Recognition of Rare Low-Moral Actions Using Depth Data.

Area of Science:

• Computer Vision
• Human Action Recognition
• Machine Learning

Background:

• Recognizing low-moral actions in public spaces is crucial but difficult.
• Low-moral actions are rare, posing challenges for machine learning models due to limited sample data.

Purpose of the Study:

• To introduce a novel dataset rich in low-moral behaviors for action recognition research.
• To evaluate the performance of classifiers using depth data and extracted skeletons on this dataset.

Main Methods:

• Developed a new dataset comprising a significant portion of low-moral human actions.
• Tested classifiers utilizing depth data and skeleton-based features.
• Applied transfer learning techniques to enhance classifier performance.

Main Results:

• Both depth and skeleton-based classifiers achieved comparable accuracy (Top-1: ~55%, Top-5: ~90%).
• Transfer learning significantly improved classification accuracy for both data types.
• The new dataset facilitated the study of action recognition with limited samples.

Conclusions:

• Depth and skeleton data are effective for recognizing low-moral actions, even with limited datasets.
• Transfer learning is a valuable technique for improving action recognition performance in low-data scenarios.
• The developed dataset serves as a valuable resource for future research in detecting subtle human behaviors.