Sequential Order-Aware Coding-Based Robust Subspace Clustering for Human Action Recognition in Untrimmed Videos

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human Action Recognition (HAR) is crucial for video analysis.
• Untrimmed videos require accurate temporal segmentation into action segments for HAR.
• Existing subspace clustering methods often overlook sequential frame information and noise effects.

Purpose of the Study:

• To propose a novel unsupervised temporal segmentation scheme for HAR.
• To address limitations in existing subspace clustering for action segmentation.
• To improve the accuracy of action segment identification in untrimmed videos.

Main Methods:

• Developed a sequential order-aware coding-based robust subspace clustering (SOAC-RSC) scheme.
• Utilized multi-layer neural networks to learn expressive code matrices from motion features.
• Proposed a robust cutting algorithm to handle noise in affinity graph segmentation.

Main Results:

• SOAC-RSC achieved state-of-the-art performance on Keck Gesture and Weizmann datasets.
• Demonstrated competitive performance on 6 additional public datasets (e.g., UCF101, URADL).
• The method accurately segments untrimmed videos into action segments.

Conclusions:

• The proposed SOAC-RSC scheme offers a robust and accurate solution for HAR.
• Sequential information and noise robustness are key improvements over existing methods.
• This approach advances unsupervised temporal segmentation for video analysis.