Action recognition algorithm based on multi-scale and multi-branch features

Number of layers	Module name	Size	Step	Input and output
1	Conv1	1×1	1	6，36
2	Conv_gcn	3×3	1	36，72
	Conv_tcn	1×1	1	72，24
		5×1	2	24，24
3	Conv_gcn	3×3	1	72，72
	Conv_tcn	1×1	1	72，24
		5×1	1	24，24
4	Conv_gcn	3×3	1	72，36
	Conv_tcn	1×1	1	36，12
		5×1	1	12，12
5	Conv_gcn	3×3	1	108，192
	Conv_tcn	1×1	1	192，64
		5×1	2	64，64
6，7	Conv_gcn	3×3	1	192，192
	Conv_tcn	1×1	1	192，64
		5×1	2	64，64
8	Conv_gcn	3×3	1	192，384
	Conv_tcn	1×1	1	384，128
		5×1	1	128，128
9，10	Conv_gcn	3×3	1	384，384
	Conv_tcn	1×1	1	384，128
		5×1	2	128，128
11	FC	-	-	384，60

Table 2. Accuracy of different feature enhancement method on NTU RGB-D 60 dataset
View table
View in Article
Table 2. Accuracy of different feature enhancement method on NTU RGB-D 60 dataset
数据归一化/% SavGol滤波/% 坐标转换/% 插帧/% 深度优先树遍历/% 应用所有方法的特征增强/%
Cross-view（CV） 89.3 90.1 92.1 90.5 92.6 95.1

Table 3. Accuracy of different feature on NTU RGB-D 60 dataset
View table
View in Article
Table 3. Accuracy of different feature on NTU RGB-D 60 dataset
Bone length+Joints/% Bone length+Velocities/% Bone length+Bone angle/% 同时输入三种特征/%
CV 93.8 93.0 92.8 95.1

Table 4. Accuracy of different feature fusion location on NTU RGB-D 60 dataset
View table
View in Article
Table 4. Accuracy of different feature fusion location on NTU RGB-D 60 dataset
k=1 k=2 k=3 k=4
CV 92.8% 94.5% 95.1% 94.1%

Table 5. Accuracy of different temporal convolution kernel on NTU RGB-D 60 dataset
View table
View in Article
Table 5. Accuracy of different temporal convolution kernel on NTU RGB-D 60 dataset
m=3 m=5 m=7 m=9
CV 94.5% 95.1% 94.7% 93.4%

Table 6. Accuracy of different action recognition methods on NTU RGB-D 60 dataset

View table

View in Article

Table 6. Accuracy of different action recognition methods on NTU RGB-D 60 dataset

Method	Cross-subject/%	Cross-view/%
TCN^［17］	74.3	83.1
Ind-RNN^［18］	81.8	88.0
ST-GCN^［12］	81.5	88.3
HCN^［8］	86.5	91.1
AS-GCN^［13］	86.8	94.2
3SCNN^［19］	88.6	93.7
SGN^［20］	86.6	94.3
PR-GCN^［21］	85.2	91.7
PeGCN^［22］	85.6	93.4
Ours	89.6	95.1

Table 7. Accuracy of different action recognition methods on NTU RGB-D 120 dataset
View table
View in Article
Table 7. Accuracy of different action recognition methods on NTU RGB-D 120 dataset
Method Cross-subject/% Cross-setup/%
Part-aware LSTM^［11］ 25.5 26.3
TSRJI^［23］ 67.9 62.8
3SCNN^［19］ 81.2 81.9
3s RA-GCN^［24］ 81.1 82.7
Gimme Signals^［25］ 70.8 71.6
SGN^［20］ 79.2 81.5
Ours 84.1 86.0

Tools

Get Citation

Copy Citation Text

Lei ZHANG, Guang-liang HAN. Action recognition algorithm based on multi-scale and multi-branch features[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(12): 1614

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: May. 25, 2022

Accepted: --

Published Online: Nov. 30, 2022

The Author Email: Guang-liang HAN (hangl@ciomp.ac.cn)

DOI:10.37188/CJLCD.2022-0176

Topics

Table 1. Network parameter table of multi-scale convolution

Table 1. Network parameter table of multi-scale convolution

Table 2. Accuracy of different feature enhancement method on NTU RGB-D 60 dataset

Table 2. Accuracy of different feature enhancement method on NTU RGB-D 60 dataset

Table 3. Accuracy of different feature on NTU RGB-D 60 dataset

Table 3. Accuracy of different feature on NTU RGB-D 60 dataset

Table 4. Accuracy of different feature fusion location on NTU RGB-D 60 dataset

Table 4. Accuracy of different feature fusion location on NTU RGB-D 60 dataset

Table 5. Accuracy of different temporal convolution kernel on NTU RGB-D 60 dataset

Table 5. Accuracy of different temporal convolution kernel on NTU RGB-D 60 dataset

Table 6. Accuracy of different action recognition methods on NTU RGB-D 60 dataset

Table 6. Accuracy of different action recognition methods on NTU RGB-D 60 dataset

Table 7. Accuracy of different action recognition methods on NTU RGB-D 120 dataset

Table 7. Accuracy of different action recognition methods on NTU RGB-D 120 dataset