Interaction-Based Human Activity Comparison.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Robotics

Background:

• Traditional motion comparison methods focus on individual character features, neglecting the semantic importance of interactions.
• Existing interaction-based features struggle with non-fixed topologies and high dimensionality, limiting their application in activity comparison.
• Understanding human activities often requires analyzing the interplay between multiple entities.

Purpose of the Study:

• To propose a unified framework for activity comparison centered on interactions.
• To develop a novel metric for evaluating interaction similarity using spatial-temporal data.
• To enable semantic comparison across diverse interaction types and improve activity recognition and retrieval.

Main Methods:

• Developed a unified framework for activity comparison from an interaction perspective.
• Adapted the Earth Mover's Distance (EMD) to a customized geometric mesh structure for spatial-temporal interaction representation.
• Created and utilized five diverse interaction databases (two-character and character-object, synthetic and real-world).

Main Results:

• The proposed metric effectively captures the semantic meaning of interactions.
• The framework demonstrates superior performance in interaction retrieval tasks compared to existing methods.
• The approach successfully aligns with human perception of interaction similarity.

Conclusions:

• The developed metric provides a robust method for comparing and understanding human activities based on interactions.
• The framework offers significant improvements for unsupervised activity detection and interaction retrieval systems.
• The publicly available databases facilitate further research in human activity analysis.