Skeleton-based Human Action Recognition via Large-kernel Attention Graph Convolutional Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Skeleton-based human action recognition is crucial for virtual reality applications due to its robustness to noise.
• Current graph convolution methods struggle to model long-range dependencies in skeleton data for action recognition.
• Existing approaches may overlook crucial action semantics embedded in long-range spatial and temporal patterns.

Purpose of the Study:

• To introduce a novel Skeleton Large Kernel Attention operator (SLKA) to enhance receptive field and channel adaptability.
• To develop a spatiotemporal SLKA module (ST-SLKA) for aggregating long-range spatial features and temporal correlations.
• To propose a new LKA-GCN network architecture and a joint movement modeling (JMM) strategy for improved action recognition.

Main Methods:

• Introduced a Skeleton Large Kernel Attention (SLKA) operator to enlarge receptive fields and improve channel adaptability.
• Developed a spatiotemporal SLKA module (ST-SLKA) for aggregating long-range spatial features and learning long-distance temporal correlations.
• Proposed a novel LKA-GCN architecture and a joint movement modeling (JMM) strategy to focus on significant temporal interactions.

Main Results:

• The proposed LKA-GCN architecture achieved state-of-the-art performance on the NTU-RGBD 60, NTU-RGBD 120, and Kinetics-Skeleton 400 datasets.
• The SLKA operator effectively captures long-range spatial and temporal dependencies without significant computational overhead.
• The joint movement modeling strategy successfully focused on valuable temporal interactions for improved recognition accuracy.

Conclusions:

• The LKA-GCN network, incorporating SLKA and JMM, represents a significant advancement in skeleton-based human action recognition.
• The developed methods demonstrate superior ability in modeling long-range dependencies crucial for understanding complex actions.
• This work sets a new state-of-the-art benchmark, highlighting the potential of large kernel attention in spatio-temporal graph networks.