CONet: Crowd and occlusion-aware network for occluded human pose estimation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human pose estimation is vital for applications like virtual reality and human-computer interaction.
• Current top-down methods struggle with multi-person pose estimation in crowded and occluded scenarios.
• Existing approaches often fail to predict poses for multiple individuals within a single bounding box.

Purpose of the Study:

• To develop a novel approach for accurate multi-person pose estimation in challenging crowded and occluded scenes.
• To address the limitations of current top-down methods in handling complex human interactions.
• To introduce a robust network capable of differentiating and estimating poses of occluding and occluded individuals.

Main Methods:

• Proposed a Crowd and Occlusion-aware Network (CONet) employing a divide-and-conquer strategy.
• Introduced a Crowd and Occlusion-aware Head (COHead) with two branches for estimating occluder and occluded poses.
• Utilized an attention mechanism for differentiated learning and proposed an interference point loss for enhanced anti-interference capabilities.

Main Results:

• CONet achieved state-of-the-art performance on the CrowdPose dataset, reaching 71.6 AP.
• Outperformed the previous state-of-the-art model by +1.6 AP.
• Demonstrated significant improvements in human pose estimation accuracy within crowded and occluded environments.

Conclusions:

• CONet offers a simple yet effective solution for multi-person pose estimation in complex scenes.
• The model's ability to handle occlusion and crowding enhances its applicability in real-world scenarios.
• Achieved state-of-the-art results highlight the potential of CONet for surveillance, sports analysis, and HCI.