• <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 ?

    Traffic flow prediction based on BILSTM model and data denoising scheme

    2022-04-12 03:45:36ZhongYuLi李中昱HongXiaGe葛紅霞andRongJunCheng程榮軍
    Chinese Physics B 2022年4期
    關(guān)鍵詞:榮軍紅霞

    Zhong-Yu Li(李中昱) Hong-Xia Ge(葛紅霞) and Rong-Jun Cheng(程榮軍)

    1Faculty of Maritime and Transportation,Ningbo University,Ningbo 315211,China

    2Jiangsu Provincial Collaborative Innovation Center for Modern Urban Traffic Technologies,Nanjing 210096,China

    3National Traffic Management Engineering and Technology Research Center Ningbo University Subcenter,Ningbo 315211,China

    Keywords: traffic flow prediction,bidirectional long short-term memory network,data denoising

    1. Introduction

    Traffic flow prediction is an important component of intelligent transportation system, which can reduce the number of accidents,improve traffic efficiency,and reduce traffic pollution. The traffic prediction is based on a large number of historical data to predict the future traffic flow.[1]Previous studies show that linear model, nonlinear model, and hybrid model are three typical traffic flow prediction techniques.[2]In the linear models the mathematical methods are used to complete the task of the traffic flow prediction. The autoregressive integrated moving average model (ARIMA) is the most commonly used linear model.[3]Chenet al.[4]proposed an ARIMA model. The ARIMA model was specified by different training data, which represent the different traffic states,according to the different periods in different days. The model from different periods would be used on different data when testing. This would make the model more special and more accurate. The data-driven approach using the ARIMA model in most of studies required sound database for building a model. The seasonal ARIMA method was proposed to predict the short-term traffic flow with limited data.[5,6]Although the ARIMA had good prediction accuracy,there were many nonlinear factors in traffic flow prediction. Therefore,the nonlinear models introduced relevant machine learning methods(including support vector regression(SVR),artificial neural networks(ANN),etc.)to achieve favorable traffic flow prediction performance.The SVR had strict theoretical and mathematical basis. Based on the principle of structural risk minimization,it has strong generalization capability and global optimality. It was a theory for small sample statistics.[7-10]Neural network had strong non-linear fitting capability and could learn rules from large samples. Moreet al.[11]predicted the future traffic volume through the ANN,and used the predicted results to control congestion. Kumaret al.[12]used the ANN to make short-term prediction of traffic volume by using past traffic data (taking the traffic volume, speed and density as inputs).The ANN can produce good results in this study. In order to better capture the spatiotemporal information of traffic flow,convolutional neural network has been proposed.[13-15]In order to better learn the historical information of data,recurrent neural network(RNN)was proposed to be able to transfer the output of the previous time to the next step as input, thereby improving the accuracy of prediction.[16]The RNN can handle certain short-term dependence, but it cannot handle the longterm dependence. To solve this problem,the LSTM was proposed.The LSTM introduced cell state,and the input gate,forgetting gate and output gate were used to maintain and control the information,which can learn long-term information.[17]It is noted that the nonlinear models have achieved great success in dealing with many traffic flow prediction tasks. However,the uncertainty of traffic flow data might have a negative effect on the nonlinear prediction model, which would reduce the performance of the model. The hybrid models were proposed to overcome the disadvantages.[18,19]For example,Tanget al.[20]combined fuzzy c-means and genetic algorithm to predict missing traffic volume data. Zhenget al.[21]used a hybrid model based on a convolutional neural network(CNN)and the LSTM.The experimental results show that the prediction accuracy of the hybrid model is higher than that of the single model.

    Traffic flow data can be obtained by many technical means, such as remote microwave, loop deductive detectors,global positioning system(GPS).[22]Historical data with high fidelity play a critical role in predicting results, but raw data always contain noise. Therefore, one introduced many noise reduction methods,[23]such as the wavelet Kalman filter model[24,25]and wavelet transform.[26,27]Yanget al.[28]selected the appropriate wavelet basis to decompose and reconstruct the traffic flow data. The results showed that prediction effect is better than the results predicted just by neural network in prediction precision and network convergence. Because the wavelet transform needed to select the appropriate wavelet base, sometimes it is difficult to find the appropriate wavelet base. One proposed the EMD,which is an adaptive decomposition method based on the original data.[29]The EMD extracts the intrinsic mode function(IMF)sets from the input data samples, which can be separated into high-frequency (HF) part and low-frequency (LF) part. The EMD is affected by mode mixing,so the EEMD was proposed.[30,31]Chenet al.[32]proposed an integrated framework based on the integrated EEMD and the ANN can predict the traffic flow at different time intervals ahead. The experimental results show that the noise reduction algorithm greatly improves the accuracy of traffic flow prediction.

    The above method combines a certain noise reduction method with the prediction method to achieve better prediction accuracy, but does not use the same set of data to verify the performance of different noise reduction algorithms. The present study intends to combine data denoising algorithms with deep learning model to predict the traffic flow and compare among prediction accuracies of various methods. The contributions of this paper are as follows:

    (i)We use denoising algorithms to process the raw data,which include WL,EMD,and EEMD,and compare the capabilities of these algorithms to suppress the outliers.

    (ii) We build a BILSTM prediction model to predict the data already processed.

    (iii)We consider the predicted effects of the model under different road scenarios(i.e.,mainline,on ramp,off ramp)and on different time scales(i.e.,5 min,10 min,and 15 min).

    The rest of this paper is organized as follows. In Section 2,the denoising methods and BILSTM are introduced. In Section 3,the source of data,the model structure,and prediction results are presented. In Sections 4 some conclusions are drawn from the present study.

    2. Method

    The methods to be used in this article are data denoising and deep learning neural networks to improve prediction accuracy. The frame of this model is shown in Fig.1.

    Fig.1. Framework of prediction model.

    2.1. Denoising methods

    2.1.1. Wavelet

    Wavelet is used to eliminate noise interference in traffic flow data.[33]The purpose of building a noise reduction model is to reduce the influence of interference factors,such as gaussian noise,on the prediction accuracy of the model in a complex,nonlinear traffic data set. One-dimensional noise model can be expressed as

    whereFtis the raw signal with noise;ytis the signal without noise;εtis Gaussian noise;nis the signal length. The principle of wavelet denoising is to suppress theεtpart ofFtand restore theytpart ofFt. Wavelet denoising model preserves the original signal characteristics after denoising. The flow chart of wavelet is shown in Fig.2. The wavelet algorithm is implemented in three steps.

    Step 1 Decomposition

    Choose a wavelet and determine the layerNfor a wavelet decomposition,then compute theN-layer wavelet decomposition for the signal.

    Step 2 Setting a threshold for high-frequency coefficients

    For the high-frequency factor of each layer from layer 1 to layerN,a threshold is selected for thresholding.

    For most of signals,the low-frequency component is very important,and it usually contains the characteristics of the signal, while the high-frequency component contains the details or differences of the signal. The low-frequency information can approximately describe the change trend of the original signal, while high-frequency information represents the detailed information of the signal. Therefore,the original signal can produce two signals (high- and low-frequencies) through two mutual filters.

    Step 3 Data reconstruction

    According to the approximate coefficients and the modified wavelet coefficients,the inverse wavelet transform is used to reconstruct the traffic data after denoising.

    Fig.2. Flowchart of wavelet.

    2.1.2. The empirical mode decomposition (EMD)

    In this paper, we use the EMD, which was first introduced by Huanget al.[34]Different applications in medical signal analysis have shown the effectiveness of this method.The principle is to adaptively decompose the given signal into frequency components,which called intrinsic mode functions(IMF).These components are obtained from the signal by an algorithm. The algorithm extracts the highest frequency oscillation of each mode from the original signal. The more detailed explanations to the EMD model are described as follows.

    (i)Decompose raw traffic flow dataR(t)into IMFs. Find out potential extreme points fromR(t) data series, then employ the cubic spline interpolation method to connect the maximum and minimum points to form the upper envelopeU(t)and lower envelopeL(t).

    (ii) Calculate the mean of the upper envelopeU(t) and lower envelopeL(t) to obtain the mean envelopeM(t) (see Eq.(2)).

    (iii) Calculate the difference betweenR(t) andM(t) to obtainI(t)(see Eq.(3)). IfI(t)satisfies the IMF’s conditions,I(t) is one of the IMFs, otherwise the above three sub-steps need repeating untilI(t) satisfies IMF’s condition. The following IMF’s conditions are satisfied:i)the number difference between extrema and the zero-crossing points is not larger than one;ii)at any point in the IMF,the mean value of the envelope defined by the local maximum and the mean value of the envelope defined by the local minimum should be equal to zero.

    (iv)Obtain the residual partr(t)by calculating the difference betweenR(t)andI(t),repeat the above three sub-steps to obtain new IMFs until the residual data distribution is monotonic or has one extreme point. Finally,R(t) is decomposed into a series of IMFs and a residual partr(t)as shown by the following equations:

    2.1.3. The ensemble empirical mode decomposition(EEMD)

    For some data with too few extreme points, EMD fails to work, so the EEMD method is proposed. The white noise is added into the EEMD to make the data meet the requirements for the EMD. Owing to the zero mean white noise,the zero mean white noise will cancel each other after many times of average calculation, so the calculation result of integrated mean can be directly regarded as the final result. The EEMD can effectively suppress the mode aliasing of EMD.The EEMD algorithm steps are as follows.

    Step 1 The white noisewk(t) of normal distribution is added into the raw traffic dataR(t)(the noise-aided traffic flow data are denoted asRk(t))(see Eq.(5)).

    Step 2 TheRk(t) are decomposed into IMFs and residual components. The data decomposition program will be executed in thek-th round until the conditionk >Enis met(Enis the ensemble number). TheRk(t)data can be decomposed into IMFsek,d(t)and corresponding residual partrk,d(t)(see Eq.(6)).

    Step 3 Repeat steps 1 and 2,add a new normal distribution white noise sequence each time.

    Step 4 Select the noise-free IMFs and residuals to reconstruct the EEMD smoothed traffic flow dataR(t)EEMDwhich are given below:

    2.2. Deep learning

    2.2.1. Long short-term memory (LSTM) network

    The LSTM is an improved neural network based on recurrent neural network (RNN). The LSTM was proposed by Hochreiter and Schmidhuber in 1997.[35]In the RNN the output of the previous moment is taken as the input of the next moment,and the RNN is often used to predict the data of time series. However,the RNN tends to disappear and explode gradients when dealing with long sequences,leading the model to be nonconvergent.To solve this problem,the input gate,forget gate, output gate, and memory cell (C(t)) are added into the LSTM.The LSTM structure is shown in Fig.3.

    Fig.3. LSTM structure.

    The three gates are connected by the Sigmoid activation function. The Sigmoid activation function is a hyperbolic tangent function, and the value field is controlled in an interval of[0,1]. Candidate memory cells(?Ct)are connected by tanh activation function,and the range was controlled in an interval of[-1,1]. The calculation steps of the model are as follows:

    whereI(t),F(t),O(t),and ?C(t)represent the input gate,forget gate, output gate, candidate memory cells, respectively;Wxi,Whi,Wx f,Whf,Wxo,Who,Wxc,andWhcare the weight matrices.bi,bf,bo, andbcare bias vectors.σis a Sigmoid activation function.

    Memory cell (Ct) is obtained fromCtand ?Ct. Then we acquireHtby combining ?Ctwith the tanh activation function as follows:

    2.2.2. Bi-directional long short-term memory network(BILSTM)

    The Long short-term memory is a forward training model,which can only capture the long-term characteristics of forward historical traffic flow data. However,the traffic flow data not only depend on past information, but also correlates with future road conditions. The BILSTM is an optimized LSTM model,which combines forward LSTM and reverse LSTM to obtain the information. Therefore, the BILSTM can process and integrate data forward and backward,which can make the prediction more accurate. The BILSTM structure is illustrated in Fig.4. The forward layer and backward layer are connected to the output layer,which contains six shared weights.

    Fig.4. Structure of BILSTM.

    In the forward layer,the forward calculation is performed from timelto timet, and the output of each forward hidden layer is obtained and saved. In the backward layer, the output of the backward hidden layer at each time is obtained and saved by backward calculation from timetto timel. Finally,the final output is obtained by combining the output results of forward layer and backward layer at each time,the calculation steps of the model are as follows:

    In Fig.4,xtrepresents the input of LSTM,Ofdenotes the output of forward LSTM,andObrefers to the output of reverse LSTM.

    3. Experiments

    3.1. Data source

    The traffic flow data in this paper are obtained from the performance measurement system (PeMS). The PeMS is an intelligent traffic management tool. It collects and stores realtime traffic data from sensors and provides online use. The PeMS data collection covers major metropolitan areas in California. We collected the traffic flow data from three detectors installed in Interstate 15 (I15) freeway segment located in San Diego, California (see in Fig. 5). The three detectors are located in mainline(denoted as detector A),on-ramp(represented as detector B), and off-ramp (marked as detector C)(see in Fig. 6) respectively. Owing to the different road conditions of these three sections, the predictions of these three sections can better illustrate the effectiveness of the model.We selected 2880 data in total from weekdays 1 March 2021 through 12 March 2021, and took the flow every 5 min. The details of the road sections are shown in Table 1.

    Fig.5. Interstate 15(I15)freeway segment.

    Fig.6. Three detectors.

    Table 1. Details of road sections.

    3.2. Experiment setups

    Firstly, the data are subjected to noise reduction. The wavelet denoising is a very common denoising method. It has many wavelet bases. Four commonly used wavelet bases are selected in this paper: WL (db4), WL (coif), WL (sym), and WL (haar). When using wavelet denoising, we set the number of decomposition layers to be 3 and the wavelet bases are assumed to be db4,coif2,haar,and sym2 respectively. In using EMD, we set the number of IMF decompositions to be 5, and the interpolation used is pchip (piecewise cubic Hermite interpolating polynomial method). In using EEMD,there are two parameters (the standard deviation ratio between the added Gaussian white noise and the amplitude of the input signal,An, and the ensemble number,En)need to be set. We setAn=0.2 andEn=1000. We use Adam gradient descent algorithm in BILSTM.The data are used four times to make a prediction once. We divide 70%of the raw traffic data as the training dataset and 30%as the test dataset,use Adam gradient descent algorithm,then train 500 times to obtain the prediction results.

    3.3. Evaluation of predictive performance of model

    The root mean square error (RMSE), the mean absolute error(MAE),and the mean absolute percentage error(MAPE)are used to evaluate the prediction performance of the model.The RMSE,MAE,and MAPE are calculated as follows:

    whereXprediction,tis the prediction results,Xreal,tis the raw traffic data,andNis the data-sampling number.

    3.4. Results

    3.4.1. Traffic flow data denoising for detector A

    We employ the wavelet model, EMD, EEMD to remove outliers from original data. In order to see clearly, the noise reduction results of the first 1000 data are shown in Fig. 7.The reliable data are provided for subsequent prediction,blue lines on the plots represent denoised data, and red lines refer to the raw data. The data of a day is 288,and we can see that these data are periodic. It can be observed that the traffic flow in a day can be divided into peak period and flat peak period.The fluctuation of the data in the peak period is significantly higher than that in the flat peak period. This is because the road occupancy rate is high in the peak period and the driving behaviors of drivers will have a great influence on the road conditions, so the noise of the traffic flow in the peak period will be much higher than that in the flat peak period. It can be seen that the noise interference in the flat hump period is less,and the noise reduction results are almost consistent with the raw data. It can be seen that the curve of data after noise reduction are significantly smoothened than that of the original data,which smoothens the fluctuations from the original data.As shown in Fig.7, WL(coif2)denoised data (see Fig. 7(a))and WL(db4)denoised data(see Fig.7(b))are better than WL(haar)denoised data(see Fig.7(c))and WL(sym2)denoised data(see Fig.7(d))in data denoising. The curve of EEMD denoised data is smoother than that of the EMD denoised data.

    Fig.7. Denoising results for detector A.

    3.4.2. Predictions for detector A

    Now, we come to acquire the smooth data and use the BILSTM model to make predictions for detector A.The traffic flow prediction schemes are denoted as follows: BILSTM,BILSTM+WL (coif2), BILSTM+WL (db4), BILSTM+WL(haar), BILSTM+WL (sym2), BILSTM+EMD, and BILSTM+EEMD. The predicted results are shown in Fig. 8.The black line represents the raw data and the red line denotes the predicted data. The results of RMSE, MAE, and MAPE are shown in Table 2.The RMSE,MAE,and MAPE measured directly from BILSTM are 30.68, 22.70, 8.76%, respectively(see Table 2). The predicted RMSE decreases by 20% after noise reduction. It can be seen that the data after noise reduction can be predicted to better learn the change trend of traffic flow and improve the predict accuracy.

    Fig.8. Results of prediction(detector A).

    Table 2. Results of different prediction models for detector A.

    Furthermore, we can see that the EEMD has the best effect. The RMSE,MAE,and MAPE of BILSTM+EEMD(see Table 2) are 3.98, 3.5, and 2.04% respectively. The performances of EEMD and EMD combined with the BILSTM are better than that of wavelet combined with the BILSTM, indicating the excellent performance of EMD with respect to EEMD in eliminating data anomalies.

    3.4.3. Traffic flow data prediction and denoising for ramp detectors B, C

    To further validate the performance of the model,we will take the data from ramp detectors(detector B,detector C)for prediction.

    Fig.9. Denoising results for detector B.

    Fig.10. Denoising results for detector C.

    Since the road situation of the ramp is different from that of the main line, we verify the predictive capability of the model in different road situations. The data denoising results for the two detectors are shown in Figs.9 and 10,respectively.The noise reduction procedure effectively suppresses outliers from the raw traffic data.

    The prediction performance for detectors B and C are shown in Tables 3 and 4. The RMSE, MAE, and MAPE for BILSTM of off ramp (see in Table 3) are 9.43, 6.76,and 24.98%, respectively. The RMSE, MAE and MAPE for BILSTM+WL(coif2)of off ramp(see Table 3)are 1.96,1.39,and 5.50%,respectively. The accuracy of noise reduction prediction from the wavelet method is improved by about 50%.Table 3 indicates that the BILSTM+EMD scheme is better than BILSTM combined with wavelet. In Table 4,we observe that the prediction distributions are the same as in Table 2.Moreover,the prediction accuracy of BILSTM+EEMD model for detector B and detector C are better than those of other hybrid models.The RMSE,MAE,and MAPE are 0.74,0.62,and 1.75 for the traffic flow data collected from detector B,and the statistical values are 0.68, 0.49, and 2.42 collected from detector C.Through the above analysis,it can be concluded that the BILSTM+EEMD model shows the predicted performance consistent with the traffic flow data collected from mainline,on-ramp, and off-ramp areas. This means that the hybrid model is suitable for different road conditions.

    Table 3. Results of different prediction models for detector B.

    Table 4. Results of different prediction models for detector C.

    3.4.4. Comparison among traffic flow data prediction for detectors A, B, and C

    In order to study the influence of noise on different roads,we will add the prediction effect of noise reduction under three road conditions: mainline, on-ramp, and off-ramp. The predicted results are shown in Tables 2-4, so we select the results of EEMD for comparisons. The comparison is shown in Table 5. It can be seen that the accuracy of the model is greatly improved after noise reduction. The prediction results of off ramp and on ramp are similar. After noise reduction,the MAPE of the ramp decreases greatly in comparison with that of the mainline. It can be seen that the noise has a greater influence on the ramp. According to the analysis of traffic flow state in reality,the ramp is only a one-way line,and there are more uncertain factors in the driving of vehicles. The noise in the data is not only increased with the increase of traffic flow,but also related to many factors, such as the weather and the working state of loop detector. The traffic flow on the ramp is much less than that on the mainline. Slight noise will have a great influence on the ramp flow prediction. However, owing to the large vehicle base on the main road,the influence of noise on the main road is not so serious as that on the ramp.

    Table 5. Comparison of traffic flow data prediction among detectors(A,B,C).

    3.4.5. Traffic flow data prediction on different time scales for detector A

    In order to verify the effect of the model on the long-term traffic flow prediction, we use the model to predict the traffic flows in different time spans,i.e., 5 min, 10 min, 15 min.we use the mainline data (detector A) to predict the traffic flow. The results are shown in Table 6. It can be observed that with the increase of time span,RMES,MAE,and MAPE all become larger. This is because with the increase of time span, the data sampling frequency turns smaller, resulting in greater noise in the sampling data,and some characteristic information of traffic flow may be confirmed. Therefore, the error of prediction results will increase. The results show that EEMD+BILSTM has the best prediction effect,the RMSEs of EEMD+BILSTM for 10 min and 15 min are 8.95 and 15.36, respectively. The prediction trend of wavelet denoising and EEMD de-noising are also consistent with that of EEMD+BILSTM. The results show that the model can also achieve good long-term forecasting results,and the model can be used for long-term prediction.

    Table 6. Statistical performance on different time scales for detector A.

    4. Conclusions

    In this study a novel traffic flow prediction framework is introduced by integrating a data denoising scheme with a deep learning model. This study adopt the traffic volume data collected from three detectors in San Diego, California, which are cited from the PeMS. To clean the data, we employ various popular and effective noise reduction models to suppress outliers from the raw traffic flow data. We chose three denoising schemes: EMD, EEMD, and Wavelet with different basis (haar, db4, sym2, coif2). Then combine with BILSTM to make predictions. Finally, we compare the prediction accuracies of different models by analyzing the RMSE, MAE,and MAPE indicators. Several interesting conclusions can be summarized as follows.

    (I) The model combines the data denoising with the model, which is better than the prediction process of without denoising strategy.

    (II) The performances of WL (db4), WL (coif2), WL(sym2) are better than that of WL (haar). In addition, the performance of WL (haar) in long-term prediction is not so satisfactory as other three wavelet methods.

    (III)The BILSTM+EEMD obtains the best performance in comparison with all the WL models and BILSTM+EMD.The RMSE,MAE,and MAPE indicators of BILSTM+EEMD are all at least 25%lower than those of other methods.

    (IV) These hybrid forecasting methods can be used in long-term prediction. The model proposed in this paper can make the prediction more accurate. It can be used to optimize traffic organization strategy,improve traffic efficiency,and reduce energy consumption.

    (V)In future,there will be several directions for more indepth study. Firstly, this paper has no predictive influence of environmental factors on the forecast,which can be taken into account later. Secondly, we can take some traffic indicators into account in traffic flow prediction,such as speed and road occupancy. Finally, we can combine deep learning with traffic flow models (e.g., car-following models) to predict traffic flow.

    Acknowledgements

    Project supported by the Program of Humanities and Social Science of the Education Ministry of China (Grant No. 20YJA630008), the Natural Science Foundation of Zhejiang Province, China (Grant No. LY20G010004), and the K C Wong Magna Fund in Ningbo University,China.

    猜你喜歡
    榮軍紅霞
    Bifurcation analysis of visual angle model with anticipated time and stabilizing driving behavior
    請(qǐng)你幫個(gè)忙
    An extended smart driver model considering electronic throttle angle changes with memory
    Stabilization strategy of a car-following model with multiple time delays of the drivers?
    《烏鴉喝水》中的“想”
    A Study of Combination of English Language Teaching and Context
    大東方(2018年1期)2018-05-30 01:27:23
    讓動(dòng)作“活”起來(lái)
    “大愛(ài)有聲 義工行動(dòng) 關(guān)愛(ài)榮譽(yù)軍人”慰問(wèn)活動(dòng) 走進(jìn)廣東省第一榮軍醫(yī)院
    大社會(huì)(2016年6期)2016-05-04 03:42:20
    “光的直線傳播”“光的反射”練習(xí)
    接力生命花季無(wú)聲
    ——記王臺(tái)中學(xué)初一158班郭榮軍同學(xué)舍己救人的英雄事跡
    在线 av 中文字幕| 1024视频免费在线观看| 亚洲精品美女久久av网站| h视频一区二区三区| 乱人伦中国视频| 人人妻,人人澡人人爽秒播| 成人av一区二区三区在线看| 黑人巨大精品欧美一区二区蜜桃| 精品亚洲成a人片在线观看| 一区二区日韩欧美中文字幕| 久久免费观看电影| av一本久久久久| 欧美国产精品va在线观看不卡| 午夜视频精品福利| 乱人伦中国视频| 国产精品免费视频内射| 人人澡人人妻人| av不卡在线播放| 国产老妇伦熟女老妇高清| 一级黄色大片毛片| 12—13女人毛片做爰片一| 97在线人人人人妻| 免费一级毛片在线播放高清视频 | 欧美亚洲 丝袜 人妻 在线| 香蕉国产在线看| 亚洲精华国产精华精| 成人手机av| 菩萨蛮人人尽说江南好唐韦庄| 欧美日韩亚洲国产一区二区在线观看 | 搡老熟女国产l中国老女人| 亚洲精品一卡2卡三卡4卡5卡| 亚洲精华国产精华精| 久久天躁狠狠躁夜夜2o2o| e午夜精品久久久久久久| 男男h啪啪无遮挡| 黄色 视频免费看| 国产91精品成人一区二区三区 | 久久久久久久久久久久大奶| 青青草视频在线视频观看| 亚洲成a人片在线一区二区| 91国产中文字幕| 美女主播在线视频| 久久午夜亚洲精品久久| 亚洲国产av影院在线观看| 精品久久蜜臀av无| 国产高清国产精品国产三级| 丰满迷人的少妇在线观看| 国产成人精品久久二区二区免费| 女人爽到高潮嗷嗷叫在线视频| 美女国产高潮福利片在线看| 成人特级黄色片久久久久久久 | 97在线人人人人妻| 午夜福利视频在线观看免费| 天天影视国产精品| 日韩 欧美 亚洲 中文字幕| 美女午夜性视频免费| 国产免费av片在线观看野外av| 国产欧美亚洲国产| 女人久久www免费人成看片| 久久中文字幕一级| av超薄肉色丝袜交足视频| 亚洲熟女精品中文字幕| 成年女人毛片免费观看观看9 | 人人妻人人澡人人看| av又黄又爽大尺度在线免费看| 大码成人一级视频| 人人澡人人妻人| 美女视频免费永久观看网站| 免费高清在线观看日韩| 一级片'在线观看视频| 757午夜福利合集在线观看| 99re在线观看精品视频| 又大又爽又粗| 久久国产精品人妻蜜桃| 婷婷成人精品国产| 久久中文字幕人妻熟女| www.精华液| 黄色视频在线播放观看不卡| 亚洲久久久国产精品| 香蕉国产在线看| 国产av国产精品国产| 午夜精品国产一区二区电影| 91av网站免费观看| 正在播放国产对白刺激| 亚洲五月色婷婷综合| kizo精华| 国产极品粉嫩免费观看在线| 老司机亚洲免费影院| 日韩视频在线欧美| 国产精品偷伦视频观看了| 黄频高清免费视频| 国产亚洲欧美精品永久| 男男h啪啪无遮挡| 久久久久网色| 99久久人妻综合| 久久国产精品男人的天堂亚洲| 在线永久观看黄色视频| 国产在线免费精品| 欧美日韩亚洲综合一区二区三区_| 俄罗斯特黄特色一大片| 制服诱惑二区| 首页视频小说图片口味搜索| 国产男女内射视频| www.精华液| 黄色毛片三级朝国网站| 在线观看66精品国产| 好男人电影高清在线观看| 国产精品麻豆人妻色哟哟久久| 亚洲精品一二三| 午夜福利,免费看| 精品久久久久久久毛片微露脸| 中国美女看黄片| 午夜老司机福利片| 国产精品一区二区在线不卡| 91九色精品人成在线观看| 99九九在线精品视频| 汤姆久久久久久久影院中文字幕| 久久精品91无色码中文字幕| 18在线观看网站| 十分钟在线观看高清视频www| 欧美另类亚洲清纯唯美| 99国产精品99久久久久| 精品少妇一区二区三区视频日本电影| 久久久久久久大尺度免费视频| 日韩制服丝袜自拍偷拍| 他把我摸到了高潮在线观看 | 日韩中文字幕视频在线看片| 啦啦啦中文免费视频观看日本| 亚洲一区中文字幕在线| 亚洲少妇的诱惑av| 亚洲av日韩精品久久久久久密| 又大又爽又粗| 2018国产大陆天天弄谢| 在线亚洲精品国产二区图片欧美| 久久中文字幕一级| 精品第一国产精品| av天堂在线播放| 亚洲精品中文字幕一二三四区 | 黄片小视频在线播放| 在线观看免费午夜福利视频| 成年人午夜在线观看视频| 热99久久久久精品小说推荐| 国产成人免费观看mmmm| 中文亚洲av片在线观看爽 | 久久久精品国产亚洲av高清涩受| 黄色视频在线播放观看不卡| 亚洲av电影在线进入| 中文字幕色久视频| 女同久久另类99精品国产91| 国内毛片毛片毛片毛片毛片| 国产一区二区 视频在线| 伊人久久大香线蕉亚洲五| 亚洲精品中文字幕一二三四区 | 亚洲av电影在线进入| 大片电影免费在线观看免费| 12—13女人毛片做爰片一| 成在线人永久免费视频| 国产aⅴ精品一区二区三区波| 飞空精品影院首页| 亚洲av第一区精品v没综合| 黑人猛操日本美女一级片| 国产无遮挡羞羞视频在线观看| 色94色欧美一区二区| 亚洲国产欧美在线一区| 黑人巨大精品欧美一区二区mp4| 国产欧美日韩精品亚洲av| 两人在一起打扑克的视频| 国产精品98久久久久久宅男小说| 91成年电影在线观看| 自拍欧美九色日韩亚洲蝌蚪91| 侵犯人妻中文字幕一二三四区| 欧美日本中文国产一区发布| 大片电影免费在线观看免费| 嫩草影视91久久| 日本a在线网址| 国产高清国产精品国产三级| 精品亚洲成a人片在线观看| 夜夜爽天天搞| 午夜免费鲁丝| 国产精品久久久久久精品古装| 国产亚洲欧美在线一区二区| 国产精品秋霞免费鲁丝片| 脱女人内裤的视频| 精品视频人人做人人爽| 国产成人一区二区三区免费视频网站| 中文亚洲av片在线观看爽 | 日韩大片免费观看网站| 精品一区二区三区视频在线观看免费 | 亚洲午夜理论影院| 亚洲视频免费观看视频| 女人精品久久久久毛片| 妹子高潮喷水视频| 久久99一区二区三区| 午夜免费鲁丝| 两个人免费观看高清视频| 国产欧美日韩综合在线一区二区| 少妇 在线观看| 国产精品九九99| 大码成人一级视频| 欧美+亚洲+日韩+国产| 亚洲av日韩在线播放| 青青草视频在线视频观看| 丝袜喷水一区| 99香蕉大伊视频| 欧美国产精品va在线观看不卡| 亚洲久久久国产精品| 国产麻豆69| 亚洲中文日韩欧美视频| 人人妻,人人澡人人爽秒播| 美女高潮喷水抽搐中文字幕| 成人手机av| 国产精品一区二区免费欧美| 999久久久精品免费观看国产| 亚洲,欧美精品.| 欧美在线一区亚洲| 久久精品国产99精品国产亚洲性色 | 老司机在亚洲福利影院| 久久人人97超碰香蕉20202| 91麻豆精品激情在线观看国产 | 999精品在线视频| 国产xxxxx性猛交| 91av网站免费观看| aaaaa片日本免费| 成人精品一区二区免费| 亚洲国产成人一精品久久久| 欧美激情高清一区二区三区| 一级毛片精品| 1024香蕉在线观看| 国产精品久久电影中文字幕 | 在线观看一区二区三区激情| 人人妻人人澡人人看| 又紧又爽又黄一区二区| 国产成人精品无人区| 99精品在免费线老司机午夜| 亚洲,欧美精品.| 狂野欧美激情性xxxx| 悠悠久久av| 欧美在线一区亚洲| 无限看片的www在线观看| 亚洲男人天堂网一区| 亚洲国产欧美在线一区| 国产日韩欧美亚洲二区| 亚洲av第一区精品v没综合| 亚洲成国产人片在线观看| 黑人欧美特级aaaaaa片| 女性被躁到高潮视频| 亚洲成人免费av在线播放| 久久久久精品人妻al黑| 精品久久久久久久毛片微露脸| 中文字幕人妻熟女乱码| √禁漫天堂资源中文www| 成人影院久久| 欧美久久黑人一区二区| 久久人人爽av亚洲精品天堂| 午夜福利影视在线免费观看| 韩国精品一区二区三区| 91成年电影在线观看| 日韩一卡2卡3卡4卡2021年| 国产av国产精品国产| 热re99久久精品国产66热6| 99在线人妻在线中文字幕 | 日日爽夜夜爽网站| 国产精品久久久久久精品古装| 久久久久精品人妻al黑| 精品熟女少妇八av免费久了| 少妇猛男粗大的猛烈进出视频| 丝瓜视频免费看黄片| 亚洲av日韩精品久久久久久密| 精品久久久久久电影网| 99久久99久久久精品蜜桃| 在线观看人妻少妇| 中文字幕另类日韩欧美亚洲嫩草| 操美女的视频在线观看| 又黄又粗又硬又大视频| 天天躁狠狠躁夜夜躁狠狠躁| 亚洲国产精品一区二区三区在线| 熟女少妇亚洲综合色aaa.| 久久久久网色| 宅男免费午夜| 亚洲综合色网址| 亚洲专区中文字幕在线| 日韩大码丰满熟妇| 亚洲欧美精品综合一区二区三区| 男女之事视频高清在线观看| 亚洲av成人一区二区三| av有码第一页| 黄色怎么调成土黄色| 亚洲国产看品久久| 狂野欧美激情性xxxx| 天天躁夜夜躁狠狠躁躁| 99香蕉大伊视频| 麻豆乱淫一区二区| av网站免费在线观看视频| 99九九在线精品视频| 久久久久精品人妻al黑| 亚洲欧美日韩高清在线视频 | 日本av手机在线免费观看| 久久久精品免费免费高清| 两人在一起打扑克的视频| 欧美日韩亚洲高清精品| 成人手机av| 国产精品免费视频内射| 男女下面插进去视频免费观看| 亚洲欧美一区二区三区久久| 亚洲av日韩精品久久久久久密| 欧美久久黑人一区二区| 欧美激情 高清一区二区三区| 久久久国产一区二区| 99久久人妻综合| 久久香蕉激情| 另类精品久久| 成年人黄色毛片网站| 亚洲一区中文字幕在线| 日本黄色日本黄色录像| av在线播放免费不卡| 亚洲欧美一区二区三区黑人| 99精品在免费线老司机午夜| 国产高清激情床上av| 美女主播在线视频| 又黄又粗又硬又大视频| 十八禁网站网址无遮挡| 大片电影免费在线观看免费| 久久久久久久大尺度免费视频| 久久久久久久精品吃奶| 丰满迷人的少妇在线观看| 欧美日韩亚洲高清精品| 大陆偷拍与自拍| 国产成人啪精品午夜网站| 不卡一级毛片| 欧美日韩中文字幕国产精品一区二区三区 | 日本欧美视频一区| 久久免费观看电影| 一级毛片女人18水好多| 女人高潮潮喷娇喘18禁视频| 精品福利永久在线观看| 日本av手机在线免费观看| 嫩草影视91久久| 99国产精品一区二区蜜桃av | 一区二区三区乱码不卡18| 久热这里只有精品99| 亚洲av日韩在线播放| 国产av精品麻豆| 精品久久蜜臀av无| 国产精品久久久久久精品电影小说| 成年动漫av网址| 国产伦理片在线播放av一区| 精品熟女少妇八av免费久了| 国产亚洲精品一区二区www | 成人国产一区最新在线观看| 一夜夜www| 亚洲欧洲精品一区二区精品久久久| 久久久精品区二区三区| 午夜福利在线免费观看网站| 色精品久久人妻99蜜桃| 国产男女内射视频| 黄色 视频免费看| 中文字幕人妻丝袜制服| 中文字幕最新亚洲高清| www.自偷自拍.com| 国产福利在线免费观看视频| 两个人看的免费小视频| av线在线观看网站| 老鸭窝网址在线观看| 超碰成人久久| 69av精品久久久久久 | 久久性视频一级片| 国产成人欧美| 最黄视频免费看| 丝袜在线中文字幕| 三级毛片av免费| 三上悠亚av全集在线观看| 最黄视频免费看| 午夜精品国产一区二区电影| 菩萨蛮人人尽说江南好唐韦庄| 日韩免费av在线播放| 99热国产这里只有精品6| 老司机在亚洲福利影院| 国产精品自产拍在线观看55亚洲 | 涩涩av久久男人的天堂| 精品欧美一区二区三区在线| 日韩欧美一区视频在线观看| 久久午夜综合久久蜜桃| 日本黄色日本黄色录像| 色播在线永久视频| 中文字幕人妻丝袜制服| 老司机在亚洲福利影院| 久久人妻熟女aⅴ| 亚洲 欧美一区二区三区| 女同久久另类99精品国产91| 亚洲精品在线美女| 午夜福利免费观看在线| √禁漫天堂资源中文www| 少妇粗大呻吟视频| 成人亚洲精品一区在线观看| 国产男女内射视频| 狠狠精品人妻久久久久久综合| 黄色视频,在线免费观看| 国产精品欧美亚洲77777| 999久久久国产精品视频| 纵有疾风起免费观看全集完整版| 啦啦啦中文免费视频观看日本| 久久久久久久国产电影| 日韩人妻精品一区2区三区| 亚洲中文日韩欧美视频| 午夜福利一区二区在线看| 黄色片一级片一级黄色片| 在线观看66精品国产| videosex国产| 欧美日韩亚洲国产一区二区在线观看 | 老司机午夜福利在线观看视频 | 两人在一起打扑克的视频| 亚洲成人免费av在线播放| 亚洲综合色网址| 考比视频在线观看| 国产成人av激情在线播放| 91精品三级在线观看| 久久国产亚洲av麻豆专区| 国产精品一区二区在线观看99| av电影中文网址| 狠狠狠狠99中文字幕| 国产精品99久久99久久久不卡| 汤姆久久久久久久影院中文字幕| 久久久久久人人人人人| 99久久精品国产亚洲精品| 少妇精品久久久久久久| 成人18禁在线播放| 国产在线视频一区二区| 精品久久久精品久久久| 天天操日日干夜夜撸| 国产日韩欧美视频二区| 99热国产这里只有精品6| 免费看a级黄色片| 少妇的丰满在线观看| av欧美777| 国产高清videossex| 中文字幕人妻丝袜一区二区| 精品卡一卡二卡四卡免费| 免费女性裸体啪啪无遮挡网站| 下体分泌物呈黄色| 国产av又大| 黑人巨大精品欧美一区二区蜜桃| 久久精品国产99精品国产亚洲性色 | 欧美在线一区亚洲| 丰满饥渴人妻一区二区三| 免费观看a级毛片全部| 亚洲伊人色综图| 一本色道久久久久久精品综合| 美女午夜性视频免费| 一级片'在线观看视频| 99久久人妻综合| 中文亚洲av片在线观看爽 | 免费高清在线观看日韩| 国产成人免费无遮挡视频| videosex国产| 午夜福利视频精品| 99在线人妻在线中文字幕 | 大片免费播放器 马上看| 高清视频免费观看一区二区| 亚洲情色 制服丝袜| av在线播放免费不卡| 少妇的丰满在线观看| 精品少妇一区二区三区视频日本电影| 母亲3免费完整高清在线观看| 亚洲欧美一区二区三区久久| 日韩熟女老妇一区二区性免费视频| 青青草视频在线视频观看| 久久天堂一区二区三区四区| 熟女少妇亚洲综合色aaa.| 美女午夜性视频免费| 一本—道久久a久久精品蜜桃钙片| 亚洲少妇的诱惑av| 一区二区av电影网| 18禁观看日本| 免费观看a级毛片全部| 每晚都被弄得嗷嗷叫到高潮| 香蕉国产在线看| 亚洲欧美激情在线| 亚洲av日韩在线播放| 搡老熟女国产l中国老女人| 俄罗斯特黄特色一大片| 一区二区三区国产精品乱码| 1024香蕉在线观看| 国产人伦9x9x在线观看| 国产精品久久久av美女十八| 一级片'在线观看视频| 久久久欧美国产精品| 色视频在线一区二区三区| 美女高潮喷水抽搐中文字幕| 男女高潮啪啪啪动态图| 日韩欧美一区视频在线观看| 亚洲成国产人片在线观看| 精品一区二区三卡| 国产麻豆69| 国产亚洲一区二区精品| 久久久久久久精品吃奶| 日韩免费高清中文字幕av| 色94色欧美一区二区| 亚洲色图综合在线观看| 亚洲精品美女久久久久99蜜臀| 最新美女视频免费是黄的| 国产免费av片在线观看野外av| 国产精品欧美亚洲77777| 国产一卡二卡三卡精品| 午夜久久久在线观看| 精品一区二区三卡| 久久久国产一区二区| 成年动漫av网址| 人人澡人人妻人| 丝袜美腿诱惑在线| 亚洲中文字幕日韩| 性少妇av在线| 国产91精品成人一区二区三区 | 亚洲七黄色美女视频| 日韩中文字幕欧美一区二区| 国产成人免费观看mmmm| 欧美激情 高清一区二区三区| 黑人操中国人逼视频| 久久久久久久久久久久大奶| 亚洲专区中文字幕在线| 欧美中文综合在线视频| 视频区欧美日本亚洲| 女性被躁到高潮视频| 成人黄色视频免费在线看| 在线观看一区二区三区激情| 人人妻人人添人人爽欧美一区卜| 免费观看av网站的网址| 国产一区二区三区视频了| 久久精品国产99精品国产亚洲性色 | av超薄肉色丝袜交足视频| 在线亚洲精品国产二区图片欧美| 大片电影免费在线观看免费| 亚洲精品粉嫩美女一区| 欧美激情久久久久久爽电影 | 精品亚洲成a人片在线观看| 国产高清videossex| 亚洲人成77777在线视频| 少妇 在线观看| 亚洲人成伊人成综合网2020| 91成人精品电影| 18禁裸乳无遮挡动漫免费视频| 国产高清视频在线播放一区| 丝袜喷水一区| 亚洲免费av在线视频| 丝袜喷水一区| 一级片'在线观看视频| 国产精品香港三级国产av潘金莲| 欧美国产精品一级二级三级| 亚洲精品一二三| 淫妇啪啪啪对白视频| 国产一区二区三区综合在线观看| 中文字幕av电影在线播放| 免费高清在线观看日韩| 欧美人与性动交α欧美软件| 交换朋友夫妻互换小说| 少妇 在线观看| 久久九九热精品免费| 午夜视频精品福利| 法律面前人人平等表现在哪些方面| 欧美日韩中文字幕国产精品一区二区三区 | 亚洲精品中文字幕在线视频| 人妻久久中文字幕网| 国产三级黄色录像| 国产精品免费视频内射| 老司机午夜十八禁免费视频| 一进一出好大好爽视频| 成人永久免费在线观看视频 | 一个人免费看片子| 亚洲精品久久午夜乱码| 少妇裸体淫交视频免费看高清 | 国产一区二区三区在线臀色熟女 | 欧美激情久久久久久爽电影 | 一区二区日韩欧美中文字幕| 色播在线永久视频| 亚洲五月婷婷丁香| 亚洲国产毛片av蜜桃av| 精品少妇久久久久久888优播| 天天操日日干夜夜撸| 欧美+亚洲+日韩+国产| 欧美日韩亚洲国产一区二区在线观看 | 国产高清激情床上av| 亚洲精品国产一区二区精华液| 亚洲视频免费观看视频| 黄色a级毛片大全视频| 99国产精品免费福利视频| 亚洲国产欧美网| 69av精品久久久久久 | 午夜福利在线观看吧| 国产一区二区三区视频了| 一进一出抽搐动态| 亚洲美女黄片视频| 欧美另类亚洲清纯唯美| 亚洲,欧美精品.| 精品国内亚洲2022精品成人 | 国产精品麻豆人妻色哟哟久久| 多毛熟女@视频| a级片在线免费高清观看视频| 一区二区三区国产精品乱码| 亚洲av欧美aⅴ国产| 久久久精品区二区三区| 黄色片一级片一级黄色片| 日韩成人在线观看一区二区三区| 国产精品电影一区二区三区 | 成人18禁高潮啪啪吃奶动态图| 18禁美女被吸乳视频| 国产欧美日韩一区二区三区在线| 亚洲精品国产色婷婷电影| 国产伦人伦偷精品视频| 男人舔女人的私密视频| 国产日韩一区二区三区精品不卡| 高清黄色对白视频在线免费看| 老熟女久久久| 亚洲情色 制服丝袜| 母亲3免费完整高清在线观看| 91成人精品电影| 免费一级毛片在线播放高清视频 |