Action Recognition Using Rate-Invariant Analysis of Skeletal Shape Trajectories

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Classifying human actions from depth sensor data is challenging due to variations in execution speed.
• Dynamical skeletons represent human body movements as trajectories on a shape manifold.
• Existing methods struggle with the high variability in action execution rates.

Purpose of the Study:

• To develop a robust framework for human action classification using depth sensor data.
• To address the problem of variable execution rates in action recognition.
• To achieve state-of-the-art performance in action classification using skeletal data.

Main Methods:

• Representing human actions as trajectories of dynamical skeletons.
• Employing a parameterization-invariant metric based on transported square-root vector fields (TSRVFs) for trajectory analysis.
• Developing computational tools for skeleton trajectory processing, including smoothing, sampling, and temporal registration.
• Extracting invertible Euclidean representations for statistical modeling.

Main Results:

• Achieved state-of-the-art action classification results on MSR Action-3D, MSR Daily Activity, and 3D Action Pairs datasets.
• Demonstrated the effectiveness of the proposed framework using Support Vector Machine (SVM) classification.
• Showcased the utility of invertible Euclidean representations for both discriminative and generative models.

Conclusions:

• The proposed method effectively classifies human actions from skeletal data, outperforming existing approaches.
• The parameterization-invariant framework successfully handles variations in action execution rates.
• Skeletal information alone is sufficient for achieving high-accuracy human action recognition.