Histogram of Oriented Principal Components for Cross-View Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Current 3D action recognition methods struggle with viewpoint variations due to reliance on viewpoint-dependent depth image features.
• Processing raw 3D pointclouds directly offers a potential solution for achieving view-invariant action recognition.

Purpose of the Study:

• To develop novel, viewpoint-invariant descriptors for 3D action recognition from pointcloud data.
• To enhance robustness against noise, scale, and action speed variations in 3D human action recognition.

Main Methods:

• Proposed the Histogram of Oriented Principal Components (HOPC) descriptor, computed by projecting local pointcloud eigenvectors onto a dodecahedron.
• Introduced spatio-temporal keypoint (STK) detection using HOPC for localized feature extraction.
• Developed a global descriptor, STK-D, based on the 4D spatio-temporal distribution of STKs.

Main Results:

• HOPC and STK-D descriptors demonstrated robustness to viewpoint, scale, noise, and action speed variations.
• The proposed methods achieved significant performance improvements compared to nine existing techniques on both cross-view and single-view datasets.
• Effective detection and description of spatio-temporal keypoints for view-invariant recognition.

Conclusions:

• Directly processing 3D pointclouds with HOPC and STK-D offers a superior approach for cross-view action recognition.
• The developed techniques provide a robust and accurate solution for 3D action recognition challenges posed by viewpoint variability.
• Future work can explore further optimizations and applications of these novel descriptors in human-computer interaction and robotics.