasets (LASIESTA, CDNet2014, BMC, SBMI2015) and real-world videos confirmed superior performance against state-of-the-art methods.

Conclusions:

• The proposed LTS method shows high potential as a general solution for moving object segmentation in real-world environments.
• The combination of temporal distribution learning and spatial correlation refinement offers a robust approach for diverse video conditions.
• The method achieves strong performance with fixed parameters across various complex natural scenes.