Human motion similarity evaluation based on deep metric learning.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Evaluating human action similarity is challenging due to variations in camera viewpoints and individual skeleton structures.
• Existing methods often struggle to provide consistent similarity metrics across different observational conditions.

Purpose of the Study:

• To propose a novel method for human action similarity evaluation that is invariant to camera viewpoint and human skeleton differences.
• To achieve more accurate and reliable action similarity assessments, aligning better with human subjective perception.

Main Methods:

• A deep metric learning approach utilizing an autoencoder-decoder neural network trained on a synthetic dataset.
• Extraction of viewpoint- and skeleton-invariant action feature vectors from 2D human skeletal key point sequences.
• Application of a sliding window technique for refined feature extraction and Dynamic Time Warping (DWT) for temporal alignment.

Main Results:

• The proposed method successfully eliminates the impact of camera viewpoint and skeleton size variations on action similarity evaluation.
• Experimental results on a homemade dataset demonstrate improved accuracy and reliability in similarity assessments.
• The method generates results that are closer to human subjective perception of action similarity.

Conclusions:

• The deep metric learning method offers a robust solution for human action similarity evaluation under varying conditions.
• This approach enhances the objectivity and reliability of action recognition systems in real-world scenarios.
• The technique holds potential for applications requiring nuanced understanding of human motion dynamics.