• <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 ?

    A non-myopic scheduling method of radar sensors for maneuvering target tracking and radiation control

    2020-04-09 18:45:20GanlinShanGongguoXuChenglinQiao
    Defence Technology 2020年1期

    Gan-lin Shan,Gong-guo Xu,Cheng-lin Qiao

    Department of Electronic and Optical Engineering,Army Engineering University,Shijiazhuang 050003,China

    Keywords:Non-myopic scheduling Radiation control PCRLB POMDP Branch and bound

    ABSTRACT In decades,the battle field environment is becoming more and more complex with plenty of electronic equipments.Thus,in order to improve the survivability of radar sensors and satisfy the requirement of maneuvering target tracking with a low probability of intercept,a non-myopic scheduling is proposed to minimize the radiation cost with tracking accuracy constraint.At first,the scheduling problem is formulated as a partially observable Markov decision process(POMDP).Then the tracking accuracy and radiation cost over the future finite time horizon are predicted by the posterior carm'er-rao lower bound(PCRLB)and the hidden Markov model filter,respectively.Finally,the proposed scheduling is implemented efficiently by utilizing the branch and bound(B&B)pruning algorithm.Simulation results show that the performance of maneuvering target tracking was improved by the improved interacting multiple model(IMM),and the scheduler time and maximum memory consumption were significant reduced by the present B&B pruning algorithm without losing the optimal solution.?2020 China Ordnance Society.Production and hosting by Elsevier B.V.on behalf of KeAi Communications Co.This is an open access article under the CCBY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).

    1.Introduction

    One of the most successful techniques to improve the performance of target tracking system is sensor scheduling[1-3].Sensor scheduling is myopic if it only performs one step ahead in the future.Although the myopic scheduling has low computational and memory cost,sometimes the performance of myopic sensor scheduling is poor.By contrast,the non-myopic scheduling which performs multi-step ahead in the future is better[4,5].A typical objective of sensor scheduling is to select sensor to minimize the tracking cost under constraints over time.In Refs.[6-8],the objective is to minimize the tracking error with dwell time constraint[6],dwell time constraint[7],power constraint[8,9]and sensor resource constraints[10],respectively.In fact,on the other hand,the tracking error is often considered as a constrain[11,12].For example,the objective is to minimize the energy consumption subject to tracking accuracy constrain in the wireless sensor network[12].

    In this paper,we consider a sensor scheduling scenario where multiple active sensors track a maneuvering target in the combat system.As the emitting energy can be captured by enemy surveillance sensors,the radiation cost of the combat system is more important than the other factors such as energy consumption[13].In the active/passive multi-sensor system,the radiation cost is calculated by the times of active sensor selection as the passive sensor cannot emit energy[14].In order to further control the radiation,the radiation cost is quantized in Ref.[15],different active sensors are set as different fixed cost,then a simulation-based Q value approximation method is utilized to obtain suboptimal solution.How ever,this approach is usually not suitable for the practice situation.Moreover,using the interception probability factor to measure the radiation cost is another common method[16-19].For example,combining with low probability of intercept(LPI),the entropy of all the sensors with no interception is denoted as the radiation cost in Ref.[16].But,as the prior know ledge of the enemy parametric windows is difficult to obtain,it is hard to compute the interception probability.To avoid this problem,the emission level impact(ELI)is used to represent the radiation cost instead of the interception probability in Refs.[20,21].How ever,the goal of the above work is to trade off the radiation cost and tracking performance that is traditionally been ill-suited for achieving specific control objectives in the combat system[22],and the balance coefficient is always given based on empirical value[15].In addition,the target is almost assumed as non-maneuvering target in the above work.In fact,to dodge tracking,the target is always maneuvering.The interacting multiple model(IMM)algorithm is simple and practical method for maneuvering target tracking.It has been applied to many sensor scheduling systems to track maneuvering target[23-25].As we know,the crucial feature of sensor scheduling is predictable,then how to obtain the predicted tracking accuracy is important.In Ref.[23],the predicted tracking accuracy is obtained without utilizing the measurement,and an adaptive resource management algorithm is proposed to reduce the LPI in radar network.Obviously,the predicted tracking error is larger than the estimated tracking error,sometimes the resource management scheme may not be an optimal scheme.To solve this issue,the predicted tracking accuracy is calculated with a model index prediction algorithm in Refs.[24,25].How ever,the studies of[23-25]are all myopic scheduling,and few researches of non-myopic scheduling for maneuvering target tracking and radiation control are exist.

    In this work,it aims to minimize the radiation cost with the tracking accuracy constraint which depends on the tactical requirement in the maneuvering target tracking.Firstly,the IMM is improved by using an adaptive Markov transition probability matrix which is modified by the error compression ratio and model probability variance ratio.Then the posterior carm'er-rao lower bound(PCRLB)is utilized to predict the tracking accuracy.Secondly,the radiation cost is represented by ELI and it is predicted by using hidden Markov model(HMM)filter.Finally,the branch and bound(B&B)pruning algorithm with the suboptimal lower bound is proposed to solve the exponential computational grow th problem.

    2.Problem formulation

    In this paper,we consider a typical target tracking scenario.There are N active sensors distributed in a sensor field to track a maneuvering target.As the system state is not directly observable,we formulate the scheduling problem as a POMDP.

    2.1.Action space and system state

    To represent the sensor assignment,the sensor scheduling action at time step k+1 is denoted by,where the vector element=1 or=0 specifies whether the sensor n is active or inactive at time step k+1.For clarity and easy of presentation,all the sensors are independent of each other and only one sensor per time step is activated to track target.

    In this paper,the system state vector Skconsists of the target state Xkand the ELI state Ek.Then it can be written as

    where[xk,yk,zk]represents the target position and[˙xk,˙yk,˙zk]is the corresponding velocity.Enkdenotes the cumulative received radiation registered by the enemy surveillance system from sensor n(n=1,…,N)till time step k[20].The superscript T represents transpose.

    For maneuvering target tracking,the target state Xkcan be estimated efficiently by IMM[25].Therefore,the dynamic state evolves according to the target motion model as

    where Fmis the state transition matrix of model m(1,…,M)and M is the number of model.wkdenotes the zero-mean white Gaussian process noise with covariance matrix Q.

    According to Ref.[20],the ELI state is quantized to a finite set{1,…,Ns},each value corresponding to a physical threat level and Nsis highest threat level.Then the ELI state can be transited by several transition matrices Tnk(n=1,…,N).If the sensor n is activated to track target,then

    2.2.Observation

    Similarly,the system observation vector Zkcan be written as

    where ZXkand ZEkare the target measurement and the ELI measurement,respectively.

    Moreover,the target measurement ZXkis produced by the sensor measurement model which can be given as

    where g denotes the measurement function of the selected sensor and vkis assumed as a zero-mean white Gaussian measurement noise.For the active sensors,the measurement includes range rk,azimuth θkand elevation φk

    2.3.Optimization function

    We denote a scheduling sequence by a H-tupleak+1,…,ak+H-1],where H is the prediction time horizon.At any given time k,the goal is to obtain the optimal scheduling sequencein order to minimize the radiation cost while maintaining a desired tracking accuracy.Obviously,there are NHpossible scheduling sequences.Moreover,we define the radiation cost function at time step k+h-1 by r(ak+h-1)which is the sum of all selected sensor ELI states.Then the optimization function is described as

    Since the system state is not directly observable,a POMDPkeeps scheduling of the belief state[3].We definewhere bXkand bEkdenote target belief state and ELI belief state,respectively.Then equation(8)can be further stated as

    where symbol Erepresents expectation.

    3.Problem solution

    Fig.1 shows the flow chart of the proposed non-myopic scheduling method.At time step k,the non-myopic scheduling process can be described as follow s:

    Step 1:Initialization.Get the target state Xkand the ELI belief state bEkat time step k.

    Step 2:Predict the tracking accuracy and the radiation cost in the future,then search the optimal sensor scheduling sequence φk:k+H-1by improved decision tree searching algorithm.

    Step 3:Select the first action akin the optimal sensor scheduling sequence φk:k+H-1to control the radar sensor.

    Step 4:k=k+1,get the system observation vector Zk+1.If the task is not finished,update the target state and ELI belief state by improved IMM algorithm and HMM,respectively,and then go to Step 1.Otherwise,go to Step 5.

    Step 5:End.

    Fig.1.Flowchart of non-myopic scheduling method.

    3.1.Adaptive Markov transition probability matrix

    As it know n,the IMM is always utilized to track maneuvering target.How ever,the fixed Markov transition probability matrix is always inaccurate and it does not fully describe the conversion between different models during the filtering.In fact,when the target motion state has a tendency to a motion model,the corresponding transition probability should be increased.According to this idea,the posterior information is utilized to revise the Markov transition probability matrix.Here,the posterior information consists of the error compression ratio[26]and the model probability variance ratio.

    We define the error compression ratio as

    Then the transition probability is modified as

    Obviously,the matching model information is magnified and the non-matching models information is omitted in(11).How ever,the modified transition probability matrix may be a non-positive matrix[25].Therefore,two necessary conditions are given as

    ·As the transition probability matrix is a diagonally-dominant matrix,the second necessary condition is given as

    Furthermore,as the error compression ratio only consider the current model information without considering the history model information,sometimes this method is not stable and accurate.To solve this issue,we utilize the model probability variance ratio to further modify the transition probability matrix.

    As the transition probability is nonnegative,we define the model probability variance ratio κjkas

    It can be seen that when the model probability of model j is increased,κjk>1.Otherwise,κjk<1.Based on this,κjkis used to further modify the transition probability matrix.It is described as

    Note that the sum of all the elements of each row in the transition probability matrix may not equal to one.Thus the normalization process is stated as

    3.2.PCRLB for tracking accuracy prediction

    As the further measurements cannot be obtained at the current time step,then the PCRLB is utilized to predict the target tracking accuracy.As it know n,PCRLB is the inverse of the Fisher information matrix(FIM)and expresses the lower bound on the tracking error.Denoting the FIM by Jk,then it is calculated recursively as

    with

    How ever,the mk+1cannot be obtained at time step k.To obtain the predicted model index,the model probabilityμkyielded by IMM algorithm is used as

    where μikrepresents the model probability of motion model i.

    Consequently,the procedure of the propose tracking accuracy prediction is described as

    Step 1:Obtain the target statecovariance matrix Pkand model probability μkfrom the target belief state bXkwhich is estimated by IMM.

    Step 2: Predict the target motion model ~mk+h=

    Step 3:Calculate the FIM Jk+hby using Eq.(17).

    Step 4:Compute the predicted tracking accuracy

    3.3.HMM for radiation cost prediction

    Specifically,to use the basic theory of HMM to derive,the system state at current time should be only related to the state at previous time.Due to the continuity of target motion,the target state and radiation state meet the above requirements.Therefore,after the ELI measurement ZEk+1is achieved by the selected sensor n,the ELI belief state can be updated as

    where symbol?denotes Hadamard product,and“1”is a Ms-dimensional unit column vector.

    As ZEk+1cannot be achieved at time step k,its distribution probability is given as

    Then the ELI belief state prediction is stated as

    Moreover,the one step radiation cost is calculated as

    where Dn=[1,…,Ns]Tis a vector w hose element denotes the nature number in{1,…,Ns}.

    Therefore,the total radiation cost over the future H time steps is stated as

    From the derivation process,unlike the commonly used interception probability method,using the HMM to derive the radiation cost is in line with the actual situation.It is not necessary to know the working parameters of enemy targets,which can be well applied to the target tracking process.

    3.4.B&B pruning algorithm for sensor scheduling

    We translate the sensor scheduling problem in Section 2.3 to a decision tree optimization problem.The depth and branching factor of the tree are equal to the time horizon length and the number of sensors,respectively.Fig.2 shows a decision tree of N=4 and H=3.We de f ine each depth-h(h=1,…,H)node aswhere is the node position.As shown in Fig.2,the decision tree is expanded by the scheduling actions[ak,…,ak+H-1],and each path from the root to a leaf node corresponds to one possible sensor scheduling sequence.

    Therefore,our objective is to find the optimal path with the lowest radiation cost such that the tracking accuracy is maintained within a desired level at each time step.Obviously,searching all the possible paths is computationally infeasible.As a result,it becomes imperative to develop an efficient search technique to reduce the search effort.

    Fig.2.An illustrative decision tree with N=4 and H=3.

    The branch and bound pruning algorithm is often utilized to prune redundant branches in the decision tree.How ever,there are two necessary conditions for its application.The first necessary condition is that each node lower bound on the cost is easier to obtain with respect to the actual cost.Another one is the one step cost must be non-negative to ensure the cost of any children of this node is greater than itself.Fortunately,the above conditions are all satisfied in our scheduling problem.By means of the lower bound,the tree is efficiently pruned.Then any node whose lower bound is smaller than the current minimum cost is expanded first.On the other hand,any node which has a bigger lower bound is pruned.

    We assume that a node is reached by utilizing the scheduling sequence φk:k+h-1= [ak,…,ak+h-1].Then the low er bound L(φk:k+h-1)of this node is stated as

    For any given ELI state transition matrix Tn(n=1,…,N),there is a total of NH-hdistinct values of R(φk+h:k+H-1).Therefore,it is difficult to obtain the optimal lower bound of R(φk+h:k+H-1),then a suboptimal lower bound is utilized to instead of it.For each possible sensor sequence φk+h:k+H-1,we have assumed that the tracking accuracy of each possible sensor sequence is maintained with the desired level at each of the future H-h time steps.Then the radiation cost R(φk+h:k+H-1)is translated to a function of Tn.Consequently,the suboptimal low er bound of R(φk+h:k+H-1)is stated as

    Moreover,the suboptimal lower bound of L(φk:k+h-1)is written as

    Obviously,there are only a total of N(H-h)nodes should be opened to obtain the suboptimal lower bound,and the calculation of tracking accuracy is not required in this process.As a result,the B&B pruning algorithm for sensor scheduling is shown in Algorithm 1.

    Algorithm 1 B&B pruning algorithm.

    4.Simulation results

    Effectiveness of the proposed sensor scheduling is validated through Monte Carlo simulations.As is shown in Fig.3,in our simulations,four active sensors are utilized to track a maneuvering target.The initial target position and velocity are(15,4,5)km and(-280,-260,0)m/s,respectively.The sampling interval is Ts=1 and the simulation duration is 100s.Without loss of generality,we assume that there are three models are employed:the nearly constant velocity(NCV)model,nearly left constant turn(NLCT)model and nearly right constant turn(NRCT)model.Then the target turns right with ω=-5°during 26 s-50 s,turns left with ω=5°during 51 s-74 s,and maintains uniform motion during the other time.Moreover,the initial model probability is μ=[0.8,0.1,0.1]Tand the initial model switching probability matrix is πij=0.025(i≠j).

    As is shown in Fig.3,sensor 1,sensor 2,sensor 3,and sensor 4 are deployed at(0,-5,0)km,(-5,0,0)km,(5,0,0)km,and(0,5,0)km,respectively.Moreover,the other working parameters of sensors can be summarized as

    Fig.3.Diagram of the simulation scenario.

    sensor 1:σr1=200 m,σθ1=σφ1=0.01 rad

    sensor 2:σr1=100 m,σθ1=σφ1=0.005 rad

    sensor 3:σr1=100 m,σθ1=σφ1=0.005 rad

    sensor 4:σr1=10 m,σθ1=σφ1=0.001 rad

    where σrn,σθnand σφnare the standard deviation of range noise,azimuth noise and elevation noise of sensor n(n=1,…,4),respectively.It can be seen that sensor 1 is more unlikely to increase the ELI but with the worst tracking performance.Sensor 4 has the best tracking performance but it is more likely to increase the ELI.Then we quantize the ELI state as{1,2,3}whose element denotes different threat level.Similarly,the ELI measurement is quantized as{1,2,3}w hose element denotes different increment in the threat level.Furthermore,the ELI state transition matrices are

    Each simulation result is averaged over 200 Monte Carlo simulation runs.

    4.1.Comparison of IMM algorithms

    To show the advantage of our proposed improved IMM(I-IMM),the standard IMM(SIMM)and IMM with the error compression ratio(EIMM)[23]are used for comparison.The root mean-square error(RMSE)curves of the target position and velocity are shown in Fig.4.Fig.5 shows the model probability of the three algorithms.

    It can be seen in Fig.4 that the target position and velocity RMSE curves of I-IMM algorithm are generally lower than EIMM and SIMM algorithms,and the improvement of the target speed is more obvious.It is because that I-IMM algorithm utilizes the error compression ratio and the variance ratio of model probability to modify the probability transition matrix.The matching probability of the model is highest during the target tracking process.Therefore,the corresponding target position and velocity RMSE are smallest.The target position and velocity RMSE curves of EIMM is only better than SIMM when the target moves with NRCT and NLCT.It is because that EIMM algorithm only utilizes the error compression ratio to modify the probability transition matrix,which is not sensitive to the change of target motion mode.Moreover,as the model probability of the three algorithms are inaccurate during the model switching,the tracking error in position and velocity of the three algorithms are comparatively large,which is in line with the actual situation.

    Fig.4.Position and velocity RMSE versus time step,(a)position RMSE and(b)velocity RMSE.

    Fig.5.Comparison of the model probability.

    Besides,it can be seen in Fig.5 that the model probability of the matching model of I-IMM is generally highest all the time step.Meanwhile,the performance of EIMM performs better than SIMM in model probability when the target moves with NRCT and NLCT.In general,the higher the matching probability is,the higher the tracking accuracy is.It is consistent with the results in Fig.4,which proves the effectiveness of the improved strategy of I-IMM.

    4.2.Comparison of pruning algorithms

    Next,we compare the pruning performance of enumerative search(ES),uniform cost search(UCS)and B&B.Table 1 summarizes the pruning statistics which include the average number of nodes opened and the maximum number of nodes stored for the case of ρth=50 m.As it know n,the memory consumption and the scheduler time are proportional to the maximum number of nodes stored and average number of nodes opened,respectively[4].It can be seen that both the average number of nodes opened and the maximum number of nodes stored of B&Bare relatively lower than UCS.Moreover,as the decision step H increases,the advantage of B&Bis more obvious.Thus the percentage of the average number of nodes opened decreases faster.Fig.6 shows the percentage of node opened versus time step for the case of decision step H=3.It is evidently that the pruning performance of B&B is better than UCS all the time.

    It is obvious in Fig.7 that the cumulative radiation cost is decreasing with H.Then the necessity of the proposed non-myopic scheduling is proved by this result.Considering the computation complexity,we choose H=3 in the next simulations.

    4.3.Comparison of scheduling policies

    In this simulation,firstly,each sensor is used separately to track the target to show differences in tracking performance of different sensors with different sensor precision.Similar results are shown in Fig.8.It can be seen that the position RMSE of the sensor 1 is the largest and the position RMSE of the sensor 4 is smallest,which correspond to the working parameters of the radar sensors.However,although the tracking error of sensor 4 is small,sensor 4 cannot be used for tracking all the time.Because the radiation cost of sensor 4 will be very large when it is used for tracking all the time.Therefore,it is necessary for the radar sensors to cooperate with each other to reduce the radiation cost and ensure high target tracking accuracy.

    Moreover,to investigate the performance of our proposed scheduling(PS),three other scheduling methods are utilized for comparison.There are closest scheduling(CS),myopic scheduling(MS)and myopic scheduling with standard IMM(SMS).

    We schedule the sensors for the desired tracking accuracy ρth=20 m,30 m,40 m,50 m,60 m,70 m and 80 m.Figs.9(a)and(b)show the position RMSE curves for the case of ρth=30 m and ρth=50 m.As the CS RMSE does not change for different desired tracking accuracy and in order to show the performance of the other scheduling methods clearly,Fig.9(b)does not show the CS RMSE curve.We note that the position RMSE curves of SMS,MS and PS are generally below the desired tracking accuracy ρthexpect during the model switching.It indicates that the scheduling methods aid in satisfying the desired tracking accuracy.By contrast,the CS RMSE curve does not meet the desired tracking accuracy.Furthermore,as the model probability is inaccurate during the model sw itching,the corresponding RMSE curves are above ρth.Similar results can also obtained for other values of ρth.

    Table 1 Comparisons of pruning algorithms.

    Fig.6.Percentage of node opened versus time step,(a)UCS and(b)B&B.

    Fig.7.Cumulative radiation cost versus decision step H.

    Fig.10 compares the cumulative radiation cost for different desired tracking accuracy.We can clearly find in Fig.10 that the cumulative radiation cost can be reduced efficiently for different desired accuracy by using SMS,MS and PS,respectively.Moreover,as the SMS utilizes the standard IMM for target tracking,the predicted tracking error is always comparatively large.Then it will activate more accurately sensor,such as sensor 4 which is more likely to increase the ELI.Thus the cumulative radiation cost of SMS is larger than MS and PS.Furthermore,it is obvious that the PS can always obtain the minimum cumulative radiation cost,as the PS can find better sensor sequence for target tracking than MS.Fig.11 shows the comparison of cumulative radiation cost and for the case of ρth=50 m.It can be seen that the cumulative ELI value is almost equal to the cumulative radiation cost.Consequently,the effectiveness of the optimization function is proved.

    Fig.8.Position RMSE versus different sensors.

    Fig.9.Position RMSE versus time step,(a)ρth=30m and(b)ρth=50m.

    Fig.10.Cumulative radiation cost versus desired tracking accuracy.

    Fig.11.Comparison of cumulative radiation cost and ELI value.

    Besides,it can be seen from the experimental results that this paper focuses on the sensor selection strategy to control the radiation cost,which does not discuss the specific working parameters of the sensor.But,the proposed sensor scheduling method is also effective when considering the specific working parameters.We only need to modify the optimization objective or the constraint of the proposed scheduling method.For example,when considering the sensor dwell time,the constraint of sensor switching frequency should be added.When considering the bandwidth and transmit power,the initial level of ELI state should be modified.Specifically,the sensor with large transmit power has a large ELI state.In the future,more detailed researches about the other working parameters will be addressed.

    5.Conclusions

    The goal in this paper is to control the sensor system radiation cost under the constrain of tracking error by using non-myopic scheduling.We formulated this sensor scheduling problem as a POMDP,and an improved IMM algorithm is presented.To reduce the computation of non-myopic scheduling problem,the B&B pruning algorithm with the suboptimal lower bound was presented.Simulation results indicated that the performance of maneuvering target tracking was improved by the improved IMM,and the scheduler time and maximum memory consumption were significant reduced by the present B&B pruning algorithm without losing the optimal solution.But,the start delay is not considered during sensor scheduling process.Therefore,how to reduce the influence of start delay on sensor scheduling is a future research direction.On the other hand,study on sensor scheduling methods in the clutter environment is another research direction.

    Declaration of competing interest

    The authors declare that they have no conflicts of interest to this work.We also declare that we do not have any commercial or associative interest that represents a conflict of interest in connection with the work submitted.

    Acknowledgement

    This work is supported by the National Defense Pre-research Foundation of China(012015012600A2203).

    日韩不卡一区二区三区视频在线| 我的老师免费观看完整版| 色婷婷久久久亚洲欧美| 我的女老师完整版在线观看| 国产探花极品一区二区| 水蜜桃什么品种好| 日韩av在线免费看完整版不卡| 成人高潮视频无遮挡免费网站| 丝袜脚勾引网站| 久久韩国三级中文字幕| 一区二区av电影网| 麻豆久久精品国产亚洲av| 熟女av电影| 久久久久性生活片| 国产精品女同一区二区软件| 欧美国产精品一级二级三级 | 在线看a的网站| 日韩电影二区| 日韩国内少妇激情av| 在线亚洲精品国产二区图片欧美 | 国产精品女同一区二区软件| 欧美精品一区二区大全| 狠狠精品人妻久久久久久综合| 91在线精品国自产拍蜜月| 亚洲欧美精品自产自拍| 激情 狠狠 欧美| 成年女人看的毛片在线观看| 欧美极品一区二区三区四区| 青青草视频在线视频观看| 肉色欧美久久久久久久蜜桃 | 精品国产三级普通话版| 免费av观看视频| 国产免费视频播放在线视频| 伦理电影大哥的女人| 女的被弄到高潮叫床怎么办| 超碰av人人做人人爽久久| 永久免费av网站大全| 欧美高清性xxxxhd video| 国产精品久久久久久精品电影| av在线亚洲专区| av国产久精品久网站免费入址| 日本色播在线视频| 人妻一区二区av| 麻豆精品久久久久久蜜桃| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 国产精品99久久99久久久不卡 | 欧美人与善性xxx| 精品久久久精品久久久| 成人国产av品久久久| 国内精品美女久久久久久| 国产亚洲av片在线观看秒播厂| 国产精品久久久久久av不卡| 免费看光身美女| 在线免费观看不下载黄p国产| 亚洲自偷自拍三级| 我的老师免费观看完整版| 国产亚洲一区二区精品| 91精品伊人久久大香线蕉| 精品人妻偷拍中文字幕| 久久精品国产鲁丝片午夜精品| 色播亚洲综合网| 日韩不卡一区二区三区视频在线| 人妻制服诱惑在线中文字幕| 99久久人妻综合| 日本av手机在线免费观看| 看非洲黑人一级黄片| 免费看不卡的av| 日韩大片免费观看网站| 国产一区有黄有色的免费视频| 久久久久网色| 国产av不卡久久| 久久久久久久久久人人人人人人| 亚洲国产欧美在线一区| 亚洲av中文字字幕乱码综合| 大又大粗又爽又黄少妇毛片口| 51国产日韩欧美| 嫩草影院新地址| 国产 一区精品| 欧美区成人在线视频| 国产v大片淫在线免费观看| 亚洲国产最新在线播放| 久久久久久久精品精品| 精品久久国产蜜桃| 久久久久久九九精品二区国产| 日韩在线高清观看一区二区三区| 色视频www国产| 欧美性感艳星| 亚洲av.av天堂| 亚洲精品自拍成人| 国产毛片a区久久久久| 老女人水多毛片| 久久99热6这里只有精品| 一个人观看的视频www高清免费观看| 亚洲一级一片aⅴ在线观看| 你懂的网址亚洲精品在线观看| 亚洲丝袜综合中文字幕| 国产黄色视频一区二区在线观看| 久久韩国三级中文字幕| 少妇人妻久久综合中文| 国产老妇女一区| 80岁老熟妇乱子伦牲交| 久久久久久久国产电影| 国产精品女同一区二区软件| www.色视频.com| 欧美日韩综合久久久久久| 国产伦在线观看视频一区| 大又大粗又爽又黄少妇毛片口| 国产精品蜜桃在线观看| 一个人观看的视频www高清免费观看| 国产亚洲91精品色在线| 少妇被粗大猛烈的视频| 国产黄片美女视频| 超碰97精品在线观看| 一级毛片久久久久久久久女| 高清日韩中文字幕在线| 日产精品乱码卡一卡2卡三| 人妻制服诱惑在线中文字幕| 在现免费观看毛片| 亚洲欧美精品专区久久| 国产精品一区www在线观看| 日本黄大片高清| 97在线人人人人妻| 午夜福利视频精品| 亚洲精品国产成人久久av| 97超视频在线观看视频| 成人黄色视频免费在线看| 在线观看人妻少妇| 麻豆久久精品国产亚洲av| 麻豆成人av视频| 精品久久久噜噜| 国产老妇伦熟女老妇高清| 久久国产乱子免费精品| 国产一区二区三区综合在线观看 | 亚洲色图综合在线观看| 国产精品国产三级国产av玫瑰| 亚洲第一区二区三区不卡| 国产欧美日韩精品一区二区| av在线老鸭窝| 老师上课跳d突然被开到最大视频| 欧美高清性xxxxhd video| 国产亚洲一区二区精品| 97精品久久久久久久久久精品| 日本一二三区视频观看| 国产成人a∨麻豆精品| 丝袜美腿在线中文| 大码成人一级视频| 搞女人的毛片| 国产一级毛片在线| 精品一区在线观看国产| 国产探花极品一区二区| 成人亚洲精品av一区二区| 我要看日韩黄色一级片| av免费观看日本| 老司机影院毛片| 老师上课跳d突然被开到最大视频| 国产黄片视频在线免费观看| 午夜福利高清视频| 赤兔流量卡办理| 特大巨黑吊av在线直播| 欧美日韩国产mv在线观看视频 | 五月天丁香电影| 成人黄色视频免费在线看| 亚洲国产高清在线一区二区三| 新久久久久国产一级毛片| 日本黄色片子视频| 18禁裸乳无遮挡免费网站照片| 嫩草影院精品99| 最新中文字幕久久久久| 国产精品成人在线| 麻豆久久精品国产亚洲av| 麻豆成人av视频| 久久这里有精品视频免费| 亚洲国产成人一精品久久久| 精品一区二区三区视频在线| 激情 狠狠 欧美| 亚洲熟女精品中文字幕| 天天一区二区日本电影三级| 国精品久久久久久国模美| 日日撸夜夜添| 精品少妇黑人巨大在线播放| 久久久久久久久久久丰满| 精品久久久久久久久av| 久久久国产一区二区| 亚洲一级一片aⅴ在线观看| 国产精品无大码| 国产欧美日韩精品一区二区| 91精品国产九色| 99久久精品国产国产毛片| 天美传媒精品一区二区| 七月丁香在线播放| 91久久精品电影网| 亚洲欧美成人精品一区二区| 欧美成人午夜免费资源| 久久精品国产亚洲网站| 禁无遮挡网站| 欧美精品人与动牲交sv欧美| 欧美区成人在线视频| 国产又色又爽无遮挡免| 精品一区二区三区视频在线| 大码成人一级视频| 国产黄频视频在线观看| 国产乱人视频| 亚洲精品aⅴ在线观看| 美女脱内裤让男人舔精品视频| 亚洲色图av天堂| 国产欧美日韩精品一区二区| 尤物成人国产欧美一区二区三区| 久久精品久久久久久久性| 黄色视频在线播放观看不卡| 久久久色成人| 亚洲欧美清纯卡通| 国产精品国产三级国产专区5o| 99久久人妻综合| 国产乱人偷精品视频| 国产亚洲av片在线观看秒播厂| 在线播放无遮挡| 看免费成人av毛片| 亚洲成人av在线免费| av在线蜜桃| 久久久久久久国产电影| 国产在线一区二区三区精| 又爽又黄a免费视频| 韩国av在线不卡| 蜜桃久久精品国产亚洲av| 亚洲av免费高清在线观看| 色播亚洲综合网| 激情 狠狠 欧美| 五月伊人婷婷丁香| 日韩欧美一区视频在线观看 | 亚洲人成网站在线播| 麻豆乱淫一区二区| 性插视频无遮挡在线免费观看| 身体一侧抽搐| 亚洲国产成人一精品久久久| 天美传媒精品一区二区| 久久99热这里只有精品18| 日韩中字成人| 国产国拍精品亚洲av在线观看| 欧美性猛交╳xxx乱大交人| 一区二区三区乱码不卡18| 国产中年淑女户外野战色| 国产91av在线免费观看| 亚洲欧美成人精品一区二区| 亚洲av中文av极速乱| 亚洲精品国产色婷婷电影| 麻豆精品久久久久久蜜桃| 久久人人爽人人爽人人片va| 成人高潮视频无遮挡免费网站| 久久久久久久久久成人| 一个人看的www免费观看视频| 国内精品美女久久久久久| av在线观看视频网站免费| 哪个播放器可以免费观看大片| 在线观看一区二区三区| 国产一区亚洲一区在线观看| 搡老乐熟女国产| av国产久精品久网站免费入址| 丝袜脚勾引网站| 极品教师在线视频| 国产色爽女视频免费观看| 免费人成在线观看视频色| 国产伦在线观看视频一区| 国产探花极品一区二区| 大陆偷拍与自拍| 亚洲一区二区三区欧美精品 | 国产欧美另类精品又又久久亚洲欧美| 激情五月婷婷亚洲| 久久韩国三级中文字幕| 国产精品久久久久久精品电影小说 | 国产乱来视频区| 亚洲精品国产av蜜桃| 亚洲精华国产精华液的使用体验| 波野结衣二区三区在线| av播播在线观看一区| 亚洲精品国产成人久久av| 亚洲最大成人手机在线| 高清毛片免费看| 中文资源天堂在线| 在线天堂最新版资源| 欧美日韩视频精品一区| 亚洲欧美日韩卡通动漫| 91精品国产九色| 白带黄色成豆腐渣| 亚洲综合色惰| 国产精品熟女久久久久浪| 亚洲丝袜综合中文字幕| 免费看光身美女| 亚洲一区二区三区欧美精品 | 国产亚洲一区二区精品| 国产日韩欧美在线精品| 国产精品99久久久久久久久| 纵有疾风起免费观看全集完整版| 免费播放大片免费观看视频在线观看| 国产综合精华液| 日韩亚洲欧美综合| 亚洲av在线观看美女高潮| 十八禁网站网址无遮挡 | 中文字幕人妻熟人妻熟丝袜美| 99久久九九国产精品国产免费| 亚洲精品乱码久久久久久按摩| 26uuu在线亚洲综合色| 国产一区二区亚洲精品在线观看| 日韩欧美精品免费久久| 美女视频免费永久观看网站| 精品久久久久久久人妻蜜臀av| 欧美日韩一区二区视频在线观看视频在线 | 波野结衣二区三区在线| 少妇人妻一区二区三区视频| 激情五月婷婷亚洲| 男人和女人高潮做爰伦理| 国产乱人偷精品视频| 色综合色国产| 街头女战士在线观看网站| 日韩强制内射视频| 欧美xxxx黑人xx丫x性爽| av线在线观看网站| 高清午夜精品一区二区三区| 国产欧美另类精品又又久久亚洲欧美| 男人爽女人下面视频在线观看| 小蜜桃在线观看免费完整版高清| 久久99蜜桃精品久久| 中国美白少妇内射xxxbb| 日韩av不卡免费在线播放| 综合色av麻豆| 日本黄色片子视频| 亚洲天堂国产精品一区在线| 伊人久久精品亚洲午夜| 亚洲成人精品中文字幕电影| 日本av手机在线免费观看| 欧美成人午夜免费资源| 成人免费观看视频高清| 精品国产三级普通话版| 一二三四中文在线观看免费高清| 国产精品蜜桃在线观看| 国产成人一区二区在线| freevideosex欧美| 国产高清国产精品国产三级 | 在线观看免费高清a一片| 晚上一个人看的免费电影| 欧美日本视频| 男的添女的下面高潮视频| 有码 亚洲区| 男人爽女人下面视频在线观看| 高清在线视频一区二区三区| 免费观看av网站的网址| 国产成人91sexporn| 禁无遮挡网站| 国产淫语在线视频| 欧美日韩一区二区视频在线观看视频在线 | 亚洲欧美一区二区三区黑人 | 国产精品久久久久久精品电影| 天美传媒精品一区二区| 欧美极品一区二区三区四区| 身体一侧抽搐| 最近中文字幕高清免费大全6| 老师上课跳d突然被开到最大视频| 久久久亚洲精品成人影院| 美女xxoo啪啪120秒动态图| 亚洲精品一区蜜桃| 亚洲av日韩在线播放| 久久久久久久午夜电影| 别揉我奶头 嗯啊视频| 精品国产露脸久久av麻豆| 久久综合国产亚洲精品| 国产日韩欧美在线精品| 天天躁夜夜躁狠狠久久av| 国精品久久久久久国模美| 丝袜喷水一区| 十八禁网站网址无遮挡 | 午夜福利视频1000在线观看| 国产中年淑女户外野战色| 91精品一卡2卡3卡4卡| 欧美日韩在线观看h| 69人妻影院| 成人鲁丝片一二三区免费| 特级一级黄色大片| 看非洲黑人一级黄片| av在线播放精品| 内地一区二区视频在线| 成人高潮视频无遮挡免费网站| 哪个播放器可以免费观看大片| 亚洲,一卡二卡三卡| 国产 一区 欧美 日韩| 亚洲成人av在线免费| 七月丁香在线播放| 亚洲最大成人手机在线| 日本一本二区三区精品| 国内少妇人妻偷人精品xxx网站| 一边亲一边摸免费视频| 国产精品久久久久久精品电影小说 | 美女xxoo啪啪120秒动态图| 三级国产精品片| 黄色配什么色好看| 少妇人妻久久综合中文| 老司机影院成人| 日日摸夜夜添夜夜添av毛片| 婷婷色综合大香蕉| 一级毛片我不卡| 日韩在线高清观看一区二区三区| 日韩av在线免费看完整版不卡| 一级毛片 在线播放| 日韩 亚洲 欧美在线| 免费大片黄手机在线观看| 日韩不卡一区二区三区视频在线| 性插视频无遮挡在线免费观看| 国产91av在线免费观看| 亚洲av二区三区四区| 国产精品av视频在线免费观看| 天天躁日日操中文字幕| 中文字幕亚洲精品专区| 国产毛片在线视频| 只有这里有精品99| 亚洲av电影在线观看一区二区三区 | 国产精品嫩草影院av在线观看| 亚洲无线观看免费| 99视频精品全部免费 在线| av免费观看日本| 天天躁日日操中文字幕| 国产精品蜜桃在线观看| 波多野结衣巨乳人妻| 亚洲人成网站在线播| 日韩国内少妇激情av| 欧美区成人在线视频| av一本久久久久| 97精品久久久久久久久久精品| 免费大片黄手机在线观看| 91aial.com中文字幕在线观看| 性插视频无遮挡在线免费观看| 国产在线一区二区三区精| 亚洲第一区二区三区不卡| 亚洲内射少妇av| 国产白丝娇喘喷水9色精品| 欧美另类一区| 国产精品国产三级国产专区5o| 在线a可以看的网站| 草草在线视频免费看| 最近手机中文字幕大全| 日日啪夜夜爽| 纵有疾风起免费观看全集完整版| 亚洲自偷自拍三级| 狂野欧美激情性bbbbbb| 天堂网av新在线| 国产一区二区三区av在线| 久久久久久久久大av| 亚洲欧美日韩卡通动漫| 99热这里只有是精品50| 亚洲av欧美aⅴ国产| 成人漫画全彩无遮挡| 黄色视频在线播放观看不卡| 日韩国内少妇激情av| 国产亚洲5aaaaa淫片| 国产片特级美女逼逼视频| 美女被艹到高潮喷水动态| 3wmmmm亚洲av在线观看| 97热精品久久久久久| 午夜免费观看性视频| 国产黄色免费在线视频| 王馨瑶露胸无遮挡在线观看| 久久久a久久爽久久v久久| av在线亚洲专区| xxx大片免费视频| 高清欧美精品videossex| 在线播放无遮挡| 视频区图区小说| 久久99热6这里只有精品| 内射极品少妇av片p| 大码成人一级视频| 国产 精品1| .国产精品久久| 777米奇影视久久| 国产精品久久久久久久久免| 日韩在线高清观看一区二区三区| 日韩不卡一区二区三区视频在线| 美女cb高潮喷水在线观看| 欧美激情国产日韩精品一区| 亚洲国产av新网站| 香蕉精品网在线| 大话2 男鬼变身卡| 国产在视频线精品| 亚洲精品一二三| 国国产精品蜜臀av免费| 久久精品久久久久久久性| 亚洲激情五月婷婷啪啪| 精品久久国产蜜桃| 亚洲不卡免费看| 两个人的视频大全免费| 美女xxoo啪啪120秒动态图| 国产淫片久久久久久久久| 人妻一区二区av| 午夜日本视频在线| 男女边吃奶边做爰视频| 亚洲最大成人手机在线| 国产免费又黄又爽又色| 高清午夜精品一区二区三区| 国产成人91sexporn| 91在线精品国自产拍蜜月| 亚洲国产色片| 日韩视频在线欧美| 内射极品少妇av片p| 日本色播在线视频| 亚洲精品国产av成人精品| av黄色大香蕉| 亚洲av中文av极速乱| 国产男女内射视频| 国产精品精品国产色婷婷| 欧美丝袜亚洲另类| 美女cb高潮喷水在线观看| av黄色大香蕉| 国产 一区精品| 好男人在线观看高清免费视频| 99精国产麻豆久久婷婷| 69人妻影院| 国产一区有黄有色的免费视频| 99久久中文字幕三级久久日本| 久久人人爽人人片av| 国产精品一区二区三区四区免费观看| 精品国产三级普通话版| 在现免费观看毛片| 人妻 亚洲 视频| 男人舔奶头视频| 亚洲精品影视一区二区三区av| 高清在线视频一区二区三区| 亚洲av国产av综合av卡| 亚洲最大成人av| 性色av一级| 麻豆久久精品国产亚洲av| 一级a做视频免费观看| 亚洲图色成人| 亚洲国产高清在线一区二区三| 舔av片在线| 国产欧美日韩一区二区三区在线 | 插阴视频在线观看视频| 伊人久久国产一区二区| 日韩欧美精品v在线| 一二三四中文在线观看免费高清| 午夜免费男女啪啪视频观看| 99久久精品热视频| 国产精品久久久久久精品古装| 亚洲av日韩在线播放| 欧美xxxx黑人xx丫x性爽| 一个人看视频在线观看www免费| 国产精品一及| 超碰97精品在线观看| 久久精品人妻少妇| 高清欧美精品videossex| 色综合色国产| 观看美女的网站| 熟女av电影| 97在线视频观看| 国产免费又黄又爽又色| 久久99热这里只有精品18| 蜜桃亚洲精品一区二区三区| 国产白丝娇喘喷水9色精品| 亚洲av一区综合| 精品视频人人做人人爽| av播播在线观看一区| 亚洲人与动物交配视频| 伊人久久精品亚洲午夜| 国产精品女同一区二区软件| 国产精品蜜桃在线观看| 日本-黄色视频高清免费观看| 嫩草影院入口| 80岁老熟妇乱子伦牲交| 国产亚洲91精品色在线| 男人狂女人下面高潮的视频| 午夜福利在线在线| 97热精品久久久久久| 国产精品.久久久| 高清av免费在线| 国内精品美女久久久久久| 欧美性猛交╳xxx乱大交人| 午夜免费观看性视频| 精品久久久久久久久亚洲| 嫩草影院新地址| 国产成人午夜福利电影在线观看| 久久久久精品久久久久真实原创| 久久精品夜色国产| 97超碰精品成人国产| 亚洲欧洲国产日韩| 在线 av 中文字幕| 91精品一卡2卡3卡4卡| 欧美日韩在线观看h| 麻豆乱淫一区二区| 亚洲国产精品999| 久久久久久久午夜电影| 欧美 日韩 精品 国产| 男人舔奶头视频| av女优亚洲男人天堂| 少妇高潮的动态图| 2018国产大陆天天弄谢| 特大巨黑吊av在线直播| 亚洲一区二区三区欧美精品 | 干丝袜人妻中文字幕| 亚洲欧美日韩东京热| 激情五月婷婷亚洲| 久久精品国产鲁丝片午夜精品| 精品国产乱码久久久久久小说| 欧美zozozo另类| 欧美成人午夜免费资源| 午夜精品一区二区三区免费看| 在线看a的网站| 国产高清有码在线观看视频| 国产毛片在线视频| av又黄又爽大尺度在线免费看| 尤物成人国产欧美一区二区三区| 国产精品一区二区性色av| 一级毛片久久久久久久久女| 精品一区二区三卡| 各种免费的搞黄视频| 亚洲国产成人一精品久久久| 日韩中字成人| 51国产日韩欧美| 久久久午夜欧美精品| 高清av免费在线| 人妻 亚洲 视频|