一种复杂场景下的人眼检测算法

崔家礼; 曹衡; 张亚明; 罗嗣梧; 李锦涛; 王华峰

doi:10.13700/j.bh.1001-5965.2019.0641

一种复杂场景下的人眼检测算法

doi: 10.13700/j.bh.1001-5965.2019.0641

崔家礼¹,
曹衡¹,
张亚明¹,
罗嗣梧²,
李锦涛²,
王华峰^{1, 2, ,}

1.
北方工业大学信息学院, 北京 100144
2.
北京航空航天大学软件学院, 北京 100083

基金项目:

国家重点研发计划 2017YFB0802300

详细信息

作者简介:
崔家礼  男, 博士, 助理研究员。主要研究方向:智能识别与数字图像处理

曹衡  男, 硕士研究生。主要研究方向:智能识别与数字图像处理

王华峰  男, 博士, 副研究员。主要研究方向:人工智能与机器人

通讯作者:
王华峰, E-mail: wanghuafeng@ncut.edu.cn

中图分类号: TP391.41;TP37
计量
- 文章访问数: 967
- HTML全文浏览量: 153
- PDF下载量: 240
- 被引次数: 0
出版历程
- 收稿日期: 2019-12-23
- 录用日期: 2020-04-03
- 网络出版日期: 2021-01-20

A human eye detection algorithm in complex scenarios

1.
School of Information Science and Technology, North China University of Technology, Beijing 100144, China
2.
School of Software, Beihang University, Beijing 100083, China

Funds:

National Key R & D Program of China 2017YFB0802300

More Information

Corresponding author: WANG Huafeng, E-mail:wanghuafeng@ncut.edu.cn

摘要

摘要:
针对复杂场景下的人眼检测问题，间接方法和直接方法具有一定的局限性。提出了一种不依赖人脸检测的直接型人眼检测算法，以解决复杂场景下多尺度尤其是小尺度人眼检测问题。算法通过减少下采样因子并加入扩张残差单元以提升小尺度人眼检测能力，且对多尺度特征相互拼接以保证多尺度人眼检测的精度。同时，算法借助于压缩特征输出通道降低了模型复杂度，使人眼检测效率得以提升。实验结果表明：所提模型可以在小尺度下有效地进行左右眼区分，并在红外数据上表现良好。经在DIF数据集上进行训练与测试，所提模型在较小尺度下人眼检测精度达到82.59%，检测效率达到30.5 fps。
- 人眼检测 /
- 深度学习 /
- 小目标检测 /
- 复杂场景 /
- 多尺度特征
Abstract:
Aiming at the problem of human eye detection in complex scenes, indirect methods and direct methods have certain limitations.A direct human eye detection method which is independent of face detection is proposed. The proposed method detects eyes under multiple scales especially small scale in complex scenarios. The improvement of the proposed method consists of improving small-scale human eye detection ability by reducing the down-sampling factor and adding extended residual units; ensuring the accuracy of multi-scale human eye detection by concatenating the multi-scale features; improving human eye detection efficiency by reducing the number of feature output channels to simplify the complexity of the model. The experimental results show that the proposed model can distinguish the left and right eyes effectively under small scale and has good performance with infrared data. The training and test on DIF dataset show that the human eye detection precision of the proposed method is 82.59%, and the detection rate is 30.5 fps.
- human eye detection /
- deep learning /
- small target detection /
- complex scenarios /
- multi-scale features

HTML全文

图 1 本文算法网络架构

Figure 1. Proposed algorithmic network architecture

下载: 全尺寸图片幻灯片

图 2 两种改进的扩张残差单元

Figure 2. Two improved expanded residual units

下载: 全尺寸图片幻灯片

图 3 骨干网络下采样

Figure 3. Backbone network downsampling

下载: 全尺寸图片幻灯片

图 4 人眼数据集实例

Figure 4. Examples of human eye datasets

下载: 全尺寸图片幻灯片

图 5 眼睛区域不可见及过度模糊等极端情况

Figure 5. Extreme cases such as invisible eye area and excessive blur

下载: 全尺寸图片幻灯片

图 6 CASIA-Iris-Distance数据集中的红外图像

Figure 6. Infrared images in CASIA-Iris-Distance dataset

下载: 全尺寸图片幻灯片

图 7 平均交并比与锚点框数量之间的关系

Figure 7. Relationship between average IoU and number of anchor boxes

下载: 全尺寸图片幻灯片

图 8 不同算法在不同尺度下的检出率

Figure 8. Statistics of detection rate of different algorithms at different scales

下载: 全尺寸图片幻灯片

图 9 四种算法的检测结果对比

Figure 9. Comparison of detection results among four algorithms

下载: 全尺寸图片幻灯片

图 10 红外图像的人眼检测结果

Figure 10. Human eye detection results of infrared image

下载: 全尺寸图片幻灯片

表 1 不同算法优劣势比较

Table 1. Comparison of advantages and disadvantages among different algorithms

算法	劣势	优势
YOLOv3	小尺度目标检测效果较差	检测效率高
Faster R-CNN	实时性差，小尺度目标检测效果差	检测精度较高
Adaboost	检测精度低，小尺度目标检测效果很差	检测效率高

下载: 导出CSV

表 2 不同尺度网络在DIF数据集上的实验结果

Table 2. Experimental results of different scale networks on DIF dataset

算法	尺度数	特征图	输入	mAP/%	F/fps
YOLOv3	3	52×26×13	416×416	76.97	33.4
Proposed-3	3	26×26×26	416×416	80.5	31.8
Proposed-4	4	52×26×26×26	416×416	82.4	30.5
Proposed-5	5	104×52×26×26×26	416×416	82.68	27.3
注：fps为帧/s。

下载: 导出CSV

表 3 人眼检测统计结果

Table 3. Human eye detection statistic results

场景	左眼检出率/%	左眼漏检率/%	右眼检出率/%	右眼漏检率/%	误检率/%
A	85.9	14.1	89.1	10.9	21.4
B	83.3	16.7	86.1	13.9	19.4

下载: 导出CSV

表 4 不同算法的评价指标对比

Table 4. Comparison of evaluation indicators among various algorithms

算法	AP_L/%	AP_R/%	mAP/%	F/fps
YOLOv3	76.71	79.23	77.97	33.4
YOLOv3-tiny	74.66	75.3	74.98	204.3
YOLOv2	60.97	65.45	63.21	45.6
Faster R-CNN	77.91	79.35	78.63	12
Proposed	84.77	80.41	82.59	30.5

下载: 导出CSV

表 5 红外人眼检测性能

Table 5. Performance of infrared human eye detection

算法	AP_L/%	AP_R/%	mAP/%	F/fps
Proposed	91.82	94.48	93.15	31.2

下载: 导出CSV

参考文献(21)

[1]	DAUGMAN J.How iris recognition works[M]//BOVIK A.The essential guide to image processing.Salt Lake City: Academic Press, 2009: 715-739.
[2]	TURK M A, PENTLAND A P.Face recognition using eigenfaces[C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway: IEEE Press, 1991: 586-591.
[3]	DUCHOWSKI A T.Eye tracking methodology[J].Theory and Practice, 2007, 328(614):2-3.
[4]	ASADA M, TANAKA T, HOSODA K.Adaptive binocular visual servoing for independently moving target tracking[C]//IEEE International Conference on Robotics and Automation.Piscataway: IEEE Press, 2000, 3: 2076-2081.
[5]	HAILIN W, HANHUI L, ZHUMEI S.Fatigue driving detection system design based on driving behavior[C]//International Conference on Optoelectronics and Image Processing.Piscataway: IEEE Press, 2010, 1: 549-552.
[6]	叶修强.基于局部特征的人眼检测[D].南昌: 南昌大学, 2018: 18-20. YE X Q.Human eye detection based on local features[D].Nanchang: Nanchang University, 2018: 18-20(in Chinese).
[7]	金花.人脸识别中眼睛定位算法的研究[D].沈阳: 东北大学, 2010: 6-7. JIN H.Research on eye location algorithm in face recognition[D].Shenyang: Northeastern University, 2010: 6-7(in Chinese).
[8]	郝叶林, 罗兵, 杨锐, 等.复杂场景图像中人员目标检测算法的改进[J].五邑大学学报, 2018, 28(1):103-108. HAO Y L, LUO B, YANG R, et al.Improvement of human target detection algorithm in complex scene images[J].Journal of Wuyi University, 2018, 28(1):103-108(in Chinese).
[9]	MERLER M, RATHA N, FERIS R S, et al.Diversity in faces[EB/OL].(2019-01-29)[2019-12-30].https: //arxiv.org/abs/1901.10436.
[10]	刘青.基于深度学习的多姿态眼睛定位算法研究及应用[D].广州: 华南理工大学, 2018: 5-7. LIU Q.Research and application of multi-pose eye positioning algorithm based on deep learning[D].Guangzhou: South China University of Technology, 2018: 5-7(in Chinese).
[11]	张杰, 杨晓飞, 赵瑞莲.基于积分投影和Hough变换圆检测的人眼精确定位方法研究[J].电子器件, 2005, 28(4):706-709. doi: 10.3969/j.issn.1005-9490.2005.04.003 ZHANG J, YANG X F, ZHAO R L.Research on human eye precise positioning method based on integral projection and Hough transform circle detection[J].Electronic Devices, 2005, 28(4):706-709(in Chinese). doi: 10.3969/j.issn.1005-9490.2005.04.003
[12]	KROON B, MAAS S, BOUGH S, et al.Eye localization in low and standard definition content with application to face matching[J].Computer Vision and Image Understanding, 2009, 113(8):921-933. doi: 10.1016/j.cviu.2009.03.013
[13]	TANG X, OU Z, SU T, et al.Robust precise eye location by Adaboost and SVM techniques[J].Lecture Notes in Computer Science, 2005, 3497:93-98.
[14]	GKIOXARI G, GIISHICK R, MALIK J.Contextual action recognition with R^*CNN[C]//Proceedings of the IEEE International Conference on Computer Vision.Piscataway: IEEE Press, 2015: 1080-1088.
[15]	REN S, HE K M, GIRSHICK R, et al.Faster R-CNN:Towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6):1137-1149. doi: 10.1109/TPAMI.2016.2577031
[16]	LIU W, ANGUELOY D, ERHAN D, et al.SSD: Single shot multibox detector[C]//European Conference on Computer Vision.Berlin: Springer, 2016: 21-37.
[17]	REDMON J, FARGADI A.YOLOv3: An incremental improvement[EB/OL](2018-04-08)[2019-12-30].https://arxiv.org/abs/1804.02767.
[18]	TAN T N, SUN Z N.CASIA iris image database[EB/OL].[2019-12-20].http://biometrics.idealtest.org/.
[19]	SCHUTZE H, MANNING C D, RAGH P.Introduction to information retrieval[C]//Proceedings of the International Communication of Association for Computing Machinery Conference.Cambridge: Cambridge University Press, 2008, 39: 234-265.
[20]	KIM C E, OGHAZ M D, FAJTL J, et al.A comparison of embedded deep learning methods for person detection[EB/OL].(2018-12-09)[2019-12-30].https//arxiv.org/abs/1812.03451.
[21]	REDMON J, FARHADI A.YOLO9000: Better, faster, stronger[EB/OL].(2016-12-25)[2019-12-30].https://arxiv.org/abs/1612.08242.