• <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 ?

    LDNet: structure-focused lane detection based on line deformation①

    2022-10-22 02:23:54ZHANGJunWANGXingbinGUOBinglei
    High Technology Letters 2022年3期

    ZHANG Jun (張 軍), WANG Xingbin, GUO Binglei

    (*School of Computer Engineering, Hubei University of Arts and Science, Xiangyang 411053, P.R.China)

    (**Institute of Information Engineering, Chinese Academy of Science, Beijing 100093, P.R.China)

    Abstract Lane detection is a fundamental necessary task for autonomous driving. The conventional methods mainly treat lane detection as a pixel-wise segmentation problem, which suffers from the challenge of uncontrollable driving road environments and needs post-processing to abstract the lane parameters. In this work,a series of lines are used to represent traffic lanes and a novel line deformation network (LDNet) is proposed to directly predict the coordinates of lane line points. Inspired by the dynamic behavior of classic snake algorithms, LDNet uses a neural network to iteratively deform an initial lane line to match the lane markings. To capture the long and discontinuous structures of lane lines, 1D convolution in LDNet is used for structured feature learning along the lane lines.Based on LDNet, a two-stage pipeline is developed for lane marking detection: (1) initial lane line proposal to predict a list of lane line candidates, and (2) lane line deformation to obtain the coordinates of lane line points. Experiments show that the proposed approach achieves competitive performances on the TuSimple dataset while being efficient for real-time applications on a GTX 1650 GPU. In particular, the accuracy of LDNet with the annotated starting and ending points is up to 99.45%, which indicates the improved initial lane line proposal method can further enhance the performance of LDNet.

    Key words: autonomous driving, convolutional neural networks (CNNs), lane detection, line deformation

    0 Introduction

    With the rapid development of high-precision optics and electronic sensors, and computing capability,autonomous driving has received much attention in both academy and industry. In these systems, camera-based lane detection plays a critical role in the semantic understanding of the world around a vehicle[1-2]. It is challenging to perform accurate lane detection for the following reasons. First, lanes are usually thin and long curves travelling through the entire scenario, and have diverse patterns, such as solid, broken, splitting and merging. Furthermore, the driving road scenarios are complex, highly variable, and uncontrollable due to lighting/weather conditions.

    Traditional methods on lane detection adopt the hand-crafted features to identify lane markings. Features like color, intensity, gradient, edge, geometric shapes, and texture are widely used to describe the segments of lane markings. As these hand-crafted features are usually based on strong assumptions (e. g.,flat ground-planes) and lack high-level semantic information, these traditional methods have difficulty in detecting the lanes in complex situations. Thanks to the strong high-level semantic representation learning ability, recent convolutional neural networks (CNNs) have pushed lane detection to a new level. Most of these methods treat lane detection as a two-stage semantic segmentation task. In the first stage, a network is designed to classify each pixel in an image if it belongs to one of lanes or not. Post-processing in the second stage usually uses some curve-fitting strategy to filter the noise points or cluster the intermittent lane segments.Although the state-of-the-art methods already achieve great progress in lane detection, there are still some important and challenging problems to be addressed.Firstly, the segmentation-based methods with square shape kernels are hard to capture the thin and long curve property of lanes[3-5]. Due to ignoring the highlevel global semantic features (or contextual information), these methods often suffer from discontinuous and noisy detection results. Secondly, the desired output for autonomous driving is control-related parameters, i. e., vehicle lateral offset, turning angle and curvature. However, the outputs of most CNN-based lane detection methods are pixel-level lane instance masks in the image view. To fill this gap, some postprocessing procedures are required, e.g., inverse perspective mapping (IPM), and lane model fitting[6].

    In this paper, a structure-focused lane marking detection network, named line deformation network(LDNet), is proposed to address the problems. LDNet can better capture the long and discontinuous structures of lane lines and predict the coordinates of lane line points in an end-to-end manner. Inspired by the dynamic behavior of previous snake methods[7-10], LDNet takes initial lane lines as input and deforms them by regressing vertex-wise offsets. PickNp= 64 points along each line as the features of the line and apply the standard 1D convolutions on the point features. The 1D convolution kernel not only captures the features of each point but also the relationship from neighboring points. This enhances feature representation and focuses on learning of the long and discontinuous structures of lane markings. Based on LDNet, a pipeline is developed for lane detection. Given a set of initial lane lines,LDNet iteratively deforms them to match the lane markings and obtain the coordinates of lane line points.In this paper, the straight line between the starting point and ending point of each lane is used as its initial lane line. Define the starting point of one lane as the lane point closest to the bottom boundary of the image.As the starting points of different lanes in an image are usually far apart from each other according to the traffic rules, the heat map based keypoint estimation method is used for starting points detection. Different from the starting points, the ending points are converged together around the farthest point of the visible lane. Inspired by VPGNet[6]and CHEVP[8], which use vanishing point as a global geometric context to infer the location of lanes, a vanishing point prediction task is designed to estimate the locations of the ending points.

    In summary, this work has the following contributions.

    (1) A novel LDNet is proposed,which focuses on the long and discontinuous structure learning of lane markings and directly predicts the coordinates of lane line points.

    (2) Based on LDNet, the lane detection method is implemented by proposing the initial lane with two branches, in which heat map and probability map are used to predict lane lines’ starting point and ending points, respectively.

    (3) The proposed method achieves comparable performance with state-of-the-art methods in terms of both accuracy and speed on the TuSimple dataset. In particular, the accuracy of LDNet with ideal starting and ending points is up to 99.45%, which indicates improved initial lane line proposal method can further enhance the performance of the method.

    1 Related work

    1.1 Deep learning based lane detection

    Due to the strong representation learning ability,CNN based approaches have been used for lane marking detection. VPGNet[6]proposed a multi-task network guided by vanishing points for lane and road marking detection. LaneNet[11]used a lane edge proposal network for pixel-wise lane edge classification,and used a lane line localization network in stage two for lane line localization prediction. Neven et al.[12]regarded lane detection as an instance segmentation problem, in which a embedding branch disentangled the segmented lane pixels into different lane instances.Zhang et al.[13]established a framework that accomplished lane boundary segmentation and road area segmentation simultaneously. Ko et al.[14]first obtained the exact lane points by a confidence branch and an offset branch, then clustered the generated points by the point cloud instance segmentation method. To get rid of the perspective effect in the image, the inverse perspective mapping (IPM) was used by several methods[11,15-18]. Although these methods already achieved great progress in lane detection, there are still some important and challenging problems to be addressed.These methods often suffer discontinuous and noisy detection results due to the thinness of traffic lanes[19-22].Furthermore, post-processing is needed to filter the noise points and get the lane line information from pixel-level segmentation results[6,12,14,18].

    To deal with the first problem, several schemes have been proposed to capture richer scene features (or contextual information). Zhang et al.[23]added one convolutional gated recurrent unit (ConveGRU) in the encoder phase to learn more accurate low-level features. Hou et al.[24]adopted a self-attention distillation(SAD) approach to allow the network to exploit attention maps within the network itself during the training stage. Zou et al.[25]combined CNN and recurrent neural network (RNN) to infer lanes from multiple frames of a continuous driving scene. SALMNet[4]used a semantic-guided channel attention module to enhance features representation to the structures of lane markings and suppresses noisy features from the background, and a pyramid deformable convolution module to capture the structures of long and discontinuous lane markings. Message passing between pixels can help capture spatial structures of objects having long structure region and could be occluded. SCNN[3]and RESA[20]utilized slice-by-slice convolutions to enable message passing between pixels across rows and columns in the feature map. Though message passing improves the performance of segmentation results, the dense pixel-wise communication-based message passing required more computational cost. To collect more information concerning a whole lane boundary, Spin-Net[26]designed a novel convolution layer to allow the convolutions to be applied in multiple directions. Line-CNN[27]utilized a line proposal unit (LPU) to proposal potential lanes, which forced the system to learn the global feature representation of the entire traffic lanes.

    To avoid post-processing and predict the coordinates of lane line points directly, Chougule et al.[19]treated the lane detection and classification problems as CNN regression task, which relaxed per-pixel classification requirement to a few points along lane boundary. Qin et al.[5]and Yoo et al.[28]translated the lane detection problem into a row-wise classification task using global features. PRNet[22]and PolyLaneNet[21]are proposed to use polynomial curves to represent traffic lanes and then use a polynomial regression network to predict the polynomial coefficients of lanes. Gansbeke et al.[29]proposed a deep neural network that predicted a weights map like a segmentation output for each lane marking and used a differentiable leastsquares fitting module to directly regress the lane parameters. Instead of representing traffic lanes by curves, PointLaneNet[30]considered lane line as a series of points, and proposed the ConvLaneNet network to predict the lane line offset, start point and confidence. RONELD[31]first extracted lane points from the probability map outputs, followed by detecting curved and straight lanes before using weighted least squares linear regression on straight lanes to fix broken lane edges resulting from fragmentation of edge maps in real images.

    1.2 Snake algorithms

    The proposed method is inspired by the dynamic behavior of classic snake algorithms[7], which have been used for contour-based instance segmentation.Snake algorithms treat the coordinates of the vertices as a set of variables. By applying proper forces at the contour coordinates, the algorithms could push the contour to the object boundary. The implementation of these algorithms contains two stages. Firstly, the contour vertices for image representation are initialized. Then, the contour is deformed to the object boundary by learningbased methods. Recently, Ling et al.[9]followed the pipeline of traditional snake algorithms and used a graph convolutional neural network to predict vertexwise offsets for contour deformation. Instead of treating the contour as a general graph, deep snake[10]leveraged the cycle graph topology and introduced the circular convolution for efficient feature learning on a contour.Wang et al.[8]proposed a B-snake based lane model to describe a wider range of lane structure. In this work,a learning-based snake algorithm LDNet is implemented to deform the initial lane line to match the lane markings. As LDNet utilizes the features along the lane line to directly predict the coordinates of lane line points, it can solve the two problems mentioned above.

    2 Methodology

    In this paper, a novel lane detection is performed by deforming the initial lane lines to match lane markings in an image. Specifically, LDNet takes initial lane lines as input and predicts per-vertex offsets pointing to the lane boundary. The vertex offsets are predicted based on the features of lane line points extracted from the input image with a CNN backbone.

    2.1 Lane representation

    In general, lanes are drawn on roads with a shape of line or curve. It is improper to represent a lane with circular contour, which is used to represent compact object[9-10]. A line or curve can be accurately represented by a series of points, which can be obtained by using spline interpolation between points. A lane line can be determined by the following elements: starting point, ending point, and lane line center points. To facilitate the operation of LDNet, the number of lane points are kept fixed. As a result, the lane lines in an image are transformed into a learnable structured representation, as illustrated in Eq.(1).

    whereNis the number of points for each lane line,Lis the number of lanes in an image, and(x,y) is the coordinate of a lane point.

    Fig.1 An overview of the network architecture

    Fig.1 illustrates the overall network architecture of the work. It contains three modules: (1) a feature extraction backbone (subsection 2.2) that takes a single image as input and provides shared intermediate feature maps for the successive modules; (2) an initial lane proposal module (subsection 2. 3) which outputs the candidate initial lane lines; (3) the learning-based snake algorithm module LDNet (subsection 2.4) which predicts vertex-wise offsets between initial lane line points and their target points. The output of the network are the coordinates of lane line points. The system is fully end-to-end trainable with stochastic gradient descent.

    2.2 Backbone

    The function of the backbone network is to extract semantically meaningful features for the successive modules. Choose stacked hourglass network as the backbone for its efficiency and effectiveness. The input images with size 512 ×256 RGB are resized to a smaller size (e.g.,64 ×32) by the resizing layer, which contains a sequence of convolution layers. The resized image is fed to multiple hourglass modules, each including one encoder, one decoder, and two output branches. The intermediate predictions and features output from the previous hourglass stage are integrated together to implement intermediate supervision. The total loss of the network is the sum of the loss on those hourglass modules.

    2.3 Initial lane proposal module

    Evidently,lane lines start from the boundary (bottom or left or right) of an image and converge together due to the perspective effect. Although lane line is not always straight, the straight line between its starting point and ending point could be used to represent its direction and range. Here the straight line between the starting point and ending point of each lane is used as its initial position. The lane line initialization task is transformed into starting points and ending points detection problems. As the starting points of different lane lines in an image are usually far apart from each other, the heatmap based keypoint estimation method can be used to predict them.

    The ending points of traffic lanes in an image are often close to each other due to perspective projection,it is difficult to accurately localize them by the same method of starting point detection. Inspired by VPGNet[6]and CHEVP[8], which use vanishing point as a global geometric context to infer the location of lanes,a vanishing point prediction task is designed to estimate the locations of the ending points. As shown in Fig.1,these two subtasks are implemented by two branches which share the input features. The pseudocode of the initial lane line proposal module is given in Algorithm 1.

    ?

    To predict starting points, a network head that consists of two 1 ×1 convolution layers is designed to transfer the feature maps into 1 channel heatmap.Peaks in this heatmap correspond to lane starting points. Training the starting point prediction branch follows Law and Den[32]. For each ground truth starting pointp, a low-resolution equivalent^p=pS」 is computed, whereSis the output stride. Then all ground truth starting points are splatted onto a heatmap by using a Gaussian kernelYxy,which is defined in Eq.(2).

    For ending point location prediction, a network head is designed that consists of two 1 ×1 convolution layers to transfer the feature maps into 1 channel probability map and estimate the location and width of vanishing point in the input image. Firstly, the lane point map is extracted by picking points more than an adaptive confidence thresholdT. This confidence threshold is selected based on the confidence point in the probability map outputs. Then,the lane point map is scanned from the first row. Here the width between the first and the last lane line points is defined as the rangeRiof the lane points in rowi. After the first row with lane points is found, assume that this row is the vanishing row

    where lanes disappear. For robustness and to exclude low-confidence noise, its rangeRvis compared with the rang of its subsequent rowRs. IfRv <Rs/2,the subsequent row is set as the vanishing row, and its range is compared with its subsequent row. This process is repeated untilRv≥Rs/2 and the estimated ending points are generated, as defined in Eq.(3).

    2.4 Learning-based snake algorithm module

    The snake algorithm module takes a list of candidate lane lines along with the feature maps of the image from the backbone network as the input and predicts the per-point offsets pointing to the lane boundary. For each candidate lane line, the feature vector for each lane line point(x,y)l,nis first constructed. The input featurefl,nfor a point (x,y)l,nis the concatenation of learning-based features and the point coordinate:[F((x,y)l,n); (x,y)l,n] whereFdenotes the feature maps. The feature mapsFare obtained by applying a CNN backbone on the input image as shown in Fig.1. The feature for one point is computed using the bilinear interpolation at the vertex coordinate(x,y)l,n.The appended point coordinate is used to encode the spatial relationship among lane points.

    The concatenated feature vectors are passed into LDNet, which implements the learning-based snake algorithm. LDNet first predicts offsets based on the initial lane line points and then deforms the initial lane lines by point-wise adding the offsets to their point coordinates. The deformed lane lines can be used for the next iteration. The impact of inference iterations will be studied in subsection 3.2.

    Following the idea from deepsnake[10], the LDNet consists three parts: a feature learning block, a fusion block, and a prediction head, as shown in Fig.2. To adapt to the long and discontinuous structure of lane lines, the circular convolution of deep snake is replaced by a 1D convolution. The feature learning block is composed of 8 ‘Conv1d-Bn-ReLU’ layers and uses residual skip connections for all layers. In all experiments, the kernel size of 1D convolution is fixed to be nine. The fusion block aims to fuse the information across all lane points at multiple scales. It concatenates features from all layers in the feature learning block and forwards them through a 1 ×1 convolution layers followed by max pooling. The fused feature is then concatenated with the feature of each point. The prediction head applies three 1 ×1 convolution layers to the point features and outputs point-wise offsets between initial lane points and the target points, which are used to deform the initial lane.

    Fig.2 Architecture of LDNet

    3 Experiments

    In this section, the accuracy and efficiency of the method are demonstrated with extensive experiments.The following sections mainly focus on three aspects:experimental settings, ablations studies of the method,and results on TuSimple dataset.

    3.1 Implementation setup

    In the experiments, the input images are resized to 512 × 256 during the data augmentation process.Then the resized image is compressed into smaller size data by a resizing layer, which contains a sequence of convolution layers and max pooling layers. The out channel of the resizing layer is 128.

    Training strategy. For starting points detection,the training objective is a penalty-reduced pixel-wise logistic regression with focal loss[32]:

    For the interactive lane deformation, the loss function is defined in Eq.(6).

    whereλk,λpandλiterare loss coefficients. Setλk=0.5,λp= 1,andλiter= 10 in the experiments. Adam with weight decay 1e-5 is used as the optimizer to train the model and the learning rate is set to be 1e-3[35].The total number of training epochs is 500 for TuSimple dataset. All models are trained with PyTorch[36].

    Dataset. To evaluate proposed approach, TuSimple lane dataset[38]is used to conduct the experiment.Tusimple dataset is collected with good and medium weather conditions in highways. It consists of 3626 training and 2782 testing images. They are recorded on 2-lane/3-lane/4-lane or more highway roads, at different daytime. For training, randomly apply simple data augmentation methods like flip, translation, and adding shadow, which contribute to a more comprehensive dataset.

    Evaluation metrics. The main evaluation metric of TuSimple is accuracy, which is calculated as the average correct number of points per image. The accuracy is defined as the average correct number of points per image.

    where,Fpredis the number of wrongly predicted lanes;Npredis the number of predicted lanes;Mpredis the number of missed ground-truth lanes andNgtis the number of all ground-truth lanes.

    3.2 Ablation study

    Effectiveness of LDNet. To prove the effectiveness of LDNet, the annotated starting points and ending points extracted from the annotation of TuSimple dataset are used to avoid the influence of staring points and ending points prediction.

    The results of LDNet model with annotated starting points and ending points are shown in Table 1. The accuracy of the proposed method is above 99.2% across all inference iterations. Fig.3 illustrates qualitative results of LDNet with two iterations on TuSimple dataset.From Fig.3, it can be seen that LDNet with annotated starting points and ending points performs well for occluded lanes (a-d), curve lanes (e-h) and lanes in non-flat plane (i-l). Both the quantitative and qualitative results indicate that LDNet has a strong ability to deform proper initial lane line to match the lane markings. Therefore, LDNet can be applied in both online and offline scenarios, such as accurate lane detection,fast interactive lane annotation,and HD map modeling.

    In LDNet, the iteration number is an important hyper-parameter, which influences the model size and speed. Table 1 also shows the evaluation results of LDNet models with various iterations. The accuracy is up to 99.45% when LDNet has two iterations. However,adding more iterations does not further improve the performance, which shows that it might be harder to train the network with more iterations. In the following experiments, the iterations of LDNet are fixed to two.

    Table 1 Results of LDNet with different iterations. Here the initial lane lines are generated with the annotated starting points and ending points directly

    Fig.3 Examples of results from the TuSimple lane detection benchmark (the initial lane lines are generated with the annotated starting points and ending points, the first row shows cases lane markings are occluded by vehicles, the second row shows curve lanes,and the third row shows lanes in non-flat ground plane)

    Effects of output strideS. Output strideSis another main hyper-parameter for LDNet, which denotes how much the feature maps are scaled down relative to the input image size. In principle, bigger output stride represents more/higher semantic but spatially coarser.

    Table 2 shows the results of LDNet with different output strideS, and the following observations. First,with the increase of output stride, the accuracy drops from 96.87% to 95.87%. Second, for a large output stride (e. g., 8), the accuracy of LDNet with initial lane generated by the annotated starting points and ending points is less than 99%. For smaller output stride(e.g.,2 and 4), the accuracy is above 99.4%. This indicates that the spatial information is much more important than the semantics information for LDNet.Third, the accuracy of LDNet with initial lane generated bySP~andEP^is much higher than initial lane generated bySP^andEP~ across all output strides. Thus,ending point location prediction is the bottleneck of implementation. More accurate ending point estimation can further improve the accuracy of the LDNet model.

    Table 2 Ablation study of output stride

    3.3 Results

    The initial lane lines generated from detected starting points(SP~) and ending points(EP~) are illustrated in Fig.4. The results indicate that the straight line between starting point and ending point can represent the direction and range of lane, especially when the lane lines are straight (Fig.4(a-d)). This figure also indicates that the heatmap based keypoint estimation method could effectively detect the starting points,even when the starting points are occluded by vehicles.Though the vanishing point-based method could provide proper ending points in most cases, it cannot deal with the situation where the ending points are not converged together as shown in Fig.4(k).

    Fig.4 Initial lane lines generated from the predicted starting points and ending points

    Table 3 reports the performance comparison of the LDNet model against the previous representative method. To show the generalization of LDNet, models with Hourglass, ResNet and DLA are used as the backbone. For a fair comparison, their lane marking detection results reported in their paper or website are used directly. The proposed models are represented as “LDNet-S-B”, where S is the output stride and B denotes the used backbone. It can be seen that the proposed models with various backbones achieve comparable performance with state-of-the-art methods.

    The final detected lane lines corresponding to the initial lane lines proposed in Fig.4 are illustrated in Fig.5. As the initial lane lines nearly match the straight lane lines (Fig.4(a -d)), the proposed method can detect straight lane lines precisely,even when the lanes are occluded by vehicles (Fig.5(a-d)). Although the initial lane lines do not match the target lane lines well(Fig.4(e -l)), our LDNet model still accurately predicts the offset between the initial lane line points andthe target boundary points for curve lanes(Fig.5(e-h))and lanes in non-flat ground plane (Fig.5(i-l)). Most notably,the proposed model precisely predicts the curves when they are occluded by vehicles (Fig.5(e,h)).Comparing with Fig.3(g), the predicted lane lines in Fig.5(g) are longer than the predicted lane lines in Fig.3(g) whose initial lane lines are generated from the annotated starting points and ending points. These results indicate that LDNet model has a strong ability in capturing the structures of lane markings.

    Table 3 Performance of different methods on TuSimple

    Fig.5 Results of LDNet with two hourglass modules and S=4

    Table 4 compares proposed approach with other methods in terms of running time. The running time of proposed method is recorded with the average time for 100 runs. For 256 ×512 input images, proposed method runs at 32 FPS on a laptop with an Intel i7 2.60 GHz and GTX 1650 GPU. The performance of GTX 1080-Ti is 2. 7 times higher than that of GTX 1650[38]. The performance of Titan X is 0.81 times higher than that of GTX 1650[39].

    Table 4 Run-time performance of different methods

    Though the proposed model achieves comparable performance with state-of-the-art methods, it cannot predict the lane lines well in complex driving road scenarios shown in Fig.6. In these situations, the lanes are not converged together, thus the vanishing point based ending point prediction method cannot propose proper ending points and initial lane lines. How to improve the ending point prediction method and the initial lane line proposal method will be the future work.

    4 Conclusion

    This paper proposes to use a series of lines to represent traffic lane and proposes a novel Line Deformation Network (LDNet) to iteratively deform an initial lane line to match the lane boundary. Heatmap based keypoint estimation method and vanishing point prediction task are used to propose the initial lane lines.

    Fig.6 Results of LDNet for lanes not converged together. The left figures show the ground truth and the right figures show the detected lanes.

    The experimental results on TuSimple lane dataset show that the proposed method achieves comparable performance with state-of-the-art methods. The accuracy of LDNet with ideal starting points and ending points is up to 99. 4%. Although the proposed initial lane lines do not match the target lane lines well,the LDNet model still accurately predicts the offset between the initial lane line points and the target boundary points.This shows that the LDNet has a strong ability in capturing the structures of lane markings.

    av一本久久久久| 中文字幕人妻熟人妻熟丝袜美| 男女无遮挡免费网站观看| 夫妻性生交免费视频一级片| 视频区图区小说| 偷拍熟女少妇极品色| 精品一区二区三区视频在线| tube8黄色片| 三级经典国产精品| 又大又黄又爽视频免费| 亚洲成人手机| 97在线人人人人妻| 欧美三级亚洲精品| 日韩伦理黄色片| 色视频在线一区二区三区| 狂野欧美白嫩少妇大欣赏| 久久国产精品男人的天堂亚洲 | 大香蕉97超碰在线| 亚洲精品456在线播放app| 日韩成人伦理影院| 欧美亚洲 丝袜 人妻 在线| 国产精品三级大全| 亚洲人与动物交配视频| 熟女av电影| 亚洲一区二区三区欧美精品| 国产国拍精品亚洲av在线观看| 一级毛片我不卡| 亚洲图色成人| 欧美日韩一区二区视频在线观看视频在线| 99久久中文字幕三级久久日本| 国产免费视频播放在线视频| 亚洲av在线观看美女高潮| 女人久久www免费人成看片| 观看av在线不卡| 国产综合精华液| 亚洲国产精品成人久久小说| 久久久精品免费免费高清| 男女免费视频国产| 又粗又硬又长又爽又黄的视频| 美女福利国产在线 | 2022亚洲国产成人精品| 黄色怎么调成土黄色| 建设人人有责人人尽责人人享有的 | 丝瓜视频免费看黄片| 日韩,欧美,国产一区二区三区| 欧美精品一区二区免费开放| 国产精品女同一区二区软件| 中文字幕制服av| 免费看光身美女| 免费人妻精品一区二区三区视频| 噜噜噜噜噜久久久久久91| 久久精品夜色国产| 91精品一卡2卡3卡4卡| 日韩av在线免费看完整版不卡| 色哟哟·www| 下体分泌物呈黄色| 中国美白少妇内射xxxbb| 汤姆久久久久久久影院中文字幕| 免费观看性生交大片5| 男人狂女人下面高潮的视频| 网址你懂的国产日韩在线| 在线观看一区二区三区| 婷婷色综合www| 亚洲国产毛片av蜜桃av| 亚洲av欧美aⅴ国产| 2018国产大陆天天弄谢| 人妻系列 视频| 国产又色又爽无遮挡免| 精华霜和精华液先用哪个| 最近中文字幕高清免费大全6| 亚洲国产精品成人久久小说| 人妻一区二区av| av播播在线观看一区| 欧美一区二区亚洲| 欧美成人午夜免费资源| 久久久久久久久久人人人人人人| 色吧在线观看| 51国产日韩欧美| 人人妻人人看人人澡| 亚洲一区二区三区欧美精品| 丰满迷人的少妇在线观看| 久久99热6这里只有精品| 亚洲,欧美,日韩| 国产精品无大码| 校园人妻丝袜中文字幕| 欧美日韩亚洲高清精品| 在线天堂最新版资源| 亚州av有码| 国产片特级美女逼逼视频| 欧美97在线视频| 永久免费av网站大全| 有码 亚洲区| 最近2019中文字幕mv第一页| 人妻一区二区av| 婷婷色综合大香蕉| 婷婷色综合www| 国产免费又黄又爽又色| 亚洲真实伦在线观看| 久久综合国产亚洲精品| 久久久久久久久久久丰满| 亚洲成人av在线免费| 十分钟在线观看高清视频www | 亚洲欧美日韩另类电影网站 | 午夜激情久久久久久久| 国产亚洲一区二区精品| 色网站视频免费| 亚洲av福利一区| 在线免费十八禁| av视频免费观看在线观看| 国产精品国产三级专区第一集| 久久人人爽av亚洲精品天堂 | 中文精品一卡2卡3卡4更新| 熟女人妻精品中文字幕| 欧美激情极品国产一区二区三区 | 亚洲av中文字字幕乱码综合| 日本色播在线视频| 男男h啪啪无遮挡| 极品少妇高潮喷水抽搐| 国产女主播在线喷水免费视频网站| 大片电影免费在线观看免费| 超碰97精品在线观看| 精品酒店卫生间| 久久99热这里只有精品18| 大香蕉97超碰在线| 一个人看视频在线观看www免费| 日韩伦理黄色片| 亚洲va在线va天堂va国产| 免费少妇av软件| 91午夜精品亚洲一区二区三区| 91久久精品国产一区二区三区| 国产69精品久久久久777片| 亚洲国产日韩一区二区| 十分钟在线观看高清视频www | 国产精品蜜桃在线观看| 亚洲欧美精品自产自拍| 国产黄片视频在线免费观看| av在线app专区| 亚洲精品aⅴ在线观看| 丰满乱子伦码专区| 内地一区二区视频在线| 国产女主播在线喷水免费视频网站| 搡老乐熟女国产| 女的被弄到高潮叫床怎么办| 成人毛片60女人毛片免费| 中国国产av一级| 免费黄网站久久成人精品| 久久人妻熟女aⅴ| 激情五月婷婷亚洲| 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 | 成年免费大片在线观看| h日本视频在线播放| 精品午夜福利在线看| 在线免费观看不下载黄p国产| 亚洲久久久国产精品| 草草在线视频免费看| 99久国产av精品国产电影| 另类亚洲欧美激情| 国产成人91sexporn| 国产淫片久久久久久久久| 国产av一区二区精品久久 | 久久久久精品性色| 看免费成人av毛片| 亚洲精品乱久久久久久| 久久99热6这里只有精品| 黄色日韩在线| 中国国产av一级| 蜜桃久久精品国产亚洲av| 制服丝袜香蕉在线| 一二三四中文在线观看免费高清| 久热这里只有精品99| 国产午夜精品久久久久久一区二区三区| 91精品国产国语对白视频| 国产色爽女视频免费观看| 日韩一本色道免费dvd| 青青草视频在线视频观看| 精品亚洲乱码少妇综合久久| 国产精品偷伦视频观看了| 王馨瑶露胸无遮挡在线观看| 午夜视频国产福利| 久久99热这里只有精品18| 人妻少妇偷人精品九色| av福利片在线观看| 国产精品女同一区二区软件| 日韩电影二区| 少妇被粗大猛烈的视频| 久久久久国产精品人妻一区二区| 日韩三级伦理在线观看| 99热全是精品| 国产av码专区亚洲av| 日韩强制内射视频| 精品亚洲成国产av| 街头女战士在线观看网站| 王馨瑶露胸无遮挡在线观看| 中国美白少妇内射xxxbb| 美女国产视频在线观看| 一级av片app| 国产精品久久久久久久电影| 最近中文字幕高清免费大全6| 在线观看一区二区三区激情| 久久久久精品久久久久真实原创| 精品人妻熟女av久视频| 国产av码专区亚洲av| 中国国产av一级| 中文在线观看免费www的网站| 婷婷色麻豆天堂久久| 亚洲aⅴ乱码一区二区在线播放| 久久国产精品男人的天堂亚洲 | 99热这里只有是精品50| 成人黄色视频免费在线看| 中文字幕精品免费在线观看视频 | 春色校园在线视频观看| 伦精品一区二区三区| 搡女人真爽免费视频火全软件| 人妻少妇偷人精品九色| 肉色欧美久久久久久久蜜桃| 免费观看无遮挡的男女| 97超碰精品成人国产| 少妇人妻久久综合中文| 欧美成人精品欧美一级黄| 三级国产精品欧美在线观看| 一区二区三区四区激情视频| kizo精华| 国产欧美另类精品又又久久亚洲欧美| 欧美高清性xxxxhd video| 午夜福利高清视频| 婷婷色综合www| 内射极品少妇av片p| 在线观看三级黄色| 尤物成人国产欧美一区二区三区| 亚洲欧美精品专区久久| 水蜜桃什么品种好| .国产精品久久| 伦理电影免费视频| 99热这里只有是精品50| 3wmmmm亚洲av在线观看| 韩国av在线不卡| 精品国产露脸久久av麻豆| 成人18禁高潮啪啪吃奶动态图 | 日本欧美国产在线视频| 亚洲欧美日韩东京热| 亚洲内射少妇av| av线在线观看网站| 日本午夜av视频| 欧美精品亚洲一区二区| 国产av精品麻豆| 偷拍熟女少妇极品色| 夜夜爽夜夜爽视频| 美女国产视频在线观看| 九九爱精品视频在线观看| 婷婷色麻豆天堂久久| 亚洲精品国产av成人精品| 99国产精品免费福利视频| 高清av免费在线| 国产91av在线免费观看| 久久99蜜桃精品久久| 男人和女人高潮做爰伦理| 亚洲国产日韩一区二区| 一级毛片久久久久久久久女| 人妻 亚洲 视频| 久久久成人免费电影| 人妻系列 视频| 大话2 男鬼变身卡| 好男人视频免费观看在线| 人人妻人人爽人人添夜夜欢视频 | 久久久久久久精品精品| 久久亚洲国产成人精品v| 国产一区亚洲一区在线观看| 新久久久久国产一级毛片| 亚洲欧美日韩无卡精品| 熟女人妻精品中文字幕| 纵有疾风起免费观看全集完整版| 2021少妇久久久久久久久久久| 日韩,欧美,国产一区二区三区| 激情五月婷婷亚洲| 国产成人精品一,二区| 久久精品久久久久久噜噜老黄| 日韩一区二区三区影片| 亚洲精品aⅴ在线观看| av国产久精品久网站免费入址| 国产毛片在线视频| 亚洲精品色激情综合| 欧美成人午夜免费资源| 欧美日韩精品成人综合77777| 亚洲欧美日韩另类电影网站 | 成人无遮挡网站| 人人妻人人爽人人添夜夜欢视频 | 国产成人免费无遮挡视频| 欧美老熟妇乱子伦牲交| 国产精品欧美亚洲77777| 老师上课跳d突然被开到最大视频| 日本色播在线视频| 久久精品久久久久久久性| 天天躁夜夜躁狠狠久久av| 午夜福利影视在线免费观看| 美女脱内裤让男人舔精品视频| 高清不卡的av网站| www.av在线官网国产| 成人18禁高潮啪啪吃奶动态图 | 在线免费观看不下载黄p国产| 天堂中文最新版在线下载| 王馨瑶露胸无遮挡在线观看| 久久女婷五月综合色啪小说| 91精品伊人久久大香线蕉| 亚洲无线观看免费| 亚洲av电影在线观看一区二区三区| 欧美97在线视频| 高清日韩中文字幕在线| 涩涩av久久男人的天堂| 色婷婷av一区二区三区视频| 国产乱人偷精品视频| 久久精品国产自在天天线| 国产亚洲一区二区精品| av播播在线观看一区| 最近的中文字幕免费完整| 寂寞人妻少妇视频99o| 精品久久久精品久久久| 人人妻人人澡人人爽人人夜夜| 精品少妇久久久久久888优播| kizo精华| 成人漫画全彩无遮挡| h视频一区二区三区| 黑人高潮一二区| av在线app专区| 日本wwww免费看| 午夜精品国产一区二区电影| 3wmmmm亚洲av在线观看| 精品99又大又爽又粗少妇毛片| 国产真实伦视频高清在线观看| 久久精品夜色国产| 亚洲精品乱码久久久久久按摩| 五月天丁香电影| 一级毛片黄色毛片免费观看视频| 我要看黄色一级片免费的| 久久国产精品大桥未久av | 国产久久久一区二区三区| 精品人妻视频免费看| 麻豆乱淫一区二区| 蜜臀久久99精品久久宅男| 日韩 亚洲 欧美在线| 日韩欧美 国产精品| 久久国内精品自在自线图片| 国产精品久久久久久久电影| 大香蕉97超碰在线| 1000部很黄的大片| 久久久久精品性色| 亚洲激情五月婷婷啪啪| 人妻系列 视频| 婷婷色麻豆天堂久久| 少妇 在线观看| 久久精品人妻少妇| 99热全是精品| 91aial.com中文字幕在线观看| 国产一级毛片在线| av国产免费在线观看| 国产69精品久久久久777片| 只有这里有精品99| 一级av片app| 少妇人妻一区二区三区视频| 纯流量卡能插随身wifi吗| 观看av在线不卡| 欧美 日韩 精品 国产| 久久99蜜桃精品久久| 国产视频内射| 少妇人妻一区二区三区视频| 亚洲av不卡在线观看| 一区在线观看完整版| 蜜桃亚洲精品一区二区三区| 少妇丰满av| 成年av动漫网址| 日韩伦理黄色片| 午夜精品国产一区二区电影| 久久久成人免费电影| 蜜桃久久精品国产亚洲av| 亚洲成人中文字幕在线播放| 国产爽快片一区二区三区| 日韩伦理黄色片| 中文字幕免费在线视频6| 国产一区二区在线观看日韩| 久久精品国产亚洲av天美| 久久久久精品久久久久真实原创| 精品亚洲乱码少妇综合久久| 国产欧美另类精品又又久久亚洲欧美| 在线观看免费视频网站a站| 亚洲国产欧美人成| 18+在线观看网站| 激情 狠狠 欧美| 在线播放无遮挡| 久久人人爽av亚洲精品天堂 | 日本一二三区视频观看| 国产免费一级a男人的天堂| 国产片特级美女逼逼视频| 亚洲av欧美aⅴ国产| 狂野欧美激情性bbbbbb| 国产精品成人在线| 极品教师在线视频| 菩萨蛮人人尽说江南好唐韦庄| 又爽又黄a免费视频| 国产精品国产三级国产av玫瑰| 久久久a久久爽久久v久久| 777米奇影视久久| 久久婷婷青草| 在线观看国产h片| 在线精品无人区一区二区三 | 寂寞人妻少妇视频99o| 国产精品久久久久久精品电影小说 | 男男h啪啪无遮挡| 国产男女超爽视频在线观看| 大香蕉97超碰在线| 男女边摸边吃奶| 日本黄色片子视频| 3wmmmm亚洲av在线观看| 亚洲经典国产精华液单| 春色校园在线视频观看| 日韩人妻高清精品专区| 久久精品久久久久久久性| 亚洲成人中文字幕在线播放| 又爽又黄a免费视频| 日韩三级伦理在线观看| 欧美变态另类bdsm刘玥| 国产综合精华液| 欧美激情极品国产一区二区三区 | 男女边摸边吃奶| 麻豆精品久久久久久蜜桃| 成人黄色视频免费在线看| 熟妇人妻不卡中文字幕| 一本—道久久a久久精品蜜桃钙片| 日本欧美视频一区| 国产高清国产精品国产三级 | 高清欧美精品videossex| 午夜福利网站1000一区二区三区| 久久精品国产亚洲网站| av国产免费在线观看| 老熟女久久久| 欧美精品国产亚洲| 欧美日本视频| 国产av国产精品国产| 国产精品一区二区在线观看99| 午夜福利影视在线免费观看| 日韩一本色道免费dvd| 热re99久久精品国产66热6| 国产一区二区在线观看日韩| 最近最新中文字幕大全电影3| 高清av免费在线| 亚洲精品一二三| 日本av手机在线免费观看| 视频中文字幕在线观看| av福利片在线观看| 亚洲欧美中文字幕日韩二区| 国产午夜精品久久久久久一区二区三区| 18禁动态无遮挡网站| 欧美bdsm另类| 一个人免费看片子| 亚洲欧洲日产国产| 亚洲久久久国产精品| 99久久综合免费| 日本黄大片高清| 少妇熟女欧美另类| 精品久久久久久久久亚洲| 久久精品国产a三级三级三级| 大码成人一级视频| 亚洲欧美成人精品一区二区| 女人十人毛片免费观看3o分钟| 九九在线视频观看精品| 黄色视频在线播放观看不卡| 国产真实伦视频高清在线观看| 老熟女久久久| 亚洲内射少妇av| 国产伦在线观看视频一区| av卡一久久| 亚洲第一区二区三区不卡| 亚洲欧美日韩另类电影网站 | 国产精品国产三级国产专区5o| 婷婷色综合大香蕉| 99热这里只有是精品50| 深爱激情五月婷婷| 亚洲经典国产精华液单| 国产 精品1| 欧美日韩视频精品一区| 国产日韩欧美亚洲二区| 夫妻午夜视频| 国产成人精品福利久久| 国产老妇伦熟女老妇高清| 日产精品乱码卡一卡2卡三| 熟妇人妻不卡中文字幕| 九色成人免费人妻av| 国产伦在线观看视频一区| 国产免费一区二区三区四区乱码| 国产永久视频网站| 欧美极品一区二区三区四区| 啦啦啦在线观看免费高清www| kizo精华| 欧美日韩国产mv在线观看视频 | 国产欧美日韩一区二区三区在线 | 国内揄拍国产精品人妻在线| 精品一区二区三卡| 观看美女的网站| 高清午夜精品一区二区三区| 人妻系列 视频| 亚洲怡红院男人天堂| 中文字幕久久专区| 观看美女的网站| 全区人妻精品视频| 麻豆乱淫一区二区| 久久久久久伊人网av| 国产午夜精品一二区理论片| 免费观看性生交大片5| 色网站视频免费| 欧美xxxx性猛交bbbb| 欧美bdsm另类| av专区在线播放| 欧美zozozo另类| 成人毛片60女人毛片免费| 一级av片app| 久久久久视频综合| 麻豆成人av视频| 国语对白做爰xxxⅹ性视频网站| 91午夜精品亚洲一区二区三区| 亚洲精品国产av蜜桃| 精品午夜福利在线看| 欧美激情国产日韩精品一区| 国内精品宾馆在线| 黄片无遮挡物在线观看| 国产大屁股一区二区在线视频| 国产精品爽爽va在线观看网站| 一个人看的www免费观看视频| 精品人妻熟女av久视频| 亚洲av成人精品一二三区| 在线观看免费视频网站a站| 久久6这里有精品| 日韩欧美 国产精品| 亚洲国产成人一精品久久久| 日本猛色少妇xxxxx猛交久久| 久久久久国产网址| 丝袜脚勾引网站| 黄片wwwwww| 亚洲丝袜综合中文字幕| 91久久精品国产一区二区成人| 亚洲av欧美aⅴ国产| 精品99又大又爽又粗少妇毛片| 亚洲内射少妇av| 少妇高潮的动态图| 国产人妻一区二区三区在| 26uuu在线亚洲综合色| 久久这里有精品视频免费| 国产成人a∨麻豆精品| 欧美精品人与动牲交sv欧美| 亚洲不卡免费看| 中文精品一卡2卡3卡4更新| 看免费成人av毛片| 精品人妻熟女av久视频| 国产伦精品一区二区三区视频9| 日韩一区二区视频免费看| 卡戴珊不雅视频在线播放| 男女啪啪激烈高潮av片| 亚洲精品一区蜜桃| 中国国产av一级| 免费观看av网站的网址| 欧美3d第一页| 国产国拍精品亚洲av在线观看| 精品酒店卫生间| 欧美xxxx性猛交bbbb| kizo精华| 国产精品久久久久成人av| 亚洲成人中文字幕在线播放| h视频一区二区三区| 99久久综合免费| 亚洲真实伦在线观看| 国产精品麻豆人妻色哟哟久久| 免费观看性生交大片5| av网站免费在线观看视频| 99热这里只有精品一区| 亚洲精品国产成人久久av| 插逼视频在线观看| 妹子高潮喷水视频| 亚洲av福利一区| 国产精品.久久久| 在线免费观看不下载黄p国产| av免费在线看不卡| 国产成人freesex在线| 乱系列少妇在线播放| 亚洲综合色惰| 内射极品少妇av片p| 人妻少妇偷人精品九色| 99热网站在线观看| 精品少妇黑人巨大在线播放| 成年人午夜在线观看视频| 亚洲欧洲国产日韩| 中文资源天堂在线| 一本一本综合久久| 亚洲av成人精品一二三区| 熟女av电影| 成年人午夜在线观看视频| 国产精品秋霞免费鲁丝片| 日日摸夜夜添夜夜添av毛片| 亚洲av免费高清在线观看| 日韩中字成人| 国产精品人妻久久久久久| 我要看黄色一级片免费的| 免费高清在线观看视频在线观看| 国产91av在线免费观看| 亚洲国产欧美在线一区| 97超碰精品成人国产| 免费观看无遮挡的男女| 日本-黄色视频高清免费观看| 三级国产精品片| 亚洲美女视频黄频| 精品久久久噜噜| 久久综合国产亚洲精品| 三级经典国产精品| 亚洲国产色片| 精华霜和精华液先用哪个| av在线老鸭窝| a级毛片免费高清观看在线播放| 内射极品少妇av片p|