Enhanced Spatial and Extended Temporal Graph Convolutional Network for Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Spatial-temporal graph convolution networks (ST-GCNs) are advanced for skeleton-based human action recognition.
• Existing ST-GCNs have limitations in comprehensive temporal cue extraction due to fixed temporal convolution kernels.
• Sequential connections of spatial and temporal graph convolution layers (GCLs) are suboptimal for performance.

Purpose of the Study:

• To propose a novel Enhanced Spatial and Extended Temporal Graph Convolutional Network (EE-GCN).
• To enhance the extraction of discriminative temporal features using multiple convolution kernels.
• To improve spatial feature representation through a novel connection paradigm.

Main Methods:

• Utilized three different sized convolution kernels for multi-term temporal feature extraction.
• Implemented a One-Shot Aggregation (OSA) and Effective Squeeze-Excitation (eSE) structure for feature concatenation and channel interdependency exploration.
• Introduced a new connection paradigm combining serial and parallel spatial GCLs to augment spatial features.

Main Results:

• The EE-GCN method demonstrated superior performance on three large-scale datasets.
• Experimental results indicate that the proposed EE-GCN surpasses current state-of-the-art methods in action recognition.
• The enhanced temporal feature extraction and spatial feature enhancement contributed to improved accuracy.

Conclusions:

• The proposed EE-GCN effectively addresses limitations in existing ST-GCNs for action recognition.
• The novel architecture enhances both temporal and spatial feature learning capabilities.
• EE-GCN represents a significant advancement in skeleton-based human action recognition.