• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Abnormal activity detection for surveillance video synopsis①

    2016-12-05 01:28:10ZhuXiaobin祝曉斌WangQianLiHaishengGuoXiaoxiaXiYanShenYang
    High Technology Letters 2016年2期

    Zhu Xiaobin (祝曉斌), Wang Qian, Li Haisheng, Guo Xiaoxia, Xi Yan, Shen Yang

    (*School of Computer and Information Engineering, Beijing Technology and Business University, Beijing 100048, P.R.China) (**Academy of Broadcasting Science, SAPPRFT, Beijing 100866, P.R.China)

    ?

    Abnormal activity detection for surveillance video synopsis①

    Zhu Xiaobin (祝曉斌)②*, Wang Qian*, Li Haisheng*, Guo Xiaoxia**, Xi Yan**, Shen Yang**

    (*School of Computer and Information Engineering, Beijing Technology and Business University, Beijing 100048, P.R.China) (**Academy of Broadcasting Science, SAPPRFT, Beijing 100866, P.R.China)

    Video synopsis is an effective and innovative way to produce short video abstraction for huge video archives, while keeping the dynamic characteristic of activities in the original video. Abnormal activity, as the critical event, is always the main concern in video surveillance context. However, in traditional video synopsis, all the normal and abnormal activities are condensed together equally, which can make the synopsis video confused and worthless. In addition, the traditional video synopsis methods always neglect redundancy in the content domain. To solve the above-mentioned issues, a novel video synopsis method is proposed based on abnormal activity detection and key observation selection. In the proposed algorithm, activities are classified into normal and abnormal ones based on the sparse reconstruction cost from an atomically learned activity dictionary. And key observation selection using the minimum description length principle is conducted for eliminating content redundancy in normal activity. Experiments conducted in publicly available datasets demonstrate that the proposed approach can effectively generate satisfying synopsis videos.

    abnormal activity detection, key observation selection, sparse coding, minimum description length (MDL), video synopsis

    0 Introduction

    Security applications have great demands on efficient technologies for fast video browsing, retrieval or analysis, facing endlessly produced surveillance videos. Therefore, how to obtain short and comprehensive video abstractions becomes an urgent task in research domain. Video synopsis is an effective method, which makes the abstraction video greatly shorter than the original one by displaying the activities from different periods simultaneously.

    Video synopsis can eliminate redundancy in the spatial-temporal domain, and generate short video abstraction. However, the existing synopsis methods[1,2]still have the following limitations: They always tend to summarize all types of activities from input videos. In video surveillance context, people mainly concern with particular activities, especially abnormal activities. The traditional synopsis video will include lots of activities people are not really interested in; They always concentrate on eliminating redundancy in the spatial-temporal domain, while neglecting the redundancy in the content domain. Too many observations for activities can make the synopsis videos chaotic and less understandable.

    To address the above issues, a novel video synopsis approach is proposed based on abnormal activity detection and key observation selection. In the proposed algorithm, based on sparse coding framework, activities are classified into two types, namely abnormal type and normal type, based on which two synopsis videos are generated separately. Synopsis with abnormal activities is usually the main concern in surveillance context, while synopsis with normal activities is a complementary video. Because adjacent observations in an activity are always similar in action and appearance, thus key observation selection is adopted to eliminate content redundancy in synopsis video for normal activities. Section 1 overviews the related works. Section 2 elaborates the methodology of this work. The experiments are given in Section 3, and this study is concluded in Section 4.

    1 Related work

    Video abstraction can be broadly divided into three categories, namely videosummarization, video skimming, and video synopsis. Video summarization techniques try to provide a summary by creating shorter video remaining descriptivesections of the original video. Typically, these techniques adopt static representations such as key-frames[3]. Although summarization based on key-frame could greatly compress theoriginal video, it loses not only the dynamic nature of video but also meaningful video contents.

    Video skimming[4,5]aims to extract informative video segments from the original video to obtain a condensed summary video. Ref.[5] adopted long-term and short-term audiovisual tempo analyses to detect valuable substories of a video and combined them for video skimming. The skimming video is generated based on the selected scene periods. Although, video skimming method can generate relatively more coherent and expressive summary video than those key-frame based ones. However, people will tend to spend large amounts of time browsing video segments with little information.

    Video synopsis methods break the previous framework through rearranging spatial-temporal location of foreground objects to generate a new efficient summary video, while keeping the dynamic nature of the original video[1,6-11]. In Ref.[1], activities were represented by space-time tubes. Then, energy function comprised of collision cost, activity cost and temporal consistency cost, etc., is minimized using annealing, MRF, or greedy optimization method, yielding an abbreviated synopsis video for fast browsing. Ref.[6]formulated the synopsis video generation problem as a maximum posterior probability (MAP)estimation problem, where video objects are chronologically rearranged in realtime without pre-computing the complete trajectory of activities. In Ref.[9], Feng, et al. adopted an online content-aware approach to achieve efficient video condensation. In the above mentioned methods, all the activities of the input video are equally treated and summarized together. The generated synopsis video tends to comprise lots of activities people are not really interested in. In addition, great redundancy in the content domain is always neglected in the above video synopsis methods, which leads to collisions,and negative impaction on subjective effect of video synopsis. In Ref.[10], space-time worms were correlated with an user-specified query to identify actions of interest, which were then condensed by optimizing their temporal shift, allowing simultaneous display of multiple instances of relevant activity. Motivated by the works in Ref.[10] and Ref.[11], a novel framework is proposed for video synopsis, which can overcome the above limitations of the current methods.

    2 The proposed algorithm

    2.1 Framework

    The proposed framework is shown in Fig.1. Firstly, background subtraction followed by a graph-based tracking[12-14]is adopted to extract moving objectactivities. Then, sparse reconstruction cost is adopted to classify normal and abnormal activities based on dense trajectory. For normal activities, key observation selection is used to eliminate content redundancy with the minimum description length (MDL) principle. Finally, two synopsis videos are generated for normal and abnormal activities, respectively.

    Fig.1 The proposed framework

    2.2 Abnormal activity detection

    After the background/foreground segmentation, a multiple hypothesis tracker is used to track blobs[15]. Then, foreground blobs are grouped into targets by employing similar techniques as in Ref.[16]. Each activity is represented by a sequence of object masks in those frames. The space-time sequences of an object is deemed as a tube, and the central point of object is taken as motion trajectory. Motion trajectories of objects are widely used in abnormal activity detection. Sparse coding framework is suitable for model high-dimensional samples. Normal samples tend to generate sparse reconstruction coefficients with a small reconstruction cost, while abnormal one is dissimilar to any of the normal basis, thus generates a dense representation with a large reconstruction cost. Recent works[17]showed the power of sparse coding in detecting abnormal activities (events). In the proposed algorithm, abnormal activity is conducted based on the sparse coding framework.

    2.2.1 Feature extraction

    In Ref.[18], typical trajectories were modelled with hierarchical clustering method for identifying abnormal behavior. One limitation of trajectory-based approaches is that the detection performance greatly relies on the accuracy of foreground object extraction and tracking. The other is that single trajectory cannot well describe the overall motion of the corresponding object across the scene[19]. Dense trajectory is extensively applied to a variety of tasks, e.g., action recognition[20], and abnormal event detection[21], etc. It is capable of well describing object activities, even in complex and crowded scenes. So, dense trajectories are extracted using particle advection[22]for abnormal activity detection in the proposed algorithm. In Fig.2(a), the dotted lines denote the dense trajectories belonging to one object (encircled in the tube) across the scene.

    To describe the motion of an object, the multi-scale histogram of optical flow (MHOF) is adopted as the feature descriptor[17], as shown in Fig.2(b). The noise motion is firstly filtered with extremely large amplitude. MHOF hasK=64 bins including four scales, for more precisely preserve motion direction information and motion energy information. The first scale uses the first 16 bins to denote 16 directions with motion magnitude r≤T1, the second scale uses the next 16 bins with motion magnitude T1≤r≤T2, the third scale uses the next 16 with motion magnitude T2≤r≤T3, and the fourth scale uses the final 16 with motion magnitude T3≤r.

    Fig.2 The proposedabnormal activity detection algorithm

    2.2.2 Dictionary learning

    Because the local trajectory may be abnormal, while global trajectory is normal, as shown in Fig.2(c), the video is splited into several sub-regions and extract MHOF features in each region to train dictionary[23,24]for sparsely representing each feature. Given a training set of feature pool as B=[b1,b2,…,bk]∈Rm×N, where each column vector bi∈Rmdenotes a feature vector withm-dimension andNdenotes the total number of feature vectors. To find a set of basis D (it can be initialized byk-means) and a matrix of mixing coefficients W, B can be reconstructed by the weighted sum of computed basis D well. More formally, this problem can be formulated as

    (1)

    The efficient sparse coding algorithm is utilized as in Ref.[23]. The objective function is not convex in terms of all the variables jointly. Therefore, it is unrealistic to expect an algorithm to easily find the global optimal solution. An alternating optimization method is adopted to solve it.

    2.2.3 Sparse reconstruction cost

    With the dictionary dat hand, a test sample y can be classified as normal activity or not. As mentioned above, the feature of a normal sample can be constructed by only a few number of bases in the dictionary D, while an abnormal sample cannot. So, the sparse representation problem can be formulated as

    (2)

    This can be solved by the gradient based method described in Ref.[23]. The l2,1-norm is adopted for W during dictionary learning. Here, l1-norm is adopted for w. The l2,1-norm is a general version of the l1-norm in nature. Since if w is a one dimension vector, then ‖w‖2,1=‖w‖1. After that, the optimal reconstruction weight vector w*is got, the sparsity reconstruction cost (SRC)[17]can be computed as

    (3)

    And the test sample y will be detected as an abnormal activity, if the following criterion is satisfied:

    S>ε

    (4)

    where ε is a parameter that is set by cross-validation. It determines the sensitivity of classifying abnormal activity.

    2.3 Key observation selection

    Video synopsis method[1]provides an effective way for fast browsing activities by spatial-temporal rearranging them into a greatly condensed video. In typical scenarios, the activities always consist of numerous observations, resulting in collision and degradation of subjective effect in synopsis video. In addition, adjacent observations may be very similar in action and appearance. In light of these factors, a few representative observations are used, namely key observations, to depict the original behavior of normal behavior, which can greatly eliminate the redundancy in the content domain, and promote the efficiency of video synopsis. In Ref.[11],k-means clustering method was adopted to select a pre-defined number of key actions. However, the number of key actions can not be fixed for different objects, even in the same scenario.

    Different from Ref.[25], the key observations is extracted from every object instead of input video. The observations which have significant action are selected as the key ones, according to the proposed criteria. In the proposed algorithm, the difference between the sampled activities and the original activities (representativeness) is tried, while to is used minimized as small number of observations as possible (compressibility), as shown in Fig.3 (the points denote key observations, and the dotted line denotes the sampled trajectory). However, the representativeness and the compressibility are contradictory to each other. For example, if all the observations of the trajectory are chosen as the key ones, then the represent ativeness is maximized. In contrast, if only the starting and ending observations of the trajectory are chosen as the key ones, the representativeness is minimized, but the compressibility is maximized. Take the representativeness and compressibility into consideration, a data-driven method is adopted to select key observations by transforming it into an MDL optimization problem[26]for exploring an optimal selection.

    Fig.3 An example of key observation selection

    The description length (DL) in the proposed algorithm is computed as: L(D,S) =L(D|S)+ L(S), where S is the learned key observation selection solution, and D is the input trajectory (consisted of observations). L(D|S) is the number of bits required for encoding the data with the help of the key observation selection, while L(S) stands for that to encode the selection solution. The optimal key observation selection solutionSis the one that minimizes L(D, S) (MDL), namely MDL.L(S) and L(D|S) is computed as follows:

    (5)

    L(D|S)=

    (6)

    As mentioned above, it is needed to search the optimal key observation selection scheme that minimizes the DL. However, it is an NP-hard problem. Therefore, an approximate method is adopted by choosing a local optimum. As shown in Fig.4, if DL(OkOk+3

    Fig.4 An example of the approximate algorithm

    2.4 Synopsis video generation

    Video synopsis can be seen as an energy minimization optimization problem, and the energy includes the cost of objects collision, the cost of objects time inconsistency, the cost of objects lost, and so on. Following Ref.[1], the concepts of collision Ec, time consistency cost Et, and compression rate cost Elare introduced for generating lossless synopsis video in the proposed algorithm. The synopsis of abnormal activities and normal activities are conducted respectively using the same energy function, generating two videos for fast browsing. The energy function can be formulated as follows:

    E=argmin?ME(M)

    (7)

    (8)

    Algorithm1:SelectingkeyobservationsintrajectorybasedonMDL.Input:N,thenumberofobservationsOutput:trajectoryO'consistofkeyobservationsData:{Os,Os+1,…,Oe}O'1=Os;p=2;for{k=1;k

    3 Experimental evaluation

    In order to evaluate the performance of the proposed abnormal activity detection based surveillance video synopsis method, experiments are conducted on three real world testing videos, captured by the equipment in outdoor scenes. The first dataset (D1) targets at pedestrian activity, which consists of 31,530 frames, and the representative images are shown in Fig.5(a). The second dataset (D2) targets at vehicle surveillance of street scenario, which consists of 43,685 frames, and the representative images are shown in Fig.5(b). The third dataset (D3) targets at vehicle surveillance of street scenario, which consists of 39,556 frames, and the representative images are shown in Fig.5(c). All video are resized to resolution 352×288, 15 FPS. The first 8minutes are selected for training dictionary.

    Fig.5 The representative images

    To demonstrate the benefit of key observation selection in video synopsis, the proposed synopsis method (denoted as Proposed) is compared with traditional method without abnormal activity detection and key observation selection[1](denoted as Method 1), cluster-based synopsis[2](denoted as Method 2), and another key observation selection based synopsis method without abnormal detection[27](denoted as Method 3). The detailed results are displayed in Table 1, Table 2 and Table 3, corresponding to D1, D2 and D3 respectively. From Table 1, it can be concluded that our method obtains 4.7% compression rate, while causing 3,784 energy loss. Method 1 obtains 7.4% compression rate, while causing 5,008 energy loss. Method 2 obtains 6.7% compression rate, while causing 4,445 energy loss. And Method 3 obtains 5.9% compression rate, while causing 4,281 energy loss. From Table 2, it can be concluded that the proposed method obtains 2.3% compression rate, while causing 2,841 energy loss. Method 1 obtains 7.2% compression rate, while causing 3,641 energy loss. Method 2 obtains 6.8% compression rate, while causing 3,577 energy loss. And Method 3 obtains 6.0% compression rate, while causing 3,302 energy loss. From Table 3, it can be conclude that the proposed method obtains 11.5% compression rate, while causing 5,682 energy loss. Method 1 obtains 15.3% compression rate, while causing 7,752 energy loss. Method 2 obtains 14.9% compression rate, while causing 7,172 energy loss. And Method 3 obtains 13.8% compression rate, while causing 6,465 energy loss. The energy loss are mainly caused by object collision and chronological mis-order, which can heavily degrade the quality of synopsis video. Obviously, the proposed method achieves lower energy loss, while preserving a high compression rate.

    Table 1 Detailed lost information in synopsis for Dataset 1

    Table 2 Detailed lost information in synopsis for Dataset 2

    Table 3 Detailed lost information in synopsis for Dataset 3

    4 Conclusion

    In this work, a novel video synopsis is proposed based on abnormal activity detection and key observation selection. In the proposed algorithm, the activities are classified into normal and abnormal ones based on sparse coding framework. For normal activities, key observation selection using MDL principle is conducted for eliminating content redundancy. Experimental results on publicly available datasets demonstrate the effectiveness of the proposed approach.

    [1] Pritch Y, Rav-Acha A l, Gutman A, et al. Webcam synopsis: Peeking around the world. In: Proceedings of the IEEE International Conference on Computer Vision, Janeiro, Brazil, 2007. 1-8

    [2] Pritch Y, Ratovitch S, Hendel A, et al. Clustered synopsis of surveillance video. In: Proceedings of the IEEE International Conference on Advanced Video and Signal Based Surveillance, Genova, Italy, 2009. 195-200

    [3] Hanjalic A, Zhang H, Vecchi M P. An integrated scheme for automated video abstraction based on unsupervised cluster-validity analysis.IEEETransactionsonCircuitsandSystemsforVideoTechnology, 1999,9(8):1280-1289

    [4] Smith M A, Kanade T. Video skimming and characterization through the combination of image and language under standing. In: Proceedings of the Content-BasedAccess of Image and Video Databases, Bombay, India, 1998. 61-70

    [5] Li Y, Lee S, Yeh C, et al. Techniques for movie content analysis and skimming tutorial and overview on video abstraction techniques.IEEESignalProcessingMagazine, 2006, 23(2):79-89

    [6] Huang C, Chung P, Yang D H, et al. Maximum a posteriori probability estimation for online surveillance video synopsis.IEEETransactionsonCircuitsandSystemsforVideoTechnology, 2014, 24(8): 1417-1429

    [7] Zhang X Y, Wang S. Bidirectional active learning, a two-way exploration into unlabeled and labeled dataset.IEEETransactionsonNeuralNetworksandLearningSystems, 2015, 28(12):3034-3044

    [8] Zhang X Y. Interactive patent classification based on multi-classifierfusion and active learning.Neurocomputing, 2014, 127(1): 200-205

    [9] Feng S, Lei Z, Yi D, et al. Online content-aware video condensation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Rhode Island, USA, 2012. 2082-2087

    [10] Rodriguez M. Cram:Compact representation of actions in movies. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, San Francisco, USA, 2010. 3328-3335

    [11] Tian Z, Xue J, Lan X, et al. Key object-based static video summarization. In: Proceedings of the ACM Multimedia, Arizona, USA, 2011. 1301-1304

    [12] Taj M, Maggio E, Cavallaro A. Multi-feature graph-based object tracking. In: Proceedings of the 1st International Evaluation Conference on Classification of Events, Activities and Relationships, Southampton, England, 2006.190-199

    [13] Su S J, Nian X H,Pan H. Trajectory tracking and formation control based on consensus in high-dimensional multi-agent systems.HighTechnologyLetters, 2012, 18(3): 326-332

    [14] Hu Z T, Fu C L. A novel multi-sensor multiple model particle filter with correlated noises for maneuvering target tracking.HighTechnologyLetters, 2014, 20(4): 355-362

    [15] Arulampalam S, Maskell S, Gordon N, et al. A tutorial on particlefilters for on line non-linear/non-gaussianbayesian tracking.IEEETransactionsonSignalProcessing, 2002, 50(2):174-188

    [16] McKenna S, Jabri S, Duric Z, et al. Trackinggroups of people.ComputerVisionandImageUnderstanding, 2000, 80(1):42-56

    [17] Cong Y, Yuan J, Liu J. Sparse reconstruction cost for abnormal event detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Colorado, USA, 2011. 3449-3456

    [18] Hu W, Xiao X, Fu Z, et al. A system forlearning statistical motion pattern.IEEETransactionsonPatternAnalysisandMachineIntelligence(TPAMI), 2006, 28(9): 1450-1464

    [19] Meng F F, Qu Z S, Zeng Q S, et al. Video objects behavior analyzing based on motion history image.HighTechnologyLetters, 2009, 15(3): 319-324

    [20] Wang H, Klaser A, Schmid C, et al. Action recognition by dense trajectories. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Colorado, USA, 2011. 3169-3176

    [21] Mehran R, Oyama A, Shah M. Abnormal crowd behavior detection using social force model. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Miami, USA, 2009. 935-942

    [22] Sand P, Teller S. Particle video: Long-range motion estimation usingpoint trajectories.InternationalJournalofComputerVision, 2008, 80(1): 72-91

    [23] Lee H, Battle A, Raina R, et al. Efficient sparse coding algorithms. In: Proceedings of the Advances in Neural Information Processing Systems, Vancouver, Canada, 2007. 3169-3176

    [24] Zhang C J, Liu J, Tian Q C, et al. Image classification bynon-negative sparse coding, low-rank and sparse decomposition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Colorado, USA, 2011. 1673-1680

    [25] Kim C, Wang J. An integrated scheme for object-based video abstraction. In: Proceedings of the ACM Multimedia, Los Angeles, USA, 2000. 303-311

    [26] Lee J, Han J. Trajectory clustering: A partitionandgroup framework. In: Proceedings of the ACM Special Interest Group on Management of Data, Beijing, China, 2007. 593-604

    [27] Zhu X, Liu J, Wang J, et al. Key observation selection-based effective video synopsis for camera network.MultimediaVisionandApplication, 2013, 25(1):145-157

    Zhu Xiaobin, born in 1982. He received his Ph.D in 2013 from Institute of Automation, Chinese Academy of Sciences. His research interests include machine learning,pattern recognition, video and image analysis, etc.

    10.3772/j.issn.1006-6748.2016.02.011

    ①Supported by the National Natural Science Foundation of China (No. 61402023), Beijing Technology and Business University Youth Fund (No. QNJJ2014-23) and Beijing Natural Science Foundation (No. 4162019).

    ②To whom correspondence should be addressed. E-mail: buddysoft@sina.comReceived on Oct. 29, 2015

    国产精品一区二区在线不卡| 黄色女人牲交| 久久中文字幕人妻熟女| 很黄的视频免费| 亚洲 欧美一区二区三区| 亚洲专区国产一区二区| 午夜福利欧美成人| cao死你这个sao货| 国产亚洲欧美精品永久| av不卡在线播放| 国产一区二区三区视频了| 在线永久观看黄色视频| 欧美日韩成人在线一区二区| 成人18禁高潮啪啪吃奶动态图| 国产精品99久久99久久久不卡| 国产精品久久久av美女十八| 久久中文字幕一级| 成人国语在线视频| 国产免费男女视频| 精品国产一区二区三区四区第35| 香蕉丝袜av| x7x7x7水蜜桃| 国产在视频线精品| 国产一区在线观看成人免费| 亚洲av欧美aⅴ国产| 亚洲一区高清亚洲精品| netflix在线观看网站| 丰满饥渴人妻一区二区三| 色在线成人网| 亚洲av欧美aⅴ国产| 成人三级做爰电影| 少妇 在线观看| 久久婷婷成人综合色麻豆| 不卡一级毛片| 老司机亚洲免费影院| 欧美日韩瑟瑟在线播放| 757午夜福利合集在线观看| 9热在线视频观看99| 一区二区三区激情视频| 黑人巨大精品欧美一区二区mp4| 中文字幕制服av| 人人妻人人爽人人添夜夜欢视频| 成年人黄色毛片网站| 欧美一级毛片孕妇| 搡老乐熟女国产| 亚洲精品av麻豆狂野| 人人妻人人添人人爽欧美一区卜| 国产精品一区二区免费欧美| 脱女人内裤的视频| 国产野战对白在线观看| 国产成人精品久久二区二区免费| 亚洲av日韩精品久久久久久密| 他把我摸到了高潮在线观看| www.999成人在线观看| 19禁男女啪啪无遮挡网站| 久久天堂一区二区三区四区| 国产一卡二卡三卡精品| 午夜福利视频在线观看免费| 天堂√8在线中文| tocl精华| 色播在线永久视频| 日韩制服丝袜自拍偷拍| 中文字幕另类日韩欧美亚洲嫩草| 国产亚洲欧美精品永久| 动漫黄色视频在线观看| 日本黄色日本黄色录像| 日韩一卡2卡3卡4卡2021年| 黑人操中国人逼视频| 久久精品国产a三级三级三级| 日本黄色日本黄色录像| 欧美丝袜亚洲另类 | tube8黄色片| 两个人免费观看高清视频| 宅男免费午夜| 色精品久久人妻99蜜桃| 久久人人爽av亚洲精品天堂| 老司机靠b影院| 亚洲熟妇熟女久久| 国产97色在线日韩免费| 97人妻天天添夜夜摸| 久久精品熟女亚洲av麻豆精品| 亚洲av熟女| 久久久久国产一级毛片高清牌| 精品久久久精品久久久| 亚洲美女黄片视频| 国产精品 国内视频| 久久久久精品国产欧美久久久| 在线观看66精品国产| 国产91精品成人一区二区三区| 久久久久国内视频| 久久草成人影院| 久久久久精品国产欧美久久久| 国产99久久九九免费精品| 久久久久国产精品人妻aⅴ院 | 大陆偷拍与自拍| 亚洲av日韩在线播放| 日韩大码丰满熟妇| 欧美黄色淫秽网站| 亚洲国产中文字幕在线视频| 成熟少妇高潮喷水视频| 高清av免费在线| 精品国产乱子伦一区二区三区| 日日摸夜夜添夜夜添小说| 日韩免费高清中文字幕av| 黄色成人免费大全| 人人妻,人人澡人人爽秒播| 丰满的人妻完整版| 99精品久久久久人妻精品| 中文字幕高清在线视频| 欧美在线黄色| 日韩大码丰满熟妇| 色综合婷婷激情| 涩涩av久久男人的天堂| 欧美日韩乱码在线| 欧美中文综合在线视频| 亚洲欧美激情在线| 久久亚洲精品不卡| 满18在线观看网站| 日韩一卡2卡3卡4卡2021年| 国产成人欧美在线观看 | 亚洲国产看品久久| 亚洲av电影在线进入| 久9热在线精品视频| 亚洲欧洲精品一区二区精品久久久| 亚洲成人手机| 欧美精品啪啪一区二区三区| 纯流量卡能插随身wifi吗| 久久久国产精品麻豆| e午夜精品久久久久久久| 又黄又爽又免费观看的视频| 欧美午夜高清在线| 天天操日日干夜夜撸| 欧美日韩福利视频一区二区| 美女扒开内裤让男人捅视频| 在线国产一区二区在线| 日韩免费av在线播放| 日韩精品免费视频一区二区三区| 99re6热这里在线精品视频| 女警被强在线播放| 美国免费a级毛片| 巨乳人妻的诱惑在线观看| 一级片免费观看大全| 中文字幕最新亚洲高清| 婷婷精品国产亚洲av在线 | 侵犯人妻中文字幕一二三四区| 国产精品二区激情视频| 精品欧美一区二区三区在线| 999久久久精品免费观看国产| 大香蕉久久网| 免费久久久久久久精品成人欧美视频| 老熟妇仑乱视频hdxx| 大陆偷拍与自拍| 欧美另类亚洲清纯唯美| 免费在线观看影片大全网站| cao死你这个sao货| 美女福利国产在线| 老司机靠b影院| 日韩欧美一区视频在线观看| 美女高潮到喷水免费观看| 国产精品.久久久| 亚洲黑人精品在线| 69精品国产乱码久久久| 成人三级做爰电影| 韩国av一区二区三区四区| 午夜视频精品福利| 天堂动漫精品| 日韩有码中文字幕| 老司机在亚洲福利影院| 曰老女人黄片| 在线天堂中文资源库| aaaaa片日本免费| 99久久99久久久精品蜜桃| 超碰97精品在线观看| 国产主播在线观看一区二区| 少妇裸体淫交视频免费看高清 | 亚洲国产精品合色在线| 亚洲专区中文字幕在线| 国产精品二区激情视频| av超薄肉色丝袜交足视频| 免费观看a级毛片全部| 人妻久久中文字幕网| 黄色怎么调成土黄色| 麻豆乱淫一区二区| 国产高清激情床上av| 精品免费久久久久久久清纯 | 在线观看www视频免费| 欧美av亚洲av综合av国产av| 丝袜人妻中文字幕| 成人永久免费在线观看视频| 精品乱码久久久久久99久播| 十分钟在线观看高清视频www| 水蜜桃什么品种好| 在线天堂中文资源库| 777米奇影视久久| 天堂俺去俺来也www色官网| 一个人免费在线观看的高清视频| 制服人妻中文乱码| 亚洲国产看品久久| av片东京热男人的天堂| 欧美亚洲日本最大视频资源| 亚洲黑人精品在线| 国产在视频线精品| 91麻豆精品激情在线观看国产 | 成人精品一区二区免费| 亚洲精品在线观看二区| 中文字幕精品免费在线观看视频| 午夜成年电影在线免费观看| 亚洲中文字幕日韩| 国产一卡二卡三卡精品| 国产精品久久久久久人妻精品电影| 色综合欧美亚洲国产小说| 成年版毛片免费区| 1024香蕉在线观看| 一级片免费观看大全| 久久人妻熟女aⅴ| 午夜福利,免费看| а√天堂www在线а√下载 | 曰老女人黄片| 他把我摸到了高潮在线观看| 国产精品偷伦视频观看了| 人妻丰满熟妇av一区二区三区 | 亚洲av美国av| 别揉我奶头~嗯~啊~动态视频| 精品国产国语对白av| 啦啦啦免费观看视频1| 电影成人av| 亚洲欧美日韩高清在线视频| 精品电影一区二区在线| 国产精品一区二区在线观看99| 18在线观看网站| 国产成人啪精品午夜网站| 欧美成人午夜精品| 国产亚洲欧美在线一区二区| 很黄的视频免费| 欧美精品啪啪一区二区三区| 久久天堂一区二区三区四区| 久久精品亚洲av国产电影网| 国产精品亚洲av一区麻豆| 午夜福利影视在线免费观看| 成人特级黄色片久久久久久久| 老司机靠b影院| 欧美日韩瑟瑟在线播放| 侵犯人妻中文字幕一二三四区| 两个人免费观看高清视频| av超薄肉色丝袜交足视频| 亚洲第一欧美日韩一区二区三区| 国产在视频线精品| 天天影视国产精品| 男人的好看免费观看在线视频 | 国产精品成人在线| 精品高清国产在线一区| 欧美+亚洲+日韩+国产| 久久人妻熟女aⅴ| av天堂久久9| 男人舔女人的私密视频| 91成年电影在线观看| 中文字幕人妻熟女乱码| 美女福利国产在线| 999久久久精品免费观看国产| 久久久精品免费免费高清| 亚洲精华国产精华精| 亚洲欧洲精品一区二区精品久久久| 麻豆乱淫一区二区| 免费在线观看完整版高清| 国产欧美日韩一区二区精品| 丝袜美足系列| 亚洲国产毛片av蜜桃av| 欧美日韩亚洲综合一区二区三区_| 亚洲综合色网址| 成人特级黄色片久久久久久久| 亚洲av日韩精品久久久久久密| 欧美丝袜亚洲另类 | 十八禁人妻一区二区| 午夜老司机福利片| 在线播放国产精品三级| 九色亚洲精品在线播放| 村上凉子中文字幕在线| 欧美激情极品国产一区二区三区| 大香蕉久久成人网| 黄色丝袜av网址大全| 色播在线永久视频| 女人被躁到高潮嗷嗷叫费观| 欧美激情极品国产一区二区三区| 视频在线观看一区二区三区| 激情视频va一区二区三区| 日本黄色视频三级网站网址 | 人人澡人人妻人| 在线国产一区二区在线| 999久久久精品免费观看国产| 亚洲成av片中文字幕在线观看| 国产精品 国内视频| 国产成人欧美在线观看 | 99精国产麻豆久久婷婷| 人人妻人人澡人人看| 一级毛片高清免费大全| 午夜视频精品福利| 性色av乱码一区二区三区2| 欧美精品av麻豆av| 老司机靠b影院| 久久人妻福利社区极品人妻图片| 少妇裸体淫交视频免费看高清 | 黄色 视频免费看| 精品一品国产午夜福利视频| 99国产极品粉嫩在线观看| 国产有黄有色有爽视频| 久久久久视频综合| 99国产精品一区二区三区| 亚洲熟妇中文字幕五十中出 | 女人精品久久久久毛片| 久久国产精品影院| 国产欧美亚洲国产| 亚洲第一欧美日韩一区二区三区| 日韩中文字幕欧美一区二区| 欧美激情极品国产一区二区三区| 校园春色视频在线观看| 一区二区三区激情视频| 国产成人精品在线电影| 午夜精品久久久久久毛片777| 黄片播放在线免费| 纯流量卡能插随身wifi吗| av视频免费观看在线观看| 亚洲色图 男人天堂 中文字幕| 亚洲色图综合在线观看| 国产一区二区三区视频了| 精品久久久久久久毛片微露脸| 免费不卡黄色视频| 老司机亚洲免费影院| 在线观看免费视频网站a站| 女警被强在线播放| 免费在线观看视频国产中文字幕亚洲| 精品人妻在线不人妻| 中文字幕人妻熟女乱码| 欧美激情久久久久久爽电影 | 老司机福利观看| 熟女少妇亚洲综合色aaa.| 国产欧美日韩一区二区精品| 真人做人爱边吃奶动态| 1024香蕉在线观看| 女警被强在线播放| 99国产精品一区二区三区| 美女午夜性视频免费| 国产成人av教育| 一边摸一边做爽爽视频免费| 日韩视频一区二区在线观看| 高清av免费在线| 一级黄色大片毛片| 黑丝袜美女国产一区| 成人18禁高潮啪啪吃奶动态图| 女人高潮潮喷娇喘18禁视频| 中国美女看黄片| 一级毛片女人18水好多| 极品少妇高潮喷水抽搐| 欧美精品人与动牲交sv欧美| 黄色怎么调成土黄色| av中文乱码字幕在线| 欧美日韩中文字幕国产精品一区二区三区 | 一区二区三区国产精品乱码| 国产蜜桃级精品一区二区三区 | 亚洲精品在线美女| 亚洲一码二码三码区别大吗| 看免费av毛片| 天天影视国产精品| 午夜老司机福利片| 建设人人有责人人尽责人人享有的| 久久中文字幕一级| 免费观看精品视频网站| 天天影视国产精品| tocl精华| 好看av亚洲va欧美ⅴa在| 美女午夜性视频免费| 亚洲熟女精品中文字幕| 精品国产国语对白av| 亚洲视频免费观看视频| 后天国语完整版免费观看| 搡老熟女国产l中国老女人| 一级a爱片免费观看的视频| 亚洲五月婷婷丁香| 亚洲人成伊人成综合网2020| 香蕉丝袜av| 少妇的丰满在线观看| 不卡一级毛片| 亚洲国产精品一区二区三区在线| 国产成人免费无遮挡视频| 亚洲美女黄片视频| 亚洲av第一区精品v没综合| 久久人妻熟女aⅴ| 亚洲av日韩在线播放| 99国产精品99久久久久| 美女午夜性视频免费| 久久久国产成人免费| 国产精品电影一区二区三区 | 午夜精品久久久久久毛片777| 亚洲午夜精品一区,二区,三区| 国产一区二区激情短视频| 在线av久久热| 亚洲欧美精品综合一区二区三区| 美女高潮喷水抽搐中文字幕| 久久久国产一区二区| 欧美日韩亚洲高清精品| 男女下面插进去视频免费观看| 窝窝影院91人妻| 欧美午夜高清在线| 很黄的视频免费| 热re99久久精品国产66热6| 久久人妻熟女aⅴ| 久久国产亚洲av麻豆专区| 国产主播在线观看一区二区| 最新在线观看一区二区三区| 18禁观看日本| 大香蕉久久成人网| 久久国产精品影院| 国产亚洲精品久久久久5区| 欧美大码av| 国产精品免费一区二区三区在线 | 国产精品电影一区二区三区 | 国产99久久九九免费精品| av欧美777| av福利片在线| 国产乱人伦免费视频| 日本精品一区二区三区蜜桃| 99久久精品国产亚洲精品| 在线十欧美十亚洲十日本专区| 亚洲av美国av| 国产成人免费观看mmmm| 美女视频免费永久观看网站| 人人妻人人澡人人爽人人夜夜| 欧洲精品卡2卡3卡4卡5卡区| 亚洲国产欧美日韩在线播放| cao死你这个sao货| 两个人看的免费小视频| 人人妻人人添人人爽欧美一区卜| 成年动漫av网址| 国产麻豆69| 国产亚洲一区二区精品| 亚洲五月天丁香| 国产精品乱码一区二三区的特点 | 狂野欧美激情性xxxx| 亚洲熟女毛片儿| 在线观看午夜福利视频| av不卡在线播放| 日韩欧美免费精品| 王馨瑶露胸无遮挡在线观看| 90打野战视频偷拍视频| 亚洲 欧美一区二区三区| 免费在线观看完整版高清| 成人影院久久| 欧美在线黄色| 亚洲全国av大片| 精品久久久久久久毛片微露脸| 看片在线看免费视频| 啦啦啦 在线观看视频| 国内久久婷婷六月综合欲色啪| 亚洲少妇的诱惑av| 一区在线观看完整版| 精品国产超薄肉色丝袜足j| av网站免费在线观看视频| 亚洲一区中文字幕在线| 免费观看人在逋| 久久国产亚洲av麻豆专区| 精品乱码久久久久久99久播| 亚洲精品自拍成人| 成人黄色视频免费在线看| 欧美激情久久久久久爽电影 | 91精品国产国语对白视频| 久久中文看片网| 夫妻午夜视频| av在线播放免费不卡| 国产97色在线日韩免费| 女人高潮潮喷娇喘18禁视频| 免费在线观看亚洲国产| 黄色 视频免费看| 后天国语完整版免费观看| 麻豆乱淫一区二区| 亚洲国产欧美一区二区综合| 9热在线视频观看99| 男女下面插进去视频免费观看| 91在线观看av| 成人免费观看视频高清| 精品国内亚洲2022精品成人 | 欧美成狂野欧美在线观看| 亚洲欧美一区二区三区久久| 精品国产乱子伦一区二区三区| 王馨瑶露胸无遮挡在线观看| 岛国毛片在线播放| 国产精品电影一区二区三区 | 亚洲免费av在线视频| 亚洲 国产 在线| 欧美日韩成人在线一区二区| 精品少妇一区二区三区视频日本电影| 天天操日日干夜夜撸| 欧美乱色亚洲激情| 国产精品av久久久久免费| 美国免费a级毛片| 国产精品av久久久久免费| 天天躁日日躁夜夜躁夜夜| 91成年电影在线观看| 久久 成人 亚洲| 亚洲熟妇中文字幕五十中出 | 一本一本久久a久久精品综合妖精| 亚洲aⅴ乱码一区二区在线播放 | 日韩 欧美 亚洲 中文字幕| 久久精品91无色码中文字幕| 国产精品一区二区精品视频观看| 欧美日韩乱码在线| 免费在线观看视频国产中文字幕亚洲| 国产精品一区二区在线观看99| 亚洲色图综合在线观看| 国产亚洲欧美在线一区二区| 精品一区二区三区av网在线观看| 黄片小视频在线播放| 最近最新中文字幕大全电影3 | 大型黄色视频在线免费观看| 国产成人影院久久av| 又大又爽又粗| 欧美大码av| 国产91精品成人一区二区三区| 中文字幕最新亚洲高清| 久久狼人影院| 黄色丝袜av网址大全| 丰满的人妻完整版| 国产高清videossex| 久久天堂一区二区三区四区| 丰满迷人的少妇在线观看| 午夜福利一区二区在线看| √禁漫天堂资源中文www| 99re在线观看精品视频| 国产精品美女特级片免费视频播放器 | 中文字幕制服av| 日本撒尿小便嘘嘘汇集6| 在线视频色国产色| 欧洲精品卡2卡3卡4卡5卡区| a级片在线免费高清观看视频| 我的亚洲天堂| 欧美老熟妇乱子伦牲交| 人妻丰满熟妇av一区二区三区 | 日韩有码中文字幕| 啪啪无遮挡十八禁网站| 两性午夜刺激爽爽歪歪视频在线观看 | 精品国产乱子伦一区二区三区| 午夜免费观看网址| 91大片在线观看| 亚洲久久久国产精品| 一级毛片精品| 国产精品.久久久| 亚洲精品国产色婷婷电影| 香蕉久久夜色| 久久人妻熟女aⅴ| 中文字幕人妻熟女乱码| 99国产精品99久久久久| 视频区欧美日本亚洲| 亚洲av片天天在线观看| 国产欧美亚洲国产| 国产伦人伦偷精品视频| 高清av免费在线| 成年动漫av网址| √禁漫天堂资源中文www| 久久中文看片网| 亚洲精品久久午夜乱码| 男女床上黄色一级片免费看| 黄色怎么调成土黄色| 国产午夜精品久久久久久| 久久 成人 亚洲| 久久香蕉国产精品| 黄色怎么调成土黄色| 99在线人妻在线中文字幕 | 国产在线观看jvid| 成年女人毛片免费观看观看9 | 69av精品久久久久久| 成人永久免费在线观看视频| 满18在线观看网站| 手机成人av网站| 电影成人av| 午夜两性在线视频| 最新的欧美精品一区二区| 亚洲精品自拍成人| 天天躁日日躁夜夜躁夜夜| 亚洲av熟女| 一边摸一边做爽爽视频免费| 两性夫妻黄色片| 999久久久国产精品视频| 最新在线观看一区二区三区| 美女午夜性视频免费| 男女高潮啪啪啪动态图| 在线免费观看的www视频| 国产人伦9x9x在线观看| 亚洲成人国产一区在线观看| 日日爽夜夜爽网站| 亚洲九九香蕉| 又黄又爽又免费观看的视频| 夜夜爽天天搞| 亚洲欧美激情在线| 99热国产这里只有精品6| 亚洲伊人色综图| 成熟少妇高潮喷水视频| av欧美777| 91大片在线观看| 国产一区二区三区综合在线观看| 亚洲男人天堂网一区| 国产精品久久电影中文字幕 | av有码第一页| 又黄又粗又硬又大视频| 69av精品久久久久久| 国产精品免费视频内射| 一区在线观看完整版| 丝瓜视频免费看黄片| 老司机午夜十八禁免费视频| 一级a爱片免费观看的视频| 亚洲精品国产色婷婷电影| 天堂动漫精品| 美女福利国产在线| 国产一区二区三区在线臀色熟女 | 少妇猛男粗大的猛烈进出视频| 人人妻人人爽人人添夜夜欢视频| 久久久久久久久免费视频了| 激情视频va一区二区三区|