Human action recognition based on enhanced data guidance and key node spatial temporal graph convolution

Chengyu Zhang

doi:10.1007/s11042-022-11947-8

Human action recognition based on enhanced data guidance and key node spatial temporal graph convolution

2022-02-02

11

PRE

AI

DOI：10.1007/s11042-022-11947-8

Original

Original request for help

Save

Abstract

En 中文

Graph convolutional networks have achieved remarkable performance in action recognition from skeleton videos. However, most of the existing GCN-based methods improve performance by increasing model parameters, which require a high amount of data. This means that they usually perform poorly on small sample learning tasks. In this paper, we propose a novel enhanced data guidance algorithm to improve the performance of the GCN-based method on small sample datasets. These enhanced data perform coordinate transformation on the skeleton to obtain robustness to scale, rotation and translation. The proposed guidance algorithm allows the target model to learn the advantages of enhanced data and reduce the complexity of the task. We also propose a new key node method, which can select key joints and frames in the spatial and temporal dimensions respectively. This removes the redundant information of the skeleton sequence and significantly reduces the computational cost. Furthermore, the combination of key nodes and enhanced data can greatly reduce the demand for training data. The recognition accuracy rates of 94.81% and 94.19% have been achieved on the public MSR Action3D and UTD-MHAD datasets, respectively. This result proves that our method is significantly better than mainstream 3D action recognition methods.

Keywords:

Action recognition

Key node

Graph convolutional networks

Enhanced data

Human action recognition based on enhanced data guidance and key node spatial temporal graph convolution

Abstract

Journal

Organization