Skeleton-Based Online Action Prediction Using Scale Selection Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Online action prediction aims to identify activities from incomplete temporal data.
• 3D skeleton sequences are a common input for human action recognition tasks.
• Existing methods face challenges with temporal scale variations and computational efficiency.

Purpose of the Study:

• To develop an effective framework for online action prediction using streaming 3D skeleton sequences.
• To address the challenges of temporal scale variations and interference from previous actions.
• To improve the efficiency and accuracy of skeleton-based action prediction.

Main Methods:

• A dilated convolutional network models temporal motion dynamics using a sliding window.
• A novel window scale selection method adapts to varying temporal scales and suppresses interference.
• An activation sharing scheme optimizes computation for overlapping time steps.
• Hierarchical dilated tree convolutions learn multi-level structured representations from skeleton joints.

Main Results:

• The proposed method demonstrates effectiveness on four challenging datasets.
• Extensive experiments validate the approach for skeleton-based online action prediction.
• The framework achieves accurate predictions even with partial and noisy action data.

Conclusions:

• The developed framework provides a robust solution for online action prediction from 3D skeleton data.
• The novel window scale selection and activation sharing mechanisms enhance performance and efficiency.
• The hierarchical dilated tree convolutions effectively capture complex spatio-temporal features for action recognition.