Moving object detection based on shape prediction

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Modeling

Background:

• Statistical analysis is common for moving object detection, typically using recent segmentations.
• Recent segmentations are often suboptimal for detecting deformable objects like pedestrians.

Purpose of the Study:

• To develop an object detection framework that selects the most suitable historical segmentations for statistical analysis.
• To improve the detection accuracy of deformable objects by utilizing learned shape evolution.

Main Methods:

• An autoregressive model is employed to learn the shape evolution of deformable objects from historical segmentations.
• The framework predicts the object's shape in the current frame based on the learned model.
• Historical segmentations with shapes similar to the predicted shape are selected for analysis.

Main Results:

• The proposed method outperforms traditional approaches that rely solely on recent segmentations.
• Experiments demonstrate significant performance improvements in detecting deformable objects.

Conclusions:

• Selecting optimal historical segmentations based on learned shape evolution enhances deformable object detection.
• The developed framework offers a more robust statistical analysis for moving object detection.