• <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 ?

    Robust SLAM using square-root cubature Kalman filter and Huber’s GM-estimator①

    2016-12-06 05:23:49XuWeijun徐巍軍JiangRongxinXieLiTianXiangChenYaowu
    High Technology Letters 2016年1期

    Xu Weijun(徐巍軍), Jiang Rongxin②, Xie Li, Tian Xiang, Chen Yaowu

    (*Institute of Advanced Digital Technology and Instrumentation, Zhejiang University, Hangzhou 310027, P.R.China)(**Zhejiang Provincial Key Laboratory for Network Multimedia Technologies, Zhejiang University, Hangzhou 310027, P.R.China)

    ?

    Robust SLAM using square-root cubature Kalman filter and Huber’s GM-estimator①

    Xu Weijun(徐巍軍)***, Jiang Rongxin②***, Xie Li***, Tian Xiang***, Chen Yaowu***

    (*Institute of Advanced Digital Technology and Instrumentation, Zhejiang University, Hangzhou 310027, P.R.China)(**Zhejiang Provincial Key Laboratory for Network Multimedia Technologies, Zhejiang University, Hangzhou 310027, P.R.China)

    Mobile robot systems performing simultaneous localization and mapping (SLAM) are generally plagued by non-Gaussian noise. To improve both accuracy and robustness under non-Gaussian measurement noise, a robust SLAM algorithm is proposed. It is based on the square-root cubature Kalman filter equipped with a Huber’s generalized maximum likelihood estimator (GM-estimator). In particular, the square-root cubature rule is applied to propagate the robot state vector and covariance matrix in the time update, the measurement update and the new landmark initialization stages of the SLAM. Moreover, gain weight matrices with respect to the measurement residuals are calculated by utilizing Huber’s technique in the measurement update step. The measurement outliers are suppressed by lower Kalman gains as merging into the system. The proposed algorithm can achieve better performance under the condition of non-Gaussian measurement noise in comparison with benchmark algorithms. The simulation results demonstrate the advantages of the proposed SLAM algorithm.

    square-root cubature Kalman filter, simultaneous localization and mapping (SLAM), Huber’s GM-estimator, robustness

    0 Introduction

    Simultaneous localization and mapping (SLAM)is a fundamental issue in the autonomous robot systems designed to realize more complex and advanced tasks, such as underground mining, planetary exploration, and disaster rescue. The objective of SLAM is to incrementally build a map of the unknown environment while concurrently using this map to localize the robot[1].

    The nonlinear discrete-time state-space model was typically formulated in the SLAM problem with Gaussian noise. The most popular filter implemented for SLAM is extended Kalman filter (EKF)[2]. However, EKF approach for SLAM tends to be inconsistent due to the accumulation of linearization error. The sigma-point Kalman filters (SPKF) which achieve second-order or higher accuracy have been proven to be far superior to standard EKF.Among the family of SPKF-class estimators, the unscented Kalman filter (UKF) and the cubature Kalman filter (CKF) have shown the capability to reduce linearization error effectively, and therefore are used in SLAM algorithms[3,4]. Especially,

    the third-order cubature rule of the CKF is claimed to be more theoretically justified and more accurate in mathematical terms than the unscented transformation of the UKF[5].

    However, the distribution of measurement noise in practical systems may deviate from the commonly assumed Gaussian model[6]. This non-Gaussian noise model is usually characterized by thick-tailed probability distributions and randomly appearing outliers, which may originate from glint noise of reflection[7]or are induced by unanticipated environment turbulence, temporary sensor failure, and incorrect modelling[8]. To deal with non-Gaussian noise model in the SLAM implementations, several methods have been proposed.For example, the Rao-Blackwellized particle filter based Fast SLAM[9]estimated the state posterior of arbitrary probability distribution by a finite number of particle samples.However, the algorithm will be computationally intensive in situations where the state vectors are high-dimensional. The H∝-filter based SLAM algorithms[4,10]are also able to estimate the state perturbed with non-Gaussian noise by treating the noise as unknown-but-bounded quantities. However the algorithms are prone to be diverged from in the presence of random outliers.

    In essence, the conventional Kalman type filters belong to the recursive minimuml2-norm or least mean-square technique, and the performance of the filters quickly degrades in the presence of outliers and thick-tailed noise. In contrast, Huber’s GM-estimator[11]is an estimation technique that gains robustness by optimizing a cost function represented in a combined minimuml1-andl2-norm. Using this estimator, the effect of the outliers is suppressed by down-weighting the normalized measurement residuals that are larger than a given threshold. Huber’s method has also shown its high robustness by integrating it into various Kalman type filters[12-14].

    Taking the advantage of square-root cubature rule’s numerical stability, a robust SLAM algorithm based on SCKF is proposed. Moreover, to accommodate the non-Gaussian measurement noise model, Huber’s GM-estimator is further introduced to improve the measurement update for each revisited landmark. Simulation results are provided to illustrate the effectiveness of the proposed algorithm in the complex scenarios with non-Gaussian measurement noise models.

    1 Problem formulation

    Consider the general discrete-time nonlinear SLAM system with the process model and measurement model

    xk=f(xk-1, uk-1)+vk-1

    zk=h(xk)+wk

    (1)

    where f(·) and h(·) are specific known nonlinear functions; xk=[xv,k, mk] is the state vector consisting of the robot pose xv,kand varying-size map of landmark mkat time step k; uk-1is the control input of the proprioceptive sensors; zkis the measurement obtained from the on-board sensors; vk-1and wkare additive process and measurement noise, respectively. The noises are assumed to be mutually independent Gaussian random variables with zero mean and covariances Qk-1and Rk, respectively.

    In this study, the third order spherical-radial cubature rule is utilized to approximate the nonlinear Gaussian integral with a set of 2N (N is the dimensionality of the state vector to be estimated) equally weighted cubature points. The set of cubature points is determined by {ξi, wi}, where ξiis thei-th element and wiis the corresponding weight factor:

    (2)

    2 SLAM based on square-root cubature rule

    The objective of the SLAM algorithm is to keep the system state estimate up to date by recursively evolving with the time update, measurement update and new landmark initialization steps. The square-root cubature rule is applied to all the SLAM steps to propagate the square-root factors of the predictive and posterior covariance directly. The complete procedure for the SCKF based SLAM(SCKF-SLAM) is derived in this section.

    2.1 Time update step

    When the robot moves according to the control signals from the proprioceptive sensors, the robot state has to be predicted based on its prior estimate of the state and the control inputs. As the process noise is non-additive, the robot state vector and its covariance squared root factor are required to be augmented as

    (3)

    The augmented state vector and its covariance squared root factor are used to determine a set of 2N1cubature points which are calculated by

    (4)

    where N1is the size of the augmented state vector. Each cubature point is propagated through the process model:

    (5)

    where uk-1is the control input. Note that the dimensionality of each propagated cubature point is the same as the original robot state vector, rather than the augmented one. The predicted robot state mean is estimated by

    (6)

    The square-root factor of the predicted robot covariance matrix is found by performing the QR decomposition

    (7)

    (8)

    2.2 Measurement update step

    Each time the robot revisits the landmarks that have already been mapped by means of its on-board sensors, the measurements are exploited to correct the estimate of both the robot state and the map of the landmarks. Landmark measurements are processed sequentially with a serial of update steps.

    The cubature points with respect to the current state are evaluated by

    (9)

    A specific landmark measurement depends only on the predicted robot pose and the particular landmark’s state, which are parts of the state vector. The propagated cubature points are evaluated with the measurement model by

    (10)

    The predicted measurement is estimated as

    (11)

    The square-root factor of the innovation matrix is found by performing the QR decomposition:

    Szz, k|k-1=qr([Zk|k-1SR,k])

    (12)

    where SR,kis the upper Cholesky factor of Rk, and Zk|k-1is a column matrix with each column calculated as

    [Zi, k|k-1]i=1,2,…,2N2

    (13)

    The cross-covariance matrix is obtained by matrix multiplying:

    (14)

    where χk|k-1is a column matrix with each column calculated as

    [χi, k|k-1]i=1,2,…,2N2

    (15)

    The Kalman gain of the SCKF is calculated by

    (16)

    The corrected state vector and corresponding square-root factor of the covariance matrix are finally obtained by

    (17)

    Sk|k=qr([χk|k-1-KZk|k-1KSR,k])

    (18)

    2.3 New landmark initialization step

    Landmark initialization happens when the robot detects a number of landmarks for the first time and decides to incorporate them into the map. The robot state vector and its covariance square-root factor are augmented with each new landmark measurement:

    (19)

    The expected position of the new landmark is calculated analogously using the cubature rule. A set of 2N3cubature points is calculated to represent the probability density of the augmented state:

    (20)

    where N3is the size of the augmented state vector. Each cubature point is propagated through the nonlinear inverse measurement model as

    (21)

    where h-1denotes the nonlinear inverse observation function, which transforms the new landmark measurements into the landmark coordinates in the global frame.

    The predicted mean of the augmented state is calculated by

    (22)

    The square-root factor of the corresponding covariance matrix is also obtained with the QR decomposition

    (23)

    (24)

    3 Huber-based SCKF SLAM

    (25)

    where δk|k-1is an unknown error vector.

    The nonlinear measurement function is also rewritten as a linear form by approximating:

    (26)

    where measurement matrix Hkis calculated by the predicted state covariance matrix and the cross-covariance matrix:

    (27)

    By combining Eqs(25) and (26) together, the linear regression equation can be obtained as

    (28)

    The error covariance matrix with respect to the far right component of the above equation is given by

    (29)

    where Skis the Cholesky factor of the error covariance.

    (30)

    (31)

    (32)

    The linear regression equation can be transformed to a compact form as

    yk=Mkxk+ξk

    (33)

    Huber’s GM-estimator is used to find the solution to this linear regression problem, by minimizing the cost function as

    (34)

    where ρ(·) is a symmetric, positive-define score function with a unique minimum at zero, dim(zk) is the size of a single landmark measurement, riis thei-th component of the residual between observation ykand its fitted value Mkxk, i.e., ri=[Mkxk-yk]. The solution of Eq.(34) is determined by the following implicit equation:

    (35)

    where the derivative ψ(ri)=dρ(ri)/dx is known as the influence function. By defining a weight function w(ri)=ψ(ri)/riand an associated diagonal weight matrix W=diag[w(ri)], it can be written in a matrix form as

    (36)

    This equation can be solved by using the iterated reweighted least-square algorithm, where the weight matrix W is recalculated in each iteration and is used in the next iteration. This process is represented as

    (37)

    (38)

    When Eq.(37) is converged, the final value of the corrected estimation of the state vector is achieved and the corresponding corrected covariance matrix is computed with the converged weight matrix:

    (39)

    Finally, the corrected square-root factor of the covariance matrix is achieved by performing the Cholesky factorization:

    Sk|k=CHOL(Pk|k)

    (40)

    The pseudo-code of the proposed Huber based SLAMalgorithm (HSCKF-SLAM) is summarized in “Algorithm 1”.

    Algorithm1 HSCKF-SLAMAlgorithmRequire:Initialrobotstatemeanx0andcovarianceP01. MainLoop:2. fork=1,2,…,Tdo3. Time-update:4. Computethepredictedrobotpose^xv,k|k-1anditsCholeskyfactorofthecovarianceSv,k|k-1via(4)-(8);5. ifnewmeasurementsreceivedthen6. Performdataassociationalgorithm;7. endif8. Measurementupdate:9. formeasurementsofrevisitedlandmarksdo10. Computethecross-covariancematrixPxz,k|k-1via(9)-(15);11. Computethecorrectedsystemstate^xk|kandthecorresponding CholeskyfactorofthecovarianceSk|kvia(25)-(40);12. endfor13. NewLandmarkinitialization:13. formeasurementsofnewlyvisitedlandmarksdo14. Performlandmarkinitializationstepvia(20)-(24);15. endfor16. endfor

    4 Simulations and results

    A series of simulations have been conducted to evaluate the performance of the proposed HSCKF-SLAM algorithm in comparison with the UKF-SLAM and the SCKF-SLAM. The publicly available UKF-SLAM simulator*https://svn.openslam.org/data/svn/bailey-slam/is modified as a benchmark platform.The other two algorithms have been implemented in Matlab R2012a on a 2.9GHz Intel Corei7-3520M Processor. As presented in Fig.1(a), the robot is assumed to move along the predefined trajectory in a rectangular plane. The robot starts from the origin of the

    global frame and detectsnearby landmarks with a laser sensor.The additive measurement noise is assumed to follow a Gaussian mixture distribution of the form:

    0≤α≤1, σ2=βσ1

    where α represents the noise model contamination, σ1and σ2are the standard deviations of the Gaussian mixture components, and β denotes the ratio between them.The process noise is 0.2m/s in wheel velocity and 2° in steering angle.Other simulation parameters are listed in Table 1.

    Table 1 Simulation parameters.

    Fig.1 The simulation results.True landmark (*) and robot path (thick solid lines), estimated landmark (diamonds)

    In order to evaluate the performance of the proposed algorithm under various measurement noise conditions,two simulation scenarios with different measurement noise models are considered: a Gaussian mixture contaminated noise model and an outlier contaminated noise model.For a fair comparison, all the SLAM algorithms are carried out with the same simulation parameters except that of the measurement noise.200 independent Monte Carlo simulation runs are conducted for each simulation scenario.

    4.1 Results in Gaussian mixture contaminated noise case

    In this simulation scenario,the measurement noise is assumed as a contaminated Gaussian model with two independent Gaussian mixtures. The standard deviation

    of the main mixture component σ1is set to 0.2m in range and 2° in bearing. Fig.1(b)~(d) show the results of a typical simulation run where the contamination and ratio parameters are set to 0.3 and 10, respectively. These plots indicate that both the robot trajectories and landmarks are estimated accurately at different time steps (k=36s, 72s, and 108s) by the proposed algorithm. The Kalman gain weights under different measurement residuals are also presented in Fig.2. It can be seen that the weights reach local minimums in the most time steps when either the range measurement residual or the bearing measurement residual is a high peak of the curve. Moreover, larger measurement residuals can bring about smaller weights. As a result, the measurement outliers are supressed to a great extent in the Kalman update process.

    Fig.2 Average Kalman gain weights under different measurement residuals

    The effects of different combinations of parameters α and β on accuracy are illustrated in Fig.3 and Fig.4. Fig.3 shows the relations between the average RMSE of the algorithms and contamination parameter α when ratio parameter β is fixed to 5. Similarly, the relations between average RMSE and ratio parameter β are shown in Fig.4, where α is set to 0.4. As observed in the figures, the HSCKF-SLAM outperforms the other algorithms in all cases. Besides, the superiority of the HSCKF-SLAM is more obvious as the parameters increase.The results indicate that the Huber based update plays a more important role when the distribution of the non-Gaussian noise has thicker tails. The SCKF-SLAM exhibits slightly better performance than the UKF-SLAM due to the reason that SCKF can approximate the nonlinear functions in higher order than UKF.The unscented transformation is no longer numerically stable and the Cholesky decomposition of the state covariance encounters troubles in cases of extremely large noise.

    Fig.3 Average RMSE under different contamination parameters

    Fig.4 Average RMSE under different ratio parameters

    4.2 Results in outlier contaminated noise case

    In this simulation scenario, the basic measurement noise follows a Gaussian distribution by setting contamination parameter to 0. This Gaussian noise model is then contaminated by a number of random measurement outliers which are induced periodically. Totally 21 measurements are selected and biased by an offset [5m, 5°]. Fig.5 depicts RMSE of the robot position and heading of the algorithms. It can be seen from the results that both the UKF-SLAM and the SCKF-SLAM suffer from estimation errors larger than the HSCKF-SLAM apparently. The position and heading RMSE are below 1m and 1.2° for HSCKF-SLAM.This proves that HSCKF-SLAM can detect all the measurement outliers and reduce their influence effectively. The average NEES of the outlier scenario are shown in Fig.6, where the two horizontal dashed lines are plotted to mark the 95% two-side confidence region.It can be seen that both SCKF-SLAM and UKF-SLAM become inconsistent for all the time steps, while HSCKF-SLAM retains consistent for more than 75 time steps.These results demonstrate that by making use of Huber’s update method, the conventional Kalman type filter is insensitive to the measurement outliers.

    Fig.5 Comparison of RMSE in outlier contaminated Gaussian measurement noise case

    Fig.6 Average NEES of the robot position in outlier case

    4.3 Computational cost

    The computational costs of these algorithms are also compared. As illustrated in Table 2, UKF-SLAM requires the minimum computational cost. SCKF-SLAM demands more running time because QR decompositions are employed to ensure numerical stability. HSCKF-SLAM takes the most computational effort due to the extra realization of robust linear regression in the measurement update stage. However, the increased average running time for one single update step with Huber’s method is of the order of a 3ms.This increase is a worthwhile price to be paid for robustness and consistency. Besides, such a level of increase is often acceptable in real-time SLAM applications.

    Table 2 Computational cost of algorithms

    5 Conclusions

    A robust SLAM algorithm based on SCKF and Huber’s GM-estimator is proposed for robot systems with non-Gaussian measurement noises. The integration of a GM-estimator doesn’t only retain the accurate merit of SCKF, but also provides an efficient way to work in non-Gaussian cases, with performance surpassing the benchmark algorithms in robustness and consistency.The influence of Huber’s GM-estimator on the convergence rate and efficiency properties with different score functions will be further studied and optimized.

    Reference

    [ 1] Dissanayake M W M G, Newman P, Clark S, et al. A solution to the simultaneous localization and map building (SLAM) problem.IEEETransactionsonRoboticsandAutomation, 2001, 17(3): 229-241

    [ 2] Durrant-Whyte H, Bailey T. Simultaneous localization and mapping: Part I.IEEERoboticsandAutomationMagazine, 2006, 13(2): 99-108

    [ 3] Martinez-Cantin R, Castellanos J A. Unscented SLAM for large-scale outdoor environments. In: Proceedings of the International Conference on Intelligent Robots and Systems, Edmonton, Canada, 2005. 328-333

    [ 4] Chandra K P B, Gu D, Postlethwaite I. A cubature H∞ filter and its square-root version.InternationalJournalofControl, 2014, 87(4): 764-776

    [ 5] Arasaratnam I, Haykin S. Cubature Kalman filters.IEEETransactionsonAutomaticControl, 2009, 54(6): 1254-1269

    [ 6] Gandhi M A, Mili L. Robust Kalman filter based on a generalized maximum-likelihood-type estimator.IEEETransactionsonSignalProcessing, 2010, 58(5): 2509-2520

    [ 7] Li X R, Jilkov V P. Survey of maneuvering target tracking. Part V. Multiple-model methods.IEEETransactionsonAerospaceandElectronicSystems, 2005, 41(4): 1255-1321

    [ 8] Zoubir A M, Koivunen V, Chakhchoukh Y, et al. Robust estimation in signal processing: A tutorial-style treatment of fundamental concepts.IEEESignalProcessingMagazine, 2012, 29(4): 61-80

    [ 9] Montemerlo M, Thrun S, Koller D, et al. FastSLAM: A factored solution to the simultaneous localization and mapping problem. In: Proceedings of the National Conference on Artificial Intelligence, Edmonton, Canada, 2002. 593-598

    [10] Ahmad H, Namerikawa T. Feasibility study of partial observability in H∞filtering for robot localization and mapping problem. In: Proceedings of the 2010 American Control Conference, Baltimore, USA, 2010.3980-3985

    [11] Huber P J. Robust Statistics. New Jersey: John Wiley & Sons, 2004. 43-48

    [12] Karlgaard C D, Schaub H. Huber-based divided difference filtering.JournalofGuidance,Control,andDynamics, 2007, 30(3): 885-891

    [13] Chang L, Hu B, Chang G, et al. Huber-based novel robust unscented Kalman filter.IETScience,Measurement&Technology, 2012, 6(6): 502-509

    [14] Wang X, Cui N, Guo J. Huber-based unscented filtering and its application to vision-based relative navigation.IETRadar,SonarandNavigation, 2010, 4(1): 134-141

    [15] Coleman D, Holland P, Kaden N, et al. System of subroutines for iteratively reweighted least squares computations.ACMTransactionsonMathematicalSoftware, 1980, 6(3): 327-336

    [16] Hampel F R, Ronchetti E M, Rousseeuw P J, et al. Robust Statistics: the Approach Based on Influence Functions. Published Online, John Wiley & Sons, 2011. 307-341

    Xu Weijun, born in 1985. He is a Ph.D. candidate in the College of Biomedical Engineering and Instrument Science of Zhejiang University. His main research fields arerobot simultaneous localization and mapping, multiple target tracking and multi-agent navigation.

    10.3772/j.issn.1006-6748.2016.01.006

    ① Supported by the National High Technology Research and Development Program of China (2010AA09Z104), and the Fundamental Research Funds of the Zhejiang University (2014FZA5020).

    ② To whom correspondence should be addressed. E-mail: rongxinj@zju.edu.cnReceived on Dec. 3, 2014

    午夜福利在线观看吧| 在线天堂最新版资源| 精品人妻视频免费看| 国产成人福利小说| 高清毛片免费观看视频网站| 欧美国产日韩亚洲一区| 国产精品国产高清国产av| 日本与韩国留学比较| 精品午夜福利在线看| 1024手机看黄色片| 淫妇啪啪啪对白视频| 国产精品1区2区在线观看.| 亚洲精品一区av在线观看| 国产成年人精品一区二区| 99久久无色码亚洲精品果冻| 精品午夜福利视频在线观看一区| 国产亚洲精品综合一区在线观看| 午夜福利高清视频| 欧美性猛交黑人性爽| 久久久久九九精品影院| 老女人水多毛片| h日本视频在线播放| 美女大奶头视频| av在线天堂中文字幕| 三级毛片av免费| 国产探花在线观看一区二区| 久久热精品热| or卡值多少钱| 在线免费观看不下载黄p国产| 精品久久久久久成人av| 欧美成人精品欧美一级黄| 欧美又色又爽又黄视频| 日本熟妇午夜| 内地一区二区视频在线| 赤兔流量卡办理| 美女cb高潮喷水在线观看| 久久久久久久久大av| 中文亚洲av片在线观看爽| 可以在线观看毛片的网站| 成人永久免费在线观看视频| 国产视频一区二区在线看| 国产乱人视频| 高清午夜精品一区二区三区 | 偷拍熟女少妇极品色| 精品久久国产蜜桃| 日本精品一区二区三区蜜桃| 男女做爰动态图高潮gif福利片| 91av网一区二区| 亚洲成人久久性| 色哟哟哟哟哟哟| 国产熟女欧美一区二区| 久久热精品热| www日本黄色视频网| 看非洲黑人一级黄片| 国产精品,欧美在线| 黑人高潮一二区| 亚洲精品国产成人久久av| 99riav亚洲国产免费| 亚洲国产精品国产精品| 国产视频一区二区在线看| 色吧在线观看| 最近在线观看免费完整版| 国内精品久久久久精免费| 欧美三级亚洲精品| 国产黄a三级三级三级人| 国产淫片久久久久久久久| 99精品在免费线老司机午夜| 国产三级中文精品| 成人漫画全彩无遮挡| 最新在线观看一区二区三区| 成人高潮视频无遮挡免费网站| 少妇人妻一区二区三区视频| 午夜亚洲福利在线播放| 成熟少妇高潮喷水视频| 三级男女做爰猛烈吃奶摸视频| 精品国产三级普通话版| 99riav亚洲国产免费| 欧美色欧美亚洲另类二区| av在线亚洲专区| 亚洲国产精品sss在线观看| 国产黄色小视频在线观看| 久久精品国产99精品国产亚洲性色| 国产精品亚洲美女久久久| 亚洲成人精品中文字幕电影| 午夜福利在线观看吧| 乱码一卡2卡4卡精品| 欧美区成人在线视频| 免费电影在线观看免费观看| 国产成人aa在线观看| 亚洲人成网站在线观看播放| 亚洲精品成人久久久久久| 两个人视频免费观看高清| 人妻久久中文字幕网| 少妇高潮的动态图| 日韩三级伦理在线观看| 国产片特级美女逼逼视频| 成人一区二区视频在线观看| 国产男靠女视频免费网站| 免费看美女性在线毛片视频| 日本在线视频免费播放| 欧美三级亚洲精品| 国产亚洲91精品色在线| 干丝袜人妻中文字幕| 搞女人的毛片| 99久久中文字幕三级久久日本| 国产精品嫩草影院av在线观看| 免费观看精品视频网站| av天堂在线播放| 亚洲av一区综合| 国产精品久久久久久久久免| 在线观看免费视频日本深夜| 久久久精品94久久精品| 亚洲av电影不卡..在线观看| 亚洲成人久久性| 中文字幕熟女人妻在线| 久久久久精品国产欧美久久久| 国产高清不卡午夜福利| 少妇的逼好多水| 午夜激情福利司机影院| 成年版毛片免费区| 国产av不卡久久| 99热6这里只有精品| av在线亚洲专区| 午夜a级毛片| 久久久久国内视频| 国产精品人妻久久久久久| 午夜福利18| 夜夜爽天天搞| 天堂av国产一区二区熟女人妻| 久久欧美精品欧美久久欧美| 国模一区二区三区四区视频| 不卡视频在线观看欧美| 国产精品一区二区三区四区久久| 亚洲欧美清纯卡通| 99精品在免费线老司机午夜| 最新中文字幕久久久久| 午夜老司机福利剧场| 99热这里只有是精品50| 搡老岳熟女国产| 久久午夜福利片| 国内精品久久久久精免费| av天堂在线播放| or卡值多少钱| 亚洲国产精品久久男人天堂| 天天躁夜夜躁狠狠久久av| 少妇裸体淫交视频免费看高清| 免费黄网站久久成人精品| 国产熟女欧美一区二区| 久久久久久国产a免费观看| 久久99热6这里只有精品| 伦理电影大哥的女人| 色吧在线观看| 亚洲精品456在线播放app| 亚洲av二区三区四区| 久久久久久久久久黄片| 免费看av在线观看网站| 永久网站在线| 中文字幕熟女人妻在线| 九九爱精品视频在线观看| 日韩精品中文字幕看吧| 一区二区三区高清视频在线| 免费看av在线观看网站| 国产aⅴ精品一区二区三区波| 国产探花极品一区二区| 深夜a级毛片| 日本a在线网址| 狂野欧美白嫩少妇大欣赏| 国产高清三级在线| 国产黄片美女视频| av国产免费在线观看| 听说在线观看完整版免费高清| 别揉我奶头~嗯~啊~动态视频| 久久人妻av系列| 小说图片视频综合网站| 菩萨蛮人人尽说江南好唐韦庄 | 在线观看午夜福利视频| 校园春色视频在线观看| 亚洲最大成人中文| 亚洲国产精品sss在线观看| 国产精品女同一区二区软件| 欧美xxxx性猛交bbbb| 啦啦啦韩国在线观看视频| 欧美三级亚洲精品| 我的女老师完整版在线观看| 六月丁香七月| 亚洲精品乱码久久久v下载方式| 大香蕉久久网| 亚洲欧美成人综合另类久久久 | 欧美日本视频| 22中文网久久字幕| 99热网站在线观看| 久久精品影院6| 日本a在线网址| 日韩 亚洲 欧美在线| 久久久久久久亚洲中文字幕| 婷婷精品国产亚洲av| 精品一区二区三区人妻视频| 美女黄网站色视频| 久久久精品欧美日韩精品| 在线免费十八禁| 久久久精品大字幕| 亚洲第一区二区三区不卡| 亚洲av.av天堂| 亚洲精品影视一区二区三区av| 免费在线观看成人毛片| 51国产日韩欧美| 国产高清三级在线| 国产老妇女一区| 菩萨蛮人人尽说江南好唐韦庄 | 久久久欧美国产精品| 日韩一本色道免费dvd| www.色视频.com| 日韩强制内射视频| 一本久久中文字幕| 久久久久精品国产欧美久久久| 亚洲最大成人中文| 丰满乱子伦码专区| 日本一本二区三区精品| 久久这里只有精品中国| 在线免费观看不下载黄p国产| 在线观看66精品国产| 成人特级av手机在线观看| 国产精品久久电影中文字幕| 亚洲欧美日韩无卡精品| 日韩 亚洲 欧美在线| av福利片在线观看| 久久6这里有精品| 国产探花在线观看一区二区| 免费电影在线观看免费观看| 欧美高清性xxxxhd video| 男女啪啪激烈高潮av片| 久久午夜亚洲精品久久| 在线观看66精品国产| 欧美性感艳星| 午夜久久久久精精品| 久久久久久久久久成人| 一个人免费在线观看电影| 少妇猛男粗大的猛烈进出视频 | 一a级毛片在线观看| 最近中文字幕高清免费大全6| 亚洲美女搞黄在线观看 | 给我免费播放毛片高清在线观看| 免费观看精品视频网站| 午夜福利在线在线| 久久久久久伊人网av| 日韩强制内射视频| 秋霞在线观看毛片| 欧美一级a爱片免费观看看| 看免费成人av毛片| 精品久久国产蜜桃| 人妻丰满熟妇av一区二区三区| 91av网一区二区| 可以在线观看的亚洲视频| 亚洲av五月六月丁香网| 久久综合国产亚洲精品| 九九久久精品国产亚洲av麻豆| 欧美性感艳星| 欧美成人免费av一区二区三区| 亚洲成人久久性| 国产精品久久视频播放| 白带黄色成豆腐渣| 男女之事视频高清在线观看| 久久久欧美国产精品| 久久久久久久久久黄片| av黄色大香蕉| 日韩制服骚丝袜av| 亚洲av电影不卡..在线观看| 成人三级黄色视频| 性色avwww在线观看| 久久久久性生活片| 美女大奶头视频| 校园人妻丝袜中文字幕| 麻豆av噜噜一区二区三区| 精品国内亚洲2022精品成人| 人人妻人人澡人人爽人人夜夜 | 悠悠久久av| 精品一区二区免费观看| 夜夜爽天天搞| 久久久久久久久久久丰满| 伦理电影大哥的女人| 亚洲自拍偷在线| 中国国产av一级| 成年免费大片在线观看| 91午夜精品亚洲一区二区三区| 美女cb高潮喷水在线观看| 白带黄色成豆腐渣| 国产成人a∨麻豆精品| 搡老熟女国产l中国老女人| 国产高清视频在线观看网站| 亚洲不卡免费看| 美女大奶头视频| 少妇熟女欧美另类| 此物有八面人人有两片| 午夜免费激情av| 欧美激情久久久久久爽电影| 日本a在线网址| 欧美日本视频| 97在线视频观看| 亚洲欧美日韩高清在线视频| 国产私拍福利视频在线观看| 一本精品99久久精品77| 日韩欧美一区二区三区在线观看| 欧美另类亚洲清纯唯美| 国产不卡一卡二| 国产视频内射| 婷婷精品国产亚洲av在线| 日韩亚洲欧美综合| 国产午夜福利久久久久久| 一个人观看的视频www高清免费观看| 国产精品99久久久久久久久| 在线观看美女被高潮喷水网站| 精品久久久久久久人妻蜜臀av| 天堂动漫精品| 亚洲在线自拍视频| 十八禁网站免费在线| 老师上课跳d突然被开到最大视频| 婷婷色综合大香蕉| 91精品国产九色| 日日啪夜夜撸| 一级黄片播放器| 日韩av不卡免费在线播放| 亚洲av免费高清在线观看| 亚洲人成网站在线播放欧美日韩| 日韩高清综合在线| 女同久久另类99精品国产91| 久久久久九九精品影院| 亚洲国产高清在线一区二区三| 精品久久久久久久末码| 国产高清不卡午夜福利| 插阴视频在线观看视频| 中文亚洲av片在线观看爽| 一区二区三区免费毛片| 小说图片视频综合网站| 午夜免费激情av| 在线观看免费视频日本深夜| 美女cb高潮喷水在线观看| 日本成人三级电影网站| 高清毛片免费观看视频网站| 欧美三级亚洲精品| 毛片女人毛片| 成人av一区二区三区在线看| 国产真实乱freesex| 国产免费一级a男人的天堂| 久久午夜亚洲精品久久| 身体一侧抽搐| 麻豆国产av国片精品| 国内精品美女久久久久久| 国产亚洲av嫩草精品影院| 在线a可以看的网站| 色播亚洲综合网| 亚洲欧美日韩东京热| 免费看美女性在线毛片视频| 少妇熟女aⅴ在线视频| 日韩欧美免费精品| 看免费成人av毛片| 天天躁日日操中文字幕| 三级男女做爰猛烈吃奶摸视频| 六月丁香七月| 你懂的网址亚洲精品在线观看 | 日日摸夜夜添夜夜爱| 久久精品国产亚洲av涩爱 | 欧美成人免费av一区二区三区| 深夜a级毛片| 久久精品国产99精品国产亚洲性色| 亚洲乱码一区二区免费版| 国产日本99.免费观看| 欧美又色又爽又黄视频| 欧美中文日本在线观看视频| 亚洲精品日韩av片在线观看| 蜜桃久久精品国产亚洲av| 少妇丰满av| 内地一区二区视频在线| 国产精品av视频在线免费观看| 亚洲国产精品久久男人天堂| 国产精品免费一区二区三区在线| 淫妇啪啪啪对白视频| 十八禁网站免费在线| 免费观看精品视频网站| 露出奶头的视频| 欧美日本视频| av视频在线观看入口| 看免费成人av毛片| 一边摸一边抽搐一进一小说| 成人特级黄色片久久久久久久| 春色校园在线视频观看| 国产精品综合久久久久久久免费| 美女 人体艺术 gogo| 国产精品久久久久久亚洲av鲁大| 久久久国产成人精品二区| 午夜福利18| 如何舔出高潮| 一进一出好大好爽视频| 国产精品一区二区免费欧美| 亚洲熟妇熟女久久| 日韩成人伦理影院| 69人妻影院| 又黄又爽又免费观看的视频| 日韩欧美精品免费久久| 国产午夜福利久久久久久| 精品一区二区三区视频在线| 亚洲精品日韩在线中文字幕 | 91狼人影院| 日本爱情动作片www.在线观看 | 亚洲国产精品成人久久小说 | 俄罗斯特黄特色一大片| 国产精品一区二区三区四区免费观看 | 国产aⅴ精品一区二区三区波| 亚洲av免费在线观看| 啦啦啦韩国在线观看视频| 一级毛片电影观看 | 国产亚洲欧美98| 国产精品无大码| 精品日产1卡2卡| 国产精品av视频在线免费观看| 中文在线观看免费www的网站| 国产精品人妻久久久影院| 成人美女网站在线观看视频| 91久久精品国产一区二区三区| 色5月婷婷丁香| 男女之事视频高清在线观看| 亚洲专区国产一区二区| 99在线视频只有这里精品首页| 又粗又爽又猛毛片免费看| 色视频www国产| 亚洲精品亚洲一区二区| 日本一二三区视频观看| 国产成人freesex在线 | 精品久久久久久成人av| av天堂在线播放| 久久草成人影院| 国产伦精品一区二区三区视频9| 中文字幕人妻熟人妻熟丝袜美| 亚洲18禁久久av| 赤兔流量卡办理| 天美传媒精品一区二区| 人人妻人人澡人人爽人人夜夜 | 色综合色国产| 久久久久九九精品影院| 日韩一本色道免费dvd| 少妇人妻精品综合一区二区 | 亚洲在线自拍视频| 婷婷精品国产亚洲av在线| 精品久久久久久久末码| .国产精品久久| 男插女下体视频免费在线播放| av在线蜜桃| 国产麻豆成人av免费视频| 亚洲精品在线观看二区| 成人亚洲精品av一区二区| 99热网站在线观看| a级一级毛片免费在线观看| 欧美性感艳星| 欧美最新免费一区二区三区| 麻豆成人午夜福利视频| 亚洲欧美日韩无卡精品| 中文字幕人妻熟人妻熟丝袜美| 少妇的逼好多水| 亚洲国产日韩欧美精品在线观看| 69av精品久久久久久| 亚洲成人av在线免费| 国产亚洲精品久久久com| 中文亚洲av片在线观看爽| 久久午夜亚洲精品久久| 亚洲乱码一区二区免费版| 亚洲美女搞黄在线观看 | 中国美白少妇内射xxxbb| 欧美zozozo另类| 免费观看在线日韩| 国产视频内射| 国产成人91sexporn| 欧美最黄视频在线播放免费| 成人漫画全彩无遮挡| 国产白丝娇喘喷水9色精品| 美女cb高潮喷水在线观看| 男人舔奶头视频| 91精品国产九色| 免费电影在线观看免费观看| 国产精品人妻久久久久久| 91久久精品国产一区二区成人| 成人特级黄色片久久久久久久| 三级毛片av免费| 欧美bdsm另类| 青春草视频在线免费观看| 一级毛片久久久久久久久女| 亚洲人成网站在线播放欧美日韩| 久久精品夜色国产| 色噜噜av男人的天堂激情| 真实男女啪啪啪动态图| 中国国产av一级| 免费av不卡在线播放| 国产精品一区www在线观看| 色在线成人网| 天堂影院成人在线观看| 国国产精品蜜臀av免费| 伦理电影大哥的女人| 欧美色视频一区免费| 亚洲自拍偷在线| 中文在线观看免费www的网站| 成年女人永久免费观看视频| 色综合色国产| 国产人妻一区二区三区在| 成人三级黄色视频| 亚洲在线观看片| 嫩草影院精品99| 国产高清不卡午夜福利| 最近2019中文字幕mv第一页| 看非洲黑人一级黄片| 中文字幕久久专区| 成人午夜高清在线视频| 亚洲欧美精品自产自拍| 男人狂女人下面高潮的视频| 国产精品,欧美在线| 亚洲国产日韩欧美精品在线观看| 亚洲精品久久国产高清桃花| 欧美+亚洲+日韩+国产| 免费黄网站久久成人精品| 变态另类丝袜制服| 亚洲国产精品合色在线| av女优亚洲男人天堂| 3wmmmm亚洲av在线观看| 黄色欧美视频在线观看| 两个人视频免费观看高清| 三级国产精品欧美在线观看| 亚洲欧美精品综合久久99| 十八禁网站免费在线| 伊人久久精品亚洲午夜| 久久午夜亚洲精品久久| 五月伊人婷婷丁香| 国产精品电影一区二区三区| 色播亚洲综合网| 日韩人妻高清精品专区| 精品99又大又爽又粗少妇毛片| 九色成人免费人妻av| 又黄又爽又免费观看的视频| 能在线免费观看的黄片| 国产精品av视频在线免费观看| 波多野结衣巨乳人妻| 18禁在线无遮挡免费观看视频 | 亚洲国产精品久久男人天堂| 成人特级av手机在线观看| 免费av不卡在线播放| 精品无人区乱码1区二区| 亚洲欧美成人综合另类久久久 | 啦啦啦观看免费观看视频高清| 免费观看人在逋| 国产成人精品久久久久久| 国产精品伦人一区二区| 国产av不卡久久| 国产精品一区二区三区四区免费观看 | 欧美不卡视频在线免费观看| 国产美女午夜福利| 色综合亚洲欧美另类图片| 免费av观看视频| 最近手机中文字幕大全| 久久国产乱子免费精品| 99精品在免费线老司机午夜| 毛片一级片免费看久久久久| 中文字幕av成人在线电影| 蜜桃久久精品国产亚洲av| a级毛片a级免费在线| av天堂中文字幕网| 最新在线观看一区二区三区| 日本精品一区二区三区蜜桃| 黄色一级大片看看| 午夜激情欧美在线| 美女黄网站色视频| 亚洲欧美日韩东京热| 能在线免费观看的黄片| 国产成人91sexporn| 国产伦一二天堂av在线观看| 1024手机看黄色片| 热99在线观看视频| 中国美白少妇内射xxxbb| 日本欧美国产在线视频| 久久精品久久久久久噜噜老黄 | 成人美女网站在线观看视频| 久久6这里有精品| 亚洲精品亚洲一区二区| 老女人水多毛片| av在线播放精品| 男女啪啪激烈高潮av片| 最近最新中文字幕大全电影3| 日韩欧美免费精品| 成人毛片a级毛片在线播放| 极品教师在线视频| av免费在线看不卡| 少妇的逼水好多| 亚洲aⅴ乱码一区二区在线播放| 精品不卡国产一区二区三区| 国产精品嫩草影院av在线观看| 97碰自拍视频| 天堂影院成人在线观看| 嫩草影院新地址| 中文字幕av成人在线电影| 乱码一卡2卡4卡精品| 嫩草影院新地址| 又粗又爽又猛毛片免费看| 国产色爽女视频免费观看| 亚洲av熟女| 人人妻,人人澡人人爽秒播| 天堂网av新在线| 亚洲五月天丁香| 2021天堂中文幕一二区在线观| 人妻丰满熟妇av一区二区三区| 性插视频无遮挡在线免费观看| 久久欧美精品欧美久久欧美| 精品一区二区三区av网在线观看| 男女视频在线观看网站免费| 国产精品福利在线免费观看| 啦啦啦观看免费观看视频高清| 18禁在线无遮挡免费观看视频 | 真实男女啪啪啪动态图| 欧美激情在线99| 2021天堂中文幕一二区在线观| 欧美最黄视频在线播放免费|