Please wait a minute...
浙江大学学报(工学版)  2020, Vol. 54 Issue (12): 2405-2413    DOI: 10.3785/j.issn.1008-973X.2020.12.015
计算机与控制工程     
动态背景下基于自更新像素共现的前景分割
梁栋1(),刘昕宇1,潘家兴1,孙涵1,周文俊2,金子俊一2
1. 南京航空航天大学 计算机科学与技术学院,江苏 南京 211100
2. 北海道大学 大学院信息科学研究科,北海道 札幌 220-0004
Foreground segmentation under dynamic background based on self-updating co-occurrence pixel
Dong LIANG1(),Xin-yu LIU1,Jia-xing PAN1,Han SUN1,Wen-jun ZHOU2,Shun’ichi KANEKO2
1. College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautic, Nanjing 211100, China
2. Graduate School of Information Science and Technology, Hokkaido University, Sapporo 220-0004, Japan
 全文: PDF(1089 KB)   HTML
摘要:

针对共现像素-支持块模型(CPB)存在的问题,提出一种新的自更新像素共现模型(SU-CPB). 引入经大规模监控场景训练的时空注意力模型(STAM),将STAM分割掩模作为指导,通过3种方法,包括像素-支持块对的动态选择,结构失效支持块的替换与前景相似度的计算,完成对支持块的在线自更新,解决CPB不具备更新能力带来的模型性能下降的问题,并使SU-CPB具备跨场景前景分割能力. 实验结果表明,该方法在所有测试场景下均优于CPB,并在未经STAM训练的Wallflower与LIMU数据集下,显著优于单纯的STAM、CPB以及其他参与对比的方法.

关键词: 前景分割像素空间关系时空注意力模型(STAM)在线自更新跨场景    
Abstract:

A new foreground segmentation method called self-updating co-occurrence pixel-block model (SU-CPB) was proposed to solve the problem of co-occurrence pixel-block model (CPB). The segmentation result of STAM was used as a reference, by introducing supervised spatio-temporal attention model (STAM) that has been trained in large-scale training data. Three methods including a pixel-block dynamic selection method, replacement of broken pairs and calculation of the foreground similarities were proposed. The pixel-block pairs were self-updated online with these methods, and the problem of the CPB model performance degradation caused by lack of updating capability was solved. The capability of foreground segmentation across scenes was possessed. Experimental results show that this method performs better than CPB model in all scenes, and is significantly better than STAM, CPB and other methods participating in comparison under the Wallflower and LIMU datasets without training by STAM.

Key words: foreground segmentation    pixel spatial relation    spatio-temporal attention model (STAM)    online self-updating    cross-scene
收稿日期: 2019-11-02 出版日期: 2020-12-31
CLC:  TP 391  
基金资助: 国家重点研发计划资助项目(2017YFB0802300);国防科技创新特区资助项目
作者简介: 梁栋(1985—),男,博士,从事模式识别与计算机视觉研究. orcid.org/0000-0003-2784-3449. E-mail: liangdong@nuaa.edu.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章  
梁栋
刘昕宇
潘家兴
孙涵
周文俊
金子俊一

引用本文:

梁栋,刘昕宇,潘家兴,孙涵,周文俊,金子俊一. 动态背景下基于自更新像素共现的前景分割[J]. 浙江大学学报(工学版), 2020, 54(12): 2405-2413.

Dong LIANG,Xin-yu LIU,Jia-xing PAN,Han SUN,Wen-jun ZHOU,Shun’ichi KANEKO. Foreground segmentation under dynamic background based on self-updating co-occurrence pixel. Journal of ZheJiang University (Engineering Science), 2020, 54(12): 2405-2413.

链接本文:

http://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2020.12.015        http://www.zjujournals.com/eng/CN/Y2020/V54/I12/2405

图 1  SU-CPB方法框架
图 2  CPB模型工作方式
图 3  STAM模型网络结构
图 4  支持块的动态选择模型流程图
图 5  模型选择效果演示
图 6  失效支持块的替换
参数 设置值
支持块数量 20
候选支持块数量 10
高斯模型阈值 2.5
相关性决策阈值 0.5
相似度决策阈值 0.8
表 1  SU-CPB中的各项参数设置
序号 算法 F-measure
BDW BSL CJT DBG IOM SHD THM TBL LFR NVD PTZ
1 SU-CPB 0.867 0.907 0.853 0.924 0.760 0.910 0.969 0.895 0.449 0.558 0.753
2 CPB[17] 0.475 0.519 0.597 0.477 0.348 0.581 0.372 0.459 0.170 0.277 0.161
3 SuBSENSE[6] 0.862 0.950 0.815 0.818 0.657 0.865 0.817 0.779 0.645 0.560 0.348
4 KDE[3] 0.757 0.909 0.572 0.596 0.409 0.803 0.742 0.448 0.548 0.437 0.037
5 GMM[2] 0.738 0.825 0.597 0.633 0.521 0.732 0.662 0.466 0.537 0.410 0.152
6 BMOG[8] 0.784 0.830 0.749 0.793 0.529 0.840 0.635 0.693 0.610 0.498 0.235
7 SGSM-BS[11] 0.856 0.950 0.820 0.848 0.819 0.890 0.850 0.850 0.750 0.510 ?
8 STAM[22] 0.970 0.989 0.899 0.948 0.916 0.966 0.991 0.933 0.668 0.710 0.865
9 DeepBS[9] 0.830 0.958 0.899 0.876 0.610 0.930 0.758 0.846 0.600 0.584 0.313
10 CascadeCNN[12] 0.943 0.979 0.976 0.966 0.851 0.941 0.896 0.911 0.837 0.897 0.917
11 DPDL[13] 0.869 0.969 0.866 0.869 0.876 0.936 0.838 0.764 0.708 0.611 0.609
12 FgSegNet[14] 0.984 0.998 0.995 0.994 0.993 0.995 0.992 0.978 0.956 0.978 0.989
表 2  不同方法的CDNet2014数据集F-measure对比
图 7  复杂场景下不同方法的检测效果对比
场景 F-measure
SU-CPB STAM[22]
PETS2006 0.957 0 0.956 3
traffic 0.835 0 0.834 9
fountain02 0.934 0 0.933 5
abandoned box 0.820 6 0.812 3
parking 0.764 1 0.763 3
表 3  特定场景下SU-CPB与STAM 的 F-measure对比
场景 F-measure
SU-CPB STAM[22] DeepBS[9] Cascade CNN[12] FgSeg-Net[14] CPB[17] SuBSENSE[6] GMM[2] PBAS[32]
Bootstrap 0.756 0 0.741 4 0.747 9 0.523 8 0.358 7 0.651 8 0.419 2 0.530 6 0.285 7
Camouflage 0.688 4 0.736 9 0.985 7 0.677 8 0.121 0 0.611 2 0.953 5 0.830 7 0.892 2
Fg Aperture 0.942 0 0.829 2 0.658 3 0.793 5 0.411 9 0.590 0 0.663 5 0.577 8 0.645 9
Light Switch 0.909 7 0.909 0 0.611 4 0.588 3 0.681 5 0.715 7 0.320 1 0.229 6 0.221 2
Time of Day 0.794 9 0.342 9 0.549 4 0.377 1 0.422 2 0.756 4 0.710 7 0.720 3 0.487 5
Waving Trees 0.666 5 0.532 5 0.954 6 0.287 4 0.345 6 0.703 3 0.959 7 0.976 7 0.842 1
Overall 0.792 9 0.682 0 0.751 2 0.541 3 0.390 2 0.671 4 0.671 1 0.644 3 0.562 4
表 4  Wallflower数据集各场景中不同方法的F-measure对比
场景 Specifity
SU-CPB STAM[22] CascadeCNN[12] FgSegNet[14] CPB[17]
Moved Object 0.997 7 0.994 9 0.773 6 0.847 0 0.892 2
表 5  Wallflower数据集Moved Object场景中不同方法的Specifity对比
场景 F-measure
SU-CPB STAM[22] CascadeCNN[12] FgSegNet[14] CPB[17]
Camera Parameter 0.748 4 0.674 2 0.102 5 0.266 8 0.654 5
Intersection 0.767 2 0.623 7 0.045 3 0.142 8 0.677 8
Light Switch 0.821 1 0.095 3 0.027 7 0.041 4 0.663 3
Overall 0.778 9 0.464 4 0.058 5 0.150 3 0.665 2
表 6  LIMU数据集各场景中不同方法的F-measure对比
图 8  不同方法在LIMU数据集不同场景中的检测效果对比
场景 F-measure
CPB[17] CPBDT SU-CPB
Camera Parameter 0.654 5 0.715 9 0.748 4
Intersection 0.677 8 0.690 8 0.767 2
Light Switch 0.663 3 0.642 5 0.821 1
Overall 0.665 2 0.683 1 0.778 9
表 7  LIMU数据集各场景中SU-CPB方法不同阶段的F-measure对比
1 VACAVANT A, CHATUAU T, WILHELM A, et al. A benchmark dataset for outdoor foreground/background extraction[C]// Asian Conference on Computer Vision. [S. l.]: Springer, 2012: 291-300.
2 STAUFFER C, GRIMSON W E L. Adaptive background mixture models for real-time tracking [C]// IEEE Computer Society Conference on Computer Vision and Pattern Recognition. [S. l.]: IEEE, 1999: 246-252.
3 ELGAMMAL A, DURAISWAMI R, HARWOOD D, et al Background and foreground modeling using nonparametric kernel density estimation for visual surveillance[J]. Proceedings of the IEEE, 2002, 90 (7): 1151- 1163
doi: 10.1109/JPROC.2002.801448
4 JODOIN P M, MIGNOTTE M, KONRAD J Statistical background subtraction using spatial cues[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2007, 17 (12): 1758- 1763
doi: 10.1109/TCSVT.2007.906935
5 BARNICH O, DROOGENBROECK M V ViBe: a universal background subtraction algorithm for video sequences[J]. IEEE Transactions on Image Processing, 2011, 20 (6): 1709- 1724
doi: 10.1109/TIP.2010.2101613
6 ST-CHARLES P L, BILODEAU G A, BERGEVIN R SuBSENSE: a universal change detection method with local adaptive sensitivity[J]. IEEE Transactions on Image Processing, 2014, 24 (1): 359- 373
7 LIANG D, KANEKO S, HASHIMOTO M, et al Co-occurrence probability-based pixel pairs background model for robust object detection in dynamic scenes[J]. Pattern Recognition, 2015, 48 (4): 1374- 1390
doi: 10.1016/j.patcog.2014.10.020
8 MARTINS I, CARVALHO P, CORTE-REAL, et al BMOG: boosted Gaussian mixture model with controlled complexity for background subtraction[J]. Pattern Analysis and Applications, 2018, 21 (3): 641- 654
doi: 10.1007/s10044-018-0699-y
9 BRAHAM M, DROOGENBROECK M V. Deep background subtraction with scene-specific convolutional neural networks [C]// 2016 International Conference on Systems, Signals and Image Processing. [S. l.]: IEEE, 2016.
10 BABAEE M, DINH D T, RIGOLL G. A deep convolutional neural network for background subtraction [EB/OL]. [2019-09-30]. https://arxiv.org/pdf/1702.01731.pdf.
11 SHI G, HUANG T, DONG W, et al Robust foreground estimation via structured gaussian scale mixture modeling[J]. IEEE Transactions on Image Processing, 2018, 27 (10): 4810- 4824
doi: 10.1109/TIP.2018.2845123
12 WANG Y, LUO Z, JODOIN P, et al Interactive deep learning method for segmenting moving objects[J]. Pattern Recognition Letters, 2017, 96: 66- 75
13 ZHAO C, CHAM T, REN X, et al. Background subtraction based on deep pixel distribution learning [C]// 2018 IEEE International Conference on Multimedia and Expo. [S. l.]: IEEE, 2018: 1-6.
14 LIM L A, KELES H Y Foreground segmentation using convolutional neural networks for multiscale feature encoding[J]. Pattern Recognition Letters, 2018, 112: 256- 262
doi: 10.1016/j.patrec.2018.08.002
15 LIM L A, KELES H Y Learning multi-scale features for foreground segmentation[J]. Pattern Analysis and Applications, 2019, 23 (3): 1369- 1380
16 QIU M, LI X A fully convolutional encoder-decoder spatial-temporal network for real-time background subtraction[J]. IEEE Access, 2019, 7: 85949- 85958
17 ZHOU W, KANEKO S, LIANG D, et al Background subtraction based on co-occurrence pixel-block pairs for robust object detection in dynamic scenes[J]. IIEEJ Transactions on Image Electronics and Visual Computing, 2018, 5 (2): 146- 159
18 ZHOU W, KANEKO S, HASHIMOTO M, et al. A co-occurrence background model with hypothesis on degradation modification for object detection in strong background changes [C]// 2018 24th International Conference on Pattern Recognition. [S. l.]: IEEE, 2018: 1743-1748.
19 ZHOU W, KANEKO S, HASHIMOTO M, et al Foreground detection based on co-occurrence background model with hypothesis on degradation modi?cation in dynamic scenes[J]. Signal Processing, 2019, 160: 66- 79
doi: 10.1016/j.sigpro.2019.02.021
20 ZHOU W, KANEKO S, SATOH Y, et al. Co-occurrence based foreground detection with hypothesis on degradation modification in severe imaging conditions [C] // Proceedings of JSPE Semestrial Meeting 2018 JSPE Autumn Conference. [S. l.]: JSPE, 2018: 624-625.
21 ZHAO X, SATOH Y, TAKAUJI H, et al Object detection based on a robust and accurate statistical multi-point-pair model[J]. Pattern Recognition, 2011, 44 (6): 1296- 1311
doi: 10.1016/j.patcog.2010.11.022
22 LIANG D, PAN J, SUN H, et al Spatio-temporal attention model for foreground detection in cross-scene surveillance videos[J]. Sensors, 2019, 19 (23): 5142
doi: 10.3390/s19235142
23 LAROCHELLE H, HINTON G. Learning to combine foveal glimpses with a third-order boltzmann machine [C]// Advances in Neural Information Processing Systems 23: Conference on Neural Information Processing Systems A Meeting Held December. [S. l.]: Curran Associates Inc, 2010: 1243–1251.
24 KIM J, LEE S, KWAK D, et al. Multimodal residual learning for visual QA [C]// Neural Information Processing Systems. [S. l.]: MIT Press, 2016: 361-369.
25 MNIH V, HEESS N, GRAVES A. Recurrent models of visual attention [C]// Neural Information Processing Systems. [S. l.]: MIT Press, 2014, 2: 2204-2212.
26 XU K, BA J, KIROS R, et al Show, attend and tell: neural image caption generation with visual attention[J]. International Conference on Machine Learning, 2015, 3: 2048- 2057
27 LI H, XIONG P, AN J, et al. Pyramid attention network for semantic segmentation [EB/OL]. [2019-09-30]. https://arxiv.org/pdf/1805.10180.pdf.
28 Liu C. Beyond pixels: exploring new representations and applications for motion analysis [D]. Cambridge: MIT, 2009.
29 GOYRTTE N, JODOIN P M, PORIKLI F, et al. Changedetection. net: a new change detection benchmark dataset [C]// 2012 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops. [S. l.]: IEEE, 2012: 1-8.
30 TOYAMMA K, KRUMM J, BRUMITT B, et al. Wallflower: principles and practice of background maintenance [C]// Proceedings of the Seventh IEEE International Conference on computer vision. [S. l.]: IEEE, 1999: 255-261.
31 Laboratory for image and media understanding [DB/OL]. [2019-09-30]. http://limu.ait.kyushu-u.ac.jp/dataset/en/.
[1] 郑守国,张勇德,谢文添,樊虎,王青. 基于数字孪生的飞机总装生产线建模[J]. 浙江大学学报(工学版), 2021, 55(5): 843-854.
[2] 张师林,马思明,顾子谦. 基于大边距度量学习的车辆再识别方法[J]. 浙江大学学报(工学版), 2021, 55(5): 948-956.
[3] 宋鹏,杨德东,李畅,郭畅. 整体特征通道识别的自适应孪生网络跟踪算法[J]. 浙江大学学报(工学版), 2021, 55(5): 966-975.
[4] 蔡君,赵罡,于勇,鲍强伟,戴晟. 基于点云和设计模型的仿真模型快速重构方法[J]. 浙江大学学报(工学版), 2021, 55(5): 905-916.
[5] 王虹力,郭斌,刘思聪,刘佳琪,仵允港,於志文. 边端融合的终端情境自适应深度感知模型[J]. 浙江大学学报(工学版), 2021, 55(4): 626-638.
[6] 张腾,蒋鑫龙,陈益强,陈前,米涛免,陈彪. 基于腕部姿态的帕金森病用药后开-关期检测[J]. 浙江大学学报(工学版), 2021, 55(4): 639-647.
[7] 郑英杰,吴松荣,韦若禹,涂振威,廖进,刘东. 基于目标图像FCM算法的地铁定位点匹配及误报排除方法[J]. 浙江大学学报(工学版), 2021, 55(3): 586-593.
[8] 雍子叶,郭继昌,李重仪. 融入注意力机制的弱监督水下图像增强算法[J]. 浙江大学学报(工学版), 2021, 55(3): 555-562.
[9] 于勇,薛静远,戴晟,鲍强伟,赵罡. 机加零件质量预测与工艺参数优化方法[J]. 浙江大学学报(工学版), 2021, 55(3): 441-447.
[10] 胡惠雅,盖绍彦,达飞鹏. 基于生成对抗网络的偏转人脸转正[J]. 浙江大学学报(工学版), 2021, 55(1): 116-123.
[11] 陈杨波,伊国栋,张树有. 基于点云特征对比的曲面翘曲变形检测方法[J]. 浙江大学学报(工学版), 2021, 55(1): 81-88.
[12] 段有康,陈小刚,桂剑,马斌,李顺芬,宋志棠. 基于相位划分的下肢连续运动预测[J]. 浙江大学学报(工学版), 2021, 55(1): 89-95.
[13] 张太恒,梅标,乔磊,杨浩杰,朱伟东. 纹理边界引导的复合材料圆孔检测方法[J]. 浙江大学学报(工学版), 2020, 54(12): 2294-2300.
[14] 晋耀,张为. 采用Anchor-Free网络结构的实时火灾检测算法[J]. 浙江大学学报(工学版), 2020, 54(12): 2430-2436.
[15] 叶刚,李毅波,马逐曦,成杰. 基于ViBe的端到端铝带表面缺陷检测识别方法[J]. 浙江大学学报(工学版), 2020, 54(10): 1906-1914.