Learning Human Actions by Combining Global Dynamics and Local Appearance.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Human action recognition is crucial for surveillance, human-computer interaction, and robotics.
• Existing methods often struggle to effectively combine global temporal dynamics and local visual features for robust recognition.

Purpose of the Study:

• To develop a novel approach for human action recognition by integrating global temporal dynamics and local visual appearance.
• To improve the accuracy and robustness of action recognition systems on various datasets.

Main Methods:

• Modeled global temporal dynamics using robust linear dynamical systems (LDSs) and employed subspace angles for distance measurement.
• Extracted local visual features using histograms of oriented gradients (HOG) on curved spatio-temporal cuboids.
• Combined global and local features using maximum margin distance learning for classification.

Main Results:

• Achieved competitive performance compared to state-of-the-art methods on multiple short and long continuous action recognition datasets.
• Demonstrated the effectiveness of combining global dynamic and local visual features for improved action recognition accuracy.

Conclusions:

• The proposed method offers a robust and effective solution for human action recognition.
• The integration of LDS-based temporal modeling and HOG-based visual features provides a powerful framework for complex action understanding.