• <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 adaptive UKF based on SVR for inertial based integrated navigation

    2020-07-02 03:16:58MengdeZhngHiDiBiqingHuQiChen
    Defence Technology 2020年4期
    關(guān)鍵詞:百度

    Meng-de Zhng , Hi-f Di ,*, Bi-qing Hu , Qi Chen

    a College of Electrical Engineering, Naval Univ. of Engineering, Wuhan 430033, China

    b Qingdao Brigade, Shandong Branch of the Armed Police Marine Police Corps, Qingdao 266033, China

    Keywords:Integrated navigation Support vector regression Unscented Kalman filter Robust filter Adaptive filter

    ABSTRACT Aiming at the problem that the traditional Unscented Kalman Filtering (UKF) algorithm can’t solve the problem that the measurement covariance matrix is unknown and the measured value contains outliers,this paper proposes a robust adaptive UKF algorithm based on Support Vector Regression (SVR). The algorithm combines the advantages of support vector regression with small samples, nonlinear learning ability and online estimation capability of adaptive algorithm based on innovation.Firstly,the SVR model is trained by using the innovation in the sliding window, and the new innovation is monitored. If the deviation between the estimated innovation and the measured innovation exceeds a given threshold,then measured innovation will be replaced by the predicted innovation, and then the processed innovation is used to calculate the measurement noise covariance matrix using the adaptive estimation algorithm. Simulation experiments and measured data experiments show that SVRUKF is significantly better than the traditional UKF, robust UKF and adaptive UKF algorithms for the case where the covariance matrix is unknown and the measured values have outliers.

    1. Introduction

    Strapdown inertial navigation System(SINS) has the characteristics of simple structure, small size, light weight, comprehensive output navigation information and good autonomy. It has become the focus of research.However,the SINS has the same problem that the navigation accuracy diverges with time as with the platform inertial navigation system.It cannot correct error relying on its own[1-3], such that it requires combined navigation. The strapdown inertial base integrated navigation system is a system that corrects navigation information using strapdown inertial navigation and another or more non-similar navigation systems [4,5]. Common Strapdown Inertial Navigation System can be applied to land navigation (GPS, odometer), underwater navigation (Doppler Velocity Log, DVL) [2,3,6,7]. However, the key technology of the integrated navigation is the filtering method [2].

    In 1960, R.E. Kalman proposed the Kalman filtering (KF) algorithm [8], which became the cornerstone of integrated navigation research in recent decades.However,the classical KF can only solve the linear state estimation problem. In 1995, S. J. Julier and J. K.Uhlmann proposed the Unscented Kalman filter (UKF) [9], which has a higher estimate and stability in nonlinear systems and become a research hotspot [10-12]. In the practical application of strapdown inertial base integrated navigation, the accuracy and real-time of estimation are both problems to be considered. The less computational complexity and higher precision of UKF make it a feasible nonlinear filtering method which is widely used in navigation[13,14].This is an important reason for its research hotspot.In 2003, Crassidis first proposed a quaternion-based UKF for spacecraft attitude estimation, known as unscented quaternion estimator (USQUE) [10]. In Ref. [15], the USQUE algorithm for quaternion expression attitude is improved, which reduces the computational complexity of USQUE. The method is applied to SINS/GPS direct integrated navigation. Zhou J, Edwan E et al. [16]studied the application of USQUE for tightly integrated MEMS and GPS navigation system.In Ref.[17],the performance of UKF,MEKF and KF in MEMS/GPS system are compared. The results show that UKF has the highest accuracy and the largest amount of calculation.

    In the practical application of strap-down inertial base integrated navigation, the SINS itself is relatively stable and is not susceptible to interference and thus affects system noise. The environmental conditions are complex and changeable, which makes it difficult to avoid pollution and interference,which in turn affects the measurement noise covariance matrix. The polluted measurements will have a greater impact on the accuracy of the UKF. The computed covariance matrix is not always guaranteed to be positive definite because of the negative sigma points which even lead to the divergence of the filter[11,18-20].For example,if a SINS/GPS system is used in a city, there may be cases where the satellite signals are blocked or interfered[21]; when the SINS/DVL system is used underwater, there may be problems such as abnormal water velocity caused by changes in water flow rate or bottom geography [22]. How to improve the anti-interference ability of the algorithm is a problem that the UKF must solve in practical applications.In Ref.[23],a robust algorithm based on the Mahalanobis distance (MD) method is proposed, which can improve the robustness of KF,but it cannot be applied to nonlinear systems.In Ref.[11],Huber’s robustness method is extended to UKF to estimate position, velocity and constant which achieves the robustness of UKF for integrated INS/GPS. In addition, an accurate priori measurement covariance matrix is often difficult to obtain,and may change with the motion state and environment. Robust methods alone do not yield optimal results. Therefore, the literature [24] proposed an adaptive UKF filtering algorithm based on sliding window, which improves the adaptive ability of the UKF algorithm, but does not consider the possibility that there may be outliers in practical applications.In Ref.[25],a robust adaptive UKF filtering method based on innovation and residuals is proposed for the simultaneous existence of outliers and uncertain covariance matrices. However, in this method, the method of reducing the weight of outliers is used. It will still be affected by the wild value and the filtering accuracy will decrease.

    Aiming at the above problems, this paper proposes a robust adaptive UKF method based on SVR, which can automatically detect outliers and replace the current innovation with real-time estimated innovation instead of reducing the weight of wild value. The algorithm enhances the anti-jamming capability of the algorithm without reducing the filtering accuracy.

    The organization of this paper is as follows. In the review of existing theories, the basic SINS equations and USQUE are introduced. In the introduce of the SVRUKF section, this section first extend the Innovation-based adaptive estimation(IAE)to nonlinear filter case within UKF framework. The principle of improving the robustness of SVR is studied. Then SVRUKF are derived in detail.Finally, simulation testing and ship-borne experiments were conducted and are summarized in the experimental section. The results are provided and the performance of the traditional methods and SVRUKF are compared. The experimental results show that SVRUKF is superior to the traditional robust algorithm and adaptive algorithm.

    2. Theoretical review

    2.1. SINS basic equations

    The quaternion kinematics equation is given by

    So, the SINS basic equations using ENU coordinate with the quaternion are summarized.

    where,

    where, ωieis the Earth’s rotation rate are given as (from WGS-84)7.292115×10-5rad/sec. and

    where, λ is the latitude a = 6378137m, e = 0.0818. The g is local gravity (from WGS-84) are given as approximately 9.8 m/s2.

    The gyro measurement model is given by

    where, ε is gyro bias, ηgvand ηguare zero-mean Gaussian whitenoise process.

    integrated navigation system is shown in Fig. 2. The accelerometer measurement model is given by

    where,?is accelerometer bias,ηavand ηauare zero-mean Gaussian white-noise process.

    Fig.1. USQUE algorithm.

    Fig. 4. The innovation process in the adaptive SVRUKF.

    2.2. USQUE

    Because UKF cannot be directly applied to the strap-down inertial base integrated navigation system with quaternion expression attitude,it is difficult to guarantee that the attitude quaternion parameter normalization constraint problem and the quaternion filtering variance matrix matching constraint problem are difficult to guarantee. USQUE solves the problem of UKF’s attitude estimation in the equation of state through“l(fā)ayered filtering”.The filtering process is divided into two layers,the inner layer filtering uses the generalized error Rodrigues parameter to update the attitude state,and the outer layer filtering uses the quaternion to perform the attitude state transfer. The quaternion and generalized error Rodrigues parameters use the attitude transformation relationship of multiplicative error quaternions as a bridge for switching. The basic idea of USQUE is shown in Fig.1, and the algorithm flow is shown in Algorithm 1-3.

    The following is the derivation of the filtering process of USQUE combined with the strapdown inertial base integrated navigation system. The equation of state has been described in detail above.Similar to MEKF [23], the state quantity is divided into two parts:attitude quantity and non-attitude quantity.The non-attitude part is seen in MEKF algorithm,and the attitude part is error Rodrigues parameteris the other states used in SINS/GPS system. The state vector is=The corresponding filter variance matrix is Pk-1.

    Fig. 3. Support vector regression diagram.

    USQUE algorithm is estimated byand Pk-1according to the state estimate. Determine the corresponding quaternion attitudeIn the measurement update,the measurement equation is nonlinear if the tight couple or super-tight couple is adopted. For the USQUE algorithm,the measurement update process is basically the same as the time update. In the loose couple method, the measurement equation is a linear equation. Since the attitude is represented by the Rodrigues parameter,the measurement update process is the same as KF. After the measurement update is completed,the attitude expressed by the Rodriguez parameter will be used by, converted to quaternion, This is the purpose of the attitude update. This paper introduces the measurement update by taking the loose couple as an example.

    Among them, depending on the measurement, Hkis different.

    When only the position is used as the external aiding information, the measurement matrix is:

    When only the velocity is used as the external aiding information, the measurement matrix is:

    When the position and velocity are used as the external aiding information, the measurement matrix is:

    Fig. 5. Simulation trajectory.

    Time Propagation:Define the filtering state Using error MRP describe attitude ^xk-1(δ?) = [^xδ?k-1;^xε k-1]Using quaternion describe attitude ^xk-1(q)=[^xqk-1;^xε k-1] and covariance Pk-1#Create sigma points using UT χk-1(i) ="χδ?k-1(i)χε k-1(i)= sigma(^xk-1(δ?),Pk-1)The χδ?k-1(i) corresponding quaternion error χδqk-1(i) = [δρk-1(i);δq4,k-1(i)]δq4,k-1(i) =-a■■■χδ?k-1(i)■■■2+f■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■f2 +(1-a2)■■■χδ?k-1(i)■■■2 f2 +■■■χδ?k-1(i)■■■2 δρk-1(i) = f-1[a +δq4,k-1(i)]χδ?k-1(i)Compute the quaternion sigma points χqk-1(i)=χδqk-1(i)?^qk-1 then χk-1(i) = [χq k-1(i);χek-1(i)]Propagate sigma points χk/k-1(i) = f(χk-1(i)) = [χq k/k-1(i);χek/k-1(i)]Compute error quaternion sigma points χδq k/k-1]-1and χδqk/k-1(i) = [δρk/k-1(i);δq4,k/k-1(i)]Compute GRP sigma points χδ?k/k-1(i)=f k/k-1(i)=χq k/k-1(i)?[x q δρk/k-1(i)a+δq4,k/k-1(i) and χk/k-1(i) = [χδ?k/k-1(i);χek/k-1(i)]The predicted mean and covariance of state^xk/k-1 = ∑2n i=0 w(i)χk/k-1(i)Pk/k-1 = ∑2n i=0 w(i)(χk/k-1(i) - ^xk/k-1)(χk/k-1(i)-^xk/k-1)T + Wk-1 Measurement update:Compute filtering gain Kk = Pk/k-1HTk[HkPk/k-1HTk +Rk]-1 Compute posterior mean and covariance,^xk =^xk/k-1 +Kk(yk-Hk^xk/k-1) and ^xk =[^xδ?k-1;^xek]Pk =(In×n-KkHk)Pk/k-1 with In×n is n-dimension unit matrix Attitude update:The ^xδ?k corresponding quaternion is shown by^xδq k (i) = [δρk(i),δqk,4(i)]T With-a■■^xδ?■■2+f■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■f2 +(1-a2)■■■^xδ?■■■2 δq4,k =k k k f2 +■■■^xδ?k■■■2 且δρk = f-1[a +δq4,k]^xδ?The update of quaternion is given by^xqk = ^xδqk ?x k/k-1 and δ?k?0,^xk(q) = [^xqk;^xεk]then enter the next filtering cycle.q

    2.3. SINS/GPS integrated navigation system

    By using accelerometer measuring force and gyroscope measuring the angular rate, SINS can output attitude, velocity and position. Then attitude, velocity and position are entered into UKF time propagation as state. The position information provided by GPS is entered into UKF measurement update. Finally, the filter estimate is obtained (see Fig. 2).

    3. Derivation of the SVRUKF

    It is known that the standard UKF performs poorly when the measurement noise deviates from the assumed Gaussian model.In order to improve the performance of the standard UKF in the case without accurate noise model and/or with outliers,this section first extend the Innovation-based adaptive estimation(IAE)to nonlinear filter case within UKF framework, forming adaptive UKF(AUKF).Then the Support Vector Regression (SVR) algorithm is used to robust the AUKF, forming robust adaptive UKF(SVRUKF).

    3.1. IAE-based UKF

    3.1.1. IAE-based EKF

    The Innovation-based adaptive estimation (IAE) algorithm is a method of calculating the system noise covariance Q,measurement noise covariance R,or simultaneously calculating Q and R from the observation information statistics. The method estimates the covariance matrix C of the observed information by the innovationin the sliding window of width n:

    This can be used to calculate the measurement noise matrix R and the system noise matrix

    3.1.2. AUKF

    For the UKF, if the observation equation is linear, then the observation update is the same as the normal KF, so the system noise and the measurement noise covariance matrix can be adaptively estimated directly using the above formula.

    If the observation equation is nonlinear,then the measurement update requires a UT transformation, and the predicted value is

    The calculation of the innovation will become

    According to the central limit theorem, the innovation satisfies the zero mean Gaussian distribution. That is E(δz-k) = 0, so

    According to Eqs.(2)and(10),the estimated covariance matrix can be obtained as

    3.2. Robustify the UKF

    Due to the influence of the environment, the measurements of the sensor may have outliers.If these outliers are not processed,the covariance matrix estimated by the adaptive will seriously deviate from the true value,resulting in a decrease in filtering performance.Therefore,this section uses support vector regression to detect and process outliers.

    3.2.1. Support vector regression

    The principle of regression analysis using the support vector machine is to project the multi-dimensional nonlinear relation sample set into the high-dimensional feature space G through the nonlinear mapping x→Φ(x) to become a linear relationship, and then perform linear regression in this high-dimensional feature space. Given n data samples {xi,yi},i = 1,2,…,n, where xiis the actual observation and yiis the expected value. Use the following formula to estimate the regression function f21:

    where,Φ(x)is the mapping function,b is the offset,ω is the vector in the high-dimensional feature space G, and Rnis the dimension real space.

    As shown in Fig. 3, the goal of SVR regression is to find a hyperplane that causes the“distance”and minimum of sample points that fall outside the red region to the hyperplane. Consider the relaxation variable and regularization, get the cost function

    where,C is a penalty factor;ξiandare the relaxation factors,ε is a loss function.

    The goal of optimization is to solve the parameters ω and b to minimize the cost function

    The above optimization model can be regarded as a quadratic programming problem, using the Lagrange dual form and introducing a kernel function expression, then equation (20) is transformed into:

    The constraint is

    where αiandare the weighting coefficients of the support vector and K is the kernel function.By solving equation(20),αiandcan be obtained, and substituting equation (22) into equation (17) can obtain the solution of SVR.

    The above process needs to satisfy the KKT condition (22)It can be seen from equation(24)that there is(C-αi)ξi=0 and αi(f(xi)-yi-ε-ξi)=0 for any one sample (xi,yi). So, after getting αi, if 0 <αi<C, there must be ξi= 0, so there is

    3.2.2. Robust AUKF based on SVR

    Let the innovation time series be {δz1,δz2,…δzt}. These time series satisfy certain changes. However, since the innovation is generally multi-dimensional,the relationship between them is not a simple linear relationship,so it is difficult to use the conventional fitting method to establish this relationship model.

    In her desire to please everyone she ceased to be sincere, and degenerated17 into a mere18 coquette; and even her lovers felt that the charms and fascinations19 which were exercised upon all who approached her without distinction were valueless, so that in the end they ceased to care for them, and went away disdainfully

    If the current time data δzt+1is predicted based on the historical data of δzt,the mapping f:Rm→R:can be established to satisfy the(23)

    Where m is the number of historical data used in the prediction function, i.e. the order of the model.

    The model of support vector regression can be obtained by learning from the sample,so that the predicted value at the current moment is:

    Calculate the difference between the predicted value and the actual measured value

    The principle of determining the wild value is:

    where, T is the threshold value. Regarding the selection of the threshold value, the literature [23] is selected according to experience and lacks theoretical basis. This paper combines the 3σ criterion to calculate the mean δ~z and standard deviation σδzof the innovation in the sliding window.

    Further determining the threshold

    If the measured value is judged to be a outlier (as long as a certain dimension in the innovation vector exceeds the threshold value and is determined to be a wild value), the predicted innovation is used to replace the original innovation, and the measurement update is completed. 百度翻譯See Fig. 4 for the whole process.

    The key to Kalman filter application in practice is to coordinate the contradiction between the algorithm in ideal state and the problems in practical application. It requires Kalman filter to solve practical problems through anti-interference processing. Robustness and adaptation are the methods to improve the anti-jamming ability of filtering.The traditional robust filter is mainly used to deal with the problem of system interference. Through the method of weighted, traditional robust filter reduce the variance of measurement noise. So, the anti-jamming ability of the filter is improved by using as little external information as possible. The essence is to reduce the filter gain matrix K.The adaptive algorithm is mainly used to deal with the degradation of filtering performance caused by inaccurate estimation of measurement noise covariance matrix. In this paper, SVRUKF is used to improve the estimation accuracy of robust filter and adaptive filter. What’s more, it can solve the problem of poor estimation effect of RUKF in the case of large outliers.The experimental part will prove the effectiveness of SVRUKF.

    4. Simulation study and experiments

    In this section,the investigated robust adaptive filtering method for the SINS integrated navigation is evaluated through simulation and ship-borne experiment. The simulation and ship-borne experiment used SINS/GPS loosely integration whose position as its measurement variables. Unlike MEMS, the most important navigation parameters of high accuracy SINS/GPS integrated navigation are yaw and position, especially in the field of marine navigation.Therefore,in this part,we will focus on the estimation error of yaw and position.From the analysis of the results,we are able to prove effectiveness of the SVRUKF (see Fig. 4).

    4.1. Simulation study

    The simulation produces a running track of the car(as shown in Fig. 5), and the simulation time is 680 s. The red pentagram indicates the initial position of the car,the blue solid circle indicates the end position of the car,the initial speed is 15 m/s,and the initial yaw is 0°. The sensor specification are shown in Table 1.

    Table 1 Sensors specifications.

    Set the initial conditions of the filter:horizontal error 0.5’, yaw error 30’,speed error in each direction 0.1 m/s,position error 10 m.The window size involved in the SVRUKF is 60 s. The CPU of computer is Intel Pentium CPU G4600 (3.60 GHz).

    4.1.1. Case 1

    The main purpose of this experiment is to test the performance of SVRUKF in inaccurate initial measurement noise covariance matrix. Set the initial covariance matrix R to 100 times the true value. The results obtained by the four methods of UKF, AUKF,HUKF, and SVRUKF, respectively, are shown in Fig. 6(a)-(c).

    It can be seen from the two graphs in Fig. 6(a)-(b) that under the condition that the initial covariance matrix has large uncertainty, the positioning error and yaw error of the traditional UKF method and the robust UKF method do not converge to zero, the estimation accuracy decreases,and the AUKF and SVRUKF methods gradually converge around 80s and finally approach zero. This is because the latter two methods succeed in estimating the more accurate measurement covariance matrix in a shorter time. Make the estimation results optimal. Through the Frobenius norm comparison curve of the difference between the covariance matrix and the real covariance matrix estimated by the two adaptive methods in Fig. 6(c), it can be seen that the estimation accuracy of the two methods for the covariance matrix is comparable, but the method proposed in this paper has a faster convergence speed than the traditional adaptive method, which is 30 s and 60 s respectively.The position errors in Figs. 6, 7 and 9 are calculated from the distance between the longitude and latitude estimates and the reference values.

    Table 2 shows the location and yaw Root mean square error(RMSE)tatistics and total computing time of the whole estimation process for different algorithms. As we all know, RMSE can well reflect the precision of measurement. The RMSE equation in Tables 2, 3 and 5 are given by:

    It can be seen from Table 2 that the accuracy of the position and yaw error of the HUKF algorithms is about 2.73 m and 4.51’. The position estimation accuracy of the AUKF and SVRUKF algorithms is 1.71 m and 1.65 m.The performance of SVRUKF is 59%higher than the ordinary UKF.The yaw error of AUKF is worse than the ordinary UKF, and the yaw error of SVRUKF is reduced to 2.1’ which has a great improvement. The above analysis shows that SVRUKF can adaptively estimate the covariance matrix under the condition of covariance uncertainty, and the performance is better than the ordinary AUKF. In terms of computing time, UKF, AUKF and RUKF are almost equivalent at around 22 s. SVRUKF is 36.0250 s. When talking about average computational time of each integrated navigation update step, UKF is 33.2 ms, RUKF is 31.3 ms, AUKF is 33.9 ms, SVRUKF is 52.9 ms. Considering the practical application,we pay more attention to the effect of integrated navigation.Therefore, GPS update frequency is used as update frequency for calculation average computational time for each integrated navigation update step.SVRUKF takes more 19.7 ms per step than UKF.Because this period of time is very short,it is completely acceptable.

    In the case of inaccurate measurement noise covariance matrix,RUKF improves anti-jamming capability by reducing filter gain matrix K. It will inevitably lead to a decline in the accuracy of estimation. SVRUKF is slightly better than AUKF. Some measurement noises can lead SVR to judging that filtering innovation is with outliers.So filtering innovation is replaced by prediction.This is the reason why SVRUKF improves the performance of adaptive filter.

    4.1.2. Case 2

    The main purpose of this experiment is to verify the performance of SVRUKF when the measurement contains large outliers and inaccurate measurement noise. Set the initial covariance matrix R to 100 times the true value, and add a field value of 1000 m every 60 s horizontal position. The results obtained by the UKF,AUKF, HUKF, and SVRUKF methods are shown in Fig. 7(a)-(b).

    It can be seen from the graphs shown in Fig. 7(a)-(b) that UKF and AUKF are completely divergent under the condition of containing the outliers and the uncertain initial measurement noise covariance matrix, and HUKF can converge slowly due to the robustness to the wild value,and the SVRUKF method proposed in this paper can converge at a faster speed and the convergence accuracy is better than HUKF.

    Table 3 shows the position RMSE,yaw RMSE statistics and total computing time of the whole estimation process for different algorithms. It can be seen from Table 3 that the positional errors of the UKF and AUKF algorithms are very large, 21 m and 16.9 m,respectively.HUKF is able to keep the position error at around 3 m,which is a big improvement.The SVRUKF method works best,with position error and yaw error of 1.93 m and 2.32’.The above analysis shows that SVRUKF can adaptively estimate the covariance matrix under the condition of outlier and covariance uncertainty,and the performance is better than the other three methods. When discussing computing time,UKF,AUKF and RUKF are almost equivalent at around 22 s. While, SVRUKF is 36.165 s. When talking about average computational time of each integrated navigation update step, UKF is 33.5 ms, RUKF is 32.2 ms, AUKF is 34.4 ms, SVRUKF is 53.1 ms.

    According to the results of the experiments,when talking about estimation of position, UKF and AUKF have no ability to tolerate position outliers. HUKF has certain capability to tolerate outliers.However, SVRUKF is the best in terms of convergence speed and estimation accuracy. It proves that SVRUKF has better robustness than RUKF.So,SVRUKF has better solved the problem that RUKF is not good at dealing with large outliers.In yaw estimation,SVRUKF has the best estimation result because it can accurately estimate the correct value of outliers and estimate measurement noise covariance. The characteristics of RUKF sacrificing accuracy to improve anti-infective ability have been verified. Similarly, AUKF also has some anti-jamming ability. UKF also appear jagged in course estimation because of poor resistance to outliers.

    Fig. 6. Comparison of different algorithm results under unknown conditions of covariance matrix.

    Fig. 7. Comparison of different algorithms with outliers and uncertain covariances.

    Table 2 Total computing time and RMSE of position and yaw.

    Table 3 RMSE of position, yaw and total computing time.

    4.2. Ship-borne experiment

    This section verifies the SVRUKF proposed in this paper based on the ship-borne experimental data. The experimental data was collected from a ship-borne experimental system. The sensor configurations used in the shipboard experiments are shown in Table 4. The ship-borne experiment was carried out in theChangJiang River. The experimental process was designed as follows: When the experimental system was turned on, the test ship remained moored for about 15 min; then the test ship sailed out and the motion was about 6.4 h. The whole motion trajectory is shown in Fig. 8.

    Table 4 Sensors specifications.

    Table 5 Total computing time and RMSE of position and yaw.

    Fig. 8. Trajectory of the ship.

    The data of 600 s was selected for experimental verification.For GPS data, add an outlier of 1000 m every 60 s horizontal position.Initial conditions:horizontal error 0.1°,yaw error 1°,using GPS data to initialize position and speed. SINS/GPS integrated navigation is implemented by four methods:UKF,AUKF,HUKF,and SVRUKF,and differential SINS/GPS is used as a benchmark. The window size involved in the SVRUKF is 60 s. The CPU of computer is Intel Pentium CPU G4600 (3.60 GHz). The results obtained are shown in Fig. 9(a)-(b).

    It can be seen from the graphs shown in Fig. 9(a)-(b) that UKF and AUKF are completely divergent under the condition of containing the outlier and the uncertain initial covariance matrix,and HUKF can be slow converge due to the robustness to the outlier.And the SVRUKF method proposed in this paper can converge at a faster speed and the convergence accuracy is better than HUKF.

    Fig. 9. Comparison of measured data filtering results.

    Table 5 shows the position and yaw RMSE statistics and total computing time of the whole estimation process for different algorithms. It can be seen from Table 5 that the positional errors of the UKF and AUKF algorithms are very large, 21.88 m and 19.74 m,respectively. HUKF is able to keep the position error at around 3.89 m. The SVRUKF method works best, with position error and yaw error of 0.48 m and 2.16’. The above analysis shows that SVRUKF can adaptively estimate the covariance matrix under the condition of outlier and covariance uncertainty, and the performance is better than the other three methods. SVRUKF takes around 50 s more computing time than UKF. It is caused by the prediction filter innovation of SVR algorithm.The SINS’s frequency update of ship-borne experimental is 200 Hz, while the SINS’s frequency update of simulation test is 10 Hz. So, the calculation time of ship-borne experimental is far longer than simulation test.As for average computational time of each integrated navigation update step,UKF is 0.880 s,RUKF is 0.882 s,AUKF is 0.884 s,SVRUKF is 0.962 s. SVRUKF does not increase the amount of calculation compared to the ordinary UKF.

    According to the results of the ship-borne experiments,we can get the same conclusion as the simulation study. In position estimation,all methods except SVRUKF are jagged due to large position outliers and measurement noise pollution. SVRUKF achieves convergence quickly and the performance of estimation is good.Although robust filter and adaptive filter have certain robust ability,large outliers and complex noise in real environment destroy the stability of filter.When talking about yaw,UKF is jagged due to poor tolerance.Although RUKF has better robustness,it has poor ability to resist complex noise in real environment. The effect of position on yaw is not particularly significant.So,the performance of AUKF is slightly better than RUKF.Because of being disturbed by outliers,the performance of AUKF is poorer than SVRUKF.The performance of SVRUKF is 58% better than AUKF and 80%better than HUKF.Because of SVR needing real-time training and output calculation,the computation time of SVRUKF is 10%longer than that of UKF.The real-time performance of UKF is not good, while SVRUKF needs more time.

    5. Conclusion

    By using support vector regression to robust adaptive UKF, this paper proposes a robust adaptive UKF inertial integrated navigation method based on SVR which named SVRUKF. SVRUKF trains the innovation model online and performs fault detection and diagnosis on the innovation model.It has the advantages of robustness and adaptability compared with HUKF, AUKF and UKF. Simulation and ship-borne experiments demonstrate the effectiveness of the algorithm.The performance of SVRUKF in computation time is not satisfactory.How to improve the real-time performance of SVRUKF is a research direction in the future.

    Declaration of competing interest

    All the authors agreed to publish the paper.We do not have any conflict of interest.

    猜你喜歡
    百度
    百度年度熱搜榜
    青年與社會(2018年2期)2018-01-25 15:37:06
    不懂就百度
    做一個“百度”不知道的老師
    百度商業(yè)模式“罪與罰”
    公民與法治(2016年6期)2016-05-17 04:10:23
    百度年度人氣萌娃
    鳳凰生活(2016年1期)2016-01-07 17:45:39
    百度遭投行下調(diào)評級
    IT時代周刊(2015年8期)2015-11-11 05:50:22
    百度醫(yī)生
    百度“放衛(wèi)星”,有沒有可能?
    太空探索(2014年4期)2014-07-19 10:08:58
    百度小說排行榜(2014年6月)
    四部委責(zé)令百度、快播停止盜版侵權(quán)
    日本免费在线观看一区| 亚洲精品国产成人久久av| 男女边摸边吃奶| 在线天堂最新版资源| 小蜜桃在线观看免费完整版高清| 秋霞在线观看毛片| 青青草视频在线视频观看| 国产午夜精品久久久久久一区二区三区| 夜夜看夜夜爽夜夜摸| 欧美激情极品国产一区二区三区 | 久久综合国产亚洲精品| 一级片'在线观看视频| 国产一区二区在线观看日韩| 男女边摸边吃奶| 免费观看a级毛片全部| 亚洲国产精品专区欧美| 亚洲一级一片aⅴ在线观看| 免费久久久久久久精品成人欧美视频 | 简卡轻食公司| 日韩 亚洲 欧美在线| 免费看av在线观看网站| 狂野欧美激情性bbbbbb| 国产中年淑女户外野战色| 狠狠精品人妻久久久久久综合| 老女人水多毛片| 尾随美女入室| av福利片在线观看| 黄色配什么色好看| 久久人妻熟女aⅴ| av福利片在线观看| 中文字幕av成人在线电影| 日本-黄色视频高清免费观看| 亚洲电影在线观看av| 91久久精品电影网| 亚洲真实伦在线观看| 街头女战士在线观看网站| 搡女人真爽免费视频火全软件| www.色视频.com| 久久人人爽人人爽人人片va| 国产黄色视频一区二区在线观看| 在线免费观看不下载黄p国产| 麻豆精品久久久久久蜜桃| 中文字幕免费在线视频6| 精品人妻一区二区三区麻豆| 另类亚洲欧美激情| 内射极品少妇av片p| 欧美区成人在线视频| 黑丝袜美女国产一区| 人人妻人人澡人人爽人人夜夜| 黄色配什么色好看| 99视频精品全部免费 在线| 熟女电影av网| 老司机影院毛片| 人人妻人人澡人人爽人人夜夜| 国产一区二区三区综合在线观看 | 极品少妇高潮喷水抽搐| 狠狠精品人妻久久久久久综合| 久久精品国产亚洲av天美| 极品少妇高潮喷水抽搐| 欧美老熟妇乱子伦牲交| 日本免费在线观看一区| 久久精品国产自在天天线| 国产亚洲欧美精品永久| 最近手机中文字幕大全| 中文字幕亚洲精品专区| 青春草视频在线免费观看| 青春草视频在线免费观看| 国产精品一区二区三区四区免费观看| 久久国产精品大桥未久av | 国产伦在线观看视频一区| 久久久久久久久久人人人人人人| 精品人妻偷拍中文字幕| 亚洲熟女精品中文字幕| 国产成人91sexporn| 一本色道久久久久久精品综合| 多毛熟女@视频| 卡戴珊不雅视频在线播放| 简卡轻食公司| 纯流量卡能插随身wifi吗| 婷婷色av中文字幕| 一级av片app| 国产色婷婷99| 久久毛片免费看一区二区三区| 岛国毛片在线播放| 国产免费福利视频在线观看| 国产高潮美女av| av在线蜜桃| 亚洲精品久久久久久婷婷小说| 久久精品人妻少妇| 亚洲一区二区三区欧美精品| 男女免费视频国产| 午夜日本视频在线| 一区二区三区精品91| 街头女战士在线观看网站| av免费观看日本| 大片电影免费在线观看免费| 精品久久国产蜜桃| 观看美女的网站| 午夜福利视频精品| 99久国产av精品国产电影| 精品人妻熟女av久视频| 日本欧美国产在线视频| 性色av一级| 美女视频免费永久观看网站| 精品久久久噜噜| 色综合色国产| 久久精品人妻少妇| 亚洲伊人久久精品综合| 舔av片在线| 高清欧美精品videossex| 少妇熟女欧美另类| 国产老妇伦熟女老妇高清| 成人二区视频| 中文在线观看免费www的网站| 国产精品麻豆人妻色哟哟久久| 在线播放无遮挡| 免费大片黄手机在线观看| 午夜福利网站1000一区二区三区| 欧美日韩在线观看h| 亚洲真实伦在线观看| 少妇人妻 视频| 爱豆传媒免费全集在线观看| 七月丁香在线播放| 久久久久视频综合| 女人十人毛片免费观看3o分钟| 毛片女人毛片| 午夜精品国产一区二区电影| 日韩欧美一区视频在线观看 | 99久久精品一区二区三区| 人妻制服诱惑在线中文字幕| 亚洲第一区二区三区不卡| 亚洲精品亚洲一区二区| 精品国产乱码久久久久久小说| 我的女老师完整版在线观看| 噜噜噜噜噜久久久久久91| 啦啦啦中文免费视频观看日本| 一本久久精品| 久久久久国产精品人妻一区二区| 极品少妇高潮喷水抽搐| 日日摸夜夜添夜夜添av毛片| 蜜桃在线观看..| 欧美激情极品国产一区二区三区 | 美女高潮的动态| 国产91av在线免费观看| av线在线观看网站| av视频免费观看在线观看| 色5月婷婷丁香| 日本黄大片高清| 视频区图区小说| 自拍欧美九色日韩亚洲蝌蚪91 | 多毛熟女@视频| 少妇猛男粗大的猛烈进出视频| 久久精品久久精品一区二区三区| 亚洲高清免费不卡视频| 水蜜桃什么品种好| 亚洲国产毛片av蜜桃av| 国产精品人妻久久久久久| 亚洲最大成人中文| 午夜福利在线观看免费完整高清在| 亚洲精品视频女| xxx大片免费视频| 毛片一级片免费看久久久久| 六月丁香七月| 大香蕉97超碰在线| 99热这里只有是精品在线观看| 日日啪夜夜爽| 99久久精品一区二区三区| 一个人看的www免费观看视频| 国产黄频视频在线观看| 亚州av有码| 久久久久久久精品精品| 国产人妻一区二区三区在| 18禁动态无遮挡网站| 最近最新中文字幕大全电影3| 亚洲最大成人中文| 日韩欧美 国产精品| 黄片wwwwww| 高清日韩中文字幕在线| 青春草亚洲视频在线观看| 午夜老司机福利剧场| 18禁动态无遮挡网站| 亚洲av不卡在线观看| 女的被弄到高潮叫床怎么办| 涩涩av久久男人的天堂| 国产色婷婷99| 亚洲欧美成人精品一区二区| 人人妻人人澡人人爽人人夜夜| 亚洲激情五月婷婷啪啪| 国产精品av视频在线免费观看| 国产精品爽爽va在线观看网站| videossex国产| 亚洲人成网站在线播| 亚洲精品一二三| 91精品一卡2卡3卡4卡| 欧美xxⅹ黑人| 国产真实伦视频高清在线观看| 色吧在线观看| 欧美成人一区二区免费高清观看| 国产乱人偷精品视频| 国产精品人妻久久久久久| 在线天堂最新版资源| 色婷婷av一区二区三区视频| 国模一区二区三区四区视频| 亚洲精品乱码久久久v下载方式| 熟女av电影| 成人毛片60女人毛片免费| 少妇人妻久久综合中文| 美女福利国产在线 | 成年免费大片在线观看| 成人午夜精彩视频在线观看| 一级毛片黄色毛片免费观看视频| 国产永久视频网站| 草草在线视频免费看| 欧美zozozo另类| 亚洲精品第二区| 国产成人精品一,二区| 一级毛片黄色毛片免费观看视频| 国产午夜精品久久久久久一区二区三区| 国产一区亚洲一区在线观看| 噜噜噜噜噜久久久久久91| 毛片女人毛片| 午夜精品国产一区二区电影| 日韩成人伦理影院| 日韩欧美 国产精品| 亚洲国产欧美在线一区| 我要看黄色一级片免费的| 伊人久久国产一区二区| 国产精品一区二区三区四区免费观看| 欧美日韩精品成人综合77777| av又黄又爽大尺度在线免费看| 激情五月婷婷亚洲| 国产精品蜜桃在线观看| 色综合色国产| 自拍欧美九色日韩亚洲蝌蚪91 | 人妻系列 视频| 久久这里有精品视频免费| 寂寞人妻少妇视频99o| 国产精品欧美亚洲77777| 国产在视频线精品| 亚洲国产毛片av蜜桃av| 亚洲第一区二区三区不卡| 日日撸夜夜添| 2021少妇久久久久久久久久久| 国产成人精品一,二区| 制服丝袜香蕉在线| 五月玫瑰六月丁香| 精品久久久精品久久久| 极品少妇高潮喷水抽搐| 国产熟女欧美一区二区| 一区在线观看完整版| 精品一区在线观看国产| 国产亚洲精品久久久com| 内地一区二区视频在线| 伦精品一区二区三区| 亚洲美女黄色视频免费看| 欧美日韩视频高清一区二区三区二| 熟妇人妻不卡中文字幕| 成人亚洲欧美一区二区av| 菩萨蛮人人尽说江南好唐韦庄| 少妇猛男粗大的猛烈进出视频| 中文字幕免费在线视频6| 久久久久国产网址| 99热网站在线观看| 热99国产精品久久久久久7| tube8黄色片| 亚洲精品亚洲一区二区| 亚洲精品国产成人久久av| 欧美3d第一页| 国产一区二区三区综合在线观看 | 亚洲国产欧美在线一区| 少妇人妻精品综合一区二区| 亚洲,欧美,日韩| 国产黄频视频在线观看| 欧美另类一区| 韩国av在线不卡| 蜜桃在线观看..| 婷婷色综合大香蕉| 亚洲精品久久午夜乱码| 国产精品国产三级国产专区5o| 欧美激情极品国产一区二区三区 | 美女福利国产在线 | 纵有疾风起免费观看全集完整版| 十分钟在线观看高清视频www | videossex国产| 又黄又爽又刺激的免费视频.| 制服丝袜香蕉在线| 少妇的逼好多水| 在线免费十八禁| 80岁老熟妇乱子伦牲交| 亚洲精品久久久久久婷婷小说| 好男人视频免费观看在线| 久久这里有精品视频免费| 欧美日韩视频精品一区| 好男人视频免费观看在线| 精品一区二区三卡| 精品人妻视频免费看| 欧美日韩国产mv在线观看视频 | 日日撸夜夜添| 最近手机中文字幕大全| 91狼人影院| 国产精品一及| 欧美精品亚洲一区二区| 日本黄色日本黄色录像| 欧美精品国产亚洲| 国产精品伦人一区二区| 亚洲第一区二区三区不卡| 大片电影免费在线观看免费| 婷婷色av中文字幕| 国产欧美亚洲国产| 日本vs欧美在线观看视频 | 在线免费十八禁| 国产精品99久久久久久久久| 在线观看一区二区三区| 亚洲色图综合在线观看| 精品亚洲成a人片在线观看 | 亚洲四区av| 少妇的逼水好多| 国产熟女欧美一区二区| 在线观看三级黄色| 亚洲国产成人一精品久久久| 精品少妇黑人巨大在线播放| 精品久久久久久电影网| 美女cb高潮喷水在线观看| 国产精品不卡视频一区二区| 久久精品久久久久久噜噜老黄| 大香蕉97超碰在线| 蜜桃在线观看..| 午夜日本视频在线| 六月丁香七月| a级一级毛片免费在线观看| 黄片wwwwww| 最近手机中文字幕大全| h视频一区二区三区| 久久精品人妻少妇| 在线观看免费高清a一片| 国产免费又黄又爽又色| 亚洲美女黄色视频免费看| 毛片女人毛片| 伦精品一区二区三区| 亚洲国产av新网站| 在线免费观看不下载黄p国产| 国产精品.久久久| 久久久国产一区二区| 妹子高潮喷水视频| 我的老师免费观看完整版| 嫩草影院新地址| 中文字幕久久专区| 国产女主播在线喷水免费视频网站| 久久综合国产亚洲精品| 美女高潮的动态| 国产白丝娇喘喷水9色精品| 免费不卡的大黄色大毛片视频在线观看| 欧美精品亚洲一区二区| 久久精品国产亚洲av涩爱| 国产成人91sexporn| 国产老妇伦熟女老妇高清| 99re6热这里在线精品视频| 亚洲欧洲国产日韩| 99久久精品国产国产毛片| 80岁老熟妇乱子伦牲交| 国产午夜精品久久久久久一区二区三区| 精品少妇久久久久久888优播| 99热这里只有精品一区| 色网站视频免费| 欧美国产精品一级二级三级 | 欧美3d第一页| 国产有黄有色有爽视频| 久久人人爽av亚洲精品天堂 | 色吧在线观看| 国产精品熟女久久久久浪| 99久久精品一区二区三区| 又大又黄又爽视频免费| 中文字幕精品免费在线观看视频 | 日日摸夜夜添夜夜添av毛片| 国产精品秋霞免费鲁丝片| 婷婷色综合www| 国产伦在线观看视频一区| 亚洲精品色激情综合| 国产视频内射| 午夜福利在线在线| 久久精品久久精品一区二区三区| 国产av一区二区精品久久 | 成人无遮挡网站| 人妻一区二区av| 中文精品一卡2卡3卡4更新| 久久精品国产亚洲网站| 性高湖久久久久久久久免费观看| 国产成人精品婷婷| 成年av动漫网址| 欧美亚洲 丝袜 人妻 在线| 国产精品熟女久久久久浪| 不卡视频在线观看欧美| 亚洲伊人久久精品综合| 国产乱来视频区| 欧美区成人在线视频| 啦啦啦视频在线资源免费观看| 久久久欧美国产精品| 欧美三级亚洲精品| 日本爱情动作片www.在线观看| 久久久久久久亚洲中文字幕| 国产午夜精品久久久久久一区二区三区| 亚洲欧美清纯卡通| 国产成人freesex在线| 久久久久精品久久久久真实原创| 免费观看无遮挡的男女| 秋霞伦理黄片| 日韩精品有码人妻一区| 2021少妇久久久久久久久久久| 国产午夜精品一二区理论片| 亚洲成人中文字幕在线播放| 春色校园在线视频观看| 高清视频免费观看一区二区| 菩萨蛮人人尽说江南好唐韦庄| 狂野欧美白嫩少妇大欣赏| 亚洲av二区三区四区| 成人漫画全彩无遮挡| av黄色大香蕉| 这个男人来自地球电影免费观看 | 天堂俺去俺来也www色官网| 精品国产露脸久久av麻豆| 美女xxoo啪啪120秒动态图| 国产 精品1| 免费看光身美女| 久久久久视频综合| 免费看日本二区| 永久网站在线| 国产成人a∨麻豆精品| 精品一品国产午夜福利视频| 七月丁香在线播放| 精品99又大又爽又粗少妇毛片| 亚洲精品日韩在线中文字幕| 在现免费观看毛片| 性高湖久久久久久久久免费观看| 国产精品一区二区在线不卡| 免费人成在线观看视频色| 十分钟在线观看高清视频www | 久久 成人 亚洲| 免费看不卡的av| 中文精品一卡2卡3卡4更新| 久久精品国产鲁丝片午夜精品| 亚洲av成人精品一二三区| 建设人人有责人人尽责人人享有的 | 国产日韩欧美亚洲二区| 国产精品无大码| 丝袜脚勾引网站| 亚洲欧美一区二区三区黑人 | 亚洲欧美日韩无卡精品| 亚洲精品第二区| 亚洲欧美一区二区三区黑人 | 久久国产乱子免费精品| 日韩欧美精品免费久久| 久久鲁丝午夜福利片| 伦理电影免费视频| 日韩电影二区| 五月伊人婷婷丁香| 天堂俺去俺来也www色官网| 婷婷色麻豆天堂久久| 久久久色成人| .国产精品久久| 丰满迷人的少妇在线观看| 一区二区三区免费毛片| 亚洲人成网站高清观看| 一区二区av电影网| 建设人人有责人人尽责人人享有的 | 久久久欧美国产精品| 国产 一区精品| 欧美丝袜亚洲另类| 人体艺术视频欧美日本| 毛片一级片免费看久久久久| 亚洲最大成人中文| 亚洲精品国产色婷婷电影| 99热这里只有是精品50| 搡女人真爽免费视频火全软件| 天堂俺去俺来也www色官网| 国产女主播在线喷水免费视频网站| 最后的刺客免费高清国语| 国产探花极品一区二区| 黑人猛操日本美女一级片| 国产亚洲av片在线观看秒播厂| 亚洲一级一片aⅴ在线观看| 国产色婷婷99| 久久久国产一区二区| av在线老鸭窝| 欧美精品国产亚洲| av视频免费观看在线观看| 91久久精品国产一区二区三区| 国产av一区二区精品久久 | 老司机影院成人| 国产av码专区亚洲av| 免费观看性生交大片5| 成人午夜精彩视频在线观看| 免费久久久久久久精品成人欧美视频 | 噜噜噜噜噜久久久久久91| 联通29元200g的流量卡| 成人国产麻豆网| 中文字幕免费在线视频6| 婷婷色麻豆天堂久久| 五月天丁香电影| 精品熟女少妇av免费看| 国模一区二区三区四区视频| 免费大片黄手机在线观看| 青春草亚洲视频在线观看| 亚洲丝袜综合中文字幕| xxx大片免费视频| av线在线观看网站| 美女主播在线视频| 国产精品蜜桃在线观看| 国产午夜精品久久久久久一区二区三区| 亚洲欧美日韩另类电影网站 | 日韩中文字幕视频在线看片 | 国产伦精品一区二区三区视频9| 日韩伦理黄色片| 妹子高潮喷水视频| 只有这里有精品99| 日韩一区二区三区影片| 国产爱豆传媒在线观看| 五月天丁香电影| 天堂中文最新版在线下载| 在现免费观看毛片| 在线看a的网站| 男人爽女人下面视频在线观看| 亚洲欧美精品自产自拍| 夜夜骑夜夜射夜夜干| 国产精品一区二区在线观看99| 99久久综合免费| 久久女婷五月综合色啪小说| 亚洲高清免费不卡视频| 人人妻人人澡人人爽人人夜夜| 亚洲aⅴ乱码一区二区在线播放| 国产毛片在线视频| a级毛片免费高清观看在线播放| 秋霞伦理黄片| 黑人猛操日本美女一级片| 久热这里只有精品99| 免费av中文字幕在线| 国产av一区二区精品久久 | 麻豆成人午夜福利视频| 永久网站在线| 国产精品99久久99久久久不卡 | 免费观看无遮挡的男女| 亚洲经典国产精华液单| 嫩草影院入口| 午夜福利在线观看免费完整高清在| 国产 精品1| av播播在线观看一区| 高清毛片免费看| 丰满人妻一区二区三区视频av| 日韩电影二区| 国产精品爽爽va在线观看网站| 国产探花极品一区二区| 国产精品久久久久久av不卡| 波野结衣二区三区在线| 午夜福利视频精品| 看十八女毛片水多多多| 男男h啪啪无遮挡| 黑丝袜美女国产一区| 国内精品宾馆在线| 精品人妻偷拍中文字幕| 亚洲欧美日韩另类电影网站 | 青春草亚洲视频在线观看| 熟女av电影| 久久精品久久精品一区二区三区| 国产精品.久久久| 午夜老司机福利剧场| 亚洲第一区二区三区不卡| a 毛片基地| 黄片无遮挡物在线观看| 久久久久久久国产电影| 久久久久久久久大av| 在线观看一区二区三区激情| 亚洲内射少妇av| 97超碰精品成人国产| 美女xxoo啪啪120秒动态图| 大香蕉久久网| 夜夜骑夜夜射夜夜干| 欧美精品人与动牲交sv欧美| 日韩一区二区三区影片| 26uuu在线亚洲综合色| 极品少妇高潮喷水抽搐| 久久久国产一区二区| 国产精品不卡视频一区二区| 91在线精品国自产拍蜜月| 精品国产一区二区三区久久久樱花 | 免费观看的影片在线观看| 久久久久久九九精品二区国产| 青春草国产在线视频| av国产免费在线观看| 蜜桃久久精品国产亚洲av| 一区二区av电影网| 老师上课跳d突然被开到最大视频| 又粗又硬又长又爽又黄的视频| 麻豆成人av视频| 一二三四中文在线观看免费高清| 肉色欧美久久久久久久蜜桃| 亚洲国产色片| av在线观看视频网站免费| 亚洲人成网站在线播| 亚州av有码| 久久亚洲国产成人精品v| 午夜福利视频精品| 寂寞人妻少妇视频99o| 干丝袜人妻中文字幕| 久久久久久伊人网av| 国产在线男女| 少妇猛男粗大的猛烈进出视频| 夜夜看夜夜爽夜夜摸| 亚洲精品成人av观看孕妇| 久久精品久久久久久噜噜老黄| av在线老鸭窝| 国产综合精华液| 美女脱内裤让男人舔精品视频| 国产淫片久久久久久久久| 国产亚洲5aaaaa淫片| 国产爽快片一区二区三区| 午夜激情久久久久久久|