Skeleton Action Recognition Based on Dense Residual Shift Graph Convolutional Network

Tao Yang; Jun Han; Haiyan Jiang

doi:10.3788/LOP220428

Laser & Optoelectronics Progress, Volume. 60, Issue 10, 1010012(2023)

Skeleton Action Recognition Based on Dense Residual Shift Graph Convolutional Network

Tao Yang^1,2、*, Jun Han^1,2, and Haiyan Jiang^1,2

Author Affiliations

¹College of Communication and Information Engineering, Shanghai University, Shanghai 200444, China

²Shanghai Institute of Advanced Communication and Data Science, Shanghai 200444, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Diagram of system block

Download full size

Fig. 2. Shift convolution operation. (a) Node 1 shift; (b) node 2 shift; (c) feature map after shift

Download full size

Fig. 3. Diagram of DRS-GCN block

Download full size

Fig. 4. Diagrams of dense residual network structure. (a) Residual connection; (b) dense connection; (c) dense residual connection

Download full size

Fig. 5. Diagram of joint-motion

Download full size

Fig. 6. Examples of DAILY dataset. (a) Walking; (b) throwing; (c) falling

Download full size

Fig. 7. Node labeling of OpenPose

Download full size

Fig. 8. Examples of skeleton diagram. (a) Walking; (b) throwing; (c) falling

Download full size

Fig. 9. Confusion matrix of multi-class

Download full size

Fig. 10. Curves of training accuracy and loss value on DAILY dataset. (a) Accuracy; (b) loss value

Download full size

Fig. 11. Curves of training accuracy and loss value on NTU60 RGB+D dataset (CS). (a) Accuracy of CS; (b) loss value of CS

Download full size

Fig. 12. Curves of training accuracy and loss value on NTU60 RGB+D dataset.(CV). (a) Accuracy of CV; (b) loss value of CV

Download full size

Table 1. Research on performance of multi-stream network on NTU60 RGB+D dataset
View table
Table 1. Research on performance of multi-stream network on NTU60 RGB+D dataset
Method CS /% CV /%
DRS-GCN（J） 88.1 95.3
DRS-GCN（B） 88.9 94.8
DRS-GCN（J-M） 86.8 93.6
DRS-GCN（B-M） 87.0 93.7
4s-DRS-GCN 90.8 96.3

Table 2. Experimental data comparison of 7 behaviors on DAILY dataset

View table

Table 2. Experimental data comparison of 7 behaviors on DAILY dataset

Class	Recall / $%$		Precision / $%$
Class	DRS-GCN	Shift-GCN	DRS-GCN	Shift-GCN
walking	73.1	75.0	73.1	69.2
sitting	90.5	85.7	84.4	80.0
standing	84.8	72.6	86.7	82.2
donning	88.6	84.1	90.7	86.1
doffing	83.3	85.0	81.4	79.1
throwing	60.0	53.2	61.4	56.8
falling	95.0	94.9	97.4	94.9

Table 3. Comparative analysis of three indicators on DAILY dataset
View table
Table 3. Comparative analysis of three indicators on DAILY dataset
Method Recall / $%$ Precision / $%$ Accuracy / $%$
Shift-GCN^［11］ 78.6 78.3 77.8
DRS-GCN 82.2 82.2 81.7

Table 4. Comparison of experimental data between DRS-GCN and advanced algorithms on NTU60 RGB+D dataset

View table

Table 4. Comparison of experimental data between DRS-GCN and advanced algorithms on NTU60 RGB+D dataset

Method	Accuracy /%		FLOPs /10⁹
Method	CS	CV	FLOPs /10⁹
VA-LSTM^［19］	79.2	87.7
ST-GCN^［9］	81.5	88.3
HCN^［20］	86.5	91.1
AS-GCN^［10］	86.8	94.2	27.0
2s-AGCN^［12］	88.5	95.1	35.8
Shift-GCN^［13］	87.8	95.1	2.5
DRS-GCN	88.1	95.3	3.7
4s-DRS-GCN	90.8	96.3	14.8

Tools

Get Citation

Copy Citation Text

Tao Yang, Jun Han, Haiyan Jiang. Skeleton Action Recognition Based on Dense Residual Shift Graph Convolutional Network[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010012

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites