ASR-GCN: Adaptive spatial information reconstruction GCN for skeleton-based action recognition

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Skeleton-based action recognition is crucial in computer vision.
• Existing methods struggle with feature extraction and dynamic learning for complex actions.

Purpose of the Study:

• To propose an innovative Adaptive Spatial Information Reconstruction Model (ASR-GCN) to overcome limitations in skeleton-based action recognition.
• To enhance feature extraction and dynamic feature learning capabilities.

Main Methods:

• Developed a Gated Reconstruction Unit (GRU) with a reweighting gating strategy for representative feature learning.
• Constructed an Adaptive Spatial Information Reconstruction Unit (ASRU) for adaptive feature contribution adjustment, extraction, reweighting, and integration.

Main Results:

• The ASR-GCN model significantly improves recognition capability with minimal parameter and computational overhead.
• Achieved state-of-the-art performance on NTU RGB+D 60 and NTU RGB+D 120 datasets.
• Reached 90.9% accuracy on cross-subject and 92.4% on cross-set tasks for NTU 120.

Conclusions:

• The proposed ASR-GCN model effectively addresses challenges in skeleton-based action recognition.
• The adaptive feature extraction approach enhances deep mining of intrinsic skeleton data features.
• The model demonstrates superior performance and efficiency in action recognition tasks.