3DMesh-GAR: 3D Human Body Mesh-Based Method for Group Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Group activity recognition is crucial for applications like crowd monitoring and video surveillance.
• Existing skeleton-based methods using graph convolutional networks face limitations in scalability, robustness, and interoperability.
• Accurate recognition requires understanding individual actions within collective group activities.

Purpose of the Study:

• To propose a novel 3D human body mesh-based approach for group activity recognition (3DMesh-GAR).
• To overcome the limitations of traditional skeleton-based methods.
• To develop a robust and scalable solution for complex, multi-person scenes.

Main Methods:

• Utilizes a 3D mesh creation method relying on body center heatmap, camera map, and mesh parameter map.
• Avoids complex and noisy 3D skeletons, offering a conceptually simple, single-stage, and bounding box-free approach.
• Effectively handles highly occluded and multi-person scenarios without increased computational cost.

Main Results:

• 3DMesh-GAR achieves state-of-the-art performance on the Collective Activity Dataset.
• The mesh-based approach demonstrates superior scalability and robustness compared to prior methods.
• The method successfully recognizes collective activities by analyzing individual actions in context.

Conclusions:

• 3DMesh-GAR presents a significant advancement in 3D human body mesh-based group activity recognition.
• The proposed method offers a more scalable, robust, and interoperable solution for video understanding tasks.
• This approach effectively addresses challenges in recognizing group actions in complex real-world scenarios.