Chengzhang Yu

Chengzhang Yu

he/him

Chengzhang Yu is a Ph.D. candidate at South China University of Technology, specializing in Large Language Models (LLMs), Multimodal Medical Artificial Intelligence, and Agent & Multi-Agent Systems (MAS). His research focuses on advancing AI-driven clinical decision-making and intelligent autonomous systems through multimodal learning and collaborative agents.

DenseGCN: A multi-level and multi-temporal graph convolutional network for action recognition

Published in IET Image Processing, 2023

Abstract

This paper presents DenseGCN, a novel multi-level and multi-temporal graph convolutional network designed for skeleton-based action recognition. The proposed method addresses the limitations of existing graph convolutional networks in capturing long-range spatial and temporal dependencies in human motion sequences.

Key Contributions

  • Multi-level Spatial Modeling: DenseGCN incorporates multiple levels of spatial graph convolutions to capture both local joint relationships and global body structure dependencies.

  • Multi-temporal Feature Learning: The network employs temporal convolutions at different time scales to model both short-term motion patterns and long-term action dynamics.

  • Dense Connectivity: Inspired by DenseNet, the method uses dense connections between different temporal scales to enhance feature reuse and gradient flow.

  • Comprehensive Evaluation: Extensive experiments on NTU RGB+D and NTU RGB+D 120 datasets demonstrate the effectiveness of the proposed approach.

Technical Innovation

The DenseGCN architecture combines:

  1. Spatial Graph Convolution Modules for modeling joint relationships
  2. Multi-scale Temporal Convolutions for capturing motion dynamics
  3. Dense Skip Connections for improved feature propagation
  4. Adaptive Graph Topology Learning for flexible joint relationship modeling

Experimental Results

The proposed DenseGCN achieves competitive performance on standard benchmarks:

  • NTU RGB+D: Significant improvements in both cross-subject and cross-view evaluations
  • NTU RGB+D 120: State-of-the-art results on the large-scale dataset
  • Computational Efficiency: Maintains reasonable computational complexity while improving accuracy

Impact and Applications

This work contributes to the field of skeleton-based action recognition by providing a more effective way to model complex human motions. The multi-level and multi-temporal approach has potential applications in:

  • Human-computer interaction
  • Video surveillance
  • Sports analysis
  • Healthcare monitoring

Recommended citation: Yu, C., et al. (2023). "DenseGCN: A multi-level and multi-temporal graph convolutional network for action recognition." IET Image Processing. 17(11), 3299-3312.
Download Paper