3D Skeletal Gesture Recognition via Hidden States Exploration.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Modeling temporal dynamics in human body gestures remains a significant challenge.
• Traditional generative models like Hidden Markov Models (HMMs) often struggle with explicit temporal structure due to fixed state anchors.

Purpose of the Study:

• To develop a novel formulation for temporal gesture modeling that captures distinct phases.
• To improve the explicit representation of temporal structures in human body gestures.

Main Methods:

• Proposed a new formulation using low-rank matrix decomposition to build temporal gesture compositions.
• Segmented gesture sequences into states based on the assumption of linearly correlated static poses.
• Utilized long short-term memory (LSTM) networks to learn emission probabilities over HMM states, incorporating temporal context.

Main Results:

• The proposed method effectively segments gesture sequences into semantically meaningful and discriminative temporal states.
• Achieved state-of-the-art performance on multiple challenging gesture datasets.
• Demonstrated effectiveness across a wide range of human body gestures.

Conclusions:

• The novel approach addresses limitations of traditional HMMs in modeling gesture temporal dynamics.
• Low-rank matrix decomposition combined with LSTM offers a powerful framework for gesture recognition and analysis.
• The method provides a more robust and accurate representation of gesture temporal structures.