Head Pose Estimation Based on Multi-Scale Convolutional Neural Network

Lingyu Liang; Tiantian Zhang; Wei He

doi:10.3788/LOP56.131003

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 13 >Page 131003 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 13, 131003 (2019)

Head Pose Estimation Based on Multi-Scale Convolutional Neural Network

Lingyu Liang^{1、2、3、**}, Tiantian Zhang^1、3, and Wei He^1、*

Author Affiliations

¹ Key Laboratory of Wireless Sensor Network and Communication, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China

² School of Information Science and Technology, ShanghaiTech University, Shanghai 200120, China

³ University of Chinese Academy of Sciences, Beijing 100049, China

show less

DOI: 10.3788/LOP56.131003 Cite this Article Set citation alerts

Lingyu Liang, Tiantian Zhang, Wei He. Head Pose Estimation Based on Multi-Scale Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(13): 131003 Copy Citation Text

show less

Fig. 1. Head pose estimation flow chart

Download full size

Deep neural network structure for head pose estimation. (a) Multi-scale convolution structure; (b) process structure after feature combination

Fig. 2. Deep neural network structure for head pose estimation. (a) Multi-scale convolution structure; (b) process structure after feature combination

Download full size

Fig. 3. Part of the experimental head posture library picture. (a) CAS-PEALR1; (b) Pointing'04

Download full size

Fig. 4. Face images after cropped

Download full size

Fig. 5. Partial head posture pictures under different interference factors. (a) Standard; (b) with mask; (c) with glasses; (d) expression; (e) weak illumination; (f) strong illumination; (g) background

Download full size

Convolution	Recognition accuracyon test set /%
Multi-scale convolution	98.9
3×3 single-scale convolution	94.3
5×5 single-scale convolution	93.1
7×7 single-scale convolution	93.2

Table 1. Multi-scale convolution vs. single-scale convolution

Algorithm	Pointing'04accuracy /%	CAS-PEAL-R1 accuracy /%	Numberofgestures
Algorithm ofthis paper	96.5	98.9	21
Cluster-glassificationBayesian network^[23]	94.8	96.2	12
Based onfacialfeature points^[24]	92.7	93.5	6

Table 2. Accuracy of different algorithms on Pointing'04 and CAS-PEAL-R1

Aftergesturesconversion	Pitch attitudebefore gesturesconversion	Yaw attitudebefore gesturesconversion /(°)
Level	PM	0,-15,+15
Level left	PM	+30,+45
Level right	PM	-30,-45
Pitch down	PD	0,-15,+15
Pitch up	PU	0,-15,+15
Left up	PU	+30,+45
Right up	PU	-30,-45
Left down	PD	+30,+45
Right down	PD	-30,-45

Table 3. Gestures Conversion relationship table

Interferencefactor	Accuracy ofthis paper /%	Accuracy ofmethod in Ref.[23] /%
Standard	98.5	96.3
With mask	92.9	86.4
With glasses	96.3	89.7
Expression	98.1	94.9
Weak illumination	95.5	91.8
Strong illumination	96.1	92.2
Background	97.7	95.1

Table 4. Different interference factors effect on recognition rate

Resolution /(pixel×pixel)	Timeof thispaper /ms	Timeof method inRef. [23] /ms	Timeof method inRef. [24] /ms
1920×1080	34.3	51.3	562.1
1360× 760	32.7	48.5	451.7
800×600	30.8	45.2	349.8

Table 5. Recognition time at different resolution

Lingyu Liang, Tiantian Zhang, Wei He. Head Pose Estimation Based on Multi-Scale Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(13): 131003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information