• <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 ?

    Aerial multi-spectral AI-based detection system for unexploded ordnance

    2023-10-09 04:29:04SeungwnChoJungmokOlegYkimenko
    Defence Technology 2023年9期

    Seungwn Cho ,Jungmok M ,Oleg A.Ykimenko

    a Republic of Korea Army,Republic of Korea/Defense Acquisition Program Administration,47,Gwanmun-ro,Gwacheon-si,Gyeonggi-do,13809,Republic of Korea

    b Department of Defense Science,Korea National Defense University,1040,Republic of Korea

    c Hwangsanbul-ro,Yangchon-myeon,Nonsan-si,Chungcheongnam-do,33021,Republic of Korea

    d Department of Systems Engineering,Naval Postgraduate School,USA

    e 1 University Circle,Monterey,CA,93943,USA

    Keywords:Unexploded ordnance (UXO)Multispectral imaging Small unmanned aerial systems (sUAS)Object detection Deep learning convolutional neural network(DLCNN)

    ABSTRACT Unexploded ordnance (UXO) poses a threat to soldiers operating in mission areas,but current UXO detection systems do not necessarily provide the required safety and efficiency to protect soldiers from this hazard.Recent technological advancements in artificial intelligence (AI) and small unmanned aerial systems (sUAS) present an opportunity to explore a novel concept for UXO detection.The new UXO detection system proposed in this study takes advantage of employing an AI-trained multi-spectral(MS)sensor on sUAS.This paper explores feasibility of AI-based UXO detection using sUAS equipped with a single(visible)spectrum(SS)or MS digital electro-optical(EO)sensor.Specifically,it describes the design of the Deep Learning Convolutional Neural Network for UXO detection,the development of an AI-based algorithm for reliable UXO detection,and also provides a comparison of performance of the proposed system based on SS and MS sensor imagery.

    1.Background

    Unexploded ordnance (UXO) is “military ammunition or explosive ordnance which has failed to function as intended”and left behind after war conflicts,weapon system testing and trainings[1].Fig.1 shows a wide range of UXO varying by size (cf.155-mm projectile vs.20-mm M55 projectile),shape,and color.

    Fig.1.Sample UXO found during Fort Ord ex-military base clean up.

    UXO poses an obvious threat to warfighters and civilians relevant to military affairs resulting in injuries and deaths of hundreds of people each year worldwide[2].Most UXO(>50%)are located on the surface of the ground and can be detected visually or photographically,but modern UXO detection systems are often equipped with subsurface UXO targeting means such as magnetometers or electromagnetic sensors [3].These UXO detection systems,designed for hand-held operations,pose several problems in terms of safety and efficiency.The systems require soldiers to walk through the hazardous areas to detect UXO;and it takes a lot of time to survey an area in question because of relatively low productivity.

    To address the problem of reliable UXO detection,researchers have tried different approaches.Most of the recent approaches involved using different sensors flown on small unmanned aerial systems (sUAS).For example,DeSmet et al.[4] designed a mine detection system using sUAS equipped with an infrared spectrum camera to detect a plastic mine-a kind of UXO that may not be detected by the electromagnetic detection system.Their study proved feasibility of the suggested approach in various conditions including different temperatures,moisture,and burial depths [4].Similarly,Qi et al.[5]attempted to detect underground UXO using the sUAS with the transient electromagnetic sensor.Their UXO detection experimentation also proved to be a good alternative to conduct dull and dangerous UXO detecting operations.

    The multi-spectral(MS)sensors have been used for a variety of applications already as well [6].According to the Defense Science Board Task Force [7],infrared spectrum data has a potential value for surface and very shallow UXO.Johnson et al.[8] suggested possibility of distinguishing between UXO and natural materials by using a spectral signature.Takumi et al.[9] studied MS object detection for autonomous vehicles in traffic and found that the detector trained with MS images featured 13% higher mean average precision compared to the detector trained with a single-spectral(SS) sensor imagery.

    While the previous studies demonstrated advantages of using sUAS for UXO detection,no attempt has been yet reported to make use of artificial intelligence (AI) including deep learning convolutional neural network (DLCNN) as applied to MS sensor [10].This paper presents the results of a preliminary study on UXO detection using AI-based detectors trained on different bands of MS sensor.And is based on[11].The paper is organized as follows.The overall concept of the sUAS-based UXO detection system and basics of the MS technology are presented in Section 2.Section 3 suggests the DLCNN architecture for MS UXO detection and provide details on the overall algorithm.Section 4 discusses the detection results using DLCNN trained on SS and MS sensor imagery.Section 5 presents the results of the flight test to demonstrate the feasibility of suggested approach in a real-world operational scenario.The paper ends with conclusions.

    2.Concept of operations and underlying MS sensing technology

    This section provides the concept of operations (CONOPS) of a sUAS-based UXO detection system and provides some background on MS sensors.

    2.1.sUAS-based UXO detection system CONOPS

    Fig.2(a)presents an example of a UXO detecting operation near Yeongpyeong shooting range in South Korea,where the Republic of Korea Army 5th Engineer Brigade and the U.S.Army 2nd Infantry Division are using a ground-based electromagnetic system.This operation is obviously dull,dirty,dangerous,and that is why using sUAS seems to be a very attractive alternative not putting humans at risk.Fig.2(b)shows a soldier testing the next-generation sUAS at Fort Benning,GA.Such a sUAS,equipped with a MS sensor can make UXO detection much more efficient.

    Fig.2.(a) Manual UXO search,vs.(b) automated search using sUAS.

    The commander can define the search area and initiate the launch of one or several sUAS to comb the specified area (Fig.3).These low-flying sUAS would use a standard pattern for detecting and identifying UXO to secure a mission area before a small tactical unit makes a maneuver.The sUAS captures footage of the mission area with onboard SS and/or MS sensor and transmits this footage to the ground control station (GCS).The image processing algorithms can reside either onboard sUAS or at GCS.The trained DLCNN is then used to run through the images(or streamed video)and mark the suspected UXO(s) with the bounding boxes (BBs),providing an operator with the coordinates of suspected UXOs.The operator assesses the UXO threats based on the detection results and makes an educated decision.

    Fig.3.CONOPS for a sUAS-based UXO detection system.

    Enabling the aforementioned capability for the operational use,as shown on the right side of Fig.4,involves several steps resulting in the development of a trained DLCNN-based UXO detector.Specifically,it involves gathering some original(relatively large)set of imagery data featuring a variety of known UXOs to be detected(as depicted on the left side of Fig.4),and then conducting several steps toward preparing data for DLCNN training,developing the detector and then validating its performance (the three blocks shown in the middle part of Fig.4).The dashed line in Fig.4 represents the process of updating the UXO imagery data set for future retraining of DLCNN as/if needed.Sections 3 and 4 describes these steps in detail,while the remainder of this section introduces the MS sensor used in this study to gather UXO imagery.

    Fig.4.Steps involved in developing a DLCNN-based UXO detector,and its further operational usage.

    2.2.MS imaging

    While the traditional single-lens SS sensors combine and store visible light spectrum (400-700 nm) as Red-Green-Blue images,and MS sensors employ multiple lenses to capture multiple images in a selected sets of wavelength bands.The latter may include wavelengths invisible to the human eye.For example,the set of wavelength bands shown in Fig.5 includes visible Blue,Green,and Red spectra,along with invisible Red Edge(RE),and Near-Infrared(NIR) spectra.As seen,RE spectrum (between 690 nm and 730 nm) is located in the very beginning of the non-visible light spectra in the region of rapid change in reflectance of vegetation,and as such can be very useful in detecting disturbed soil features caused by changes in the physical environment and in chemical properties.NIR wavelengths fall into the invisible part of the light spectrum between 700 nm and 1200 nm.The MS sensors have already proved to be a very useful tool in agriculture,forestry and land management applications;hence it is natural that this study assesses their effectiveness for UXO detection.

    Fig.5.SS vs.MS imaging.

    One example of a relatively inexpensive MS sensor is Parrot Sequoia which features Green,Red,RE and NIR bands plus a standard 16Mpix Red-Green-Blue (RGB) lens for color mapping (while the former four lenses use a global shatter,the RGB lens uses a rolling shatter so that the quality of RGB images is not as good as those of the individual bands).Parrot Sequoia manufacturer offers an integration kit with a Parrot Bluegrass Parrot Disco-Pro AG sUAS[12].

    Another popular (a bit more expensive) sensor is MicaSense RedEdge sensor(Fig.6(a))with five narrow bands centered around Blue(475 nm),Green(560 nm),Red(668 nm),RE(717 nm),and NIR(840 nm)[13].Compared to Parrot Sequoia,the addition of the Blue component allows compositing RGB bands to create a nice crisp color maps).This sensor became very popular so that many leading drone manufacturers offer integration kits with their sUAS(Figs.6(b) and 6(c)).Having such a sensor integrated with a small multirotor aerial platform,like the one shown in Fig.6(b),allows covering up to 0.3 km2within about 15 min,i.e.,during a single sortie.It is this sensor that was used in this research.

    Fig.6.(a) MicaSence RedEdge MX sensor,(b) integrated with DJI Inspire 2 and (c) in the nose bay of Quantum Trinity F90+.

    Fig.7 shows an example of MicaSense RedEdge MS imagery recorded by each individual spectrum lens.While all images look somewhat the same,there may be slightly different sets of scene features captured by each of them,thus complementing each other.The images are stored in the Tag Image File Format (TIFF) files for each spectrum band.

    Fig.7.Example of five-band MS sensor imagery.

    3.AI-based algorithm for MS data processing

    This section presents the developed approach to process MS sensor imagery to detect UXO.The key is that each MS band is processed separately using the corresponding pretrained detector(different for each spectrum band).The detection results from all bands are then combined to make a final determination.

    3.1.MS-based UXO detection

    All five simultaneously obtained images shown in Fig.7 feature the same UXO seen in the different spectrum bands.As suggested in one of the previous research efforts involving MS imagery,each spectral image features different characteristics and as such the best approach to process MS imagery would be to detect the object(s) of interest in each spectrum separately and then combine the detection results [9].Following this suggestion,the data flow for the developed algorithm looks like presented in Fig.8.

    Fig.8.MS-based UXO detection algorithm data flow.

    Five same time-stamp MS images are processed using five different detectors developed (trained) for each spectrum.The detection results are then merged using the two-step procedure.The data flow depicted in Fig.8 (developing five complementary detectors as opposed to just one) essentially enhances the three blocks shown in the middle part of Fig.4.

    The following two sections describe the developed algorithms,essence of detector training and detection results merging procedure.

    3.2.Multi-stage DLCNN architecture

    For UXO detection in each MS band,this research employed a deep learning convolutional neural network (DLCNN) as shown in Fig.9(deep refers to the number of hidden layers,i.e.depth of the neural network).The three main pieces of the developed DLCNN architecture include feature training subnetwork,feature transfer layer,detection subnetwork,and output layer.

    Fig.9.DLCNN architecture for UXO detection.

    Rather than building a feature training subnetwork from scratch,the pre-trained CNN structure for feature extraction was used.After experimenting with DarknetReference,Darknet19,ResNet50,and Desnet201,the ResNet50,a variant of ResNet model,was finally chosen.

    ResNet50 refers to the Residual Network with 50 layers based on residual learning [13].Theoretically,the performance of CNN increases with the increase in the number of layers because these additional layers learn more complex features.However,there is a limit to the number of layers,so that when it is reached,the accuracy of the model starts to saturate and then degrades(due to the so-called vanishing gradient problem) [14].In other approaches,each layer feeds to its next layer and the same goes on subsequently.In a residual network,each layer feeds to its next layer and directly to the 2-3 layers below it (not shown in Fig.9).Using ResNet drastically improves the CNN performance despite having more layers (to be able to identify features of smaller objects like UXO shown in Fig.7).Each,convolution layer is followed by a batch normalization layer and a Rectified Linear Unit (ReLU) activation function,f(x)=max(0,x).(Without an activation function the model would behave like a linear regression model with a limited learning capacity.)

    As shown in Fig.9,using the trial-and-error approach the ResNet50 CNN was actually limited to just 40 layers which provided a good balance between UXO detection performance and training time.As such,ReLU 40 activation function transforms the final detected feature data to the detection/classification subnetwork.

    After experimenting with three popular CNN models for object detection -Faster R-CNN [15],SSD [16],and YOLOv2 [17] -the latter one was chosen for this particular application,proving the fastest learning speed yet again [18].

    3.3.Essence of the YOLOv2 algorithm

    The key advantage of YOLOv2 in improving detection/classification speed and efficiency is the use of the anchor boxes(ABs).The ABs are a set of predefined boxes of different sizes defined to capture the scale and shape (aspect ratio) of specific object class(es).Instead of predicting one BB per pixel,YOLOv2 divides a reduced-size input image into a square grid and predicts several BBs per grid.As an example,Fig.10(a)shows a 5 by 5 grid with a set of three fixed ABs tiled across the image (for simplicity,Fig.10(a)visualizes just four grid cells filled with the ABs).These 5 × 5 × 3=75 ABs represent the initial BB guesses.An AB-based object detector processes an entire image at once as opposed to scanning an image with a sliding window(as some other methods do) and computing a separate prediction at every potential position.

    Fig.10.(a)Example of a 5×5 grid populated with three different-shape Abs;(b)Computation of an adjusted AB parameters[17];(c)Mean IoU for UXO detection vs.the number of ABs per grid cell.

    In reality,YOLOv2 uses the input image of size 416 pix×416 pix divided into a 13 × 13 grid enabling to downsample it 32x (five times)and execute a multiscale training to find both large and finer objects(YOLOv2 architecture randomly chooses image dimensions for every 10 batches).

    Fig.10(a) also shows a hypothetical ground truth box (GTB)created during data labeling (this GTB contains the object of interest in it).The ABs of the grid cell where the center of GTB resides are responsible for detecting this object.(One grid cell may be responsible for predicting several objects and that is another reason to use multiple ABs per cell.)

    The BB predictions are made by varying (predicting) five parameters per each involved AB.Four of them,tw,th,tx,andty,define the size of the adjusted AB (BB prediction) relative to the GTB’s width and height,pwand ph,as shown in Fig.10(b),as well as the location of its center relative to the upper-left corner of the corresponding grid cell(σ(x)denotes a sigmoid function to constraint its output between 0 and 1 [19]).The fifth varied parameter is the confidence score(CS)reflecting how confident the model is that the BB region (predicted by DLCNN) contains an object and also how accurate it thinks the predicted BB(pBB)is.It is defined as a product

    whereP(Object),the probability of example belonging to a given class (just one class,UXO,in this study),would be 1 or 0,as each sample in the training dataset is a known example from the domain,and IoU(AB,GTB),the intersection over union(IoU)for the predicted BB (pBB) and GTB defined as Ref.[20].

    For the specific case shown in Fig.10(a),for all but one grid cell,P(Object)=0(since the center of the GTB does not fall in them)and CSs are zero.For the ABs residing at the fourth raw (counted from the top of the image)and the second column of the grid cell,CSs are identical to the IoU values for those three ABs(sinceP(Object)=1).

    During CNN training,we only want one AB(BB predictor)to be responsible for each object.We assign one predictor to be“responsible”for predicting an object based on which prediction has the highest current IoU with the GTB.This leads to specialization between the BB predictors.Each predictor gets better at predicting certain sizes,aspect ratios,or classes of an object,improving overall ability of the detector to find all relevant objects.In the situation shown in Fig.10(a),out of three pBBs the one with highest IoU will be selected.

    In the case of multiple objects in the image (and even within a single grid cell),the number of pBBs may be adjusted by establishing a threshold on IoU.Decreasing the threshold reduces the number of selected BBs.Of course,decreasing the threshold too much,leads to elimination of pBBs that represent objects close to each other in the image.This procedure is referred to as nonmaximum suppression (NMS).

    In the specific application this study dealt with,all UXOs are considered to be of a single class,but theoretically,the proposed DLCNN could be trained to recognize several UXO classes since there may be a need to differentiate between the different types of UXOs (see Fig.1).Fig.10(c) shows how the number of ABs (within each grid cell) affects the average IoU.Choosing a large number of prior boxes allows for greater overlap between ABs and GTB(higher IoU values),however,using more ABs can increase the training time and lead to overfitting and therefore worsening detector performance.In this study it was chosen to use nine ABs(ensuring a mean IoU value of about 0.8),since after that the mean IoU plateaus(typically,any threshold greater than 0.5 should suffice).

    During an object detector training,the AB positions and sizes are refined to minimize the so-called loss function,which is a weighted sum of squared error and consist of classification loss(1-CS)and BB localization error.The DLCNN training process (the process of updating network parameters θ,weights and biases) employs a stochastic gradient descent with momentum optimizer.In this case,the parameter vector θ is updated as follows [21]:

    wherelis the iteration number,α is the learning rate,?denotes the gradient,E(θ)is the loss function and γ determines the contribution of the previous gradient step to the current iteration.Usually,it is not possible to fit the whole available training dataset(hundreds of images)in the machine’s memory and that is why the second term in Eq.(3) is computed using a minibatch data processing.

    wherenis the number of minibatches andEi(θ) is the i-th minibatch loss.A batch or minibatch refers to equally sized subsets of the dataset over which the gradient is calculated,and weights updated.(Mini-batch sizes are commonly called “batch sizes”for brevity.The term batch itself is ambiguous,however and can refer to either batch gradient descent or the size of the minibatch.)If the total number of available data (images) in N and the size of a minibatch is k,then nk ≤N.If the dataset does not divide evenly by the batch size,the final batch has fewer samples than other batches.

    The number of batches is also referred to as a training epoch.One epoch means that each sample in the training dataset has had an opportunity to update the internal model parameters.If the total number of available data(images)inNand the size of the batch isN,then the parameter vector θ is only updated once per epoch(which in this case is the same as iteration).Otherwise,it is updatedm=ceil(N/k)times.Minibatch gradient descent reduces the risk of getting stuck at a local minimum,since different batches will be considered at each iteration,granting a robust convergence.

    The UXO detection DLCNN ends with the YOLOv2 output layer(Fig.9).The output layer provides the locations of refined BB enclosing all found UXOs.As an example,Fig.11 shows these BBs for MS images similar to those of Fig.7.

    Fig.11.BBs for a detected UXO as seen on MS images.

    3.4.Merging of the MS detection results

    As seen from Fig.8,after five MS band detectors find (or not)UXO(s),the detection results for multiple detectors need to be merged.The first step is simple merging that shows the detection results from five detectors on a single image(which can be one of the band images or composed RGB image).For instance,if only two out of five detectors detect UXO,the simple merging would result in two BBs showing up on the image.

    The simple merging step for BBs shown in Fig.11 (in the case when the corresponding detectors detect UXO in each spectrum),is shown in Fig.12(a)featuring five BBs around the same single UXO.

    Fig.12.(a) Simple merging;(b) Suppression of week detection results.

    The final merging step (Fig.8) combines whatever number of BBs was found for each UXO into a single one.This is done by using the NMS method,described in Section 3.3.The only difference is that IoU values (see Eq.(2)) are computed using a pairwise comparison between all found BBs.If it exceeds a certain threshold,the weaker detections are suppressed.

    To prevent the actual overlapped detection result from being recognized as not overlapped,this NMS IoU threshold is set as low as 0.1.Hence,the redundant detection results can be effectively suppressed for which IoU is higher than 0.1,except for the best detection result.The reason for setting the low IoU threshold relates to the characteristic of the MS specific sensor used in this research (Fig.6(a)).MicaSence RedEdge MX sensor has five physically different lenses for five spectra,and the different physical locations of these lenses cause some misalignment of MS data.Even though each spectrum detector successfully detects UXO,the simply merged BBs are misaligned and with a high IoU threshold this makes the system think they detected different objects.Fig.12(b)shows an example of NMS method as applied to the situation shown in Fig.12(a)leaving the only BB with the highest CS among the five detection results.

    4.Algorithm performance evaluation

    This section describes imagery collected in the close to actual operational environment to assess feasibility of the proposed UXO detection approach,followed by describing the process of detector training and finally evaluating the overall performance of the developed and trained MS UXO detector as compared to the detector based on a standard SS sensor imagery.

    4.1.UXO imagery gathering and preprocessing

    This study took advantage of using nine different types of UXO shown in Fig.13(cf.Fig.1).These included mortar projectile,handgrenade,bullets,etc.All of these UXOs have a different shape,color,and size.(As mentioned earlier,this study aimed at assessing the overall concept of detecting UXO using MS sensor and was not intended to attempt to classify different types of UXO.) To obtain diverse data,these nine UXOs were randomly placed on the ground.Some were placed close to/under the bushes to mimic the operational environment (Fig.14).

    Fig.13.A set of UXO used to train DLCNN.

    Fig.14.Examples of UXO placement to mimic the operational environment.

    A total of 4075 still images -815 photos for each spectrum -were collected using the MicaSense RedEdge MX sensor as visualized in Fig.15(a).These images had 1280 pix ×960 pix resolution.To assess advantages of using an MS sensor to detect UXO,the corresponding imagery was also taken using a SS sensor.Specifically,1225 still images and 59 videos (each about ten to 20 s in length each)were taken using the Sony Alpha 6000 digital camera(Fig.15(b)).The still images had 6000 pix × 4000 pix resolution.

    Fig.15.Data collected with the MS (a) and SS (b) sensors.

    Both sets of still images were preprocessed using the same fourstep routine.

    (1) All collected data were resized to 416 pix × 416 pix resolution,which happens to be the best image size for YOLOv2 CNN model performance(Fig.16(a))(Figs.16 and 17 show the four-step routine as applied to the SS images),

    Fig.16.(a) Image resizing;(b) Labeling.

    (2) To establish ground truth data,the resized images were labeled using the Image Labeler app provided by MATLAB(Fig.16(b)),

    (3) Labeled data were randomly divided into three groups-70% for training the UXO detector,15% for validation (for selfcorrection at regular intervals during training to improve the detector’s performance),and 15% for testing data (for evaluation of the trained detector)For the SS sensor it meant 857:184:184 split (Fig.17(a)),and for the MS sensor -571:122:122 split,

    Fig.17.(a) Division of collected imagery into the training/validation/testing subsets;(b) Examples of labeled image augmentation.

    In order to increase the size of the ground truth data pool(to add more variety to the training data without actually having to increase the number of labeled training samples),the training subset of labeled data was augmented by applying rotation,changing brightness and contrast (Fig.17(b)).(Validation and testing data should be representative of the original data and as such is left unmodified for unbiased evaluation.) As a result,the number of training subset of images was increased fourfold:up to 3428 for the SS sensor and up to 2284 for the MS sensor.

    4.2.Detector training

    All DLCNN training was conducted using a MATLAB (interpretative) development environment on a generic laptop (Table 1).

    Table 1 Laptop specifications.

    First,the SS detector was trained as explained in Section 3.3.Step 1 of the four-step routine of Section 4.1(1225 images resizing)takes 2 min,while Step 2(image labeling)happens to be the most time consuming -on the order of 2 h.

    As far as a training procedure itself,after several tries with the different training options,the number of the ABs was set as nine,maximum epochs as 20,and mini-batch sizes as eight(Fig.18).One epoch represents one backward and one forward pass through the network for all the training data;the batch size is the number of training data in one forward and backward pass;the iteration is the number of batches for one epoch.This setup seems to deliver the optimal training performance as applied to the collected UXO imagery.Using the aforementioned parameters,the training of a visual spectrum UXO detector took about 2 h.

    Fig.18.SS imagery training setup.

    The results of applying the trained SS detector to the several images from the testing set are shown in Fig.19.The UXO detector puts BB around the predicted UXO in the testing image along with the detection CS.Specifically,Fig.19 shows three examples of the successful UXO detection and one example of unsuccessful detection.Unsuccessful means that the detection CS happened to be below the threshold (of 0.5) and as such -BB does not show up.

    Fig.19.Examples of three successful and one non-successful UXO detection.

    The UXO detector training for each of five MS sensor spectra was conducted next(Fig.15(a)).Similar to the SS detector training data preparation,Step 1 takes about 2 min and Step 2 takes an hour and half.

    The same number of ABs(nine)and the same(eight)mini-batch size were used,but the number of epochs was reduced to five.This was done to keep the training time about the same as for the SS detector (with no significant loss of performance).As a result,it took only 20 min to train each of five detectors,1 h 40 min in total.

    Fig.20 shows a typical example of applying five separately trained detectors to five individual spectrum band images (see Fig.7).As expected,UXO is not necessarily detected in each spectrum band.In this particular example,only the Blue and Green detectors were successful while three other were not.In some other cases,only the NIR detector detected UXO while other detectors failed.Indeed,having five detection results rather than just one for a SS detector increases the overall reliability of UXO detection.Fig.20 also shows the merging step of the Blue and Green detector BBs.

    Fig.20.Results of detection in multiple spectra and the final merging BB.

    It should be noted that as opposed to the SS detector,for all five MS detectors,the IoU threshold was set to a slightly lower value(0.4 as opposed to 0.5).The reason is that individual spectrum band images for MS data are not aligned with each other (Section 3.4).The misalignment of data causes a phenomenon in which a True Positive-correct detection result-is recognized as a False Positive-wrong detection result.Setting a lower threshold for the MS detectors can prevent the actual True Positive from being evaluated as False Positive.

    4.3.Overall performance assessment

    As a metric to evaluate the effectiveness of the trained detector,the average precision (AP) is commonly used [22,23].The average precision provides a single number that incorporates the ability of the detector to make correct classifications (precision) and the ability of the detector to find all relevant objects (recall).

    The first step to compute AP,is to build a confusion matrix(CM)that is a very convenient tool to visualize algorithm performance(Fig.21).The CM rows represent the test outputs,while its columns indicate an actual state.Positive or Negative is determined by whether the UXO detector detects UXO in a given image.True or False is determined by whether the value of IoU is larger than a certain IoU threshold,indicating how much the detection results and the actual UXO location overlap.Precision measures how accurate the UXO detector's predictions are based on the ground truth data-in this case,actual UXO location.Hence,if the IoU between the prediction and the ground truth is more than or equal to a IoU threshold,the detection is classified as True Positive;if IoU is less than a IoU threshold,the detection is classified as False Positive;and if a ground truth is present in the picture and the detector failed to detect the object,it is classified as False Negative.

    Fig.21.CM for UXO detector evaluation.

    Next,the values of precision,p,and recall,r,are computed.Precision measures a fraction of Positives within all detections

    Recall measures how well the detector finds all Positives based on the ground truth data

    Finally,after the trained UXO detector detects UXO in a given set of test data,the detection results are ranked in the descending order of the predicted CS (see example in the left-hand side of Fig.22).To visualize thepandrvalues of the detection results,the Precision-Recall (PR) curve is plotted (right-hand side of Fig.22).This curve shows the precision of the trained detector at different recalls [23].

    The AP is then computed as an average precision value among all recalls,i.e.,as an area under the PR curve.By construction,the calculated AP is between 0 and 1,and it is presumed that the higher the AP,the better the detector.

    Following the PASCAL Visual Object Classes(VOC)Challenge,in order to reduce the impact of the“wiggles”in the PR curve,caused by small variations in the ranking of examples,the AP value is calculated by averaging precision at a set of 11 evenly-spaced recall values [0,0.1,…,1] [24,25].

    The precision at each recall level is interpolated as

    What is shown in Fig.22 is the PR curve computation and the 11-point-interpolation AP evaluation for the trained SS detector,which resulted in detecting 193 UXOs within 184 test images.As shown,the AP was evaluated to be 0.774.

    For MS detection,the respective detectors were evaluated with the 122 test images for each spectrum.The AP for the respective detectors were computed as being between 0.484 and 0.592 as shown in Fig.23.

    Fig.23.The PR curve for individual spectrum detectors.

    As seen,the Blue,Green,and Red detectors feature a(0,0)point meaning that they have no detection result in some images.On the contrary,the RE and NIR detectors do not have this point meaning they found a candidate UXO in every image,even if there were no UXO.

    Data shown in Fig.23 were further integrated into a combined PR curve using the two-step merging procedure described in Section 3.4 (Fig.7).By doing so,the five detection results were integrated into a single one (the integrated MS detection results were evaluated on the basis of Red spectrum test data).The integrated MS detector PR curve is shown in Fig.24.Compared to Fig.22,the integrated MS detector features a 12.5% higher AR value as opposed to the SS detector (0.871 as opposed to 0.774) which matches the results of the earlier study [9].

    Fig.24.The PR curve for the integrated MS detector.

    While the evaluation results of respective detectors were all less than the evaluation of the RGB-spectral detection,when integrated they feature da better effectiveness of the trained UXO detector(Table 2).

    Table 2 Comparison of SS and MS UXO detectors.

    Fig.25 illustrates an attempt to increase the AP value even further by using a tighter GTBs when labeling (Fig.16(b)).While being much more time consuming,this approach resulted in only 1% improvement of the tight-fitted detector as compared to the loose-fitted detector using the same dataset (AP=0.904 and AP=0.895,respectively).(It should be noted that because of randomization in splitting the original data set into three subsets(Fig.17(a)),the initial value in Fig.25 (0.895) differs from that of Fig.24(0.891).)

    Fig.25.Effect of tightness while data labeling [26].

    5.CONOPS assessment

    To prove the overall sUAS-based UXO detection CONOPS,one more set of tests was conducted.The idea was to check if the developed DLCNN architecture and training procedure could be used to detect other objects rather than UXOs and if the detection could be performed using a steamed video as opposed to still images in a real-time.Specifically,in this series of tests,four small firearms-two rifles and two pistols-were used as the objects of interest as shown in Fig.26(a).Fig.26(b) shows a sUAS integrated with a Zenmuse X5 sensor used in this case (there were no MS sensors that would provide a steamed video available).Fig.27 illustrates the operational scenario.

    Fig.26.(a) Typical small firearms,and (b) sUAS integrated with an SS sensor.

    Fig.27.Operational scenario used in the CONOPS evaluation flight test.

    Specifically,the mission area was established as a size of a basketball court 28.65×15.24 m2(94×50 ft2)or 440 m2(4700 ft2),sUAS executed a serpentine search pattern flying 3 m(10 ft)above the ground level at an average walking speed of about 5 km/h(1.4 m/s or 3 mph).The complete sweep of the area took 18 passes,about 30 s each,or approximately 10 min in total.Therefore,it only took one sortie to sweep the area.

    In total,18-30-s-long video clips,were collected.These raw 18 video clips were recorded at 30 frames per second(fps).They were then split into still frames.Further processing included only one of every three consecutive images 5400 images in total),thus effectively reducing the rate down to 10 fps.The SS detector training was conducted using the same DLCNN structure and training procedure as outlined in Sections 3 and 4.Specifically,twelve video clips(equivalent to 1812 RGB images)were used for small arm detector training,three(389 images)-for validation and three(389 images)-for testing.These images had 3,840 pix × 2,160 pix resolution.

    The training procedure employed eight ABs.Similar to what was shown in Fig.9(c),having eight ABs ensured about the same mean IoU value(of about 0.8)as nine ABs for UXO detection.The DLCNN training options were also similar to those for the SS and MS UXO detector training.To decrease the training time,a minibatch size was set to eight,and the maximum number of epochs-to five.The training took an hour and 10 min.

    The trained DLCNN detector was then applied to process three testing video clips.Fig.28(a) shows four snapshots capturing a BB around each type of the small firearms used in the experiments.Visually,there were no dropouts or false detection during the detector testing,suggesting that the AP value of the small firearms detector is close to 1.This is confirmed by data presented in Fig.28(b)featuring the AP value of 0.966,which is 25% higher than that of the SS UXO detector and 11% higher than the MS UXO detector (Table 2).Hence,due to a specific shape of small firearms,even a SS detector does almost a perfect job in reliably (precisely)finding all objects in questions.

    Fig.28.(a) Examples of successful firearms detection and PR curve for the SS-based firearms detector.

    Even though the trained DLCNN detector was applied to the video clips in the interpretive environment of MATLAB,each frame was processed in 0.1 s,which suggests that the algorithm should work in real time (at least for the rate of 10 fps and less).

    This field flight proves the feasibility of the trained detector within the actual operational flying environment.Plus,this implementation explores the potential to expand he UXO detection system for other articles.For the operational assessment,the test result shows that the one sUAS trained with 1812 images can cover the area of operation 440 m2(4700 ft2)operational within 30 s with 96% reliability in the condition of the open field and the 3 m(10 ft)flight altitude.If the sUAS fly higher,it could cover the wider area,or it could sweep the area much faster.In consideration of the trade-off between the reliability and the flight altitude,the future study is suggested to find the optimal highest flight height to cover the operational environment as quick as possible.

    6.Conclusions

    This paper presented the concept of using a small multirotor sUAS equipped with SS or MS sensor to take advantage of AI-based algorithms for small object(UXO and small firearms)detection.The DLCNN detectors developed and trained on realistic imagery proved overall feasibility of the concept and demonstrated quite a promising performance.With only about 1000 images (per each spectrum band) needed to train a SS detector or MS detector,it takes on the order 2 h to prepare(resize,label,augment)data and about 1-2 h to actually train the detector.Once trained,the detector can process streamed video at the 10fps rate(in the MATLAB interpretative environment).For the compiled code this rate is expected to be at least 100 times higher.As far an average precision,even a standard single (visual) spectrum camera allows detecting UXO and small firearms with 0.774 and 0.966 precision,respectively,which are the high values.Employing a MS sensor with multiple spectrum bands,especially Red Edge and NIR,allows capturing different sets of features so that while the AP values for individual spectrum bands are lower than that of the SS sensor,when combined (integrated) they ensure much better detection results.A typical small multirotor UAS equipped with the SS or MS sensor featuring about 30 min flight duration can effectively sweep an area equal to three basketball courts in just one sortie.Future research will attempt the UXO detection in different weather conditions and partially under the ground.Another extension will be attempting UXO classification along with UXO detection.

    Authors’contributions

    SC designed and coordinated this research,carried out experiments,and drafted the manuscript.JM conceived of publishing the paper and helped to draft the manuscript.OY participated in the design of the study research coordination and revised the manuscript.The authors read and approved the final manuscript.

    Funding

    The authors declare that no funds,grants,or other support were received during the preparation of this manuscript.

    Code availability

    Not available publicly.

    Ethics approval

    Not applicable.

    Consent to participate

    Not applicable.

    Consent for publication

    Not applicable.

    Declaration of competing interest

    The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

    Acknowledgements

    The authors would like to thank the Office of Naval Research for supporting this effort through the Consortium for Robotics and Unmanned Systems Education and Research.They would also like to thank the personnel of the California National Guard post at Camp Roberts,CA,Manager of the NPS Field Laboratory at McMillan Airfield,Mr.Greg Arenas,and NPS Base Police LT Supervisor Edward Macias.

    国产精品98久久久久久宅男小说| 99热国产这里只有精品6| 高清在线国产一区| 一边摸一边抽搐一进一出视频| 好男人电影高清在线观看| 他把我摸到了高潮在线观看 | 一区在线观看完整版| 99热网站在线观看| 黄色丝袜av网址大全| 国产精品国产高清国产av | 久久中文字幕人妻熟女| 99久久精品国产亚洲精品| 中文字幕av电影在线播放| 黑人巨大精品欧美一区二区mp4| 美女午夜性视频免费| 国产男女超爽视频在线观看| 日本wwww免费看| 国产日韩一区二区三区精品不卡| 精品国产超薄肉色丝袜足j| 午夜福利一区二区在线看| 大型av网站在线播放| 国产一区二区在线观看av| 亚洲精品av麻豆狂野| 999久久久精品免费观看国产| 午夜福利一区二区在线看| 亚洲av欧美aⅴ国产| 777米奇影视久久| 又黄又粗又硬又大视频| 国产不卡一卡二| 免费观看人在逋| 亚洲人成伊人成综合网2020| 欧美日韩国产mv在线观看视频| 嫩草影视91久久| 国产在线视频一区二区| 国产精品秋霞免费鲁丝片| 99精品在免费线老司机午夜| 每晚都被弄得嗷嗷叫到高潮| 欧美日韩中文字幕国产精品一区二区三区 | 成人18禁在线播放| 欧美成人午夜精品| 欧美黑人精品巨大| 一区二区三区精品91| 国产人伦9x9x在线观看| 亚洲精品国产区一区二| 岛国毛片在线播放| 精品久久久久久电影网| 两性午夜刺激爽爽歪歪视频在线观看 | 99九九在线精品视频| 老熟妇仑乱视频hdxx| 亚洲精品美女久久av网站| 国产精品成人在线| 后天国语完整版免费观看| 国产在线精品亚洲第一网站| 三上悠亚av全集在线观看| 国产一区二区三区综合在线观看| 欧美成狂野欧美在线观看| 中文字幕最新亚洲高清| 美女视频免费永久观看网站| 91成年电影在线观看| 免费在线观看黄色视频的| 啦啦啦在线免费观看视频4| 成人av一区二区三区在线看| 人人妻人人澡人人爽人人夜夜| 国产极品粉嫩免费观看在线| 一区二区三区乱码不卡18| 国产日韩欧美亚洲二区| 欧美在线一区亚洲| 久久热在线av| 好男人电影高清在线观看| 久久久国产精品麻豆| 国产极品粉嫩免费观看在线| 国产成人精品久久二区二区免费| 在线观看一区二区三区激情| 麻豆av在线久日| av天堂久久9| 国产精品1区2区在线观看. | 精品国产一区二区三区久久久樱花| 日本撒尿小便嘘嘘汇集6| 99九九在线精品视频| 免费看a级黄色片| 精品国产乱码久久久久久男人| 首页视频小说图片口味搜索| 亚洲欧美精品综合一区二区三区| 日本黄色日本黄色录像| 在线观看免费午夜福利视频| 又黄又粗又硬又大视频| 1024香蕉在线观看| 欧美日韩一级在线毛片| 亚洲va日本ⅴa欧美va伊人久久| 日韩视频在线欧美| 大香蕉久久成人网| 色播在线永久视频| 国产精品香港三级国产av潘金莲| 嫁个100分男人电影在线观看| 精品第一国产精品| 午夜精品国产一区二区电影| 免费高清在线观看日韩| 午夜久久久在线观看| 午夜激情av网站| 久久久欧美国产精品| 两性夫妻黄色片| 精品少妇久久久久久888优播| 三上悠亚av全集在线观看| 国产男女超爽视频在线观看| √禁漫天堂资源中文www| 麻豆成人av在线观看| 亚洲成av片中文字幕在线观看| 欧美人与性动交α欧美精品济南到| 中文字幕av电影在线播放| 精品人妻1区二区| 性色av乱码一区二区三区2| 一级片'在线观看视频| 水蜜桃什么品种好| 日本黄色日本黄色录像| 大码成人一级视频| 不卡一级毛片| 欧美激情高清一区二区三区| 成人三级做爰电影| 777久久人妻少妇嫩草av网站| 蜜桃在线观看..| 日本av免费视频播放| 啦啦啦在线免费观看视频4| 十八禁高潮呻吟视频| 黄频高清免费视频| 天堂8中文在线网| 男人舔女人的私密视频| 啪啪无遮挡十八禁网站| 国产一区二区 视频在线| 在线亚洲精品国产二区图片欧美| 日韩视频一区二区在线观看| 成人18禁在线播放| 亚洲伊人色综图| 久久人妻熟女aⅴ| 夫妻午夜视频| 伊人久久大香线蕉亚洲五| 中文字幕另类日韩欧美亚洲嫩草| 国产精品一区二区在线不卡| 男女午夜视频在线观看| 在线观看免费日韩欧美大片| 香蕉久久夜色| 日韩欧美三级三区| 成人国语在线视频| 中文亚洲av片在线观看爽 | 日韩有码中文字幕| 国产精品1区2区在线观看. | 中亚洲国语对白在线视频| 久久人人爽av亚洲精品天堂| 成人国产一区最新在线观看| av超薄肉色丝袜交足视频| 国产精品一区二区免费欧美| 久久精品国产a三级三级三级| 亚洲欧美一区二区三区黑人| xxxhd国产人妻xxx| 午夜福利乱码中文字幕| 18在线观看网站| 777米奇影视久久| 黄色丝袜av网址大全| 国产av国产精品国产| www.精华液| 窝窝影院91人妻| 久久精品国产综合久久久| 97在线人人人人妻| 一边摸一边抽搐一进一小说 | 999久久久国产精品视频| 日韩欧美免费精品| 男女高潮啪啪啪动态图| av天堂在线播放| 夜夜爽天天搞| 成人精品一区二区免费| a级片在线免费高清观看视频| 日本一区二区免费在线视频| 少妇 在线观看| 亚洲中文av在线| 一本—道久久a久久精品蜜桃钙片| 国产精品99久久99久久久不卡| 91精品三级在线观看| 欧美黄色淫秽网站| 悠悠久久av| 男女之事视频高清在线观看| 亚洲色图综合在线观看| 考比视频在线观看| 亚洲精品乱久久久久久| 飞空精品影院首页| 亚洲精华国产精华精| xxxhd国产人妻xxx| 考比视频在线观看| 久久久久视频综合| 12—13女人毛片做爰片一| 大陆偷拍与自拍| 亚洲avbb在线观看| 乱人伦中国视频| 久久免费观看电影| 午夜91福利影院| 国产成人影院久久av| 精品人妻1区二区| 一本一本久久a久久精品综合妖精| 久久人妻福利社区极品人妻图片| 日本av免费视频播放| 一级毛片女人18水好多| 女同久久另类99精品国产91| 精品一品国产午夜福利视频| 久久国产精品大桥未久av| 国产熟女午夜一区二区三区| 久久午夜综合久久蜜桃| 亚洲五月色婷婷综合| 久久狼人影院| 国产亚洲精品久久久久5区| 国产单亲对白刺激| 亚洲自偷自拍图片 自拍| 女人高潮潮喷娇喘18禁视频| 黄色视频不卡| 亚洲一卡2卡3卡4卡5卡精品中文| 国产亚洲午夜精品一区二区久久| 99九九在线精品视频| av天堂在线播放| 叶爱在线成人免费视频播放| e午夜精品久久久久久久| 亚洲av片天天在线观看| 欧美+亚洲+日韩+国产| a级片在线免费高清观看视频| 久久天躁狠狠躁夜夜2o2o| www.熟女人妻精品国产| 久久久久网色| 大型黄色视频在线免费观看| 国产欧美日韩精品亚洲av| 大香蕉久久成人网| 91老司机精品| 亚洲三区欧美一区| 日本精品一区二区三区蜜桃| 精品福利永久在线观看| 亚洲成a人片在线一区二区| 成人精品一区二区免费| 国产深夜福利视频在线观看| 日本黄色日本黄色录像| 成人手机av| 青青草视频在线视频观看| 超碰97精品在线观看| 他把我摸到了高潮在线观看 | 9色porny在线观看| 欧美成人午夜精品| 18禁国产床啪视频网站| 欧美日韩成人在线一区二区| 在线播放国产精品三级| 亚洲色图av天堂| 国产欧美日韩精品亚洲av| 欧美另类亚洲清纯唯美| 91成人精品电影| 最新的欧美精品一区二区| 国产精品久久久久成人av| 91精品三级在线观看| 免费久久久久久久精品成人欧美视频| 波多野结衣av一区二区av| 最黄视频免费看| 亚洲欧美精品综合一区二区三区| 精品国产乱码久久久久久男人| 怎么达到女性高潮| 国产伦理片在线播放av一区| 免费久久久久久久精品成人欧美视频| 国产精品麻豆人妻色哟哟久久| 免费观看a级毛片全部| 国产精品99久久99久久久不卡| 成在线人永久免费视频| 啦啦啦在线免费观看视频4| 国产精品国产av在线观看| 国产成人免费观看mmmm| 亚洲 国产 在线| 黄片大片在线免费观看| 一二三四在线观看免费中文在| 黑丝袜美女国产一区| 青青草视频在线视频观看| 在线观看免费视频网站a站| 菩萨蛮人人尽说江南好唐韦庄| 建设人人有责人人尽责人人享有的| 满18在线观看网站| 日韩免费高清中文字幕av| 国产在线视频一区二区| 国产成人av激情在线播放| 少妇精品久久久久久久| 日韩一卡2卡3卡4卡2021年| 在线观看www视频免费| 男女下面插进去视频免费观看| 亚洲av片天天在线观看| 亚洲成人国产一区在线观看| 久久午夜亚洲精品久久| 色视频在线一区二区三区| 中文字幕制服av| 大香蕉久久网| 首页视频小说图片口味搜索| 国产日韩欧美亚洲二区| 免费不卡黄色视频| 女人高潮潮喷娇喘18禁视频| 亚洲伊人久久精品综合| 大型黄色视频在线免费观看| 成年人黄色毛片网站| 亚洲专区国产一区二区| 亚洲国产欧美一区二区综合| 免费日韩欧美在线观看| 中文字幕人妻丝袜制服| 欧美日韩国产mv在线观看视频| 欧美激情极品国产一区二区三区| 99精品在免费线老司机午夜| 久久久水蜜桃国产精品网| 国产人伦9x9x在线观看| 亚洲视频免费观看视频| www.熟女人妻精品国产| 日韩欧美一区二区三区在线观看 | 欧美久久黑人一区二区| 国产欧美日韩一区二区三| 亚洲精品国产一区二区精华液| 免费黄频网站在线观看国产| 十分钟在线观看高清视频www| 嫩草影视91久久| 午夜精品久久久久久毛片777| 黄网站色视频无遮挡免费观看| 欧美精品人与动牲交sv欧美| 九色亚洲精品在线播放| 久久天躁狠狠躁夜夜2o2o| 在线十欧美十亚洲十日本专区| 一个人免费在线观看的高清视频| 啦啦啦在线免费观看视频4| 人人妻人人爽人人添夜夜欢视频| 欧美久久黑人一区二区| 中文字幕人妻丝袜一区二区| 亚洲精品久久午夜乱码| 精品少妇黑人巨大在线播放| 精品一区二区三区av网在线观看 | 欧美一级毛片孕妇| 亚洲成人免费av在线播放| 黄色片一级片一级黄色片| 少妇粗大呻吟视频| 亚洲一卡2卡3卡4卡5卡精品中文| 久久久久久人人人人人| 99久久国产精品久久久| 99国产极品粉嫩在线观看| 日韩有码中文字幕| 亚洲av成人不卡在线观看播放网| 菩萨蛮人人尽说江南好唐韦庄| 亚洲全国av大片| 久久久国产成人免费| 热re99久久精品国产66热6| 18在线观看网站| 在线观看66精品国产| 亚洲国产欧美一区二区综合| 欧美精品亚洲一区二区| 在线看a的网站| 一本大道久久a久久精品| 久久中文字幕人妻熟女| 两性午夜刺激爽爽歪歪视频在线观看 | 亚洲av日韩精品久久久久久密| 久久人人爽av亚洲精品天堂| 捣出白浆h1v1| 99国产精品99久久久久| 国产精品自产拍在线观看55亚洲 | 国产无遮挡羞羞视频在线观看| 亚洲av国产av综合av卡| 亚洲成人免费电影在线观看| tocl精华| 久久久久久久国产电影| 一区二区av电影网| 精品亚洲成国产av| 国产成人欧美在线观看 | 18禁美女被吸乳视频| 日韩免费高清中文字幕av| 丝袜美腿诱惑在线| 欧美人与性动交α欧美软件| cao死你这个sao货| 精品少妇一区二区三区视频日本电影| 丝袜人妻中文字幕| 中文亚洲av片在线观看爽 | 黑人欧美特级aaaaaa片| 日本精品一区二区三区蜜桃| 搡老乐熟女国产| 国产不卡av网站在线观看| 男女高潮啪啪啪动态图| 国产xxxxx性猛交| 国产精品美女特级片免费视频播放器 | 男女高潮啪啪啪动态图| 欧美老熟妇乱子伦牲交| 免费在线观看日本一区| 午夜福利免费观看在线| 一本久久精品| 肉色欧美久久久久久久蜜桃| 五月天丁香电影| 国产成人精品无人区| 国产精品国产高清国产av | 久久久久精品国产欧美久久久| 久久亚洲精品不卡| 国产高清国产精品国产三级| 亚洲成av片中文字幕在线观看| 怎么达到女性高潮| 露出奶头的视频| 女人爽到高潮嗷嗷叫在线视频| 欧美午夜高清在线| 日韩中文字幕视频在线看片| 少妇精品久久久久久久| √禁漫天堂资源中文www| av视频免费观看在线观看| 亚洲精品国产色婷婷电影| 欧美激情 高清一区二区三区| 精品一区二区三区av网在线观看 | 制服人妻中文乱码| 在线天堂中文资源库| 成人精品一区二区免费| 免费女性裸体啪啪无遮挡网站| 在线 av 中文字幕| 日韩大码丰满熟妇| 久久久久久人人人人人| 操美女的视频在线观看| 日韩 欧美 亚洲 中文字幕| √禁漫天堂资源中文www| 亚洲精品一二三| 天堂8中文在线网| 欧美激情极品国产一区二区三区| 欧美乱码精品一区二区三区| 国产福利在线免费观看视频| 午夜激情av网站| 欧美亚洲 丝袜 人妻 在线| 婷婷成人精品国产| 午夜老司机福利片| 亚洲第一av免费看| 国产人伦9x9x在线观看| 久久久水蜜桃国产精品网| 丝袜美足系列| 国产成人精品久久二区二区91| 一级a爱视频在线免费观看| 大香蕉久久成人网| www.自偷自拍.com| 久久久久国产一级毛片高清牌| 午夜福利乱码中文字幕| 黄色a级毛片大全视频| 亚洲中文av在线| 一本综合久久免费| 亚洲黑人精品在线| 成人特级黄色片久久久久久久 | 国产日韩欧美视频二区| 亚洲 欧美一区二区三区| 国产三级黄色录像| 天天躁日日躁夜夜躁夜夜| 最近最新中文字幕大全免费视频| 黑人巨大精品欧美一区二区mp4| 国产欧美日韩一区二区精品| 国产男靠女视频免费网站| 大码成人一级视频| 久久中文看片网| 日日摸夜夜添夜夜添小说| 国产有黄有色有爽视频| 少妇 在线观看| 波多野结衣一区麻豆| 老熟女久久久| 国产激情久久老熟女| 久久av网站| 国产免费视频播放在线视频| 免费观看人在逋| 久热爱精品视频在线9| 制服诱惑二区| 亚洲成人免费av在线播放| 国产成人啪精品午夜网站| 宅男免费午夜| 成人精品一区二区免费| 国产福利在线免费观看视频| 亚洲精品在线美女| 亚洲熟女精品中文字幕| 亚洲精品国产区一区二| 99精品久久久久人妻精品| 91成人精品电影| 日韩大片免费观看网站| 91成人精品电影| 久久久久久亚洲精品国产蜜桃av| 日韩欧美一区二区三区在线观看 | netflix在线观看网站| 在线看a的网站| 日韩中文字幕欧美一区二区| 亚洲人成伊人成综合网2020| 久久精品aⅴ一区二区三区四区| 91老司机精品| 久久人妻福利社区极品人妻图片| 亚洲欧美日韩另类电影网站| 亚洲人成电影免费在线| 欧美精品高潮呻吟av久久| 久久ye,这里只有精品| 国产三级黄色录像| a级毛片黄视频| 日本av免费视频播放| 精品人妻1区二区| 法律面前人人平等表现在哪些方面| www.999成人在线观看| 女人爽到高潮嗷嗷叫在线视频| av超薄肉色丝袜交足视频| 亚洲国产欧美日韩在线播放| 久久久久国产一级毛片高清牌| 欧美老熟妇乱子伦牲交| 精品国产国语对白av| 母亲3免费完整高清在线观看| 精品国产亚洲在线| 日本撒尿小便嘘嘘汇集6| 久久精品国产亚洲av香蕉五月 | 国产深夜福利视频在线观看| 法律面前人人平等表现在哪些方面| 精品人妻熟女毛片av久久网站| 美国免费a级毛片| 午夜福利在线观看吧| 少妇粗大呻吟视频| 午夜激情久久久久久久| 国产在线视频一区二区| 欧美国产精品va在线观看不卡| 汤姆久久久久久久影院中文字幕| 亚洲精品在线观看二区| 国产亚洲午夜精品一区二区久久| 五月开心婷婷网| 岛国在线观看网站| 99精品久久久久人妻精品| 亚洲第一av免费看| 一级毛片女人18水好多| 高清黄色对白视频在线免费看| 欧美日韩黄片免| 国产单亲对白刺激| 国产精品二区激情视频| 不卡av一区二区三区| 欧美成狂野欧美在线观看| 亚洲美女黄片视频| 黄片播放在线免费| 中文字幕人妻丝袜制服| 不卡一级毛片| 大码成人一级视频| 亚洲熟妇熟女久久| 五月天丁香电影| 久久久国产成人免费| 欧美在线黄色| 啦啦啦在线免费观看视频4| 国产成人精品无人区| 12—13女人毛片做爰片一| 精品午夜福利视频在线观看一区 | av天堂久久9| 亚洲性夜色夜夜综合| 日韩一区二区三区影片| 热99re8久久精品国产| 久久久精品免费免费高清| 男男h啪啪无遮挡| 国产精品久久久久久人妻精品电影 | 亚洲黑人精品在线| 欧美+亚洲+日韩+国产| 一级毛片电影观看| 黄色视频,在线免费观看| 黄网站色视频无遮挡免费观看| 国产伦理片在线播放av一区| 美国免费a级毛片| 在线观看免费日韩欧美大片| 一个人免费看片子| 午夜免费鲁丝| 99国产精品一区二区蜜桃av | 色94色欧美一区二区| 一级a爱视频在线免费观看| 麻豆国产av国片精品| 丁香欧美五月| 啦啦啦在线免费观看视频4| av网站免费在线观看视频| 自线自在国产av| 黑人猛操日本美女一级片| 国产成人精品在线电影| 又紧又爽又黄一区二区| 最新在线观看一区二区三区| 亚洲久久久国产精品| 女性被躁到高潮视频| 最黄视频免费看| 国产av一区二区精品久久| 少妇猛男粗大的猛烈进出视频| 日韩人妻精品一区2区三区| 国产在线一区二区三区精| 伦理电影免费视频| 国产有黄有色有爽视频| 黄色 视频免费看| www.精华液| 黑丝袜美女国产一区| 亚洲精品在线美女| 2018国产大陆天天弄谢| 国产精品一区二区在线观看99| 在线观看免费视频日本深夜| 国产精品亚洲av一区麻豆| 另类亚洲欧美激情| 91九色精品人成在线观看| 国产精品成人在线| 国产又色又爽无遮挡免费看| 国产成人精品久久二区二区免费| 黄片大片在线免费观看| 欧美在线黄色| 国产成人精品久久二区二区免费| 国产精品偷伦视频观看了| 欧美在线黄色| 亚洲中文日韩欧美视频| 国产黄频视频在线观看| 日韩视频在线欧美| 亚洲国产欧美网| 免费人妻精品一区二区三区视频| 丁香欧美五月| 一区二区三区国产精品乱码| 超色免费av| 国产不卡一卡二| 黑人操中国人逼视频| 青草久久国产| 午夜福利乱码中文字幕| 一区二区三区乱码不卡18| 久9热在线精品视频| 午夜福利乱码中文字幕| 精品久久久精品久久久| 99香蕉大伊视频| 国产精品国产av在线观看| 欧美 亚洲 国产 日韩一| 美女主播在线视频| 蜜桃在线观看..| 五月天丁香电影| 日本撒尿小便嘘嘘汇集6| 捣出白浆h1v1| 黑人巨大精品欧美一区二区mp4| 亚洲国产欧美在线一区| 国产免费现黄频在线看|