Analysis of Teachers&#39; Actions Using Feature Dense Computation and Fusion Algorithm

Xiaolong Zhang; Jianfei Liu; Luguo Hao

doi:10.3788/LOP56.161503

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 16 >Page 161503 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 16, 161503 (2019)

Analysis of Teachers' Actions Using Feature Dense Computation and Fusion Algorithm

Xiaolong Zhang¹, Jianfei Liu^1、*, and Luguo Hao²

Author Affiliations

¹ School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 300401, China

² School of Information Engineering, Guangdong University of Technology, Guangzhou, Guangdong 510006, China

show less

DOI: 10.3788/LOP56.161503 Cite this Article Set citation alerts

Xiaolong Zhang, Jianfei Liu, Luguo Hao. Analysis of Teachers' Actions Using Feature Dense Computation and Fusion Algorithm[J]. Laser & Optoelectronics Progress, 2019, 56(16): 161503 Copy Citation Text

show less

Fig. 1. Schematic of TSDCFN structure. (a) Overall structure of TSDCFN; (b) structure of densely connected module (in which, w/o represents without)

Download full size

Fig. 2. Spatiotemporal pyramid pooling model

Download full size

Fig. 3. Scheme of spatiotemporal and non-local feature computation

Download full size

Fig. 4. Relationship between recognition accuracy and Epoch No.

Download full size

Fig. 5. Examples of recognition effects of action videos

Download full size

Layer	Output size	3D DenseNet
3D Convolution	D×H×W	3×3×3 conv
Dense block 1	D×H×W	$\begin{matrix} [\begin{matrix} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{matrix}] \end{matrix}$ ×10
Transition layer 1	D×H×W	1×1×1 conv
Transition layer 1	D× $\begin{matrix} \frac{H}{2} \end{matrix}$ × $\begin{matrix} \frac{W}{2} \end{matrix}$	1×2×2 max pooling
Dense block 2	D× $\begin{matrix} \frac{H}{2} \end{matrix}$ × $\begin{matrix} \frac{W}{2} \end{matrix}$	$\begin{matrix} [\begin{matrix} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{matrix}] \end{matrix}$ ×10
Transition layer 2	D× $\begin{matrix} \frac{H}{2} \end{matrix}$ × $\begin{matrix} \frac{W}{2} \end{matrix}$	1×1×1 conv
Transition layer 2	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	2×2×2 max pooling
Dense block 3	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	$\begin{matrix} [\begin{matrix} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{matrix}] \end{matrix}$ ×10
Transition w/opooling layer 1	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	1×1×1 conv
Dense block 4	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	$\begin{matrix} [\begin{matrix} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{matrix}] \end{matrix}$ ×10
Transition w/opooling layer 2	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	1×1×1 conv
Dense block 5	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	$\begin{matrix} [\begin{matrix} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{matrix}] \end{matrix}$ ×10
Transition w/opooling layer 3	$\begin{matrix} \frac{D}{2} \end{matrix}$ × $\begin{matrix} \frac{H}{4} \end{matrix}$ × $\begin{matrix} \frac{W}{4} \end{matrix}$	1×1×1 conv
Classification layer		fully connected
Classification layer		softmax and prediction

Table 1. Parameters of modified 3D densely connected convolutional neural network

Method	d	k	θ	L	Data Aug	Accuracy /%
3D DenseNet	30	24	0.5	32	-	85.62
Modified 3D DenseNet	58	48	1.0	16	-	91.44
Modified 3D DenseNet with non-localfeature computation block	58	48	1.0	32	-	93.02
Modified 3D DenseNet with spatio-temporally pyramid pooling	58	48	1.0	-	-	93.76
TSDCFN with non-local feature computationblock and spatio temporal pyramid pooling	58	48	1.0	-	-	96.87
	58	48	1.0	-	√	98.13

Table 2. Test results of networks with different configuration parameters on dataset of teachers' actions (in which, Data Aug represents Data Augmentation)

Xiaolong Zhang, Jianfei Liu, Luguo Hao. Analysis of Teachers' Actions Using Feature Dense Computation and Fusion Algorithm[J]. Laser & Optoelectronics Progress, 2019, 56(16): 161503

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information