• <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 ?

    Parking Availability Prediction with Coarse-Grained Human Mobility Data

    2022-08-23 02:15:42AuroraGonzalezVidalFernandoTerrosoenzandAntonioSkarmeta
    Computers Materials&Continua 2022年6期

    Aurora Gonzalez-Vidal,Fernando Terroso-Sáenzand Antonio Skarmeta

    1University of Murcia,Murcia,30003,Spain

    2Universidad Católica de Murcia(UCAM),Murcia,30107,Spain

    Abstract: Nowadays,the anticipation of parking-space demand is an instrumental service in order to reduce traffic congestion levels in urban spaces.The purpose of our work is to study,design and develop a parking-availability predictor that extracts the knowledge from human mobility data,based on the anonymized human displacements of an urban area, and also from weather conditions.Most of the existing solutions for this prediction take as contextual data the current road-traffic state defined at very high temporal or spatial resolution.However,access to this type of fine-grained location data is usually quite limited due to several economic or privacy-related restrictions.To overcome this limitation, our proposal uses urban areas that are defined at very low spatial and temporal resolution.We conducted several experiments using three Artificial Neural Networks: Multilayer Perceptron, Gated Recurrent Units and bidirectional Long Short Term Memory networks and we tested their suitability using different combinations of inputs.Several metrics are provided for the sake of comparison within our study and between other studies.The solution has been evaluated in a real-world testbed in the city of Murcia(Spain)integrating an open human-mobility dataset showing high accuracy.A MAPE between 4%and 10%was reported in horizons of 1 to 3 h.

    Keywords:Parking space;human mobility mining;recurrent neural networks;prediction

    1 Introduction

    At the dawn of the smart cities era,municipalities are still facing a lack of proper transportation policies for the efficient management of human mobility.As a result,problems related to road traffic congestion have become challenging issues within urban settlements.In economic terms,it is estimated that traffic jams cost$87 millions in the USA in 20191https://inrix.com/scorecard/.Also,traffic congestions have a direct impact on the air quality of a region since petrol vehicles are one of the most important sources of air pollutants[1].

    The development of services for parking space management is a sensible solution to reduce trafficcongestion,since roughly 40%of road traffic in cities is caused by drivers searching for parking space2https://www-03.ibm.com/press/us/en/pressrelease/35515.wss.In that sense,the anticipation of future parking-space demand would allow operators to develop shortterm and long-term counteractions to avoid mobility problems in urban regions.

    For that prediction,a paramount contextual factor to be considered is the road-traffic conditions of the urban areas.In that sense,a large course of action for parking availability prediction leverages the endless deployment of new and more powerful sensors across cities thanks to the Internet of Things(IoT)[2].More in detail,mechanisms based on static infrastructure sensors[3],mobile ones[4]or the crowdsensing paradigm[5]have already been proposed.In general terms,these solutions rely on highresolution mobility data that captures human or vehicle displacements at quite fine-grained temporal or spatial scales.For example,a solution based on mobile sensors mounted on floating vehicles(e.g.,taxis or buses)would allow monitoring the state of each road segment.

    However, high-resolution human mobility data is quite sensitive in terms of privacy.Its access is rather limited because of regulatory and economic policies [6].At the same time, the open data movement has promoted the release of an increasing number of human-mobility datasets [7–9].However, they are usually filtered and pre-processed for anonymization.Hence, open mobility data provides a much lower temporal and spatial resolution than ad-hoc mobility feeds.The development of intelligent systems leveraging such coarse-grained mobility data is still scarce in the domain.

    In this context,the present work introduces a novel mechanism to forecast the number of available spaces of a set of urban parking lots in the city of Murcia(Spain).Our system relies on two datasources.

    On the one hand, the predictor is timely fed with the sheer number of free spaces in the target parking lots.This is provided by a set of counter sensors installed in the parking lots which are,in turn,part of a wide IoT infrastructure deployed in the city.On the other hand,the system also considers the anonymized human trips co-occurring within the city.Such trips are extracted from a nationwide open mobility dataset released by the Spanish Ministry of Transportation(SMT)considering mobile-phone location data from several telephone operating carriers3https://www.mitma.es/ministerio/covid-19/evolucion-movilidad-big-data/opendata-movilidad.To ensure users’privacy,trips are defined on a spatial tessellation of the country based on large geographical areas.

    These two sources are used to train an Artificial Neural Network (ANN) able to anticipate the values from the IoT counters several hours ahead.Given the sequential nature of the inputs,two types of Recurrent Neural Networks(RNNs)have been evaluated,a LSTM[10]and a GRU[11]one.Both types of architectures can learn long and short-term patterns from a given input sequence but some slight differences in terms of complexity exist between them.Nevertheless,both have been successfully applied to the parking availability prediction problem [12].Hence, the present work also relies on the hypothesis that the latent human movement of a region provides valuable information about its parking occupancy level so such knowledge should be integrated as part of the prediction system.

    It is important to note that the inputs are defined at different spatial scales and sources.The occupancy of the parking lots is extracted from an IoT infrastructure,and,in spatial terms,such data is related to very particular spots within the city.However, the human trips dataset only provides a region-based view of the actual human mobility.Therefore, our solution studies whether human mobility data defined at a very high level gives insight into road-traffic behaviour related to city locations captured by the parking lots’sensors.Some parking-space predictors incorporate in their pipelines data defined at a coarse spatial granularity such as weather conditions.However,to the best of the authors’knowledge,there is a scarcity of proposals making use of coarse-grained feeds related to human movement information.This targets an important research gap in the parking availability prediction domain.

    Furthermore,our work also brings three important operational benefits with respect to existing solutions.Firstly,we can deploy it in cities without an IoT infrastructure able to capture fine-grained mobility data.Secondly,as a side effect,it can be regarded as a cost-effective mechanism.Finally,our approach relies on pre-anonymized human data so it will limit the privacy concerns among end-users.

    The rest of the paper is structured as follows.Section 2 provides an overview of current approaches for parking space prediction from different perspectives.Next,Section 3 puts forward the proposed forecasting mechanism and the main results of the performed experiments are described in Section 4.Finally,the main conclusions and the future work are summed up in Section 5.

    2 Related Work

    This section provides an overview of existing approaches for parking space prediction.We gathered three dimensions of the solutions, their input sources, applied algorithms and the urban service that they provide.Last, we also briefly review the main findings within the human mobility mining discipline.

    2.1 Input Sources

    Sensors,parking meters and crowdsensing are common sources for collecting real-time parking availability.

    Regarding static sensors, the mechanisms in [4,13,14] use parking meter data for its prediction.Counting data from parking lots at each moment is used in [15–21] where the available slots are collected every 5–15 min.Infrared and ultrasonic sensors can be used to identify a car in a particular spot[2], as well as Radio-Frequency Identification(RFID) [22] for single vehicles that are provided with a radio frequency tag.Furthermore,ferromagnetic sensors installed under the asphalt are used in[23].Others works focus on the ticket data collected from a smart-parking system[24].In terms of communication protocols,some proposals use an energy efficient LoRaWAN channel to transmit the collected data to a centralized repository[25].

    Concerning mobile sensors,most works rely on vehicle-mounted devices sensing traffic contextual features[13].A similar approach is followed in[4]where the GPS traces of the fleet of taxis are used for parking availability monitoring.More generalist ways to collect data use machine vision[26]and car count in order to know the number of free spaces in real-time.In this scenario, crowdsensing focuses mainly on developing smartphone apps to ensure the accuracy of parking information and location[5].

    Lastly, weather conditions might affect parking demand at different degrees.For that reason,several works have incorporated this data as part of their input [13,15,20].Other contextual urban feed that has also used in some works has been Point of Interest(POI)data indicating the density of certain venues like restaurants or bars or schools[20,21],since the POI distribution of a region strongly affects its human mobility patterns[27].

    The present work takes as input count-based parking data and weather conditions as other proposals in the literature.Our approach also considersmobiledata as input in the form of human mobility flows but its spatial and temporal granularity is much larger than the sources of previous works [4,13].As a matter of fact, the temporal resolution of the traffic dataset in [4] was fixed to 10 min instead of 1 h as in our case as it is pointed out in Sec.3.2.1.Furthermore,whilst works based on mobile data rely on street-based reports [4,13] our solution considers the incoming or outgoing human flows of large-scale geographical urban areas.For estimating human flows,other works rely on social media publications(geo-tagged information)[28,29].Unlike previous proposals,our approach combines datasources of different nature.Those are raw sensor data from the counters installed in the parking lots and transmitted through a IoT infrastructure(see Sec.3.1).and open data capturing the human flows in the target city at very low resolution.As we have seen,this fusion of raw sensor data and an open mobility feed for parking availability prediction has not been fully explored in the literature.

    2.2 Applied Methods

    Different ANN models have been applied for parking demand forecasting.Since this can be regarded as a time-series prediction problem,most works have used different types of Recurrent Neural Networks(RNNs),such as Long short-term memory(LSTM)[14,17,18]models.Moreover,the work in [13] combines a Graph Neural Network (GNN) and a LSTM model so as to also consider the underlying road topology of the target area.Hierarchical GNNs have been also used to predict the availability of different parking lots since they capture well the spatial correlations among different entities[21].Apart from that,Multi-layer Perceptrons(MLPs)have been also proposed and used as a baseline to compare with other algorithms,especially deep learning algorithms[23,30,31].GRU and LSTM have been proven to be the best RNN architectures for solving a wide set of sequential data problems[32].In[12],the GRU architecture achieves better results in nearly all scenarios(4 countries and different exogenous variables) compared to the LSTM version for 6 h-ahead prediction.GRU have also been successfully applied to capture dynamic temporal autocorrelations between several parking lots [33].In [25], authors proposed a network architecture combining three types of layers,1D convolutional, a LSTM and three fully connected ones for the prediction task.Also, generative adversarial networks (GANs) that merge spatio-temporal and weather data from different urban datasources have been used[20].

    Furthermore,researchers in[15]make use of Markov chain analysis as they assume the Markov property in the context of parking space prediction.Similarly,a continuous-time Markov queue is put forward in[16].A likelihood maximization approach is applied to deal with the heterogeneity of crowdsensed data is presented in[5].Other methods like Kalman filters have been also studied[4].In addition to that,some proposals have used ensemble-learning algorithms like Random Undersampling Boosts to perform the prediction task [24].In a similar way, [19] proposes an ensemble model where three different algorithms,Random Forest regressor,Gradient boosting regressor and Adaboost regressor are combined by means of bagging and boosting techniques to provide a final prediction outcome.

    In terms of applied models,our proposal has evaluated three different models,MLP,GRU and bi-LSTM ones in order to find the most suitable hand for the task in hand as it is described in Sec.4.4.2.

    2.3 Provided Urban Services

    Different parking scenarios:research works deal with improving efficiency at closed parking lots which are paid parking lots(both indoors and outdoors),andpen parking lots or ticketing.

    As long as the first scenario is concerned,a mechanism to predict the occupancy of three private and public parking lots in the next 60 min in Geneva (Switzerland) is proposed in [15].A set of 14 parking lots in San Francisco (USA) are used as a use case in [16].In this case, the future available parking places area provided a time horizon of up to 12 min.In the same city, the proposal in[4] controls on-street parking spaces of 579 road segments.The work in [17] provides a prediction mechanism for a single parking lot in Shenzhen (China) with a prediction horizon of 30 min.In[21]the proposed system was able to provide predictions for a set of different parking lots at Beijing and Shenzhen 15, 30 and 45 min ahead.A holistic IoT infrastructure that able to make predictions of 30 parking locations of Birmingham (UK) is proposed in [18].Unlike previous approaches, the predictions are given in hour intervals instead of minutes and the same dataset is used for 30-minute time horizons in[19].

    Regarding the second group of on-street parking spaces,the work in[13]applies its predictor to a downtown area in Pittsburgh(USA).In[20],the prediction targets an outdoor regulated parking zone in Santander (Spain).In both cases, the prediction horizon is up to 30 min.In [5], the crowd-based prediction mechanism is tested in seven on-street public parking spaces in London(UK).In[14],3000 on-street parking slots were controlled by the solution to provide predictions up to 30 min ahead.The problema can be shifted from regression to classification[24].This way,the solution is able to classify the level of service at different on-street parking areas in Madrid(Spain).In[25],the predictor focuses on a street in the city of Montova(Italy)providing parking lots prediction from 1 to 8 h ahead.

    Our work forecasts the available parking places of two public garages with a larger time granularity(hourvs.minutes)and a wider time horizon of the prediction,as is stated in Sec.4.4.2.

    2.4 Human Mobility Mining

    Human mobility repeats over time(e.g.,people tend to go to work at the same time every working day)so it is usually modeled as a time pattern highly predictable at some extend[34].The prediction of where and when people are going to move is an instrumental tool in many domains[35].

    Nowadays it is possible to capture location data at multiple spatial and temporal scales from a large palette of technologies like GPS, Bluetooth or WiFi.Based on these technologies, a variety of machine learning models have been proposed for human mobility forecasting [36].These solutions basically rely on the raw spatio-temporal trajectories from movingentities, like taxis or individuals,equipped with the location technologies.This allows collecting their current location every few minutes or even seconds.

    In this context,human mobility patterns have already been successfully extracted from such open feeds and used in different predictive problems[37].However,there is a scarcity of proposals leveraging such mobility coarse-grained feeds as reliable input sources to improve other predictive tasks.

    3 Setting Overview

    This section describes the urban scenario for testing our approach.In that sense,we describe the available data sources and their suitability as input features for parking availability prediction.

    3.1 General Description

    The feasibility of our solution has been evaluated in Murcia, a city in the South-East of Spain.Fig.1 depicts the target urban area within this city.The city of Murcia is involved in a holistic smartcity project called MiMurcia4https://mapamurcia.inf.um.es/that is developing a platform that interconnects data and services to provide citizens with personalised and updated information about the status of the city and to make it smart.Murcia is a city with a great trajectory in the use of IoT data for providing smart services such as water [38] and energy management [39].Within this platform, several data is published:irrigation, problems reported by citizens (with pictures) and the parking and weather information that our solution uses.

    Figure 1:Map of the target city.Each colored polygon is a particular Mobility area from the mobility study whereas the dots indicate the locations of the two target parking lots

    Our solution would eventually enrich this ecosystem with a decision-support system for road traffic management.This way,it would provide drivers with real-time information about the available parking spaces in their destination areas on the basis of their time of arrival to such destinations.

    3.2 Target Datasets

    Given the urban infrastructure described in the previous section,we focused on the following data sources related to the city to perform the prediction task.

    3.2.1 Low-Resolution Human Mobility

    As it was mentioned in Sec.1, we used the nationwide human mobility dataset released by the SMT in December 2020.This dataset covers a 9-month period from February 29th to November 30th,2020,and it indicates the number of trips among 3216 ad-hoc administrative areas(hereby Mobility Areas,MA)per hour in Spain both in its peninsular and insular extension.In that sense,a single trip stands for the spatial displacement of an individual with a distance above 500 meters.Consequently,this dataset can be regarded as a set of tuples taking the form〈date,hour,morigin,mdest,ntrp〉,reporting that there was ntrphuman trips from the MA moriginto mdestduring the indicateddateandhour.

    This data has been collected through Call Detail Records (CDRs) from 13 million users of an unspecified mobile-phone carrier [40].Once anonymised, representative mobility statistics at the nation-level of the population of Spain were inferred and made publicly available open data.In its raw form, the dataset comprises 830, 450, 300 trips among MAs.To filter the trips related to the target city,we focused on the 3 particular MAs that spatially cover its geographical extension.Fig.1 shows this set of three MAs,M=〈mA,mB,mC〉with areas of 5.08,3.04 and 3.66 km2 and populations of 54,279 people,47,249 and 46,738respectively.Next,we extracted the time series for each target area,comprising the time-ordered sequence of incoming trips ih towards m ∈M per hour h.This way,the mobility profile of each MA of the city was composed in terms of its incoming flows of people.Given the geographical and population magnitudes of these MAs,their profiles provide much more coarse-grained mobility information than one based on road segments as much of the actual state of the art relies on.

    3.2.2 Parking Space

    We are collecting also the occupancy information from two underground parking lots in the area that operate 24/7,La Vega and Libertad parking.The actual locations of both of them are depicted in Fig.1 as colored dots.In particular,the number of free spaces is collected every two minutes for each of them.This will constitute the target variable of our system.

    The two parking lots are very close toThe Circular Squareof Murcia,which is one of the most important squares in the city.Very crowded streets link to this place, including theGran Vía, full of commerce and bustle.La Vega Hospitalis also in the area.The data was aggregated into hourly intervals in order to make it coincide with that of mobility flows and the tables were joined.Adding the number of parking slots from both parking we count with 642 spaces.Last,the quality of the data from this source has already been tested in previous studies[11].

    3.2.3 Weather Conditions

    Wind (strength and direction), rainfall, daylight, relative humidity and temperature have been proven to have an impact on the choice of travel modes(walk,public transportation and vehicles)[41]and therefore,we have also studied the suitability of weather data as input for the target predictor.Since people would normally make a decision based on current weather,we decided not to incorporate future weather information into the system.We used the IMIDA (The Research Institute of Agriculture and Food Development of Murcia) service in order to obtain the historical weather information5https://www.imida.es/.They provided us with an hourly historical set of data including the following variables:temperature(mean, min and max) (?C), humidity (mean, min and max) (%), radiation (mean and max) (w/m2),wind speed (mean and max) (m/s2), wind direction (mean)(degrees), precipitations (mm), dew point(?C)and vapour pressure deficit(kPa)from the closest weather station to the studied zone.

    3.3 Correlation Study

    Given these input sources, we carried out a correlation study to assess whether the extracted mobility flows and the weather parameters are reliable sources for parking-space prediction.In that sense, statistical correlation measures the relationship between two random variables, whether it is causal or not.Although it could refer to any statistical association, is commonly related to linear correlation.

    In our scenario, we can assume that our data is normal, which means that the distribution of each variable fits a Gaussian distribution.As the sample becomes larger, the sample’s distribution approximates to a normal distribution according to the Central Limit Theroem[10.Therefore,given our scenarios,we can use Pearson’s correlation coefficient to summarize the correlation between the variables.

    Pearson’s correlation coefficient is the covariance of the two variables divided by the product of their standard deviations.Given paired data{(x1,y1).,...,(xn,yn)}consisting of n pairs,it is written as follows,

    Pearson’s correlation coefficient is a number between-1 and 1 that describes a negative or positive correlation respectively.A value of zero indicates no correlation.

    Fig.2 shows the correlation heatmap between the free parking spaces (free), the considered mobility flows〈ImA,ImB,ImC〉namedflow1,flow2andflow3in the figure and the weather data.From this,we chose the first two flows since they show a greater negative correlation with the output variable than the third.We also chose the mean temperature(tmed),mean radiation(radmed)and mean wind speed(vvmed).The rest of the weather variables are highly correlated amongst themselves and would add noise to the model.Also, even though from this correlation analysis we find that precipitations(prec)might not be crucial,this could be due to the lack of rain in the studied region.In that sense,we decided to include it nevertheless because it is an important factor when choosing transportation.

    Figure 2:Correlation plot of all considered input variables

    Regarding the mobility flows, we can see from Fig.2 that the three of them have a negative correlation with the parking-space variable.However,flows ImAand ImBhad a slightly higher negative correlation(-0.642,-0.719)than the flow for the area mC(-0.437)so we opted for just including the two former flows to the predictor.This difference in the coefficients makes sense as area mCis the one more distant to the target parking lots location(see Fig.1).Hence,it associated latent mobility had a lower impact on the actual demand of the parking lots.

    For the sake of clarity,Fig.3 shows the two mobility flows selected from the analysis,ImAand ImB.Both flows are depicted along with the evolution of available parking spaces during the whole period of study.

    Figure 3: The two human mobility flows finally used for the prediction task along with number of available parking spaces during the period of study (a) Available parking spaces and the ImA flow during the period of study(b)Available parking spaces and the ImB flow during the period of study

    Furthermore,there are many phenomena in which the past influences the present in the sense that they can be used to foretell what will happen.When such phenomena are represented as a time series,they are said to have an auto-regressive property.The correlation for univariate time series observations with previous time steps,called lags,is an autocorrelation.A plot of the autocorrelation of a time series by lag is called the AutoCorrelation Function(ACF).In this case,instead of correlation between two different variables,the autocorrelationat lagkis between two values of the same variable at timesxiandxi+k.The previous formula is then transformed to the following,

    Second, the heroine is accomplishing a difficult task, removing a settled stain from clothing. Psyche, in Cupid and Psyche, has to perform three impossible tasks to prove her devotion to Cupid.

    Finally,partial autocorrelation is a summary of the relationship between an observation in a time series with observations at prior time steps with the relationships of intervening observations removed.The autocorrelation for an observation and some prior time step of itself is comprised of both the direct correlation and indirect correlations.These indirect correlations are a linear function of the correlation of the observation,with observations at intervening time steps.

    Fig.4 shows that parking spaces have a positive autocorrelation in the first 7 lags, reaches its maximum negative autocorrelation in lag 12 and then again the best positive correlation is present 24 h later.Regarding the partial autocorrelation, we see that there are many values statistically significant.From this we can conclude that the free parking presents an autorregresive behaviour and that including lagged values as inputs of the predictive models will likely imply a greater accuracy.

    Figure 4:The visualization of the autocorrelation and partial autocorrelation of free parking spaces(a)Autocorrelation parking availability(b)Partial autocorrelation parking availability

    4 Proposed Solution for Parking Space Prediction

    In the light of the available datasources in the urban scenario and their suitability as independent variables for the prediction task at hand, we are now able to properly define a solution for parking space prediction based on them.

    Our system takes as input three sources:historical data of available spaces in the target parking lots, the human mobility flows at low resolution and the weather conditions in the parkings’surrounding areas.

    4.1 Problem Formulation

    The parking space prediction problem that we focus on can be formulated as follows:

    Given the hour h ∈〈0,...,23〉,the number of available parking spaces during the last hprev hours from the two target parking lots,Ph=(ph,ph-1,..,ph-hprev),the number of incoming trips of a set of MAsand the weather conditions of the region Wh=〈wh,wh-1,..,wh-hprev〉,Find a mapping function F(Ph,Ih,Wh)→ph+T,where ph+Tis the sheer number of available parking spaces in the parking lots in the next T hours.

    4.2 Candidate Prediction Methods

    4.2.1 MLP Model

    This is one of the most popular and successful ANNs architectures that for a wide range of regression problems[42].Its inner architecture basically comprises an input layer,one or more hidden layers and an output one where each layer is connected to another one by non-linear functions.Each layer comprises a set of neurons connected to all neurons of the next layer but not to each other.In computational terms,the output of a MLP layer be defined as h(W,b)(X)=φ(XW+b),where W is the connections’weight matrix,X is the matrix with the input features,b is the vector with the bias terms andφis the activation function.

    4.2.2 GRU Model

    Given the sequence nature of the two validated input sources of our approach, we have used a GRU model for the prediction task [43].This is a foremost variant of Recurrent Neural Networks(RNN)able to learn short-term and long-term patterns in sequences of data.Unlike other RNNs,like LSTM,a GRU model has a slightly simpler structure which makes it faster to train[44].

    Fig.5a shows the general structure of a GRU network whereas Fig.5b depicts the inner structure of a GRU cell.As we can see,a GRU model just follows the composition of a regular RNN.Concerning the structure of the cells,they make use of a gated mechanism to memorize long-term patterns in the target sequence.Thus,a cell receives as input the current input vector x(t)and the previous state vector h(t- 1).Then,the cell generates the associated output y(t)which is also the state vector h(t)of the next cell.

    Figure 5: Illustrative architecture of the GRU model (a) A GRU model with three layers unrolled thought time(b)Gated structure of a GRU cell.FC stands for Fully Connected

    More in detail, the cell comprises three different gates, the update z(t), the reset r(t) and the memory-content g(t).The computations of each gate are as follows,

    where W{z,r,g}are the weight matrices for the input xt,U{z,r,g}are the weight matrices for the connections to the previous short-term state h(t-1)and b{z,r,g}are the bias terms of each layer.

    4.2.3 Bi-LSTM Model

    Last, we have evaluated a Long-Short Term Memory (LSTM) model.This is also a wellestablished RNN variant[45].We have evaluated a bidirectional LSTM model[46].Unlike atraditionalLSTM network, a bidirectional approach allows the model not only to learn “causal” patterns by means of its past and present inputs but also to“l(fā)ook ahead”into the future.As Fig.6a depicts,this is done by concatenating two RNNs one reading the input from left to right and the other from right to left.Then,the output of each cell is combined with its counterpart.

    Figure 6:Illustrative architecture of the bidirectional LSTM model.(a)A bidirectional LSTM layer with six cells(b)Gated structure of a LSTM cell.FC stands for Fully Connected

    The inner structure of a LSTM cell is depicted in Fig.5b,where h(t-1)indicates the short-term state at time instant and x(t)is the current input vector at instant t.Concerning the cell outputs,y(t)is the predicted availability at instant t whereas c(t)is the long-term state that traverses the network from left to right.Focusing on the four inner gates of the cell that modulates the outputs of the model,they can be formulated as follows:

    where x(t)is the parking availability at time instant t,W{f,i,o,g}are the weight matrices for this feature,U{f,i,o,g}are the weight matrices for the connections to the previous short-term state h(t-1)and b{f,i,o,g}are the bias terms of the four gates.

    As we can see,the key difference between the GRU and the LSTM cell is that the former has three layers that are reset r(t),update z(t)and memory-content g(t)gates while the latter has four layers that are the input i(t),output o(t)and forget f(t)gates along with an output layer g(t).Thus,GRU is less complex than LSTM because it has a smaller number of gates.

    4.3 Generation of the Models

    4.3.1 Data Pre-Processing

    The chosen period of analysis consisted of two ranges: 13th of July-31st of July and 29th of September-19th of October.We avoided August and September because in Murcia activities are reduced given the temperatures and holidays.Also,COVID-19 restrictions were more severe in those periods,so we decided to let them out of the analysis in order not to contaminate the sample.In the end,we considered 927 data points in roughly 2 months.

    The raw parking data and the raw weather data are collected with more granularity than the mobility flows,that is 4 and 15 min respectively.However,we homogenised all data in hourly intervals,which was the mobility flows’granularity.Given that we had to resample to a wider time frame,we used down-sampling and the aggregation function was the mean.After downsampling,due to failures in the communication system in the parking set,a 6%of the sample was missing.In order to impute the values,we used linear interpolation.

    The performance of Deep Learning models and regression models is improved when scaling[47].Scaling the inputs and outputs used in the training of the model is meaningful since small weights and errors in prediction values are used.Using unscaled inputs may slow or destabilize the learning process,while un-scaled target variables in regression problems may cause gradient explosions,leading to failure of the learning process.For that purpose, we have used Standardization (Z-score Normalization) in all our inputs.Feature standardization makes the values of each feature in the data have zero-mean(when subtracting the mean in the numerator)and unit variance.The formula is

    4.3.2 Models’Configuration

    All models had a training rate of 70 %, batch size 32 and learning factor 0.01, Mean Squared Error (MSE) as loss function and the Adam optimizer was used (α= 0.001,β1=0.9,β2=0.999,epsilon=10-7).The hyperparameters of the models were selected using a grid search with Tab.1 configuration.MLP used the ReLU activation function with 3 layers,50 cells per layer and 50 epochs.GRU used the Hyberbolic tangent function for activation with 2 layers, 50–20 cells in each of them and 100 epochs.Finally,LSTM used the same activation function with 3 layers,50–20–20 cells in each of them and 400 epochs.

    Table 1: Configuration of the grid search for the models’hyperparameters

    Fig.7 shows the learning curves of the models with the provided configuration.As we can see,we avoid overfitting all models with the provided datasets.

    Figure 7:Learning curves of the three target models in the experiment.(a)MLP model(b)GRU model(c)Bi-LSTM model

    Last,the number of parameters of these three models when they are fed with the incoming human flows,available parking spaces and weather conditions are 8801 in the case of the MLP,13611 for the GRU model and 33321 for the bi-LSTM model.

    4.4 Evaluation of the Approach

    4.4.1 Metrics

    4.4.2 Obtained Results

    Tab.2 shows the metric values obtained by the models.This table also includes, as a baseline, a naive method that just returns the last value phin each series as the predicted value p∧h+T.Surprisingly,this type of straightforward mechanism is sometimes quite difficult to outperform.

    Table 2:Metrics values for the evaluated models for different time horizons T(h).The parameters in brackets of each model indicate its input sources.The best value per metric is shown in bold

    Table 2:Continued

    The table includes the results for 3 different hourly horizons (T): 1, 2 and 3 h ahead parking prediction.Obviously, a decrease in accuracy is observed the further we try to estimate, especially in the combinations(model and input)that provide better results.As was highlighted in bold in the table,GRU model is clearly better than the other alternatives in all cases.However,there are certain differences between metrics and scenarios when including or not the weather in the inputs.

    In 1 h-ahead forecasting, when penalizing the points further away from the mean, the best combination of inputs includes previous parking observations and mobility flows: RMSE=20.36 parking spaces and CVRMSE=4.39%.However, when not penalizing it, the best combination includes also the weather information:MAE=15.32 parking spaces and MAPE=4.04%.The choice between the two depends on the objectives of parking spaces.We consider that if we are mistaken by 10 spaces should count more than just two times being mistaken by 5 spaces since precision is very important.Therefore,we would choose not to include the weather.

    In 2 h-ahead and 3 h-ahead the better results are achieved when not considering the weather as input for all metrics,having a CVRMSE=8.5%and CVRMSE=10%respectively.For the 3 h-ahead scenario,the Bi-LSTM model achieved slightly better results than the GRU model.Nevertheless,it is important to remark that this improvement comes at the cost of using a much complex neural network as a bi-LSTM not only uses more complex cells than a GRU model but also its layers include extra connections(parameters)for the in-reverse flow of information as it was put forward in Sec.4.2.3.

    Fig.8 shows the actual available spaces in the target parking lots in the test interval along with the predictions of the generated models for a time horizon T=1 h.As we can see,the actual availability of the target parking lots follows a quite strong daily seasonal pattern.However, this figure also shows that only the GRU model was able to correctly predict the abnormal availability of the parking lots during the interval between the 80 and 120 h.This clearly shows a better matching between the prediction and the actual free parking when using GRU compared to MLP.

    Figure 8: Time series of the available parking spaces (ph) and the prediction outcome of the models(.(a)MLP(I,P,W)(b)GRU(I,P,W)(c)Bi-LSTM(I,P,W)

    Finally, Tab.3 shows a comparison between the results obtained by our solution and previous approaches in the literature for availability prediction in closed parking lots(Sec.2.3).

    From this table, it is important to remark that results in [13–17] are given with respect the occupancy rate of the car parks.If we compare such values with the computed rate-based metrics of our approach(0.0233 MAE and 0.0311 RMSE for 1 h time horizon),we can see that our approach obtained slightly more accurate results.Furthermore,the time horizon is much larger than the baseline approaches as they range from a few minutes to 1 h.Regarding the proposal in[18],the absolute values of the metrics are also quite similar with a much larger time horizon.

    All in all, results state that the usage of the human mobility dataset has clearly improved the accuracy of the predictor.However,such a dataset reflects all types of human displacements without considering the adopted means of transport.Since the parking lots are only used by motorized vehicles,the impact of using the SMT feed also reflects that the human mobility contained in this source is mainly defined by private-car trips.Consequently, this also confirms the fact that such vehicles are still the dominant means of transport for regular trips in most urban settlements.

    Table 3:Comparison of our approach with existing approaches of the literature.In brackets the time horizon of each solution

    5 Conclusions

    The management of road traffic in large cities is one of the most challenging problems for authorities and operators.Since many urban traffic jams are caused by drivers searching for parking spaces,the accurate prediction of available spaces is an instrumental tool to enable efficient road-traffic policies.

    However, most of the existing solutions generally assume that such contextual data is easily available through a large palette of location sensors.This limits the usability of current solutions because the access to location data in most real-world scenarios is restricted due to many regulations.

    In this context,the present work proposes a novel mechanism for parking space prediction that leverages area-based human mobility data as the primary contextual source.The key goal was to evaluate whether such a coarse-grained mobility feed was useful to detect latent movement behaviors in particular locations within a city.To do so,we tested our approach in a real-world setting by making use of different ANNsto predict the available spaces of two downtown parking lots.Results have shown that enriching the models with the flows from the open dataset meaningfully improved their prediction accuracy.

    In the future, with regards to research, our system can be used to feed fine-grained systems to improve their generalisation and also to study general parking patterns between cities to develop transfer the extracted knowledge between the similar ones using transfer learning.With regards to the development of new services,our parking availability prediction can be part of a system that includes route suggestions to drivers according to users’preferences with regards to time,pollution,money,and availability to use other transport services in conjunction with their private car.Also,a new business model can emerge,such as the automatic,hourly,and renting of private parking spaces according to real-time needs.

    Code availability:The source code of the project is available at https://github.com/fterroso/mobility_murcia_umu_ucam.

    Funding Statement:This work has been sponsored by UMU-CAMPUS LIVING LAB EQC2019-006176-P funded by ERDF funds,by the European Commission through the H2020 PHOENIX(grant agreement 893079)and DEMETER(grant agreement 857202)EU Projects.It was also co-financed by the European Social Fund(ESF)and the Youth European Initiative(YEI)under the Spanish Seneca Foundation(CARM).

    Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.

    成年人午夜在线观看视频| 日韩三级伦理在线观看| 一区二区三区精品91| 成人亚洲精品一区在线观看| 99久久人妻综合| 亚洲成人手机| 极品人妻少妇av视频| 久久久久久久久久人人人人人人| 中文字幕制服av| 丝袜脚勾引网站| 一级片'在线观看视频| 中文欧美无线码| 大陆偷拍与自拍| 日本欧美视频一区| 侵犯人妻中文字幕一二三四区| 一级a做视频免费观看| 只有这里有精品99| 色网站视频免费| 久久这里有精品视频免费| 亚洲人与动物交配视频| 综合色丁香网| 日本vs欧美在线观看视频| 国产av码专区亚洲av| 久久精品aⅴ一区二区三区四区 | 国产精品秋霞免费鲁丝片| 在线天堂最新版资源| 午夜福利乱码中文字幕| 免费看av在线观看网站| 国精品久久久久久国模美| 九色成人免费人妻av| 男女啪啪激烈高潮av片| 久久久久久久久久久免费av| 午夜福利视频在线观看免费| 国产日韩欧美视频二区| 9191精品国产免费久久| 91久久精品国产一区二区三区| 男人操女人黄网站| 国产高清不卡午夜福利| 男女无遮挡免费网站观看| 蜜桃国产av成人99| av国产久精品久网站免费入址| 亚洲成国产人片在线观看| 视频在线观看一区二区三区| 女性生殖器流出的白浆| 免费人成在线观看视频色| 超碰97精品在线观看| 一区二区三区乱码不卡18| 欧美国产精品一级二级三级| 日本vs欧美在线观看视频| 18在线观看网站| 你懂的网址亚洲精品在线观看| 热99久久久久精品小说推荐| 久久精品国产鲁丝片午夜精品| 国产片特级美女逼逼视频| 最近中文字幕高清免费大全6| 七月丁香在线播放| 国产一区二区在线观看av| 日本色播在线视频| 黑丝袜美女国产一区| 国产精品人妻久久久久久| 欧美成人精品欧美一级黄| 春色校园在线视频观看| av一本久久久久| 色婷婷av一区二区三区视频| 亚洲精品视频女| 久久久久国产网址| 午夜福利乱码中文字幕| 国产精品麻豆人妻色哟哟久久| 少妇熟女欧美另类| 精品久久久精品久久久| 视频中文字幕在线观看| 亚洲欧洲日产国产| 久久久久视频综合| 观看av在线不卡| 国产精品 国内视频| 亚洲成色77777| 国产精品国产av在线观看| 久久精品国产鲁丝片午夜精品| 秋霞在线观看毛片| 美女内射精品一级片tv| 国产男女内射视频| 国产黄频视频在线观看| 全区人妻精品视频| 少妇被粗大的猛进出69影院 | 欧美变态另类bdsm刘玥| 亚洲少妇的诱惑av| 亚洲国产欧美日韩在线播放| 人妻少妇偷人精品九色| 国产成人av激情在线播放| 国产精品人妻久久久影院| 免费在线观看黄色视频的| 男的添女的下面高潮视频| 国产在线视频一区二区| 国产av一区二区精品久久| 国产在线一区二区三区精| 欧美3d第一页| 久久久久国产精品人妻一区二区| 午夜老司机福利剧场| 国精品久久久久久国模美| 国产亚洲最大av| 少妇被粗大猛烈的视频| 久久久亚洲精品成人影院| 色网站视频免费| 亚洲国产日韩一区二区| 男女无遮挡免费网站观看| 久久久a久久爽久久v久久| 久热久热在线精品观看| 国产日韩欧美视频二区| 人妻少妇偷人精品九色| 热99国产精品久久久久久7| 黄色毛片三级朝国网站| 久久午夜综合久久蜜桃| 有码 亚洲区| 久久久久久伊人网av| av免费观看日本| 久久久亚洲精品成人影院| 日本-黄色视频高清免费观看| 国产一区二区三区综合在线观看 | 人人妻人人澡人人看| 婷婷色综合大香蕉| 国产熟女午夜一区二区三区| 国产一区二区在线观看日韩| 午夜福利网站1000一区二区三区| 国产成人精品在线电影| 国产精品人妻久久久久久| 赤兔流量卡办理| 久久久精品免费免费高清| 巨乳人妻的诱惑在线观看| 久久久久久久亚洲中文字幕| 免费黄频网站在线观看国产| 在线观看www视频免费| 欧美精品亚洲一区二区| 最后的刺客免费高清国语| 久久 成人 亚洲| 在线观看国产h片| 欧美国产精品一级二级三级| 精品久久国产蜜桃| 亚洲欧洲精品一区二区精品久久久 | 亚洲精品一二三| 蜜桃在线观看..| 赤兔流量卡办理| 国产毛片在线视频| 欧美激情 高清一区二区三区| 久久久欧美国产精品| h视频一区二区三区| 亚洲精华国产精华液的使用体验| 在线天堂中文资源库| 考比视频在线观看| 赤兔流量卡办理| 热99久久久久精品小说推荐| 日韩成人伦理影院| 亚洲久久久国产精品| 午夜免费男女啪啪视频观看| 少妇被粗大的猛进出69影院 | 国产有黄有色有爽视频| 国产精品久久久久久久电影| 日本免费在线观看一区| 视频区图区小说| 乱码一卡2卡4卡精品| 人妻一区二区av| 日本-黄色视频高清免费观看| 日本黄大片高清| 日韩人妻精品一区2区三区| 男女啪啪激烈高潮av片| 亚洲天堂av无毛| 精品久久久久久电影网| 色哟哟·www| 黄色怎么调成土黄色| 婷婷色综合www| av女优亚洲男人天堂| 欧美3d第一页| 久久精品国产a三级三级三级| 日本欧美视频一区| 亚洲久久久国产精品| 精品国产一区二区三区久久久樱花| 大话2 男鬼变身卡| 一区二区三区精品91| 亚洲美女视频黄频| 亚洲精品国产色婷婷电影| 青青草视频在线视频观看| 18禁在线无遮挡免费观看视频| 亚洲国产精品成人久久小说| 一级黄片播放器| 亚洲国产欧美日韩在线播放| 男女午夜视频在线观看 | 中文字幕人妻熟女乱码| 亚洲色图 男人天堂 中文字幕 | 亚洲内射少妇av| 激情五月婷婷亚洲| 一边亲一边摸免费视频| 久久人人爽人人爽人人片va| 欧美另类一区| 久久99热这里只频精品6学生| 国产色婷婷99| freevideosex欧美| 中文欧美无线码| 国产欧美日韩综合在线一区二区| 街头女战士在线观看网站| 日本wwww免费看| 亚洲国产av新网站| 亚洲精品色激情综合| xxx大片免费视频| 亚洲国产欧美在线一区| av一本久久久久| 黄色怎么调成土黄色| 91精品国产国语对白视频| 黄色毛片三级朝国网站| 国产免费一区二区三区四区乱码| 欧美日韩国产mv在线观看视频| 国产老妇伦熟女老妇高清| 国产精品久久久久久精品电影小说| 2021少妇久久久久久久久久久| 国产片特级美女逼逼视频| 狂野欧美激情性bbbbbb| 国产精品久久久久久精品古装| 亚洲国产精品国产精品| 男人添女人高潮全过程视频| 少妇被粗大的猛进出69影院 | 夫妻性生交免费视频一级片| 欧美日韩综合久久久久久| 国产免费福利视频在线观看| 90打野战视频偷拍视频| 免费观看在线日韩| 国产极品天堂在线| av福利片在线| 国产精品国产三级国产av玫瑰| 亚洲内射少妇av| 免费观看性生交大片5| 春色校园在线视频观看| 熟女人妻精品中文字幕| 秋霞在线观看毛片| 999精品在线视频| 91成人精品电影| 赤兔流量卡办理| 亚洲一码二码三码区别大吗| 成人国语在线视频| videossex国产| 亚洲欧美清纯卡通| 一二三四在线观看免费中文在 | 女人被躁到高潮嗷嗷叫费观| 老女人水多毛片| 久久久欧美国产精品| 欧美日韩综合久久久久久| 只有这里有精品99| 亚洲精品456在线播放app| a级毛色黄片| 国产日韩欧美亚洲二区| 亚洲人成网站在线观看播放| 精品少妇黑人巨大在线播放| 成人免费观看视频高清| 制服丝袜香蕉在线| 91aial.com中文字幕在线观看| 最近最新中文字幕免费大全7| 亚洲av电影在线观看一区二区三区| 91在线精品国自产拍蜜月| 夜夜骑夜夜射夜夜干| 欧美亚洲日本最大视频资源| 免费少妇av软件| 亚洲中文av在线| 永久免费av网站大全| 26uuu在线亚洲综合色| 亚洲欧洲日产国产| 中文字幕av电影在线播放| 男人爽女人下面视频在线观看| 涩涩av久久男人的天堂| 18+在线观看网站| 老司机影院毛片| 国产黄色免费在线视频| 国产国语露脸激情在线看| 精品久久国产蜜桃| 看免费av毛片| 亚洲四区av| 亚洲精品中文字幕在线视频| 精品亚洲成a人片在线观看| 国产不卡av网站在线观看| av有码第一页| 午夜福利在线观看免费完整高清在| freevideosex欧美| 色吧在线观看| 亚洲色图综合在线观看| 亚洲国产精品成人久久小说| 欧美日韩亚洲高清精品| 一区二区日韩欧美中文字幕 | 久久久久久久久久成人| 波野结衣二区三区在线| 久久久亚洲精品成人影院| 国产熟女午夜一区二区三区| 我的女老师完整版在线观看| 精品一区在线观看国产| av电影中文网址| 18禁观看日本| a级片在线免费高清观看视频| 国产成人精品久久久久久| 成人国产麻豆网| 美女xxoo啪啪120秒动态图| 热99国产精品久久久久久7| 久久人人97超碰香蕉20202| 免费看不卡的av| 狠狠婷婷综合久久久久久88av| 精品久久久久久电影网| 国产精品女同一区二区软件| 免费日韩欧美在线观看| 久久久久久人人人人人| 国产伦理片在线播放av一区| 亚洲国产精品一区二区三区在线| 亚洲熟女精品中文字幕| 国产精品三级大全| 国内精品宾馆在线| 亚洲图色成人| 日韩成人伦理影院| 纵有疾风起免费观看全集完整版| 精品国产一区二区久久| 91久久精品国产一区二区三区| 夜夜骑夜夜射夜夜干| 久久99热6这里只有精品| 国产午夜精品一二区理论片| 午夜日本视频在线| 久久ye,这里只有精品| 久久久久精品性色| 亚洲精品久久成人aⅴ小说| 成人毛片60女人毛片免费| 十分钟在线观看高清视频www| 一本久久精品| 新久久久久国产一级毛片| 在线精品无人区一区二区三| 亚洲天堂av无毛| 亚洲精品久久午夜乱码| 久久久精品区二区三区| 中国美白少妇内射xxxbb| 国产精品国产三级专区第一集| 欧美bdsm另类| 国产精品三级大全| 亚洲欧美一区二区三区国产| 精品亚洲成a人片在线观看| 亚洲成人av在线免费| 国产国语露脸激情在线看| 国产av国产精品国产| 香蕉精品网在线| 久久亚洲国产成人精品v| 国产黄色免费在线视频| 亚洲av电影在线观看一区二区三区| 免费黄色在线免费观看| 80岁老熟妇乱子伦牲交| 22中文网久久字幕| 国产亚洲欧美精品永久| 亚洲国产精品成人久久小说| 亚洲精品美女久久久久99蜜臀 | 中文字幕人妻丝袜制服| 少妇人妻 视频| 亚洲色图综合在线观看| 欧美成人午夜免费资源| 久久国产精品大桥未久av| 爱豆传媒免费全集在线观看| av女优亚洲男人天堂| 秋霞伦理黄片| 欧美亚洲 丝袜 人妻 在线| 天美传媒精品一区二区| 久久久久久人人人人人| 看免费成人av毛片| 国产极品粉嫩免费观看在线| 久久99蜜桃精品久久| 国产成人精品婷婷| 在线 av 中文字幕| kizo精华| 亚洲av国产av综合av卡| 国产亚洲最大av| 麻豆精品久久久久久蜜桃| 国产成人午夜福利电影在线观看| 亚洲精品视频女| 成人漫画全彩无遮挡| 亚洲国产成人一精品久久久| 少妇高潮的动态图| 中国三级夫妇交换| 色婷婷久久久亚洲欧美| 午夜视频国产福利| 在线观看免费日韩欧美大片| 综合色丁香网| 日日啪夜夜爽| 大话2 男鬼变身卡| 欧美少妇被猛烈插入视频| 欧美 日韩 精品 国产| 人人妻人人澡人人爽人人夜夜| 日本wwww免费看| 夜夜骑夜夜射夜夜干| 成人午夜精彩视频在线观看| 2022亚洲国产成人精品| 久久综合国产亚洲精品| 免费看av在线观看网站| 黄色配什么色好看| 精品国产一区二区久久| 99久久精品国产国产毛片| 亚洲第一av免费看| 2022亚洲国产成人精品| 色视频在线一区二区三区| 日本欧美视频一区| 大香蕉久久成人网| 国语对白做爰xxxⅹ性视频网站| 国产日韩欧美在线精品| 亚洲精品久久久久久婷婷小说| 综合色丁香网| 黑人巨大精品欧美一区二区蜜桃 | 亚洲情色 制服丝袜| 男人舔女人的私密视频| 久久国产精品大桥未久av| 国产欧美另类精品又又久久亚洲欧美| 老司机影院成人| 熟女av电影| 看非洲黑人一级黄片| 婷婷色综合www| 秋霞伦理黄片| 成人黄色视频免费在线看| av国产久精品久网站免费入址| 国产精品国产av在线观看| 日本色播在线视频| 国语对白做爰xxxⅹ性视频网站| 午夜免费观看性视频| 精品国产乱码久久久久久小说| 高清欧美精品videossex| 亚洲欧美成人精品一区二区| 日本wwww免费看| 99久国产av精品国产电影| 久久久久久久久久成人| 一区二区av电影网| 欧美人与性动交α欧美软件 | 爱豆传媒免费全集在线观看| 国产乱来视频区| 欧美激情极品国产一区二区三区 | 亚洲精品国产av蜜桃| 亚洲成人av在线免费| av一本久久久久| 新久久久久国产一级毛片| 丝袜人妻中文字幕| av卡一久久| 人妻 亚洲 视频| 亚洲欧洲国产日韩| 在线 av 中文字幕| 成年人免费黄色播放视频| 最近最新中文字幕免费大全7| 亚洲精品第二区| 黄色配什么色好看| 亚洲精品色激情综合| 999精品在线视频| 人成视频在线观看免费观看| 久久国产亚洲av麻豆专区| 日韩一区二区视频免费看| www.色视频.com| 热re99久久精品国产66热6| 精品熟女少妇av免费看| 久久午夜综合久久蜜桃| tube8黄色片| av在线观看视频网站免费| 丝袜美足系列| kizo精华| 亚洲国产欧美在线一区| 亚洲精品国产av蜜桃| 日韩av在线免费看完整版不卡| 王馨瑶露胸无遮挡在线观看| 国产成人午夜福利电影在线观看| 美女xxoo啪啪120秒动态图| 精品少妇内射三级| 国产又爽黄色视频| 日韩伦理黄色片| 国产精品熟女久久久久浪| 亚洲成av片中文字幕在线观看 | 少妇高潮的动态图| 午夜久久久在线观看| 自线自在国产av| 91午夜精品亚洲一区二区三区| 精品国产国语对白av| 尾随美女入室| 日韩中文字幕视频在线看片| 妹子高潮喷水视频| 亚洲欧美一区二区三区黑人 | 国产 精品1| 久久99一区二区三区| 亚洲综合精品二区| 国产高清不卡午夜福利| xxxhd国产人妻xxx| 亚洲欧美成人精品一区二区| 99久久中文字幕三级久久日本| 国产一级毛片在线| 纵有疾风起免费观看全集完整版| 精品少妇内射三级| 人妻一区二区av| 777米奇影视久久| 中文字幕亚洲精品专区| 纵有疾风起免费观看全集完整版| av卡一久久| 亚洲久久久国产精品| 男女边吃奶边做爰视频| 精品人妻一区二区三区麻豆| 91aial.com中文字幕在线观看| 婷婷色综合www| 成年人免费黄色播放视频| 男女边吃奶边做爰视频| 新久久久久国产一级毛片| 久久久久久久大尺度免费视频| 亚洲成人av在线免费| 少妇猛男粗大的猛烈进出视频| 99国产精品免费福利视频| 国产男人的电影天堂91| 久久久久久久国产电影| 亚洲精品日韩在线中文字幕| 国内精品宾馆在线| av有码第一页| 久久青草综合色| 尾随美女入室| 亚洲内射少妇av| 在线观看一区二区三区激情| 久久久久精品久久久久真实原创| 一边摸一边做爽爽视频免费| 永久免费av网站大全| av电影中文网址| av有码第一页| 久久人人爽av亚洲精品天堂| 18禁国产床啪视频网站| 日韩av免费高清视频| 涩涩av久久男人的天堂| 国产极品粉嫩免费观看在线| 亚洲成人一二三区av| 1024视频免费在线观看| 精品一区二区三区视频在线| 永久网站在线| 好男人视频免费观看在线| 亚洲国产日韩一区二区| 精品第一国产精品| 国产成人午夜福利电影在线观看| 黑人高潮一二区| av线在线观看网站| 欧美人与性动交α欧美精品济南到 | 国产色婷婷99| 亚洲国产精品专区欧美| 婷婷成人精品国产| 视频在线观看一区二区三区| 在线 av 中文字幕| 欧美xxⅹ黑人| 国产免费一区二区三区四区乱码| 99国产精品免费福利视频| 国产视频首页在线观看| 青春草视频在线免费观看| 桃花免费在线播放| 99久久人妻综合| 日本午夜av视频| 性色av一级| 亚洲av中文av极速乱| 精品一区二区三区视频在线| 欧美日韩综合久久久久久| 国精品久久久久久国模美| 亚洲成av片中文字幕在线观看 | 日本免费在线观看一区| 一级毛片我不卡| 久久久久国产网址| 成人漫画全彩无遮挡| 国产成人午夜福利电影在线观看| 大片电影免费在线观看免费| 亚洲欧美色中文字幕在线| 亚洲av福利一区| 久久国产精品男人的天堂亚洲 | 欧美成人精品欧美一级黄| 久久99精品国语久久久| 九色亚洲精品在线播放| 在线看a的网站| 精品一区二区三区视频在线| 满18在线观看网站| 国产精品嫩草影院av在线观看| 天堂中文最新版在线下载| 日本欧美国产在线视频| 国产有黄有色有爽视频| 最后的刺客免费高清国语| videosex国产| 日日摸夜夜添夜夜爱| 日韩,欧美,国产一区二区三区| 视频中文字幕在线观看| 日韩,欧美,国产一区二区三区| 日日摸夜夜添夜夜爱| 久久精品人人爽人人爽视色| 亚洲成人手机| 亚洲经典国产精华液单| 大片免费播放器 马上看| 90打野战视频偷拍视频| 午夜福利视频精品| 亚洲经典国产精华液单| 啦啦啦中文免费视频观看日本| 欧美激情国产日韩精品一区| 观看美女的网站| 国产探花极品一区二区| 在线观看美女被高潮喷水网站| 久久亚洲国产成人精品v| av天堂久久9| 蜜桃在线观看..| 国产成人av激情在线播放| 午夜福利,免费看| 欧美性感艳星| 在线精品无人区一区二区三| 最近中文字幕高清免费大全6| 捣出白浆h1v1| 在线观看免费高清a一片| 亚洲中文av在线| 丝袜美足系列| 一级毛片黄色毛片免费观看视频| 亚洲四区av| 亚洲av日韩在线播放| 水蜜桃什么品种好| 日韩av不卡免费在线播放| 久久人人爽人人片av| 午夜av观看不卡| 自线自在国产av| 精品久久久久久电影网| 国产日韩欧美亚洲二区| 中国三级夫妇交换| 免费少妇av软件| √禁漫天堂资源中文www| 少妇被粗大的猛进出69影院 | 一本久久精品| 大话2 男鬼变身卡|