Extended multi-stream temporal-attention module for skeleton-based human action recognition (HAR)

TL;DR

This paper proposes an extended multi-stream temporal-attention module to enhance skeleton-based human action recognition using graph convolutional networks, addressing limitations in static graph structures across layers.

Contribution

It introduces a novel multi-stream temporal-attention mechanism that dynamically adapts graph structures for improved HAR performance.

Findings

Enhanced recognition accuracy on benchmark datasets

Dynamic graph adaptation improves model flexibility

Outperforms existing GCN-based HAR methods

Abstract

Graph convolutional networks (GCNs) are an effective skeleton-based human action recognition (HAR) technique. GCNs enable the specification of CNNs to a non-Euclidean frame that is more flexible. The previous GCN-based models still have a lot of issues: (I) The graph structure is the same for all model layers and input data.