Modeling Geometric-Temporal Context With Directional Pyramid Co-Occurrence for Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visual action recognition is crucial for understanding human activities in videos.
• Existing methods like bag-of-visual-words (BOVW) have limitations in capturing complex spatio-temporal information.

Purpose of the Study:

• To develop a new geometric-temporal representation for enhanced visual action recognition.
• To improve the descriptive power and accuracy of action recognition models.

Main Methods:

• Proposed a modified covariance descriptor using the log-Euclidean Riemannian metric for spatio-temporal cuboids.
• Introduced a directional pyramid co-occurrence matrix (DPCM) to capture geometric-temporal context and feature distribution.
• Developed a DPCM matching kernel for video similarity measurement.

Main Results:

• Achieved state-of-the-art performance on six public datasets.
• Significantly outperformed BOVW models, e.g., 98.78% accuracy on KTH dataset vs. 88.06% for BOVW.
• Reached 100% accuracy on Weizmann and UCF CIL datasets.

Conclusions:

• The proposed geometric-temporal representation and DPCM method offer superior performance for visual action recognition.
• The method effectively captures spatio-temporal relationships, leading to substantial accuracy improvements.