A multi-context fusion-aware graph modelling for group activity recognition using pose-conditioned spatial encoding...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Group activity recognition is challenging due to complex inter-player and scene dependencies.
• Traditional methods often overlook holistic scene context and relational information.

Purpose of the Study:

• Develop a robust model for group activity recognition.
• Integrate individual actor poses with scene context for enhanced relational reasoning.

Main Methods:

• Extracted pose features using mmPose and encoded scene context via pose-conditioned spatial feature aggregation.
• Constructed Actor Relation Graphs (ARGs) using Zero Normalized Cross Correlation (ZNCC) for robustness.
• Utilized Graph Convolutional Networks (GCNs) to model actor relationships and group activities.

Main Results:

• Achieved high classification accuracies of 95.02% on the Collective Activity dataset (CAD) and 94.81% on the Volleyball dataset (VD).
• Demonstrated efficient processing with an average time of 0.2 seconds per video clip (41 frames) on a TITAN-XP GPU.
• Showcased the effectiveness of combining pose and scene context features for graph-based relational learning.

Conclusions:

• The proposed framework significantly improves group activity recognition by explicitly combining pose-level and scene-level contextual features.
• This integrated approach offers a more holistic understanding compared to methods relying solely on appearance features.
• The model provides a robust and efficient solution for complex real-world activity recognition tasks.