• <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 ?

    TinyML-Based Fall Detection for Connected Personal Mobility Vehicles

    2022-08-24 03:31:14RamonSanchezIborraLuisBernalEscobedoJoseSantaandAntonioSkarmeta
    Computers Materials&Continua 2022年5期

    Ramon Sanchez-Iborra,Luis Bernal-Escobedo,Jose Santaand Antonio Skarmeta

    1University Center of Defense,General Air Force Academy,San Javier,30720,Spain

    2University of Murcia,Murcia,30100,Spain

    3Technical University of Cartagena,Cartagena,30202,Spain

    Abstract: A new wave of electric vehicles for personal mobility is currently crowding public spaces.They offer a sustainable and efficient way of getting around in urban environments,however,these devices bring additional safety issues,including serious accidents for riders.Thereby,taking advantage of a connected personal mobility vehicle, we present a novel on-device Machine Learning(ML)-based fall detection system that analyzes data captured from a range of sensors integrated on an on-board unit (OBU) prototype.Given the typical processing limitations of these elements, we exploit the potential of the TinyML paradigm,which enables embedding powerful ML algorithms in constrained units.We have generated and publicly released a large dataset,including real riding measurements and realistically simulated falling events,which has been employed to produce different TinyML models.The attained results show the good operation of the system to detect falls efficiently using embedded OBUs.The considered algorithms have been successfully tested on mass-market low-power units, implying reduced energy consumption, flash footprints and running times, enabling new possibilities for this kind of vehicles.

    Keywords: Personal mobility;machine learning;on-board unit;C-ITS;IoT

    1 Introduction

    The arrival of eco-efficient mobility solutions is a reality in both dense urban areas and city outskirts[1].Electric devices such as mopeds,segways,or scooters are currently massively employed for both transportation and leisure activities (Fig.1).Nevertheless, the integration of these vehicles within the Cooperative-Intelligent Transportation Systems (C-ITS) ecosystem is far from being accomplished.This would enable the development of novel services and applications devoted to this specific vehicle segment, e.g., device monitoring, route planning, or traffic safety, among others.Regarding the latter,the number of accidents in which personal mobility devices are involved is notably increasing every year[2].Key statements such as per 100,000 e-scooter trips,20 people will end up with some sort of injury[3],evidence the importance of this new type of traffic accidents.This may be highly dangerous for riders, as they can suffer severe falls [4].Recent studies reveal that these vehicles are proner to falls than regular bicycles,due to body dynamics and wheel size[5].According to a recent study in [6], the type of injuries mainly involves the head (52%), arms (40%) and legs (20%), which indicate that these accidents can involve critical consequences.In fact,22.5%of the injures types were identified as fractures.Currently, these vehicles are not provided with devices detecting any kind of accident and their connection with an external system is limited to a tracking system when they are rented.In this line, from the authors’perspective, automatic fall detection (AFD) systems may be critically useful for noticing these events and triggering an alarm to the corresponding emergency service using wireless communications.

    Figure 1:Personal mobility devices ecosystem in urban environments supported by mobile computing devices

    On-board units (OBUs) are processing elements that are usually attached to vehicles aiming at enriching them with computing capabilities.This permits to run applications for supporting driving tasks and enhancing user experience[7].OBUs embedded in traditional vehicles(i.e.,cars or vans)do not usually present neither computation nor energy constraints, as they are powered by the vehicle battery.However, OBUs attached to personal mobility vehicles suffer from energy limitations, since they can have a built-in battery or be connected to the limited vehicle’s power source.For that reason,both processing and memory capacities should be carefully saved in this case.This is the reason why hardware embedded in these units and application/system software should be designed to maintain a low operation profile.This is an issue for artificial intelligent algorithms needed by smart AFD systems, which are based on complex Machine Learning (ML) techniques requiring intensive data processing.

    At this point is where the TinyML approach emerges[8].This novel paradigm proposes to adapt ML models generated by widely-adopted frameworks such as TensorFlow or Scikit-learn, aiming at making them operational on constrained end-devices.Several alternatives are already publicly available in the market, e.g., Google’s TensorFlow Lite or Microsoft’s Embedded Learning Library(ELL), among many others.In this work, we propose an AFD mechanism to be embedded in an energy-constrained OBU.To this end,firstly,a large dataset has been generated,which includes data collected from real e-scooter rides, including provoked fall events.This dataset has been employed for training a set of ML algorithms,which have been integrated within an embedded OBU prototype following the TinyML paradigm for enabling on-device inference.Once a fall event has been detected,an early alert may be triggered by means of reliable and long-range communication interfaces such as those making use of Low Power Wide Area Network(LPWAN)technologies[9].Therefore,the main contributions of this paper are the following:

    ·A public dataset for automatic detection of fall events in personal mobility devices.

    ·A discussion focused on the TinyML paradigm and its application to the sustainable mobility ecosystem.

    ·An ML-based fall detection system for personal mobility devices.

    ·A detailed study in terms of accuracy and footprint of a series of TinyML mechanisms integrated within a real OBU.

    The contents of the paper are distributed as follows.Section 2 frames the work in the literature.Section 3 discusses the applicability and benefits of TinyML for providing smart services in the field of personal mobility vehicles.Section 4 describes the reference OBU used to gather data to generate ML models and later perform the evaluation.The TinyML-based fall detection system is explained in Section 5.The attained results are presented and discussed in Section 6.Finally,the paper is concluded in Section 7,summarizing the main findings and indicating future directions.

    2 Background

    The application of optimized ML-based solutions within the scope of eco-efficient mobility has been scarcely studied in the related literature.Given that we present a real implementation of an ondevice fall detection system,in the following we explore the fields of human activity and fall detection and the use of embedded ML mechanisms for classification tasks.

    2.1 Detection of Human Activity and Falls

    Human Activity Recognition (HAR) and fall detection have been hot research topics since the advent of the IoT together with the integration of different types of low-cost sensors within smartphones and other everyday gadgets.Work in [10], focused on elderly people monitoring, proposed using floor vibrations and sound to detect falls.The presented system uses signal processing and a pattern recognition algorithm to discriminate between falls and other events, gathering training data by mimicking falls using human dolls.In turn, more recent works made use of smart-phone’s embedded sensors.For example,authors of[11]employed a simple algorithm employing acceleration measurements for detecting falls as well as long lie periods in order to reduce false positives.Work in [12] presented a more sophisticated detection method based on different ML models employing accelerometer data as well, hence demonstrating the usefulness of these low-cost and widespread smart-phone sensors for HAR and fall detection applications.Another approach using two independent devices for fall detection can be found in [13].In this case, authors made use of a smart-phone and smart-watch for simultaneously tracking the status of the user by means of their embedded accelerometers and gyroscopes.The results showed a clear decrease on false positives while maintaining the effectiveness of the detection decisions.Other works in this field focused on video and image processing can be found in the survey work presented in[14].

    In[15],authors demonstrated that model personalization,for example by using similarity between people related to physical aspects,notably improves the accuracy of HAR systems.Thus,a thorough study of the available data and user features is crucial for obtaining satisfactory results.Hence, the dataset is a key piece aiming at recognizing different patterns and situations.In the use-case under study in this work,obtaining a realistic dataset may be dangerous or hardly reproducible under laboratory conditions.Although different public and comprehensive HAR datasets can be found,e.g.,[16],they focus on daily-live movement traces and not on riding activities.To tackle this issue,we have generated our own dataset collected from real rides (including falls) in order to produce different optimized and well-adjusted models for detecting falls in the specific use-case of electric personal mobility.As explained later,this dataset is publicly available for contributing to the HAR,ITS,and IoT research communities.

    2.2 Embedded ML

    ML is an area of high importance nowadays for the IoT ecosystem, given that many consumer applications can benefit from built-in advanced computing capabilities as they permit processing sensor data directly on the end-device.This leads to energy saving in communications and a reduction in application latency[17].The integration of ML algorithms within IoT constrained elements implies extending the edge computing concept to the end-device/user itself,leading to the fog and mist computing paradigms[18].However,ML-based mechanisms usually imply intensive computing tasks and,so far,they have required their execution on high-performance computers or clusters,sometimes aided by high-throughput GPU computing or application-specific integrated circuits(ASICs).As detailed in the following section, the TinyML paradigm permits to embed intelligence (ML mechanisms) in resource-constrained end-devices, hence, opening a plethora of opportunities for developing truly intelligent IoT units.Given the novelty of this concept[8],there is not any related literature addressing the application of TinyML models to the HAR or fall detection problems.However, some efforts proposing ML-based solutions to tackle these issues have been done,as these mechanisms have shown better performance than simpler threshold-based algorithms[19].

    Work in [20] evaluated a set of ML classification algorithms, namely, k-Nearest Neighbors (k-NN, Naive Bayes, Artificial Neural Network (ANN), Support Vector Machine (SVM), and Least Squares Method (LSM), for the task of fall detection.Good accuracy over 85% was attained when distinguishing between daily activities and fall events,but authors did not present any results in terms of computing performance or memory footprint.Another approach can be found in [21], where an accident detection and classification system was developed by using data gathered from smart-phone’s built-in sensors.Here a set of ML algorithms were implemented and evaluated within the smart-phone with good accuracy results,leveraging the superior processing power of these devices in comparison with common IoT units.Work in [22] presented a fuzzy inference system that, by gathering data collected from multiple sensors in an IoT environment,is able to infer that a fall has occurred.Authors emphasized the reliability of the solution given that multiple information sources are employed.Given the great range of available options to tackle the fall detection problem with ML-based solutions,works in[23,24]deeply explored the performance of different algorithms’and models’configurations in terms of accuracy but obviating a thorough evaluation of the computation and memory impact of the considered models.

    Although additional works addressing ML-based automated fall and accident detection can be found in the literature [25], none of them permits to obtain an optimized mechanism that could be executed by an independent resource-constrained processing device.This is a critical aspect for our studied use-case,as the intelligent unit should be an embeddable solution for light and small vehicles such as electric scooters or segways not requiring external computation support.Thus, we propose to adopt the TinyML paradigm to build an intelligent and independent unit, capable of collecting the necessary data as well as conducting on-device ML inference for fall event detection.Besides,we also provide a comprehensive comparison of different TinyML models’footprint, performance and accuracy,extracted from a real proof of concept framed within the personal urban mobility ecosystem.

    3 Smart Personal Mobility Vehicles and Machine Learning

    The adoption of electric mobility solutions in our daily lives is a reality thanks to the great range of benefits brought by these devices.On the one hand, eco-efficient vehicles can replace traditional transportation means in cities due to their reduced carbon footprint as well as their mobility flexibility in crowded scenarios.Recent access restrictions to the city center in Europe, such as the cases of London,Paris or Madrid,for regular combustion vehicles,lead to the adoption of sustainable vehicles that complement regular bicycles.In addition, devices such as electric scooters and bikes, segways,etc.are also being employed for outdoor recreational activities.As it has been done for regular vehicles,the attachment of OBUs to personal mobility vehicles will allow their integration within the C-ITS ecosystem.This will boost the development of novel services for this vehicle segment.Besides,embedding intelligence in these processing units will make them evolve from simply connected elements to truly smart“things”,hence improving the eco-mobility experience and riders safety.To this end,the TinyML paradigm will fill the gap that currently hinders the implementation of ML-based mechanisms within these constrained computing platforms.

    3.1 TinyML

    The recent emergence of the TinyML paradigm [8] has enabled the integration of powerful ML schemes within constrained processing devices such as those powered by microcontrollers [26].Although previous advances have been achieved in this field, the attained solutions are usually adhoc implementations addressing specific use-cases for a given processing board,which clearly limits the generalization capability of this approach.As depicted in Fig.2,TinyML allows to convert ML models generated by well-known toolkits,e.g.,TensorFlow,Keras,PyTorch,Scikit-learn,etc.,in order to make them executable in constrained platforms.Once the pre-trained model is deployed on the resource-limited unit,the ML algorithm can be executed to cover the target use case.

    From a technical perspective,embedding intelligence in end-devices will bring interesting features to IoT deployments.Firstly,current constrained devices notably depend on their connection with the cloud in order to send collected data and/or receive commands.As communication activities can be more energy demanding that moderate processing tasks in these nodes,performing local computation for making smart decisions can be more adequate than relying on remote processing.This aspect is not negligible given that a great part of IoT devices is powered by batteries.Besides,reducing raw data transmissions further diminish cybersecurity risks and reduce computing impact of cryptography operations with their associated energy consumption.

    Figure 2:Machine learning conversion workflow to pass from a general-purpose model to a TinyML model for constrained devices

    Some TinyML frameworks are already publicly available in the market.The interest of technological companies in this paradigm is clear,as Google,Microsoft,or ARM have released different toolkits.Google’s TensorFlow Lite and Microsoft’s ELL permit to transform Deep Learning (DL) models from TensorFlow and the Microsoft Cognitive Toolkit(CNTK),respectively,to C++code that can be compiled for 32-bit processors such as the ARM Cortex-M series.Devoted to this processor as well,ARM has released the Cortex Microcontroller Software Interface Standard-NN(CMSIS-NN)framework,which is interoperable with models from TensorFlow and Caffe.Besides,ARM has also developed the ARM-NN library,which widens the range of supported boards towards ML-oriented Ethos neural processing units such as ARM Mali GPUs and ARM Cortex-A CPUs.

    As shown in Tab.1, others interesting TinyML frameworks are also available for enabling interoperability with additional ML libraries and constrained boards.For example,emlearn1https://github.com/emlearn/emlearnpermits to port a set of ML models, namely, Multi-Layer Perceptron (MLP), Decision Tree (DT), Random Forest (RF), or Gaussian Naive Bayes (GNB) generated by Scikit-learn or Keras, for making them runnable in 8-bit platforms such as the widely-adopted Arduino Uno.Also compatible with this development board, the MicroMLGen framework ports Support Vector Machine (SVM) models produced by Scikit-learn.As a result,a variety of models generated by robust ML libraries are now at our disposal to be integrated within constrained processing units.This will foster the development of a vibrant TinyML community,which will grow until forming a rich and productive ecosystem.

    In the light of the previous discussion,the TinyML paradigm,driven by the available frameworks,will permit to turn current end-devices on smart elements, therefore evolving the traditional IoT landscape to the novel Internet of Intelligent Things (IoIT).Thanks to this revolution, a new wave of applications will be designed and implemented.In the following we explore the range of potential services to be developed in the area of sustainable personal mobility.

    Table 1:TinyML frameworks reviewed

    3.2 Enabled Services

    Traffic safety will be one of the most benefited fields of application with the integration of intelligent schemes in constrained OBUs.One of the most critical services within this scope is automatic accident detection, which is deeply investigated in the following sections.A quick detection of falls is crucial for a fast reaction in order to notify the event and the accident location to the closest emergency service.Although some efforts have been devoted to the development of AFD systems for pedestrians[27]and bike riders[28],the new range of electric personal mobility vehicles has not been investigated so far, as studies such as the one in [29] reveal.Besides this application, the intelligent interconnection of personal mobility devices to the C-ITS ecosystem will boost the development of vehicle-to-vehicle (V2 V) collision avoidance systems such as those proposed for two-wheelers[30], and the implementation of advanced interaction schemes of personal vehicles with the traffic infrastructure[31].

    Another area of application is on-board intelligent vehicle diagnosis [32].By introducing ML mechanisms that could inspect different vehicle parameters,it will be able to perform self-inspections for prompt fault detection.This feature will permit to prevent mechanical failures and auto-adjust the device in case of software-solvable issues.In this line, intelligent device status tracking will be also enabled;hence,vehicle telemetry will be accessible for analysis from users’devices such as smartphones or tablets.Offline route planning is another interesting service for providing riders with tailored routes according to learned habits,cached digital maps,and special characteristics of personal mobility vehicles.Thus, this application could indicate, for example, green tracks or alternative safer routes that avoid dense traffic situations.Once parked, intelligent services related to device security will be powered by ML mechanisms.A great number of these vehicles are stolen every day, therefore the implementation of intelligent e-locks relaying on user’s biometric data will be of great interest.Traditional mechanical locks will be replaced by smart systems that may block the device upon manipulation.Besides, real-time alerts can be triggered to the vehicle’s owner if a steal-attempt is detected.

    Advanced real-time scenario monitoring will be also useful for warning riders against unhealthy pollution conditions.Continuous city sensing will help smart city platforms to collect pre-processed data from streets.Hence,a plethora of recommendation services will be enabled for improving citizens well-being,at the same time that communication resources and battery are saved.Finally,aligned with the recreational perspective of these vehicles,smart infotainment applications will be also powered by the use of ML,e.g.,music recommendations,audio guides and books,or contextual advice.

    4 Reference on-Board Unit

    In previous work carried out by authors, a functional and efficient OBU for personal mobility vehicles was developed [1].This OBU has been considered in this work to initially get training data and later validate and test the proposed TinyML-based AFD system.Its main features are depicted in Fig.3.It is a low-power consumption device provided with an off-the-shelf single board computer(SBC)equipped with Low-Power Wide-Area Network(LPWAN)capabilities.Concretely,Long Range Wide-Area Network(LoRaWAN)and Narrow-Band IoT(NB-IoT)technologies are used for the sake of minimizing energy consumption and obtaining long-range and reliable communication links.Using either LoRaWAN or NB-IoT,the OBU can access cloud services through the Internet,which can feed web applications or smart-phone apps,for instance,offering Cooperative,Connected and Automated Mobility(CCAM)advances.

    Figure 3:Architecture and prototype of the OBU used for e-scooters.(a) OBU architecture design with sensors, processing unit, basic I/O and communication transceivers (b) Functional prototype developed for validation and testing purposes

    The prototype implemented (right side of Fig.3) includes a Raspberry Pi Zero W SBC, with a System on Chip(SoC)powered by a 1 GHz ARM11 CPU coupled with 512 MB of SDRAM.The top white part is a communication board integrating a GPS receiver, a LoRaWAN transceiver (Murata CMWX1ZZABZ-093)and a NB-IoT modem(Quectel BC95).A 5 Ah lithium battery(RS PRO PBA5200)is used to reach running times over 48 h reporting data continuously.The unit case has been manufactured using a 3D printer and its dimensions are 15×7×5 cm.Among the various sensors installed in the unit,the most important one in the studied use case is the 9 degrees of freedom(DOF)inertial unit ST LSM9DS1.It includes a three-axis accelerometer, a three-axis gyroscope and threeaxis magnetometer.Hence,it provides angular velocity,acceleration and heading,which are the input of the developed AFD system.

    Given the wide adoption of Raspberry solutions in embedded electronics and,particularly,when an agile prototyping is necessary in research tasks in Academia,we have opted for a Pi Zero platform.It provides good performances under low power consumption,good I/O capabilities and the market is plenty of sensory solutions for Raspberry.Additionally,it is able to execute an embedded operating system,which makes it convenient to deploy several applications and,in our case,different TinyML models.Although microcontroller-based designs could be developed to run services and further reduce energy consumption,their flexibility for software upgrading and multi-task purposes would be much more limited.

    5 TinyML-Based Fall Detection

    This section presents a specific TinyML demonstrator in which we address the AFD issue in personal mobility vehicles by developing different ML models that are adapted to constrained devices.In our particular evaluation, we consider the OBU described above to perform an experimental evaluation of the solution.

    5.1 Dataset

    The first step to develop the proposed AFD system is creating a rich and large dataset for training and testing the generated models.To this end, a sampling campaign has been carried out with the reference OBU recording about three hours of urban e-scooter riding and 50 provoked falls in a controlled environment, including left and right overturns and front collisions at moderate speed.Fig.4 includes real photos taken during the trials, for the cases of left falls and forward collision.It is indicated the initial position of the scooter,the direction of the movement and its final status after the trial.For the forward collision trials,the scooter was manually propelled at 10 km/h to collide with a mat.The scooter used was an Outsider E-Volution 8,5 Phoenix, from Cecotec.The user was the same along the regular sampling process,covering different road types and speed/turning conditions.All trials have been carried out in the city center of Cartagena(Spain),driving on regular roads and cycle lanes.

    Data were recorded from the inertial sensor unit integrated in the OBU in order to obtain threeaxis acceleration(axi,ayi,azi)and angular velocity(θxi,θyi,θzi)at a sampling rate of 10 Hz.Considering these features,the obtained raw data vector presents six elements.We have considered the three axis components given that falls in personal mobility vehicles may happen in any direction[5].After the sampling campaign,the obtained dataset was examined and missing or corrupted samples were fixed.For the sake of experiment repeatability and reproducibility as well as for enabling further related studies,the resulting dataset has been publicly released8https://gitlab.com/scooter-project/ml-dataset.

    Besides, aiming at evaluating the performance of the considered models with different input feature vectors,we also compute the corresponding signal magnitude vector(SMV)of both monitored parameters as indicated in Eqs.(1)and(2).

    Figure 4:Procedure followed to obtain dynamics data from falls with the e-scooter.(a) Fall to left,initial status(b)Fall to left,final status(c)Frontal collision,initial status(d)Frontal collision,final status

    Given that captured measurements may suffer isolated sudden changes due to shocks caused by speed bumps or road imperfections,we finally aggregate every second,i.e.,10 samples,the captured raw data and the corresponding computed SMVs and calculate their variance.Thus,the final feature vectors that are passed to the ML algorithms are the following:(σ2(axi),σ2(ayi),σ2(azi),σ2(θxi),σ2(θyi),σ2(θzi))or(σ2(ai),σ2(θtilt i)),depending on the selected size of the input vector,namely,six or two elements,respectively.With these input vectors,the decision algorithm outputs if a fall detected or not.

    5.2 TinyML Models

    In order to evaluate the adequacy of different ML algorithms for the AFD task,we have followed the workflow of Fig.2 previously described.We have made use of the well-known Scikit-learn toolkit9https://scikit-learn.orgto train and generate a variety of ML models,namely,Support Vector Machine(SVM),Multi-Layer Perceptron (MLP), Decision Tree (DT), Gaussian Naive Bayes (GNB), and Random Forest (RF).Given the processing and memory constraints of the OBU presented above, these models have been optimized by using different TinyML frameworks.Concretely,MicroMLGen has been used to adapt the SVM models, while the versatile emlearn library has been employed to port the rest of models generated by Scikit-learn.The dataset generated as described above has been split using stratification into three subsets:70% for training and 20% for validation in Scikit-learn, and 10% for testing in the OBU.As performance metrics,we have chosen the accuracy level of the generated models,their memory footprint,the required processing time and energy consumed once deployed in the OBU.

    5.3 Working Flow

    The overall operation of the system using any of the adopted TinyML models is depicted in the flow chart included in Fig.5.After beginning the ride,the software starts collecting data from inertial sensors continuously at a sampling rate of 10 Hz,as aforementioned.Once the samples collected fill the reference window size(10 samples in our experiments),data are aggregated to be processed by the TinyML model.This module is in charge of detecting a fall.If this is not the case,a new data collection cycle is started;whereas if a fall is detected,an alarm message is transmitted using the communication module of the OBU towards a given alert management system in the cloud.Under normal operation with no falls,the system would be running and continuously gathering data and executing the TinyMLbased algorithm in a continuous cycle.

    Figure 5:Flow chart of the operation of the TinyML-based fall-detection solution

    6 Results

    6.1 Overall Comparison of TinyML Models

    The obtained results in terms of the identified figures of merit are shown in Tab.2.Different configurations have been tested for each of the algorithms although we only present the most adequate ones for the sake of clarity.As can be seen in the last column, the accuracy achieved by all the considered algorithms is similar and their performance is notably successful(around 0,96).For most of the cases,it can be seen that the computing time is noticeably low,in the order of tens of microseconds,except for the case of the SVM, which is the slowest algorithm.This is explained by its greater complexity in comparison with the other algorithms,also evidenced by the size of the generated binary file due to the kernels are defined as float random variables.SVM also provides the lowest accuracy levels when using the Sigmoid kernel,so its usage is not recommended.

    It is noticeable the high accuracy offered by RF(0,97)when using five estimators and six input features, and requiring an execution time of only 11±14 μs.Regarding MLP models, the different evaluated network configurations provide similar accuracy levels, although the processing time and the required RAM increase with the number of neurons forming the MLP architecture because,differently from the SVM implementation,these values are stored in volatile memory.Finally,GNB presents greater accuracy than DT but longer processing times with both two or six input features.As mentioned above,the accuracy detecting falls of all of the evaluated algorithms is remarkable.

    Aiming at providing insights regarding the validity of the results,the F1-score attained for each model is also included in Tab.2.Observe that the obtained F1 scores are aligned with the accuracy values,hence indicating the good performance of the models avoiding false positives or negatives.This is of prominent importance in the application under study given that a non-detected fall may lead to serious issues for the injured driver.

    Table 2:Execution time,on-device footprint,accuracy,and F1-score of the evaluated TinyML models

    Table 2:Continued

    Regarding the footprint of the implementations on the target processing board, i.e., Raspberry Pi Zero in our case,it can be seen the low impact on both RAM and flash memories,hence making feasible to run all the tested ML classifiers with real time sensor data on such a constrained OBU.The moderate RAM required permits the generated models to coexist with other simultaneous functions running in the unit.Besides,the limited size of the binary files built for the target platform makes them suitable to be embedded in the considered unit as well as other more resource-limited boards, e.g.,microcontroller-based architectures,without the risk of consuming a significant part of the available flash memory.

    Given the relevance of power consumption in the target environment, indicative values are also given in Tab.2.They have been obtained taking as reference the consumption of the unit under 100%CPU usage together with sensor reads at a rate of 1 Hz.The value obtained after a set of consecutive execution of ML algorithms was 294,17±14,49 mA, not observing differences when executing the different models.Hence,considering the execution time of each model,it has been computed the power consumption(in nanoamperes per hour-nAh).Results in Tab.2 showcase the limited energy needed by most of the cases except for SVM,due to the aforementioned algorithm complexity.

    Considering the discussion above,GNB,DT,or RF algorithms would be the most recommended options to implement the embedded fall detection mechanism, as they present very high accuracy with a reduced impact in terms of processing latency,memory requirements,and power consumption.Additionally,SVM and MLP produce slower and heavier models as compared to GNB,DT,and RF,without providing a better performance in terms of accuracy.

    6.2 Impact of Number of Input Features

    Fig.6 depicts the models’footprint and performance variation depending on the number of inputs(two or six).Concretely,we show the relative difference of the evaluated metrics when the classifiers receive feature vectors of two or six elements,taking the former as the baseline.We find that models with six inputs present an almost negligible accuracy improvement, however, this is achieved at the expense of increasing the algorithms’running time.It can be observed that in every case,models with six input features require much more processing time than their analogous models with two inputs.This fact is especially remarkable for GNB, MLP, and SVM, with more than 20% increase in the algorithm’s computation time.

    Flash memory usage especially increases with six input features in SVM.This confirms the previous discussion regarding memory requirements of support vectors.Models based on GNB and MLP do not suffer a noticeable variation in their memory footprint depending on the number of input features.In fact, all MLP models have an identical executable file size and similar RAM footprints.This is due to the binary file padding and how the code works in this implementation (by iterating through a list of neurons), so the binary files only differ in the quantity of biases and weights that describe the network, which are really small, e.g., less than one kB for a 10+10+5 neurons model and less than 100 bytes for the two neurons model.These small differences are further reduced by the executable file padding,used to align different memory blocks in accordance with the target computer architecture.Additionally,optimizing memory accesses and code makes RAM usage to be almost the same on the small networks we have used.

    Figure 6:Models’footprint and performance variation according to the number of input features(6 vs.2),taking as a reference the results obtained with two inputs

    The cases of DT and RF are worth further analysis.In both of them a decrease in the required flash memory is obtained when increasing the number of inputs,especially in the case of RF.This behavior can be explained by the fact that the six inputs models are able to find a solution with lower tree or forest complexity in comparison with models receiving feature vectors of two elements.Regarding the impact on RAM, a similar behavior is attained for RF, while a small increase in volatile memory demands is detected for DT.In the light on these and previous outcomes, together with the high accuracy achieved,we recommend the use of RF with five estimators to build a precise and efficient ML-based AFD system.

    Overall,from the obtained results it can be extracted that the TinyML paradigm opens a plethora of opportunities for integrating ML-based mechanisms on end-devices with resource limitations.Besides,the low footprint detected enables interesting features such as over the air(OTA)updates of ML models,thus offloading the heavy task of training and upgrading the ML system onto a cloud or Multi-Access Edge Computing(MEC)node;and even further distribute training tasks by exploiting federated learning capabilities while reducing communication requirements.

    7 Conclusion

    The traditional traffic landscape consisting of fuel-powered vehicles is being enriched with a plethora of eco-efficient devices that provide users with more sustainable mobility options.However,given the novelty of personal mobility vehicles such as e-scooters,segways,or electric wheels,among others, there are still different aspects related to their tracking, connectivity, and safety that should be promptly addressed.In this work, we have explored the improvement of driver safety under falls or accidents in this segment by proposing a fall detection system powered by ML.Concretely, the state-of-the-art TinyML paradigm is exploited to develop an ML-based efficient and accurate fall detection system.To this end,a riding/fall dataset has been produced(and publicly released)to build a range of TinyML models considering different algorithms,namely,RF,DT,SVM,GNB and MLP,considering a range of configurations.These data have been obtained from real trials with an electric scooter in the city of Cartagena (Spain), and they have been used for training, validation and tests.Hence, these models have been deployed on a lightweight and connected OBU prototype in order to analyze the attained accuracy when detecting a fall, as well as evaluating the models’footprint on the host processing unit.The results validate the good operation of the proposal,as the attained classifiers’performance permits to quickly and accurately detect fall events,while the limited memory requirements and energy consumption of the produced models ensure their integration in a plethora of constrained processing units.As future work,we plan to generate the TinyML models for even more limited computing boards as well as explore innovative ML paradigms such as federated learning to speed-up algorithm generation and sharing.Besides,further analysis on the sensed data may permit to extract more sophisticated outcomes such as accident type or severity.

    Funding Statement:This work has been supported by the Spanish Ministry of Science,Innovation and Universities, under the Ramon y Cajal Program (ref.RYC-2017-23823) and the projects ONOFRE 3(ref.PID2020-112675RB)and Go2Edge(ref.RED2018-102585-T);by the European Commission,under the 5G-MOBIX(ref.825496)project;and by the Spanish Ministry for the Ecological Transition and the Demographic Challenge, under the MECANO project (ref.PGE-MOVES-SING-2019-000104).Ramon Sanchez-Iborra would like to thank Prof.Sanchez-Mu?oz for his patience and valuable comments during the preparation of this paper.

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

    av在线老鸭窝| 国产不卡一卡二| 日本黄色视频三级网站网址| 两性午夜刺激爽爽歪歪视频在线观看| 国产精品一区二区性色av| 美女免费视频网站| 久久久精品欧美日韩精品| videossex国产| 干丝袜人妻中文字幕| 尤物成人国产欧美一区二区三区| 在线观看66精品国产| 色噜噜av男人的天堂激情| 舔av片在线| 99在线视频只有这里精品首页| 99久久成人亚洲精品观看| 久久精品国产亚洲av涩爱 | 内地一区二区视频在线| 真实男女啪啪啪动态图| 内地一区二区视频在线| 直男gayav资源| 国产在视频线在精品| 日本在线视频免费播放| 尤物成人国产欧美一区二区三区| 女的被弄到高潮叫床怎么办 | 人人妻,人人澡人人爽秒播| 波多野结衣高清无吗| 亚洲精品一区av在线观看| 日韩欧美国产一区二区入口| 亚洲av中文av极速乱 | 最新在线观看一区二区三区| 国产伦一二天堂av在线观看| 黄片wwwwww| 欧美一区二区亚洲| xxxwww97欧美| 波多野结衣巨乳人妻| 免费在线观看影片大全网站| 久久久久久国产a免费观看| 亚洲av免费高清在线观看| 日韩欧美国产在线观看| 免费av不卡在线播放| 日韩中字成人| 欧美成人a在线观看| 如何舔出高潮| 国产精品亚洲美女久久久| 国产精品久久久久久亚洲av鲁大| 乱人视频在线观看| 91麻豆av在线| 不卡视频在线观看欧美| 午夜a级毛片| 日韩av在线大香蕉| 亚洲国产精品sss在线观看| 国产av不卡久久| 成年人黄色毛片网站| 久久人妻av系列| 国产视频一区二区在线看| 亚洲综合色惰| 久久久久久久精品吃奶| 特大巨黑吊av在线直播| 婷婷亚洲欧美| www.色视频.com| 夜夜爽天天搞| av黄色大香蕉| 色综合色国产| 色吧在线观看| 成人性生交大片免费视频hd| 在线国产一区二区在线| eeuss影院久久| 91在线精品国自产拍蜜月| 欧美中文日本在线观看视频| 少妇高潮的动态图| 精品乱码久久久久久99久播| 精品一区二区三区视频在线| 日韩欧美在线乱码| 美女免费视频网站| 欧美在线一区亚洲| 午夜久久久久精精品| 国产高清不卡午夜福利| 少妇的逼好多水| 久9热在线精品视频| 国产精品野战在线观看| 欧美日本亚洲视频在线播放| 女人十人毛片免费观看3o分钟| 蜜桃亚洲精品一区二区三区| 看黄色毛片网站| 波多野结衣高清无吗| 极品教师在线免费播放| 男人狂女人下面高潮的视频| 国产白丝娇喘喷水9色精品| avwww免费| 少妇猛男粗大的猛烈进出视频 | 非洲黑人性xxxx精品又粗又长| 国产在线男女| 美女大奶头视频| 女人十人毛片免费观看3o分钟| 内地一区二区视频在线| av.在线天堂| 精品午夜福利在线看| 亚洲四区av| 99热精品在线国产| 亚洲欧美日韩无卡精品| 久久天躁狠狠躁夜夜2o2o| 午夜久久久久精精品| 999久久久精品免费观看国产| 亚洲自偷自拍三级| 最新在线观看一区二区三区| 国产毛片a区久久久久| 黄色女人牲交| 亚洲最大成人手机在线| 欧美高清成人免费视频www| 最好的美女福利视频网| 欧美成人a在线观看| 国产综合懂色| 国产一区二区三区视频了| 亚洲国产欧美人成| 人妻久久中文字幕网| 久久久久国产精品人妻aⅴ院| 男人舔女人下体高潮全视频| 一边摸一边抽搐一进一小说| 日韩一区二区视频免费看| 又粗又爽又猛毛片免费看| 黄色视频,在线免费观看| 极品教师在线视频| 成人特级黄色片久久久久久久| 国产亚洲91精品色在线| 日本撒尿小便嘘嘘汇集6| 熟妇人妻久久中文字幕3abv| 久久久久久伊人网av| 精品欧美国产一区二区三| 天堂av国产一区二区熟女人妻| 国产三级中文精品| 男女视频在线观看网站免费| 成人欧美大片| 亚洲精品成人久久久久久| 亚洲精华国产精华精| 精品一区二区三区av网在线观看| 无遮挡黄片免费观看| 色播亚洲综合网| 精品久久久久久久久久久久久| 身体一侧抽搐| 黄色视频,在线免费观看| 女的被弄到高潮叫床怎么办 | 国产成人av教育| 欧美绝顶高潮抽搐喷水| 香蕉av资源在线| 亚洲四区av| 久久久久久久亚洲中文字幕| 成人美女网站在线观看视频| 欧美日韩国产亚洲二区| 欧美激情国产日韩精品一区| 久久热精品热| 日韩精品有码人妻一区| a级毛片免费高清观看在线播放| 亚洲av熟女| 国产美女午夜福利| 国产色爽女视频免费观看| 欧美激情国产日韩精品一区| 我的老师免费观看完整版| 亚洲国产精品sss在线观看| 亚洲国产欧美人成| 午夜精品在线福利| 黄色丝袜av网址大全| 国产亚洲精品综合一区在线观看| 久久久久久久久大av| 欧美色视频一区免费| 日韩在线高清观看一区二区三区 | 热99re8久久精品国产| 看免费成人av毛片| АⅤ资源中文在线天堂| 99热这里只有精品一区| 一进一出抽搐gif免费好疼| 少妇裸体淫交视频免费看高清| 久久精品国产鲁丝片午夜精品 | 老司机午夜福利在线观看视频| 女同久久另类99精品国产91| 听说在线观看完整版免费高清| 亚洲一区高清亚洲精品| av天堂中文字幕网| 最近视频中文字幕2019在线8| 欧美另类亚洲清纯唯美| 波多野结衣高清无吗| 亚洲国产精品成人综合色| 国产 一区 欧美 日韩| 22中文网久久字幕| 色精品久久人妻99蜜桃| 少妇的逼水好多| 精品久久久久久久久av| 国产精品永久免费网站| 午夜福利在线在线| 春色校园在线视频观看| 久久久久国产精品人妻aⅴ院| 免费不卡的大黄色大毛片视频在线观看 | 校园人妻丝袜中文字幕| 欧美一级a爱片免费观看看| 人人妻人人澡欧美一区二区| 久久国内精品自在自线图片| 亚洲最大成人中文| 国产一区二区激情短视频| 如何舔出高潮| 日韩欧美国产一区二区入口| 毛片一级片免费看久久久久 | 丰满人妻一区二区三区视频av| 成人性生交大片免费视频hd| 成年版毛片免费区| 国内毛片毛片毛片毛片毛片| 精品久久久噜噜| 亚洲乱码一区二区免费版| 亚洲五月天丁香| 国产在线男女| 深夜精品福利| 一卡2卡三卡四卡精品乱码亚洲| 免费观看在线日韩| 男女视频在线观看网站免费| 国产成人一区二区在线| 国产男人的电影天堂91| 国产精品永久免费网站| 久久久午夜欧美精品| 日本 欧美在线| 窝窝影院91人妻| 黄片wwwwww| 成人av在线播放网站| 日本熟妇午夜| a级毛片a级免费在线| 99热这里只有精品一区| 村上凉子中文字幕在线| 亚洲四区av| 亚洲欧美日韩东京热| 亚洲18禁久久av| 欧美丝袜亚洲另类 | 国产真实乱freesex| 国内精品一区二区在线观看| 国产精品自产拍在线观看55亚洲| 国产高清三级在线| avwww免费| 三级毛片av免费| 白带黄色成豆腐渣| 天美传媒精品一区二区| 午夜日韩欧美国产| 变态另类丝袜制服| 午夜久久久久精精品| 久久人人精品亚洲av| 在现免费观看毛片| 波野结衣二区三区在线| 欧美成人性av电影在线观看| 国产精品人妻久久久久久| 91麻豆精品激情在线观看国产| 成年免费大片在线观看| 国内少妇人妻偷人精品xxx网站| 99久久无色码亚洲精品果冻| 国产探花极品一区二区| 色哟哟·www| 亚洲第一电影网av| 91av网一区二区| 能在线免费观看的黄片| 国产精品1区2区在线观看.| 噜噜噜噜噜久久久久久91| 精品人妻一区二区三区麻豆 | 他把我摸到了高潮在线观看| 免费看av在线观看网站| 久久欧美精品欧美久久欧美| 性色avwww在线观看| 内射极品少妇av片p| 亚洲熟妇中文字幕五十中出| 最近最新中文字幕大全电影3| 亚洲欧美日韩无卡精品| 在线看三级毛片| 亚洲成人免费电影在线观看| 最近最新免费中文字幕在线| 国产精品一区二区三区四区免费观看 | 国产精品自产拍在线观看55亚洲| 日日摸夜夜添夜夜添小说| 99热网站在线观看| 久久精品国产亚洲av天美| 国产男靠女视频免费网站| 99久久精品国产国产毛片| 日韩国内少妇激情av| 欧美一区二区国产精品久久精品| 淫妇啪啪啪对白视频| 日本 av在线| 国产日本99.免费观看| 日本黄色视频三级网站网址| 久久久久国内视频| av天堂中文字幕网| 啪啪无遮挡十八禁网站| 亚洲欧美日韩东京热| 淫秽高清视频在线观看| 欧美一级a爱片免费观看看| 热99在线观看视频| 性欧美人与动物交配| av在线蜜桃| 97热精品久久久久久| 日本熟妇午夜| 精品免费久久久久久久清纯| 欧美激情国产日韩精品一区| 国产一区二区在线av高清观看| 国产精品久久久久久久久免| 婷婷精品国产亚洲av| 国产三级在线视频| 九九热线精品视视频播放| 午夜激情欧美在线| 联通29元200g的流量卡| 日韩精品青青久久久久久| 黄色视频,在线免费观看| 桃色一区二区三区在线观看| 亚洲无线观看免费| 色在线成人网| 免费搜索国产男女视频| 欧美中文日本在线观看视频| 亚洲精华国产精华液的使用体验 | 99热精品在线国产| 午夜福利在线在线| 一区二区三区四区激情视频 | or卡值多少钱| 一卡2卡三卡四卡精品乱码亚洲| 男女啪啪激烈高潮av片| 欧美xxxx黑人xx丫x性爽| 亚洲国产日韩欧美精品在线观看| 久久婷婷人人爽人人干人人爱| 国产成人aa在线观看| 国产成人一区二区在线| 国产精品精品国产色婷婷| av专区在线播放| 国产精品人妻久久久影院| 国产久久久一区二区三区| www日本黄色视频网| 99热这里只有是精品在线观看| 22中文网久久字幕| 亚洲人与动物交配视频| 亚洲av中文av极速乱 | www日本黄色视频网| 最新在线观看一区二区三区| 亚洲乱码一区二区免费版| 美女 人体艺术 gogo| 三级国产精品欧美在线观看| 麻豆成人午夜福利视频| 999久久久精品免费观看国产| 一区二区三区免费毛片| 一个人看视频在线观看www免费| 听说在线观看完整版免费高清| 真人做人爱边吃奶动态| a级毛片免费高清观看在线播放| 联通29元200g的流量卡| 亚洲精品粉嫩美女一区| 日日夜夜操网爽| 亚洲自偷自拍三级| 联通29元200g的流量卡| 国产精品三级大全| 国产在线男女| 最近视频中文字幕2019在线8| 中文字幕高清在线视频| 最近视频中文字幕2019在线8| 午夜视频国产福利| 国国产精品蜜臀av免费| 在线国产一区二区在线| 欧美性猛交╳xxx乱大交人| 中文字幕av在线有码专区| 久久精品影院6| 国产精品一区二区免费欧美| 国产男人的电影天堂91| av专区在线播放| 国产成人福利小说| 国产激情偷乱视频一区二区| 国产精品日韩av在线免费观看| 午夜精品一区二区三区免费看| 搡女人真爽免费视频火全软件 | 久久国产精品人妻蜜桃| 国产精品一区二区三区四区久久| 欧美潮喷喷水| a级毛片a级免费在线| 色综合婷婷激情| 露出奶头的视频| 午夜免费男女啪啪视频观看 | 97超级碰碰碰精品色视频在线观看| 国产精品久久久久久亚洲av鲁大| 日本黄色片子视频| 一级a爱片免费观看的视频| 国产精品伦人一区二区| avwww免费| 亚洲国产精品久久男人天堂| 99久久中文字幕三级久久日本| 亚洲天堂国产精品一区在线| 波多野结衣高清作品| 国产高清三级在线| 亚洲内射少妇av| 国产激情偷乱视频一区二区| 欧美性猛交╳xxx乱大交人| 久久久久免费精品人妻一区二区| 18禁在线播放成人免费| 欧美日韩瑟瑟在线播放| 少妇猛男粗大的猛烈进出视频 | 不卡视频在线观看欧美| 99久久无色码亚洲精品果冻| 久久香蕉精品热| 国产亚洲精品久久久com| 欧美日韩黄片免| 精品一区二区三区视频在线| 国产三级中文精品| 欧美中文日本在线观看视频| 久久精品夜夜夜夜夜久久蜜豆| 欧美日韩亚洲国产一区二区在线观看| www.色视频.com| 国产又黄又爽又无遮挡在线| АⅤ资源中文在线天堂| 国产精品人妻久久久久久| 亚洲性夜色夜夜综合| 亚洲国产日韩欧美精品在线观看| 我要搜黄色片| 国产黄片美女视频| 国产成人影院久久av| 在线观看免费视频日本深夜| 午夜精品在线福利| 成人国产一区最新在线观看| 又爽又黄无遮挡网站| eeuss影院久久| 日本三级黄在线观看| 精品一区二区免费观看| 亚洲欧美日韩高清在线视频| 色综合站精品国产| 老司机深夜福利视频在线观看| 99国产精品一区二区蜜桃av| 国产亚洲精品久久久久久毛片| 嫁个100分男人电影在线观看| 在线天堂最新版资源| 1024手机看黄色片| 97碰自拍视频| 婷婷丁香在线五月| 高清日韩中文字幕在线| 可以在线观看毛片的网站| 露出奶头的视频| 伦精品一区二区三区| 日本爱情动作片www.在线观看 | 最新中文字幕久久久久| 久久国产精品人妻蜜桃| 国产精品日韩av在线免费观看| 国产亚洲精品综合一区在线观看| 亚洲七黄色美女视频| 国产精品98久久久久久宅男小说| 在现免费观看毛片| 精品人妻1区二区| 女人被狂操c到高潮| 一区福利在线观看| 九色国产91popny在线| 亚洲av成人av| 国产午夜精品论理片| 最近中文字幕高清免费大全6 | 俄罗斯特黄特色一大片| 白带黄色成豆腐渣| 成年人黄色毛片网站| 亚洲 国产 在线| 桃色一区二区三区在线观看| 国产高清激情床上av| 国产精品一区二区免费欧美| 日韩强制内射视频| 成人国产综合亚洲| 国产精华一区二区三区| 欧美又色又爽又黄视频| 精品久久久久久成人av| 亚洲国产欧洲综合997久久,| 色综合婷婷激情| 91在线精品国自产拍蜜月| 在线a可以看的网站| 好男人在线观看高清免费视频| 国产真实乱freesex| 亚洲成人久久爱视频| 欧美区成人在线视频| 成人特级av手机在线观看| 亚洲欧美激情综合另类| 成年女人看的毛片在线观看| 一区二区三区免费毛片| 国产精品1区2区在线观看.| 亚洲av熟女| 欧美日韩瑟瑟在线播放| 校园人妻丝袜中文字幕| 欧美+日韩+精品| 亚洲av日韩精品久久久久久密| 国产av麻豆久久久久久久| 国产成年人精品一区二区| 亚洲精品456在线播放app | 欧美激情在线99| 国产精品99久久久久久久久| 欧美一区二区国产精品久久精品| 欧美一级a爱片免费观看看| 色吧在线观看| 尤物成人国产欧美一区二区三区| 最新中文字幕久久久久| 中出人妻视频一区二区| 色在线成人网| 亚洲欧美激情综合另类| 大型黄色视频在线免费观看| 欧美色视频一区免费| 精品一区二区三区人妻视频| 桃红色精品国产亚洲av| 女同久久另类99精品国产91| 亚洲精华国产精华精| 国语自产精品视频在线第100页| 在线观看美女被高潮喷水网站| 国产伦一二天堂av在线观看| 乱人视频在线观看| 看黄色毛片网站| 久久久久久大精品| 大又大粗又爽又黄少妇毛片口| avwww免费| 如何舔出高潮| 免费看a级黄色片| 国产日本99.免费观看| 久久欧美精品欧美久久欧美| 国产精品美女特级片免费视频播放器| 精品一区二区三区视频在线观看免费| 色精品久久人妻99蜜桃| 一个人看的www免费观看视频| 欧美zozozo另类| 国产毛片a区久久久久| 国产高清有码在线观看视频| 亚洲国产精品合色在线| 91在线精品国自产拍蜜月| 97碰自拍视频| 亚洲人成伊人成综合网2020| 欧美又色又爽又黄视频| 很黄的视频免费| 日韩一本色道免费dvd| 久9热在线精品视频| 全区人妻精品视频| 日本一二三区视频观看| 麻豆久久精品国产亚洲av| 一级av片app| 亚州av有码| 国产伦一二天堂av在线观看| 桃红色精品国产亚洲av| 久久人人精品亚洲av| 午夜久久久久精精品| 久久久精品欧美日韩精品| av天堂中文字幕网| 很黄的视频免费| 国产精品久久久久久精品电影| 久久久久久久久中文| 亚洲五月天丁香| 国产三级中文精品| 尤物成人国产欧美一区二区三区| 国产一区二区三区在线臀色熟女| 欧美日韩综合久久久久久 | 欧美区成人在线视频| 国产日本99.免费观看| 亚洲av.av天堂| 国产精品不卡视频一区二区| 久久久久九九精品影院| 亚洲自拍偷在线| 久久久久久久亚洲中文字幕| 搡老妇女老女人老熟妇| 国产极品精品免费视频能看的| 97超级碰碰碰精品色视频在线观看| 99热网站在线观看| 91狼人影院| 一本久久中文字幕| 精品人妻视频免费看| 真实男女啪啪啪动态图| 超碰av人人做人人爽久久| 成年女人看的毛片在线观看| 国产v大片淫在线免费观看| 久久人人爽人人爽人人片va| 色综合色国产| 国产成人av教育| 免费av毛片视频| 国产午夜精品论理片| 亚洲最大成人手机在线| 五月伊人婷婷丁香| 免费一级毛片在线播放高清视频| 日本 欧美在线| 精品人妻1区二区| 十八禁国产超污无遮挡网站| 国产精品女同一区二区软件 | 国产亚洲欧美98| 欧美黑人欧美精品刺激| 久久久久国产精品人妻aⅴ院| 能在线免费观看的黄片| 午夜a级毛片| 国产色爽女视频免费观看| 国产精品三级大全| 欧美日韩国产亚洲二区| 男女视频在线观看网站免费| 国产极品精品免费视频能看的| 亚洲人与动物交配视频| 美女大奶头视频| 国产av一区在线观看免费| 免费高清视频大片| 在线播放无遮挡| 狠狠狠狠99中文字幕| 婷婷精品国产亚洲av| 日韩亚洲欧美综合| 此物有八面人人有两片| 婷婷丁香在线五月| 色吧在线观看| 国产在线男女| 69人妻影院| 国产视频一区二区在线看| 别揉我奶头~嗯~啊~动态视频| av在线观看视频网站免费| 人妻久久中文字幕网| 国产精品免费一区二区三区在线| 中文字幕久久专区| a级毛片a级免费在线| 久久久久国产精品人妻aⅴ院| 俺也久久电影网| 又粗又爽又猛毛片免费看| 亚洲欧美精品综合久久99| www.色视频.com| 成人欧美大片| 日本a在线网址| 麻豆成人av在线观看| 国产精品一区www在线观看 | 久久精品国产99精品国产亚洲性色| 久久精品国产亚洲av香蕉五月| 日韩大尺度精品在线看网址| 99九九线精品视频在线观看视频| 国产乱人伦免费视频| 最近在线观看免费完整版| 国产精品福利在线免费观看|