Multipe/single-view human action recognition via part-induced multitask structural learning.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Human action recognition is crucial for understanding human behavior in various applications.
• Existing methods often struggle with multi-view data and generalizing across different action categories.
• The need for robust and accurate action recognition frameworks that leverage body structure cues is evident.

Purpose of the Study:

• To propose a unified framework for both single-view and multiple-view human action recognition.
• To introduce a novel hierarchical partwise bag-of-words representation for encoding visual saliency.
• To formulate action recognition as a part-regularized multitask structural learning (MTSL) problem.

Main Methods:

• Developed a hierarchical partwise bag-of-words representation based on body structure for visual saliency.
• Formulated human action recognition as a part-regularized multitask structural learning (MTSL) problem.
• Introduced two new datasets: TJU (single-view multimodal) and MV-TJU (multiview multimodal).

Main Results:

• The proposed MTSL framework preserves consistency between body-based and part-based classifications.
• It discovers action-specific and action-shared feature subspaces, enhancing generalization.
• The method demonstrated superior performance on RGB, depth, and multi-view datasets compared to existing methods.

Conclusions:

• The unified framework effectively addresses multiple/single-view human action recognition challenges.
• MTSL with part-based regularization offers significant advantages in model learning and feature selection.
• This work is the first to apply MTSL with part-based regularization to multi-view action recognition in RGB and depth modalities.