基于光流与多尺度上下文的图像序列运动遮挡检测

冯诚; 张聪炫; 陈震; 李兵; 黎明

doi:10.16383/j.aas.c210324

基于光流与多尺度上下文的图像序列运动遮挡检测

doi: 10.16383/j.aas.c210324 cstr: 32138.14.j.aas.c210324

冯诚^1,,
张聪炫^{1, 2,},
陈震^1,,
李兵^2,,
黎明^3,

1.
南昌航空大学测试与光电工程学院南昌 330063
2.
中国科学院自动化研究所模式识别国家重点实验室北京 100190
3.
南昌航空大学信息工程学院南昌 330063

基金项目: 国家重点研发计划(2020YFC2003800), 国家自然科学基金(61866026, 61772255, 62222206), 江西省杰出青年人才计划(20192BCB23011), 江西省自然科学基金重点项目(20202ACB214007), 江西省优势科技创新团队(20165BCB19007)资助

详细信息

作者简介:
冯诚：南昌航空大学测试与光电工程学院硕士研究生. 主要研究方向为计算机视觉. E-mail: fengcheng00016@163.com

张聪炫：南昌航空大学测试与光电工程学院教授. 2014年获得南京航空航天大学博士学位. 主要研究方向为图像处理与计算机视觉. 本文通信作者. E-mail: zcxdsg@163.com

陈震：南昌航空大学测试与光电工程学院教授. 2003年获得西北工业大学博士学位. 主要研究方向为图像处理与计算机视觉. E-mail: dr_chenzhen@163.com

李兵：中国科学院自动化研究所模式识别国家重点实验室研究员. 2009年获得北京交通大学博士学位. 主要研究方向为视频内容理解, 多媒体内容安全. E-mail: bli@nlpr.ia.ac.cn

黎明：南昌航空大学信息工程学院教授. 1997年获得南京航空航天大学博士学位. 主要研究方向为图像处理, 人工智能. E-mail: liming@nchu.edu.com

计量
- 文章访问数: 844
- HTML全文浏览量: 464
- PDF下载量: 86
- 被引次数: 0
出版历程
- 收稿日期: 2021-04-15
- 录用日期: 2021-07-02
- 网络出版日期: 2021-08-31
- 刊出日期: 2024-09-19

Occlusion Detection Based on Optical Flow and Multiscale Context

FENG Cheng^1
,,
ZHANG Cong-Xuan^{1, 2
,},
CHEN Zhen^1
,,
LI Bing^2
,,
LI Ming^3
,

1.
School of Measuring and Optical Engineering, Nanchang Hangkong University, Nanchang 330063
2.
National Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190
3.
School of Information Engineering, Nanchang Hangkong University, Nanchang 330063

Funds: Supported by National Key Research and Development Program of China (2020YFC2003800), National Natural Science Foundation of China (61866026, 61772255, 62222206), Outstanding Young Scientist Project of Jiangxi Province (20192BCB23011), Natural Science Foundation of Jiangxi Province (20202ACB214007), and Advantage Subject Team of Jiangxi Province (20165BCB19007)

More Information

Author Bio:
FENG Cheng　Master student at the School of Measuring and Optical Engineering, Nanchang Hangkong University. His main research interest is computer vision

ZHANG Cong-Xuan　 Professor at the School of Measuring and Optical Engineering, Nanchang Hangkong University. He received his Ph.D. degree from Nanjing University of Aeronautics and Astronautics in 2014. His research interest covers image processing and computer vision. Corresponding author of this paper

CHEN Zhen　Professor at the School of Measuring and Optical Engineering, Nanchang Hangkong University. He received his Ph.D. degree from Northwestern Polytechnical University in 2003. His research interest covers image processing and computer vision

LI Bing　Professor at the National Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences. He received his Ph.D. degree from Beijing Jiaotong University in 2009. His research interest covers video understanding and multimedia content security

LI Ming　Professor at the School of Information Engineering, Nanchang Hangkong University. He received his Ph.D. degree from Nanjing University of Aeronautics and Astronautics in 1997. His research interest covers image processing and artificial intelligence

摘要

摘要: 针对非刚性运动和大位移场景下运动遮挡检测的准确性与鲁棒性问题, 提出一种基于光流与多尺度上下文的图像序列运动遮挡检测方法. 首先, 设计基于扩张卷积的多尺度上下文信息聚合网络, 通过图像序列多尺度上下文信息获取更大范围的图像特征; 然后, 采用特征金字塔构建基于多尺度上下文与光流的端到端运动遮挡检测网络模型, 利用光流优化非刚性运动和大位移区域的运动检测遮挡信息; 最后, 构造基于运动边缘的网络模型训练损失函数, 获取准确的运动遮挡边界. 分别采用MPI-Sintel和KITTI测试数据集对所提方法与现有的代表性方法进行实验对比与分析. 实验结果表明, 所提方法能够有效提高运动遮挡检测的准确性和鲁棒性, 尤其在非刚性运动和大位移等困难场景下具有更好的遮挡检测鲁棒性.
- 图像序列 /
- 遮挡检测 /
- 深度学习 /
- 多尺度上下文 /
- 非刚性运动
Abstract: In order to improve the accuracy and robustness of occlusion detection under non-rigid motion and large displacements, we propose an occlusion detection method of image sequence motion based on optical flow and multiscale context. First, we design a multiscale context information aggregation network based on dilated convolution which obtains a wider range of image features through multiscale context information of image sequence. Then, we construct an end-to-end motion occlusion detection network model based on multiscale context and optical flow using feature pyramid, utilize the optical flow to optimize the performance of occlusion detection in areas of non-rigid motion and large displacements region. Finally, we present a novel motion edge training loss function to obtain the accurate motion occlusion boundary. We compare and analysis our method with the existing representative approaches by using the MPI-Sintel datasets and KITTI datasets, respectively. The experimental results show that the proposed method can effectively improve the accuracy and robustness of motion occlusion detection, especially gains the better occlusion detection robustness under non-rigid motion and large displacements.
- Image sequence /
- occlusion detection /
- deep learning /
- multiscale context /
- non-rigid motion

HTML全文

图 1 上下文网络结构示意图

Fig. 1 Structure diagram of context network

下载: 全尺寸图片幻灯片

图 2 常见的感受野扩张网络结构示意图

Fig. 2 Structure diagram of common receptive field expansion

下载: 全尺寸图片幻灯片

图 3 多尺度上下文信息聚合网络结构示意图

Fig. 3 Structure diagram of multiscale context information aggregation network

下载: 全尺寸图片幻灯片

图 4 遮挡检测网络结构示意图

Fig. 4 Structure diagram of occlusion detection network

下载: 全尺寸图片幻灯片

图 5 基于光流和多尺度上下文信息的遮挡检测模型结构

Fig. 5 The structure of the occlusion detection model based on optical flow and multiscale context information

下载: 全尺寸图片幻灯片

图 6 本文方法和IRR-PWC方法遮挡检测结果对比

Fig. 6 Comparison of occlusion detection results between our method and IRR-PWC method

下载: 全尺寸图片幻灯片

图 7 MPI-Sintel数据集非刚性运动与大位移序列遮挡检测结果对比图. 从左至右分别是:alley_2、ambush_2、market_6以及temple_2序列

Fig. 7 Comparison results of occlusion detection between non-rigid motion and large displacement sequences on MPI-Sintel dataset. From left to right are alley_2, ambush_2, market_6, and temple_2 sequence

下载: 全尺寸图片幻灯片

图 8 各个遮挡检测方法在KITTI数据集上的遮挡检测结果对比图. 从左至右分别是输入图像和Unflow、Back2Future、MaskFlownet、IRR-PWC以及本文方法的运动遮挡检测图

Fig. 8 Comparison of occlusion detection results of each occlusion detection method on KITTI dataset. From left to right are the input image, Unflow, Back2Future, MaskFlownet, IRR-PWC and our method

下载: 全尺寸图片幻灯片

图 9 利用光流真实值生成的运动遮挡掩膜部分示例图$(N=3) $

Fig. 9 Examples of motion occlusion mask generated by ground truth of optical flow $(N=3 )$

下载: 全尺寸图片幻灯片

图 10 各消融模型可视化结果对比图

Fig. 10 Comparison of visualization results of each ablation model

下载: 全尺寸图片幻灯片

表 1 MPI-Sintel数据集平均F1分数对比结果

Table 1 Comparison of average F1 score on MPI-Sintel dataset

对比方法	多帧	类型	clean	final
Unflow^[24]		传统方法	0.28	0.27
Back2Future^[25]	√	无监督学习	0.49	0.44
MaskFlownet^[27]		无监督学习	0.37	0.36
IRR-PWC^[26]		监督学习	0.71	0.67
本文方法		监督学习	0.75	0.72

下载: 导出CSV

表 2 MPI-Sintel数据集平均漏检率与误检率对比结果(%)

Table 2 Comparison of average omission rate and false rate on MPI-Sintel dataset (%)

对比方法	clean		final
对比方法	OR	FR	OR	FR
Unflow^[24]	1.96	18.32	1.94	20.51
Back2Future^[25]	5.03	2.75	5.08	2.96
MaskFlownet^[27]	5.77	1.37	5.76	1.72
IRR-PWC^[26]	1.98	0.96	2.84	1.29
本文方法	1.85	0.83	2.31	1.08

下载: 导出CSV

表 3 非刚性运动与大位移图像序列运动遮挡检测平均F1分数对比结果

Table 3 Comparison of average F1 scores of motion occlusion detection between non-rigid motion and large displacement image sequences

对比方法	clean				final
对比方法	alley_2	ambush_2	market_6	temple_2	alley_2	ambush_2	market_6	temple_2
Unflow^[24]	0.414 9	0.431 3	0.433 0	0.324 3	0.405 7	0.392 0	0.449 9	0.312 0
Back2Future^[25]	0.681 6	0.588 8	0.629 0	0.271 2	0.675 6	0.519 9	0.623 9	0.268 3
MaskFlownet^[27]	0.505 7	0.540 3	0.466 0	0.383 8	0.503 9	0.408 5	0.473 5	0.350 8
IRR-PWC^[26]	0.870 9	0.917 2	0.815 5	0.740 4	0.877 0	0.780 9	0.802 3	0.690 5
本文方法	0.881 1	0.921 6	0.830 4	0.774 7	0.876 4	0.795 9	0.810 6	0.710 3
注: 加粗字体表示各列最优结果.

下载: 导出CSV

表 4 不同方法的时间消耗对比

Table 4 Comparison of time consumption of different methods

对比方法	多帧输入	类型	运行时间(s)
Unflow^[24]		传统方法	0.13
Back2Future^[25]	√	无监督学习	0.13
MaskFlownet^[27]		无监督学习	0.10
IRR-PWC^[26]		监督学习	0.18
本文方法		监督学习	0.19
注: 加粗字体表示评价最优值.

下载: 导出CSV

表 5 MPI-Sintel全图像序列平均F1分数对比

Table 5 Comparison of average F1 scores of whole image sequence on MPI-Sintel

模型类型	MPI-Sintel 训练数据集
模型类型	clean	final	运行时间(s)	训练时间(d)
全模型	0.75	0.72	0.19	13
去除多尺度上下文网络	0.72	0.68	0.18	12
去除边缘损失函数	0.74	0.71	0.19	13
注: 加粗字体表示评价最优值.

下载: 导出CSV

表 6 MPI-Sintel全图像序列在不同运动边界区域内的平均F1分数对比

Table 6 Comparison of average F1 scores of whole image sequence in different motion boundary regions on MPI-Sintel

模型类型	MPI-Sintel 训练数据集
	clean				final
	$N=1 $	$N=3 $	$N=5 $	$N=10 $	$N=1 $	$N=3 $	$N=5 $	$N=10 $
全模型	0.63	0.67	0.69	0.71	0.59	0.62	0.64	0.67
去除多尺度上下文网络	0.59	0.62	0.65	0.67	0.55	0.59	0.61	0.63
去除边缘损失函数	0.60	0.64	0.67	0.69	0.56	0.60	0.62	0.64
注: 加粗字体表示评价最优值.

下载: 导出CSV

参考文献(34)

[1]	张世辉, 何琦, 董利健, 杜雪哲. 基于遮挡区域建模和目标运动估计的动态遮挡规避方法. 自动化学报, 2019, 45(4): 771−786 Zhang Shi-Hui, He Qi, Dong Li-Jian, Du Xue-Zhe. Dynamic occlusion avoidance approach by means of occlusion region model and object motion estimation. Acta Automatica Sinica, 2019, 45(4): 771−786
[2]	Yu C, Bo Y, Bo W, Yan W D, Robby T. Occlusion-aware networks for 3D human pose estimation in video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, South Korea: IEEE, 2019. 723−732
[3]	张聪炫, 陈震, 熊帆, 黎明, 葛利跃, 陈昊. 非刚性稠密匹配大位移运动光流估计. 电子学报, 2019, 47(6): 1316−1323 doi: 10.3969/j.issn.0372-2112.2019.06.019 Zhang Cong-Xuan, Chen Zhen, Xiong Fan, Li Ming, Ge Li-Yue, Chen Hao. Large displacement motion optical flow estimation with non-rigid dense patch matching. Acta Electronica Sinica, 2019, 47(6): 1316−1323 doi: 10.3969/j.issn.0372-2112.2019.06.019
[4]	姚乃明, 郭清沛, 乔逢春, 陈辉, 王宏安. 基于生成式对抗网络的鲁棒人脸表情识别. 自动化学报, 2018, 44(5): 865−877 Yao Nai-Ming, Guo Qing-Pei, Qiao Feng-Chun, Chen Hui, Wang Hong-An. Robust facial expression recognition with generative adversarial networks. Acta Automatica Sinica, 2018, 44(5): 865−877
[5]	Pan J Y, Bo H. Robust occlusion handling in object tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Minneapolis, USA: IEEE, 2007. 1−8
[6]	刘鑫, 许华荣, 胡占义. 基于 GPU 和 Kinect 的快速物体重建. 自动化学报, 2012, 38(8): 1288−1297 Liu Xin, Xu Hua-Rong, Hu Zhan-Yi. GPU based fast 3D-object modeling with Kinect. Acta Automatica Sinica, 2012, 38(8): 1288−1297
[7]	张聪炫, 陈震, 黎明. 单目图像序列光流三维重建技术研究综述. 电子学报, 2016, 44(12): 3044−3052 doi: 10.3969/j.issn.0372-2112.2016.12.033 Zhang Cong-Xuan, Chen Zhen, Li Ming. Review of the 3D reconstruction technology based on optical flow of monocular image sequence. Acta Electronica Sinica, 2016, 44(12): 3044−3052 doi: 10.3969/j.issn.0372-2112.2016.12.033
[8]	Bailer C, Taetz B, Stricker D. Flow fields: Dense correspondence fields for highly accurate large displacement optical flow estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(8): 1879−1892 doi: 10.1109/TPAMI.2018.2859970
[9]	Wolf L, Gadot D. PatchBatch: A batch augmented loss for optical flow. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 4236−4245
[10]	Li Y S, Song R, Hu Y L. Efficient coarse-to-fine patch match for large displacement optical flow. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 5704−5712
[11]	Menze M, Heipke C, Geiger A. Discrete optimization for optical flow. In: Proceedings of the German Conference on Pattern Recognition (GCPR). Aachen, Germany: Springer Press, 2015. 16−28
[12]	Chen Q F, Koltun V. Full flow: Optical flow estimation by global optimization over regular grids. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 4706−4714
[13]	Guney F, Geiger A. Deep discrete flow. In: Proceedings of the Asian Conference on Computer Vision (ACCV). Taipei, China: Springer Press, 2016. 207−224
[14]	Hur J, Roth S. Joint optical flow and temporally consistent semantic segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). Amsterdam, The Netherlands: Springer, 2016. 163−177
[15]	Ince S, Konrad J. Occlusion-aware optical flow estimation. IEEE Transactions on Image Process, 2008, 17(8): 1443−1451 doi: 10.1109/TIP.2008.925381
[16]	Sun D Q, Liu C, Pfister H. Local layering for joint motion estimation and occlusion detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Columbus, USA: IEEE, 2014. 1098−1105
[17]	Sun D Q, Sudderth E B, Black M J. Layered image motion with explicit occlusions, temporal consistency, and depth ordering. In: Proceedings of the 24th International Conference on Neural Information Processing Systems (NIPS). Vancouver, Canada: Curran Associates Inc., 2010. 2226−2234
[18]	Vogel C, Roth S, Schindler K. View-consistent 3D scene flow estimation over multiple frames. In: Proceedings of the European Conference on Computer Vision (ECCV). Zurich, Switzerland: Springer Press, 2014. 263−278
[19]	Zanfir A, Sminchisescu C. Large displacement 3D scene flow with occlusion reasoning. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: IEEE, 2015. 4417−4425
[20]	Zhang C X, Chen Z, Wang M R, Li M, Jiang S F. Robust non-local TV-L¹ optical flow estimation with occlusion detection. IEEE Transactions on Image Process, 2017, 26(8): 4055−4067 doi: 10.1109/TIP.2017.2712279
[21]	张聪炫, 陈震, 汪明润, 黎明, 江少锋. 基于光流与Delaunay三角网格的图像序列运动遮挡检测. 电子学报, 2018, 46(2): 479−485 doi: 10.3969/j.issn.0372-2112.2018.02.030 Zhang Cong-Xuan, Chen Zhen, Wang Ming-Run, Li Ming, Jiang Shao-Feng. Motion occlusion detecting from image sequence based on optical flow and Delaunay triangulation. Acta Electronica Sinica, 2018, 46(2): 479−485 doi: 10.3969/j.issn.0372-2112.2018.02.030
[22]	Kennedy R, Taylor C J. Optical flow with geometric occlusion estimation and fusion of multiple frames. In: Proceedings of International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition (EMMCVPR). Hong Kong, China: IEEE, 2015. 364−377
[23]	Yu J J, Harley A W, Derpanis K G. Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In: Proceedings of the European Conference on Computer Vision (ECCV). Amsterdam, The Netherlands: Springer, 2016. 3−10
[24]	Meister S, Hur J, Roth S. UnFlow: Unsupervised learning of optical flow with a bidirectional census loss. In: Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI). San Francisco, USA: AAAI, 2017. 7251−7259
[25]	Janai J, Güney F, Ranjan A, Black M, Geiger A. Unsupervised learning of multi-frame optical flow with occlusions. In: Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: Springer, 2018. 713−731
[26]	Hur J, Roth S. Iterative residual refinement for joint optical flow and occlusion estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE, 2019. 5747−5756
[27]	Zhao S Y, Sheng Y L, Dong Y, Chang E I C, Xu Y. MaskFlownet: Asymmetric feature matching with learnable occlusion mask. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Virtual Event: IEEE, 2020. 6277−6286
[28]	Yang M K, Yu K, Zhang C, Li Z W, Yang K Y. DenseASPP for semantic segmentation in street scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City, USA: IEEE, 2018. 3684−3692
[29]	Mehta S, Rastegari M, Caspi A, Shapiro L, Hajishirzi H. ESPNet: Efficient spatial pyramid of dilated convolutions for semantic segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: Springer, 2018. 561−580
[30]	Szegedy C, Liu W, Jia Y Q, Sermanet P, Reed S, Anguelov D, et al. Going deeper with convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Boston, USA: IEEE, 2015. 1−9
[31]	Chen L C, Zhu Y K, Papandreou G, Schroff F, Adam H. Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). Munich, Germany: Springer, 2018. 833−851
[32]	Yu F, Koltun V. Multi-scale context aggregation by dilated convolutions [Online], available: https://arxiv.org/abs/1511.07122, Apr 30, 2016
[33]	Butler D J, Wulff J, Stanley G B, Black M J. A naturalistic open source movie for optical flow evaluation. In: Proceedings of the European Conference on Computer Vision (ECCV). Florence, Italy: Springer, 2012. 611−625
[34]	Menze M, Geiger A. Object scene flow for autonomous vehicles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Boston, USA: IEEE, 2015. 2061−3070