Multi-Person Pose Estimation Using an Orientation and Occlusion Aware Deep Learning Network.

Area of Science:

• Computer Vision
• Multimedia Analysis
• Human Behavior Understanding

Background:

• Skeleton estimation (pose estimation) is crucial for human behavior analysis in computer vision.
• Existing deep learning methods primarily focus on joint features, which are insufficient for multi-person or occluded scenarios.
• Accurate pose estimation requires integrating diverse human body information beyond just joint locations.

Purpose of the Study:

• To propose a novel multi-task framework for robust multi-person pose estimation.
• To enhance pose estimation by integrating joint features, body boundary, orientation, and occlusion information.
• To improve performance by organizing multi-task learning in a serial rather than parallel architecture.

Main Methods:

• Developed a multi-task framework based on Mask Region-based Convolutional Neural Networks (R-CNN).
• Integrated joint features, body boundary, body orientation, and occlusion conditions within the framework.
• Employed a serial multi-task model architecture for improved information integration.
• Augmented the Common Objects in Context (COCO) dataset with ground truths for body orientation and occlusion masks.

Main Results:

• Achieved 84.6% Percentage of Correct Keypoints (PCK) for multi-person pose estimation.
• Obtained an 83.7% Correct Detection Rate (CDR).
• Demonstrated superior performance in body orientation estimation compared to existing methods.
• Showcased a reduction in over-detection rates through comparative analysis.

Conclusions:

• The proposed serial multi-task framework effectively enhances multi-person pose estimation accuracy.
• Integrating diverse body information (joints, boundary, orientation, occlusion) is vital for challenging scenarios.
• The method offers a significant advancement in human activity understanding and computer vision applications.