A Bayesian Dynamical Approach for Human Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Action recognition from 3D skeletal data is crucial for human-computer interaction.
• Existing methods often neglect the complex temporal dynamics inherent in human motion.
• There is a need for models that can capture nonlinear inter-dependencies and provide interpretable motion primitives.

Purpose of the Study:

• To introduce a novel generative Bayesian switching dynamical model for robust action recognition in 3D skeletal data.
• To explicitly model temporal transitions and nonlinear inter-dependencies for improved motion understanding.
• To provide visual and quantitative interpretations of motion primitives underlying action classes.

Main Methods:

• Developed a generative Bayesian switching dynamical model.
• Encoded correlated skeletal data into low-dimensional switching temporal processes.
• Utilized a switching deep autoregressive prior to handle multimodal and higher-order nonlinearities.
• Employed a nonlinear second-order dynamical transition model.

Main Results:

• Achieved 6.3% higher action classification accuracy compared to state-of-the-art methods.
• Improved predictive error by 3.5% using the nonlinear dynamical transition model.
• Demonstrated superior performance on two large-scale 3D skeletal datasets.
• Enabled parsing of meaningful intrinsic states in skeletal dynamics.

Conclusions:

• The proposed dynamical deep generative latent model significantly advances action recognition in 3D skeletal data.
• Explicitly modeling temporal dynamics and nonlinearities leads to enhanced performance and interpretability.
• The model offers a new perspective on understanding motion primitives through sequences of intrinsic states.