• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Traffic Sign Recognition for Autonomous Vehicle Using Optimized YOLOv7 and Convolutional Block Attention Module

    2023-12-12 15:50:06KuppusamySanjayDeepashreeandIwendi
    Computers Materials&Continua 2023年10期

    P.Kuppusamy,M.Sanjay,P.V.Deepashree and C.Iwendi

    1School of Computer Science and Engineering,VIT-AP University,Andhra Pradesh,522237,India

    2School of Creative Technology,University of Bolton,Manchester,BL3 5AB,UK

    ABSTRACT The infrastructure and construction of roads are crucial for the economic and social development of a region,but traffic-related challenges like accidents and congestion persist.Artificial Intelligence(AI)and Machine Learning(ML)have been used in road infrastructure and construction,particularly with the Internet of Things(IoT)devices.Object detection in Computer Vision also plays a key role in improving road infrastructure and addressing trafficrelated problems.This study aims to use You Only Look Once version 7(YOLOv7),Convolutional Block Attention Module(CBAM),the most optimized object-detection algorithm,to detect and identify traffic signs,and analyze effective combinations of adaptive optimizers like Adaptive Moment estimation (Adam),Root Mean Squared Propagation(RMSprop)and Stochastic Gradient Descent(SGD)with the YOLOv7.Using a portion of German traffic signs for training,the study investigates the feasibility of adopting smaller datasets while maintaining high accuracy.The model proposed in this study not only improves traffic safety by detecting traffic signs but also has the potential to contribute to the rapid development of autonomous vehicle systems.The study results showed an impressive accuracy of 99.7%when using a batch size of 8 and the Adam optimizer.This high level of accuracy demonstrates the effectiveness of the proposed model for the image classification task of traffic sign recognition.

    KEYWORDS Object detection;traffic sign detection;YOLOv7;convolutional block attention module;road sign detection;Adam

    1 Introduction

    Infrastructure and construction of roads in any geographical area play a pivotal role in the economic and social development of the region,as it connects people to business and allows the movement of locomotives and services.One of the present-day primary challenges relating to road infrastructure is accidents,and other traffic-related concerns like traffic congestion,restricted infrastructure capacity,low maintenance of roads,etc.[1,2].Classically,human conception and past experiences have guided the progress of road infrastructure.However,as technology has become ubiquitous,and owing to advancements in automobile-related technologies such as self-parking systems,self-driving cars,fully autonomous systems,etc.,all of which are essentially categorized under the umbrella of Autonomous Driving Systems(ADS).There has been a significant increase in the usage of AI,and its sub-domains in accomplishing some cardinal tasks in ADS.An evaluative study on Deep Neural Networks(DNN)for Traffic Sign Detection (TSD),throws some light on how the detection of traffic signs is an indispensable study because these detection systems encompass anchor components required for safety and support in ADS [3].The IoT devices are utilized to gather the data from the environment,and ML analyses the data to solve the challenges in traffic management systems.The traffic management system contains three layers such as data acquisition,network transmission,and application.The data acquisition is done via sensors,cameras,video monitoring,and online monitoring.The collected data is transmitted over the network using Bluetooth,Wi-Fi (Wireless Fidelity),mobile network,etc.Finally,AI and ML play a major role in the analysis,visualize the analyzed outputs,and derive the systems based on the outputs like ADS [4].The study using ML showcased the latest and most advanced techniques for monitoring construction progress,including methods for collecting data,retrieving information,estimating progress,and presenting the results visually.Along similar lines,AI and ML are used in many more traffic-related issues[5].A review of traffic congestion prediction using AI described the probabilistic reasoning models like fuzzy logic,Hidden Markov Model(HMM),the Bayesian network,Support Vector Machine(SVM),Artificial Neural Networks(ANN),Decision Trees,etc.,Deep Learning(DL)algorithms like Convolutional Neural Networks(CNN),Long Short-Term Memory(LSTM),are used for short-term traffic congestion prediction[6].AI and ML-based incident detectors in Road Transport Systems(RTS)discussed dire problems and plausible solutions for reducing traffic accidents that enhanced the automatic incident system detectors[7].

    The technology development in computer vision plays a key role with goals revolving around improving road infrastructure like road accidents,traffic congestion,etc.Object detection is a subfield of computer vision that uses various DL architectures for recognizing and classifying objects.A comparative analysis of CNN-based object detection algorithms shows YOLOv3 is the fastest,and performs best overall outperforming Single-Shot Detector(SSD),and Faster Region-based CNN(RCNN) [8].However,it is also highlighted that the choice of the algorithm may be dependent upon the specific situation or problem that needs to be solved.For instance,R-CNN works best for small datasets that do not require real-time video outputs,whereas YOLO works best for object detection in the live environment.YOLOv4 runs twice as fast as EfficientNet,with an Average Precision(AP)of 14%more than YOLOv3.The YOLOv7 algorithm surpasses all the well-known real-time objectdetection algorithms concerning AP at 56.8%,and speed with a maximum range of 160 FPS[9].So far,research on the detection of traffic signs has been done using several versions of YOLO,and other object-detection algorithms.

    This study aims to use the fairly latest version of the most optimized object-detection algorithm YOLOv7 to detect and identify traffic signs.This study also tries to dive deep into analyzing effective combinations of adaptive optimizers like Adam and SGD along with YOLOv7.SGD has solid theoretical and mathematical support,along with an exhibition of enhanced stability and generality[10].In most applications,the Adam optimizer is recommended as the default optimization method because it usually generates better results,is faster to compute,and requires fewer tuning parameters than conventional optimization methods [11].Batch sizes 8 and 16 are used for the task of TSD.A portion of the German traffic signs is used for the training purpose.This study also explores the feasibility of adopting smaller datasets while keeping high accuracy to modify the application domain.Fig.1 shows the various traffic sign classes that are pointed by red arrows for human reference.

    Traffic sign recognition is a primary factor for autonomous cars to make safe travel.However,traffic sign recognition system contains more challenges due to limitations that are shown by recent incidents involving autonomous vehicles[12].Conventional traffic sign recognition encounters numerous challenges,such as occlusion,lighting conditions,and the existence of several neighboring traffic signs[13].

    Figure 1:Input images of each class for prediction

    1.1 Motivation

    The motivation for this research work is as follows:

    ? Traffic sign recognition is essential for autonomous cars to navigate safely and efficiently.But there are severe worries regarding the limitations of traffic sign recognition systems and the methods they employ,as shown by recent incidents involving autonomous vehicles and research connected to recognition system failure.Therefore,it becomes even more crucial to create powerful algorithms that can get beyond these constraints and provide precise and trustworthy traffic sign detection to improve the performance and safety of autonomous cars.Effective traffic sign identification is essential for maximizing traffic flow and raising overall road safety in addition to lowering the likelihood of accidents.

    ? Traditional techniques of traffic sign recognition,on the other hand,encounter various problems,such as occlusion,fluctuating lighting conditions,complicated backdrops,and the presence of multiple signs nearby.Due to these challenges,improved approaches must be proposed to manage these situations and provide precise and dependable traffic sign detection and identification.

    1.2 Contributions

    The contributions of this research work are as follows:

    ? This research intends to improve the accuracy and speed of traffic sign detection by incorporating the CBAM into the YOLOv7 framework.The CBAM’s potent attention mechanism enables the model to effectively acquire,and highlight key spatial and channel-wise information,enabling reliable detection of traffic signs even under difficult conditions like occlusion or complicated backdrops.

    ? Investigate and compare the effectiveness of the proposed model with different optimizers and batch sizes.

    ? The enhanced model proposed in this study exhibits improved feature representation,higher detection accuracy,and resilience by combining the characteristics of YOLOv7 and three CBAM modules in a complementary way,advancing the development of autonomous driving technology.

    ? The significance of this study is to achieve high accuracy on a small-sized real-time dataset thereby applying the model proposed in this study to a larger,and more diverse dataset for real-time applications in autonomous vehicles.

    The remainder of this study is organized as follows: Section 2 gives an overview of existing literature on traffic sign recognition for autonomous vehicles,highlighting strengths,limitations,and extensions of current knowledge.Section 3 focuses on the description of the YOLOv7 with the CBAM framework,including its working principle,architecture,and loss function.The dataset used for training and evaluation is described in Section 4.Section 5 describes the evaluation metrics,hyperparameters,and hardware/software configurations used in the experiments.Section 6 presents a detailed analysis of the results,including performance comparisons and visual representations of the model’s capabilities.Finally,the conclusion and future scope are discussed in Section 7.

    2 Literature Review

    There has been a positive trend toward applications in computer vision resulting in a substantial amount of research on TSD using various object-detection algorithms.Relevant to this study,an indepth inspection and analysis of various machine vision-based traffic detection models divided into 5 categories viz.color,shape,color and shape,ML,and Light Detection and Ranging(LiDAR)based models[14].A TSD system based on novel DL architectures used the YOLOv3,and Xception models along with Adam and RMSprop optimizers.These models are designed using the dataset with 3 classes such as “Yellow,diamond-shaped pedestrian crossing sign”,“Yellow,diamond-shaped other traffic signs”and“others”.However,this study is processed with a lower frame rate of 4.5 fps,which could be increased to improve the processing time and performance accuracy[15].

    A study specifically focused on the detection of Indian traffic signs using YOLOv3,and CNN over 5 classes,and attained an accuracy of 87%.However,the authors have not used a real-time traffic detection system to predict each frame in a video[16].A study proposed a cascaded R-CNN to obtain the multi-scale features for TSD that resulted in an accuracy of 99.7%.Additionally,the study also proposed a multi-scale attention mechanism to improve the detection of true traffic signs and reduce false detections[17].The YOLOv5 model is implemented on 8 classes of datasets viz.“No U-turn”,“Road bump”,“Road works”,“Watch for children crossing”,“Crosswalk ahead”,“Give way”,“Stop”,and “No entry”,along with a thorough comparison between the YOLOv5 and SSD.The own dataset used for the model displayed an accuracy of 97.70%.The future scope of the study is to expand the existing dataset and apply newly developed models like Mask R-CNN,CapsNet,and Siamese Neural Networks[18].An improved YOLOv5 model is implemented for real-time multi-scale TSD over a massive size of 182 classes.Data augmentation,and Adaptive Feature Fusion Pyramid Network(AF-FPN)methods were implemented to increase the performance of the standard YOLOv5 model,which indeed increased the accuracy from 60.18%to 62.67%.The performance of the model is low due to the blurring of images captured by the high-speed motion of a vehicle[19].An indigenous CNN architecture is used for TSD with the dataset having 16 classes viz.“green light”,“speed limit”,“no parking”,“bicycle and pedestrians only”,“crossroad 1”,“red light”,“crosswalk 1”,“straight ahead or left turn permitted”,“crossroads 2”,“traffic division”,“no overtaking”,“no turns”,“stop”,“one-way street”and “yellow light”.This approach outperforms YOLOv2,and Fast R-CNN,with an average accuracy of 90% in all types of weather conditions.However,authors have developed a model with less training data that could be increased to improve performance in more environments[20].A combination of Faster R-CNN,and Extreme Learning Machines(ELM)is used over 3 classes.However,the accuracy and performance of the model are not discussed quantitatively,but qualitatively it is stated that combining CNN with ELM increases the accuracy[21].

    A study on TSD and classification in the wild constructed a benchmark dataset “Tsinghua—Tencent 100K” covering real-world conditions.The study trained two models CNN and Fast RCNN which resulted in an accuracy of 88%,and 50%,respectively.The study had been implemented with a minimum number of traffic sign classes that rarely appear in benchmark datasets [22].Another study presented the YOLOv3 model in detecting temporary traffic control detection for road construction projects.The mentioned study used a dataset containing 8 classes viz.“construction cones”,“l(fā)ooper cones”,“construction barrels”,“construction barricades”,“end construction signs”,“road construction ahead signs”,“right lane reduction signs”,“right lane closed ahead signs”.The training resulted in a mean Average Precision(mAP)of 90.82%.The proposed model in the mentioned study recognized more than 98% of the temporary traffic signs correctly and approximately 81% of temporary traffic control devices correctly [23].A design for real-time TSD was implemented with CNN on 50,000 traffic-sign images and reached an accuracy of 97.3%.This model is designed by considering more traffic sign classes,and possible weather conditions affecting the visibility of the signs[24].The“WAF-LeNet”(an upgraded version of LeNet)is developed to recognize and identify traffic signs for autonomous vehicles.The accuracy attained in the study was 96.4%among 43 classes[25].Though there is a fairly small amount of research studies revolving around TSD using YOLOv7,research work was carried out to collect,and label the road damage data using Google Street View.The YOLOv7 model is trained with the collected data and results in an F1 score of 81.7%[26].A study focused on improving the performance of YOLOv5 for the detection of traffic signs in bad weather conditions made use of the Global Context (GC) block,which combined with YOLOv5’s results in an accuracy of 79.2% [27].A study quantitatively demonstrates that the combination of YOLOv7 with a lightweight convolution-based Spatial Pyramidal Pooling Fusion (SPPF) module leads to a significant improvement in model accuracy.The study reports a precise increase of 6.7%in accuracy when incorporating the SPPF module into the YOLOv7 framework[28].A portable image-based ADS system was developed using the YOLOv5 algorithm and Tesla P100 Graphics Processing Unit(GPU)system.It achieved a remarkable speed of 43.59 frames per second[29].Multiple studies have utilized a pre-trained model for TSD on large datasets,and have fine-tuned the respective models by using various optimizers [30–32].Some studies have implemented multi-task learning to simultaneously detect objects like pedestrians and bicycles [33,34].The use of LiDAR and Radar sensors has come up as one of the ways to increase the accuracy of models for TSD in challenging conditions like low lighting[35].A unique method is described in a study for analyzing Global Positioning System(GPS)trajectory data to detect vehicle turns,which involves converting the data to image-based data,postconversion,a personalized CNN model is designed[36,37].

    Previous approaches to TSD have used models like YOLOv5[27],YOLOv7[28],and CNN[36],which are popular and efficient models.However,these models do not have an attention mechanism,which can limit their performance.The model proposed in this study uses YOLOv7 with CBAM,which is an attention mechanism that helps to improve the model’s performance.Specifically,CBAM helps to focus the model’s attention on the most important features in an image,which can lead to better object detection,especially in cases where the objects in an image are small or have low contrast.

    3 Traffic Sign Detection Using YOLOv7 with CBAM

    YOLOv7 is the latest and state-of-the-art object detection model in the family of YOLO singleshot object detection models.YOLOv7 is currently the fastest and best-performing object detection model.YOLOv7 significantly enhances real-time object detection accuracy while lowering inference costs.By cutting around 40% of the parameters and 50% of the processing speed,YOLOv7 effectively beats other well-known object detectors with faster inference speeds,and higher recognition accuracy[38].

    3.1 Working Principle

    The four components that the YOLO algorithm uses to operate are residual blocks,bounding box regression,Intersection Over Union(IOU),and Non-Maximum Suppression(NMS).The initial component of the residual block divides the original image(A)into N equal-sized grid cells,where N is a hyperparameter.Localizing and determining the object’s class using the probability/confidence value is the responsibility of each grid cell.Bounding box regression is the second element that identifies the bounding boxes that correspond to rectangles highlighting all the objects in the image.There can be as many bounding boxes as there are objects within a given image.YOLO uses a single regression module to compute the characteristics of these bounding boxes.Y is the final vector of each bounding box as given in Eq.(1).

    where,Pcis the grid’s probability score for the cell that contains the object.The bounding box’s center’s x and y coordinates in relation to the surrounding grid cell are represented by bx,by.The height,and the width of the bounding box are represented by bh,bw,respectively.The four classes namely prohibitory,dangerous,mandatory,and others are represented by C1,C2,C3,and C4,respectively.Despite not all of them being significant,a single object in an image might frequently have many grid box possibilities for prediction.Such grid boxes are to be discarded in order to retain the relevant grid boxes using the third component IOU.IOU always ranges from 0 and 1.The IOU selection threshold is initially set at 0.5.Fig.2 shows the intersection area divided by the union area which is then calculated for each grid cell by YOLO.Finally,it considers grid cells with an IOU>threshold rather than those predicted to have an IOU ≤threshold.

    Figure 2:Intersection over union

    The final part NMS algorithm is a post-processing technique to remove duplicate and overlapping detections of the same object.When an object is detected,the YOLO algorithm generates multiple bounding boxes with confidence scores indicating the likelihood of an object being present in each box.However,some of these boxes may overlap or contain the same object,resulting in multiple detections for the same object.To address this issue,NMS is used to suppress all but the most confident detection of each object.The algorithm works by first sorting the detected bounding boxes by their confidence scores.Then for each box,it compares its overlap with all other boxes.If the overlap exceeds a certain threshold,the box with the lower confidence score is suppressed.The process is repeated until all boxes have been considered.The generated output helps to improve the overall performance and accuracy of the object detection algorithm.Establishing an IOU threshold is not always adequate since an item may contain several overlapping boxes.Noise might be included if many boxes are overlapped based on an IOU that exceeds the threshold and all those boxes are left unclosed.NMS can be used in these circumstances to keep only the boxes with the highest likelihood of being identified.Hence,the algorithm is designed by initializing the confidence threshold,and IOU threshold values.Then the bounding boxes are organized according to decreasing confidence.If any bounding box contains a confidence threshold 0 that is eliminated.The rest of the bounding boxes are iterated through in a loop beginning with the greatest confidence,and the IOU of the current box with every remaining box that belongs to the same class is calculated.If the IOU of the 2 boxes>IOU_Threshold,then the box with lower confidence is removed from the list of boxes.This operation is repeated until all the boxes are processed in the list.Here is an outline of the code for YOLO,a popular object detection algorithm.Table 1 shows the pseudocode of the steps involved in implementing YOLO.

    3.2 Architecture

    YOLOv7 can be used in many applications other than object detection,like instance segmentation,pose estimation,etc.In comparison to YOLOv4,YOLOv7 utilizes 36% less processing,reduces the number of parameters by 75%,and generates 1.5%higher AP.When compared to the edge-optimized version,YOLOv4-tiny,and YOLOv7-tiny reduce the number of parameters by 39%and computation by 49% while keeping the same AP.Hence,it can be stated that YOLOv7 is more optimized.A YOLO architecture is made up of various components,including a head,neck,and backbone.For the inference speed,the effectiveness of the YOLO network’s backbone is essential.The full YOLOv7 architecture can be seen in Fig.3.

    Figure 3:Proposed architecture of YOLOv7 with CBAM

    The Extended Efficient Layer Aggregation Network (E-ELAN) helps the model learn better while preserving its original gradient path.To increase the speed and accuracy of the model,EELAN considers several variables,including memory cost,input-output channel ratio,element-wise operation,activations,gradient routes,etc.[39].The CSPDarknet53 serves as the backbone network for the YOLOv7 architecture that makes up the Efficient Layer Aggregation Network(ELAN)model.CSPDarknet53 was created to increase the precision and effectiveness of object detection models.On the other hand,E-ELAN is another YOLOv7 architecture that uses EfficientNet as the backbone network.A series of CNNs called EfficientNet is created to attain cutting-edge accuracy while keeping the model’s computing cost to a minimum.The main difference between these two models is the backbone network,i.e.,ELAN uses CSPDarknet53,and E-ELAN uses EfficientNet.EfficientNet is more computationally efficient,but it may sacrifice some accuracy compared to CSPDarknet53.YOLOv7 uses an optimized compound model scaling approach that modifies the characteristics to produce suitable models for various application requirements.For instance,model scaling can improve the resolution of the model,the size of the input image,the depth,or the number of stages,and the width,or the number of channels.The compound scaling technique can keep the model’s original design characteristics.

    After training,one way to improve the model is by re-parameterizing it.The inference process takes longer,but the outcomes are more substantial.The two forms of ensemble re-parameterizations used to complete models are model level and module level.Model level re-parameterization can be done in two ways.In the first method,distinct sets of data are used to train several models with the same architecture,and then average their weights to get the final model.The second method is to take the average of a model’s weight at different epochs.But recently,module-level re-parameterization has been used in a lot of research works.The YOLOv7 contains several heads,including the Lead Head,which is accountable for all the output,and the Auxiliary Head,which helps with training middle layers.To enhance deep network training,a Label Assigner method was created that assigns soft labels after considering ground truth and network prediction results.Reliable soft labels employ optimization techniques to raise the standard and distribution of prediction output in addition to the accuracy of the prediction.However,conventional label assignment generates hard labels based on predetermined norms by directly referencing reality.The YOLOv7 architecture shown above uses kernel sizes such as 3×3,and 1×1 in all its convolution layers with padding of 1 and 2.

    Fig.4 shows a crucial component Cross-Branch Scalability (CBS).It is designed with a convolution layer,Batch Normalization(BN)layer,and a Sigmoid Linear Unit(SILU)activation function to extract images at various scales.Based on the CBS module,which makes up the upper and lower divisions,the MP1 module adds the max-pooling layer.Using max-pooling and the CBS module,the upper division reduces the image’s length and width in half.The lower division uses the first CBS module to reduce the image channel in half,the second CBS layer reduces the image’s width and length in half,and finally,the Concatenation(CAT)operation is used to combine the features retrieved from the top and lower branches,enhancing the network’s ability for feature extraction.The upsampling and CBS modules make up the UP module.

    Figure 4:CBS,MP1,UP modules

    In Fig.5,the ELAN module is made up of numerous CBS modules that have been piled on top of one another while maintaining the same input and output feature sizes.The learning capacity of the network is increased without deviating from the initial gradient path by directing the computing units of various feature groups to learn more diverse features.The Spatial Pyramid Pooling Concat Spatial Convolutional(SPPCSPC)module shown in Fig.6,ELAN-H(Extreme Low-latency Architecture for Network Heads) module,and UP module makes up the majority of the Path Aggregation Feature Pyramid Network(PAFPN)structure that makes up the Neck component of YOLOv7.The bottomup approach makes it simple to move bottom-level data up to the top level,allowing for the effective fusion of various hierarchical aspects.The CBS module,CAT module,and max-pooling module make up the majority of the SPPCSPC module.SPPCSPC uses different pooling kernel sizes such as 5 × 5,9 × 9,and 13 × 13.These modules obtain various perception fields through max-pooling.To predict confidence,category,and anchor frame,Head uses Re-parameterization Visual Geometry Group Block (RepVGG) structure to adjust the number of image channels for the output of Neck at three distinct scales and then passes through 1×1 convolution.The model proposed in this study addresses the scale problem in TSD by utilizing the SPPCSPC module in the last layer of the proposed model.Spatial Pyramid Pooling(SPP)allows capturing features at different scales without reducing the input resolution,while Cross Stage Partial (CSP) connections reduce the number of parameters in the proposed model.By incorporating these modules,the model can effectively handle the large variations in object scales commonly encountered in TSD tasks,improving the accuracy of predictions and enhancing the overall performance of the proposed model.

    Figure 5:ELAN module

    Figure 6:SPPCSPC module

    CBAM is a module used to enhance the performance of CNNs by incorporating spatial and channel attention mechanisms.It focuses on capturing both local and global context information from input feature maps,allowing the network to prioritize relevant image regions while suppressing irrelevant ones.The module consists of two components such as spatial attention and channel attention.The spatial attention module captures spatial dependencies among different channels by modeling interdependencies between spatial locations.This enables the network to focus on relevant regions and suppress background regions.The channel attention module captures interdependencies among channels by assessing the importance of each channel in conveying discriminative information.It emphasizes informative channels while suppressing less informative ones.The spatial and channel attention maps are combined to generate an attention map that captures both spatial and channel-wise information.This attention map is used to weigh the feature maps,allowing the network to selectively attend to relevant features.The YOLOv7 model is trained using the sum of the squared error between the predicted bounding boxes and the actual boxes,along with the cross-entropy loss for the class predictions.Its combination of a lightweight backbone network,effective neck,and multi-scale head make it a powerful tool for a variety of computer vision applications.The technical contribution of this study lies in the integration of three CBAM units before the three outputs of YOLOv7,a model that already detects objects at three different scales.By incorporating the CBAM module,weights are assigned to channel and spatial features of the feature map,which effectively increases the importance of useful features while suppressing irrelevant ones.This attention mechanism enables the proposed model to focus on target regions containing important information that improves accuracy in detecting objects of various sizes.

    3.3 Loss Function

    The loss function used in YOLOv7 is a mixture of different components,including:

    3.3.1 Localization Loss(LL)

    This component of the loss measures the difference between the predicted bounding box coordinates and the actual bounding box coordinates.It uses the Mean Squared Error(MSE)loss function to calculate the loss.

    3.3.2 Confidence Loss

    This component of the loss measures how confident the model is in its predictions.It calculates the difference between the predicted confidence score and the actual confidence score.The confidence score indicates whether the bounding box contains an object or not.The binary cross-entropy loss function is used to calculate this confidence.

    3.3.3 Classification Loss

    This component of the loss measures the difference between the predicted class probabilities and the actual class probabilities.The cross-entropy loss function is used to calculate this classification loss.

    3.3.4 Total Loss

    The overall loss function is a weighted sum of these three components.The weights are hyperparameters that are tuned during training to balance the contributions of the different components.The loss function’s ultimate goal is to reduce the difference between predicted,and ground truth bounding boxes,confidence scores,and class probabilities.

    4 Dataset Description

    The images in the dataset are part of the famous German traffic sign dataset and were preprocessed to ensure consistency in size,resolution,and color.The dataset consists of 741 images of traffic signs that are divided into three subsets like training set of 592 images(79.8920%),a validation set of 99 images (13.3606%),and a test set of 50 images (6.7476%) in a stratified method,ensuring that each subset had a proportional representation of each class.Table 2 shows the dataset with four classes such as prohibitory,dangerous,mandatory,and others with a total of 1,213 appearances of traffic signs.The prohibitory class(class 0)has 731 appearances of traffic signs accounting for 45.89%of the dataset.This class includes traffic signs that prohibit certain actions such as no trucks,speed limit,no traffic both ways,and no overtaking.The dangerous class(class 1)has 268 appearances of traffic signs accounting for 18.04%of the dataset.This class includes traffic signs that warn drivers of potential hazards or dangers,such as construction,priority at next intersection,bend left,bend right,bend,uneven road,slippery road,the road narrows,traffic signal,pedestrian crossing,school crossing,dangerous,cycles crossing,animals and snow.The mandatory class(class 2)has 211 appearances of traffic signs accounting for 13.44%of the dataset.This class includes traffic signs that indicate actions that drivers must take,such as a roundabout,go straight,go right,go left,go left or straight,go right or straight,keep right,and keep left.The other class (class 3) has 345 appearances of traffic signs accounting for 22.63% of the dataset.This class includes traffic signs that do not fall into the prohibitory,dangerous,or mandatory categories such as no entry,stop,give way,priority road,and restriction ends.

    Table 2:Description of dataset

    5 Experimental Setup

    The objective of this study is to achieve high accuracy while keeping the model size and computational complexity low,making it suitable for deployment on embedded systems.The dataset contains 741 images of traffic signs with varying lighting conditions,occlusions,and backgrounds.The pre-processed version of the dataset is used,where images were cropped and resized to 416×416×3 pixels and annotated properly.The dataset was divided into training(79.89%),validation(13.36%),and test (6.747%) sets.The latest version YOLOv7 object detection model is used.The model has three components that predict the class,location,and confidence of the traffic sign detected.It was first trained using the SGD optimization algorithm with batch sizes 8 and 16.It had the following values of hyperparameter with a learning rate of 0.001,weight decay of 0.0005,and momentum of 0.937.The model was trained for 100 epochs,and the total training time was 1.868,and 1.845 h for batch sizes 8,and 16,respectively.Then the model was trained using the Adam optimization algorithm with batch sizes 8,and 16.It had the following values of hyperparameter with a learning rate of 0.001,weight decay of 0.0005,and momentum of 0.937.The model was trained for 100 epochs,and the total training time was 1.916,and 1.862 h for batch sizes 8,and 16 respectively.At last,the model was trained using the AdamW optimization algorithm with batch sizes 8,and 16.It had the following values of hyperparameter with a learning rate of 0.001,weight decay of 0.0001,and momentum of 0.937.The model was trained for 100 epochs,and the total training time was 1.942,and 1.857 h for batch sizes 8,and 16,respectively.

    The evaluation metric mAP is used to measure the accuracy of the model in detecting traffic signs of different sizes,and aspect ratios.Precision,recall,and F1 score are used as secondary evaluation metrics [50,95].The training was conducted on a single Tesla K80 GPU which is available for free version of Google Colab.The calculation of mAP involves calculating the AP for each class of the detected object and then averaging those AP values across all classes.AP is the area under the precisionvs.recall curve.

    The precision-recall curve shows how the precision and recall of the algorithm vary with the detection threshold.The precision is the fraction of detected objects that are correct.The recall is the fraction of true positive predictions among all the real positive cases.Precision and recall are calculated as follows:

    The F1 score is the harmonic mean of precision and recall,calculated as follows:

    A good model should have a high F1 score,high recall,high accuracy,and high precision.Precision and recall,however,typically trade off against one another.To determine the ideal balance between precision and recall,use the F1 score.The precision-recall curve shows how accuracy and recall are traded off for different thresholds.While a low false negative rate is related to great recall,a low false positive rate is related to superior accuracy.A large area under the precision-recall curve indicates that the precision and recall are high.

    6 Results and Discussion

    The approach used in this study is evaluated using the standard mAP metric which measures the accuracy of object detection by computing the AP over all possible levels of recall.Table 3 shows the YOLOv7 model’s performance under different training configurations using four optimization algorithms such as Adam,SGD,AdamW,and RMSProp with batch sizes of 8,and 16.The outcomes demonstrated that both during training and testing,the YOLOv7 model was able to obtain high precision and recall values.This proved that the model has a low rate of FP and FN when identifying objects in images.The mAP metric measures the model’s ability to detect objects at different IOU thresholds.In particular,mAP50 evaluates the AP at an IOU threshold of.5 or 50 percent,while mAP[50,95]measures the AP across all IOU thresholds from.5 to.95,with increments of.05 or 5 percent.The YOLOv7 model achieves high mAP values for both mAP50 and mAP[50,95]indicating that the model can accurately detect objects at different IOU thresholds.

    Table 3:Results of all the combinations of YOLOv7 for TSD

    The results show that Adam performs consistently better than SGD in terms of precision,recall,and mAP50,regardless of batch size.On the other hand,AdamW is similar to Adam in most cases but shows slightly lower performance in terms of mAP50 for batch size 8.In Table 3,it can be observed that the batch size significantly impacts the performance of the YOLOv7 model.For instance,SGD shows lower precision,recall,and mAP values compared to the other configurations for batch size 16.This suggests that SGD is less effective in handling larger batch sizes,possibly due to its inherent instability in noisy,and high-dimensional optimization spaces.

    Fig.7 shows the predicted images for each class including dangerous,other,prohibitory,and mandatory signs,and each image is accompanied by the corresponding confidence level.Fig.7a inferred that the model accurately predicted dangerous class signs with high confidence levels,indicating its ability to detect,and classify potentially hazardous situations on the road.Fig.7b shows that the model was able to predict other class signs with high confidence levels,indicating its ability to accurately classify signs that do not fit into well-defined categories.Fig.7c tells that the model accurately predicted prohibitory class signs with high confidence levels,indicating its ability to detect and classify signs that restrict certain behaviors on the road.Similarly,Fig.7d indicates that the model was able to predict mandatory class signs with high confidence levels,indicating its ability to accurately detect and classify signs that require specific actions from drivers.

    Figure 7:Predicted image for each class

    Figs.8a and 9a illustrate that the YOLOv7 model performs well as it has high recall at low confidence,and as the confidence increases the recall values decrease,and finally confidence becomes 1 when recall becomes zero.Figs.8b and 9b illustrate that the model is performing well as it has high precision at high confidence,and as the confidence threshold decreases,the precision also decreases.Figs.8c and 9c prove that the curve near the upper right corner indicates the increase in recall,the reduction in precision is not immediately apparent,and the overall performance of the model is better.Figs.8d and 9d show that the model has a high F1 score at high confidence,and as the confidence threshold decreases,the F1 score also decreases.It proves that the model becomes less conservative and makes more predictions.Both the models Adam with batch sizes 8,and 16 are performing well in the dataset.However,the model with batch size 8 is performing slightly better than the one with 16.It can be inferred that the precision and the mAP50 of Adam with batch size 8 is more than the Adam with batch size 16.

    Fig.10a illustrates that the model performs well as it has high recall at low confidence,and as the confidence increases the recall values decrease,and finally confidence becomes 1 at recall is zero.Fig.10b shows that the model is performing well but the performance is less than the Adam models as the class-wise prediction is deviating from the overall prediction.By comparing Figs.10c,8c and 9c exhibited that the Adam optimizer performs better than the SGD optimizer.Fig.10d shows that the model has a high F1 score at high confidence,although not as high as the models with Adam,and as the confidence decreases,the F1 score also decreases.It showed that the model becomes less conservative.

    Figure 8:Adam optimizer with batch-size 8

    Figure 9:Adam optimizer with batch-size 16

    Figure 10:SGD optimizer with batch-size 8

    Fig.11 conveys that the model is performing very poorly compared to the previous Adam models,and SGD with batch size 8.For instance,the precisionvs.recall curve in Fig.11c depicts much below at the top corner indicating the model has not learned properly and is struggling to predict the test images and in Fig.11d the model has a very low F1 score at high confidence,and as the confidence decreases,the F1 score also decreases.

    Figure 11:SGD optimizer with batch-size 16

    Figs.12 and 13 indicate that the models AdamW with batch-size 8 and 16 perform well but not as well as models with Adam.Although these models perform better than the SGD optimizer for this particular dataset.The AdamW with batch size 8 performs better than the AdamW with batch size 16.This results that the model with the less batch size performs better for this dataset.

    Figure 12:AdamW optimizer with batch-size 8

    Figure 13:AdamW optimizer with batch-size 16

    Figs.14 and 15 illustrate that the models RMSProp with batch-size 8 and 16 perform well but not as well as models with other optimizers.In comparison between the different batch sizes,RMSProp with batch size 16 performs better than the RMSProp with batch size 8.This results that the model with the more batch size performs better for this dataset.However,its performance is lesser than other all optimizers.

    Figure 14:RMSProp optimizer with batch-size 8

    Figure 15:RMSProp optimizer with batch-size 16

    The results proved that Adam is a more effective optimization algorithm for training the YOLOv7 model and that the performance of the model varies with batch size.This information can be useful for selecting the best configuration for their specific use case.Table 4 shows the comparison of the proposed model with existing models.

    Table 4:Comparison of results

    Overall,the results illustrate the effectiveness of the YOLOv7 model in object detection tasks and provide valuable insights into its performance under different training configurations.This study concluded that Adam with batch size 8 is the most effective of all the combinations above for the TSD use case.

    Considering the proposed model’s effectiveness,it is important to acknowledge its potential limitations in extreme weather conditions and low-lighting scenarios.Since the model is trained on images captured under specific conditions,its performance may be compromised when exposed to diverse and challenging environments.To enhance the model’s generalizability and improve its performance,a more diverse range of input images can be included that encompass varied conditions during the training process.This will enable the model to learn and adapt to different environmental factors,ultimately enhancing its robustness and reliability in real-world applications.

    7 Conclusion and Future Scope

    The proposed study is implemented with four optimization algorithms,namely Adam,SGD,AdamW,and RMSProp with different batch sizes for TSD using YOLOv7.The integration of CBAM improved the model’s performance by focusing on the spatial and channel regions in the input.The evaluation has been performed on both the training,and testing datasets by considering four different metrics namely Precision,recall,mAP50,and mAP[50,95].The experimental results have shown that the Adam optimizer with batch-size 8 and 16 achieves the highest accuracy in terms of all four metrics for both training and testing datasets.Specifically,the precision and recall rates of the Adam optimizer are very high.It proved that the model can correctly identify traffic signs in the input images with high accuracy.Moreover,the results indicate that the choice of batch size can also have a substantial effect on the accuracy of the model.In general,the smaller batch size can lead to better performance,but it also increases the training time.Hence,the trade-off between accuracy,and training time should be considered while selecting the batch size.

    The results obtained in this research provide a strong basis for future work in the field of traffic sign detection.There are several directions in which this work can be extended.One prominent direction for future work is optimizing the performance of YOLOv7 specifically for real-time traffic sign detection on embedded systems like Raspberry Pi and NVIDIA Jetson.By enhancing the efficiency and speed of the model,its applicability in resource-constrained environments can be significantly expanded,enabling practical implementation in various settings.Another crucial aspect that holds great potential is the interpretability of deep learning models.In recent years,there has been a growing interest in understanding and visualizing the decision-making processes of such models.Hence,future research can delve into exploring the interpretability of YOLOv7 for traffic sign detection.This can involve visualizing saliency maps or activation patterns of the model to gain insights into how it makes predictions.These directions hold immense potential for advancing traffic sign detection systems,paving the way for improved road safety,efficient traffic management,and safer autonomous vehicles in the future.

    Acknowledgement:The authors would like to thank the editors and reviewers.

    Funding Statement:The authors received no specific funding for this study.

    Author Contributions:conceptualization,validation,review and editing: Kuppusamy Pothanaicker;supervision: Kuppusamy Pothanaicker and Celestine Iwendi;methodology,formal analysis,result analysis,writing:Sanjay Mythili and Deepashree Pradeep Vaideeswar.All authors reviewed the results and approved the final version of the manuscript.

    Availability of Data and Materials:Data used in this study is available from Traffic Signs Dataset in YOLO format,Version 1.Retrieved December 26,2022 from https://www.kaggle.com/datasets/valentynsichkar/traffic-signs-dataset-in-yolo-format.

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

    亚洲在线观看片| 国产色婷婷99| 久久99热6这里只有精品| 欧美丝袜亚洲另类| 一本久久中文字幕| 性色avwww在线观看| 日日摸夜夜添夜夜爱| 国产精品人妻久久久影院| 男女啪啪激烈高潮av片| 亚洲国产高清在线一区二区三| 欧美日韩一区二区视频在线观看视频在线 | 熟妇人妻久久中文字幕3abv| 免费av观看视频| 91狼人影院| 婷婷亚洲欧美| 大香蕉久久网| 国产精品不卡视频一区二区| 精品人妻一区二区三区麻豆 | 大香蕉久久网| 国产探花极品一区二区| 麻豆久久精品国产亚洲av| 丰满的人妻完整版| 欧美日韩乱码在线| 日本黄色片子视频| 特级一级黄色大片| 国产一区二区在线观看日韩| 欧美成人a在线观看| 99久国产av精品| 97超碰精品成人国产| 成人毛片a级毛片在线播放| 国产精品av视频在线免费观看| 校园人妻丝袜中文字幕| 一卡2卡三卡四卡精品乱码亚洲| 亚洲性久久影院| 日本免费一区二区三区高清不卡| 亚洲18禁久久av| 男插女下体视频免费在线播放| 又爽又黄无遮挡网站| 小蜜桃在线观看免费完整版高清| 日韩人妻高清精品专区| 国产精品日韩av在线免费观看| 亚洲最大成人手机在线| 天堂影院成人在线观看| 不卡一级毛片| 国产精品国产三级国产av玫瑰| 嫩草影院入口| 午夜福利成人在线免费观看| 中文资源天堂在线| 久久人人精品亚洲av| 成人特级av手机在线观看| 日本一二三区视频观看| 精品久久久噜噜| 国产老妇女一区| 搡女人真爽免费视频火全软件 | 日日摸夜夜添夜夜添小说| 国内精品美女久久久久久| 久久久久久伊人网av| 91久久精品国产一区二区成人| 最新在线观看一区二区三区| 久久久久久久久久成人| eeuss影院久久| 天美传媒精品一区二区| 国产成人aa在线观看| 日韩大尺度精品在线看网址| 97碰自拍视频| 九色成人免费人妻av| 日韩亚洲欧美综合| av黄色大香蕉| 国产免费男女视频| 久久亚洲精品不卡| 欧美成人一区二区免费高清观看| 中文字幕人妻熟人妻熟丝袜美| 亚洲av二区三区四区| 成人国产麻豆网| 欧美zozozo另类| 一级av片app| 亚洲最大成人中文| 国产成人91sexporn| 最近中文字幕高清免费大全6| 夜夜夜夜夜久久久久| 亚洲成av人片在线播放无| 日韩欧美 国产精品| 国产精品一区二区三区四区久久| 久久精品夜色国产| 大又大粗又爽又黄少妇毛片口| 亚洲色图av天堂| 国内揄拍国产精品人妻在线| 亚洲高清免费不卡视频| 男女那种视频在线观看| 久久久久久九九精品二区国产| 女生性感内裤真人,穿戴方法视频| 国产人妻一区二区三区在| 免费观看精品视频网站| 日韩精品青青久久久久久| 欧美性感艳星| 成年免费大片在线观看| 久久久久国产网址| 国内揄拍国产精品人妻在线| 亚洲最大成人手机在线| 丰满的人妻完整版| 18禁裸乳无遮挡免费网站照片| 九九爱精品视频在线观看| 一个人看的www免费观看视频| 日韩制服骚丝袜av| 国产亚洲精品久久久com| 久久久久久久亚洲中文字幕| 国内少妇人妻偷人精品xxx网站| 免费不卡的大黄色大毛片视频在线观看 | 亚洲最大成人av| 麻豆国产97在线/欧美| 亚洲性夜色夜夜综合| 嫩草影院入口| 免费看av在线观看网站| 国产成人a区在线观看| 成人综合一区亚洲| 欧美+亚洲+日韩+国产| av在线天堂中文字幕| 亚洲一区高清亚洲精品| 亚洲专区国产一区二区| 女生性感内裤真人,穿戴方法视频| 麻豆成人午夜福利视频| 观看免费一级毛片| 午夜久久久久精精品| 久久久精品94久久精品| 亚洲av成人av| 日韩三级伦理在线观看| 亚洲精品国产成人久久av| а√天堂www在线а√下载| 在现免费观看毛片| 99热全是精品| 午夜免费男女啪啪视频观看 | 亚洲婷婷狠狠爱综合网| 99热精品在线国产| 国产精品福利在线免费观看| 一区二区三区免费毛片| 联通29元200g的流量卡| 亚洲乱码一区二区免费版| 国产精品女同一区二区软件| 在线播放国产精品三级| 国产免费一级a男人的天堂| АⅤ资源中文在线天堂| 99热6这里只有精品| 色5月婷婷丁香| 一级a爱片免费观看的视频| 免费大片18禁| 中文字幕免费在线视频6| 国产高清不卡午夜福利| 国国产精品蜜臀av免费| 成人二区视频| 看十八女毛片水多多多| 天堂√8在线中文| 亚洲人与动物交配视频| 欧美绝顶高潮抽搐喷水| 亚洲18禁久久av| 久久精品夜色国产| 成人永久免费在线观看视频| 国产爱豆传媒在线观看| 少妇丰满av| 久久中文看片网| 97在线视频观看| 久久天躁狠狠躁夜夜2o2o| 欧美另类亚洲清纯唯美| 国产av麻豆久久久久久久| 国产毛片a区久久久久| 国产极品精品免费视频能看的| 村上凉子中文字幕在线| 亚洲av.av天堂| 观看美女的网站| 最近手机中文字幕大全| 免费看av在线观看网站| 国内久久婷婷六月综合欲色啪| 日韩欧美在线乱码| av在线蜜桃| 男女下面进入的视频免费午夜| 成人午夜高清在线视频| 最近的中文字幕免费完整| 99热这里只有是精品50| 国产毛片a区久久久久| 欧美日韩精品成人综合77777| 99久久无色码亚洲精品果冻| 啦啦啦韩国在线观看视频| 亚洲真实伦在线观看| 搡老妇女老女人老熟妇| 啦啦啦啦在线视频资源| 国产 一区 欧美 日韩| 亚洲无线观看免费| 在线a可以看的网站| 成人毛片a级毛片在线播放| 丰满人妻一区二区三区视频av| 久久天躁狠狠躁夜夜2o2o| 日本精品一区二区三区蜜桃| 欧美日韩精品成人综合77777| 天堂√8在线中文| 热99re8久久精品国产| 久久亚洲国产成人精品v| 18禁黄网站禁片免费观看直播| 日本a在线网址| 亚洲图色成人| 日本-黄色视频高清免费观看| 3wmmmm亚洲av在线观看| 精品久久久久久久久av| 国产成人a∨麻豆精品| 国产精品久久电影中文字幕| a级一级毛片免费在线观看| 日本五十路高清| 精品免费久久久久久久清纯| 久久久a久久爽久久v久久| 国产精品综合久久久久久久免费| 亚洲七黄色美女视频| 久久人妻av系列| 嫩草影视91久久| 一个人免费在线观看电影| 久久精品国产鲁丝片午夜精品| 日本黄大片高清| 日本免费a在线| 国产在线精品亚洲第一网站| 国产一区二区三区av在线 | 极品教师在线视频| 91久久精品电影网| 最近中文字幕高清免费大全6| 亚洲不卡免费看| 两个人的视频大全免费| 久久久久久大精品| 久久久久精品国产欧美久久久| 男女啪啪激烈高潮av片| 欧美激情在线99| 国产色婷婷99| 午夜影院日韩av| 午夜老司机福利剧场| 久久久欧美国产精品| 此物有八面人人有两片| 国产黄a三级三级三级人| 久久午夜亚洲精品久久| 亚洲久久久久久中文字幕| 卡戴珊不雅视频在线播放| 午夜精品在线福利| 十八禁国产超污无遮挡网站| 久久精品国产亚洲av涩爱 | 国产探花在线观看一区二区| 三级经典国产精品| 精品国产三级普通话版| 国产成人aa在线观看| 在线播放无遮挡| 国产乱人视频| 国产一区二区亚洲精品在线观看| 啦啦啦啦在线视频资源| 亚洲欧美精品自产自拍| 日日干狠狠操夜夜爽| 亚洲在线自拍视频| 2021天堂中文幕一二区在线观| 99九九线精品视频在线观看视频| 亚洲精品粉嫩美女一区| 网址你懂的国产日韩在线| 22中文网久久字幕| 久久99热这里只有精品18| 欧美一区二区亚洲| 久久精品国产亚洲网站| 精品福利观看| 男女边吃奶边做爰视频| 午夜久久久久精精品| 亚洲真实伦在线观看| 插阴视频在线观看视频| 国产精品一区二区三区四区免费观看 | 99国产精品一区二区蜜桃av| 别揉我奶头 嗯啊视频| 黄色一级大片看看| 国产精品综合久久久久久久免费| 国产精品人妻久久久影院| 在线观看午夜福利视频| 久久久久精品国产欧美久久久| 日韩av不卡免费在线播放| 一夜夜www| 中文字幕久久专区| 日韩制服骚丝袜av| 99久国产av精品| 亚洲电影在线观看av| 国产男人的电影天堂91| 男插女下体视频免费在线播放| 国产精品电影一区二区三区| 亚洲综合色惰| 麻豆精品久久久久久蜜桃| 特级一级黄色大片| 免费看a级黄色片| 丝袜美腿在线中文| 久久久久精品国产欧美久久久| 插逼视频在线观看| 欧美在线一区亚洲| 淫妇啪啪啪对白视频| 一级黄色大片毛片| 亚洲精品一区av在线观看| 欧美性猛交黑人性爽| 晚上一个人看的免费电影| 成人av一区二区三区在线看| 午夜免费激情av| 精品久久久久久久久久久久久| 亚洲成av人片在线播放无| 精华霜和精华液先用哪个| 一区二区三区四区激情视频 | 非洲黑人性xxxx精品又粗又长| 老女人水多毛片| 国产极品精品免费视频能看的| 久久久精品94久久精品| 韩国av在线不卡| 国产亚洲精品久久久久久毛片| 国产 一区精品| 美女xxoo啪啪120秒动态图| 久久久国产成人精品二区| 久久鲁丝午夜福利片| 国产真实伦视频高清在线观看| 亚洲aⅴ乱码一区二区在线播放| 内地一区二区视频在线| 欧美日本亚洲视频在线播放| 欧美激情在线99| 又爽又黄a免费视频| 人妻夜夜爽99麻豆av| 女生性感内裤真人,穿戴方法视频| 男人狂女人下面高潮的视频| 全区人妻精品视频| 淫秽高清视频在线观看| 日本欧美国产在线视频| 免费观看精品视频网站| 久久久久久九九精品二区国产| 亚洲国产高清在线一区二区三| 亚洲av.av天堂| 给我免费播放毛片高清在线观看| 国产av麻豆久久久久久久| 91午夜精品亚洲一区二区三区| 欧美在线一区亚洲| 免费看光身美女| 亚洲国产欧美人成| 久久久国产成人免费| av专区在线播放| 91久久精品电影网| 日韩成人av中文字幕在线观看 | 日本免费一区二区三区高清不卡| 亚洲精品日韩av片在线观看| 俄罗斯特黄特色一大片| 蜜桃亚洲精品一区二区三区| 免费观看在线日韩| 高清毛片免费观看视频网站| 国产老妇女一区| 日本爱情动作片www.在线观看 | av国产免费在线观看| 嫩草影院精品99| 久久久久国内视频| 久久人妻av系列| 亚洲精品日韩在线中文字幕 | 十八禁国产超污无遮挡网站| 国内少妇人妻偷人精品xxx网站| 日本免费a在线| 精品国内亚洲2022精品成人| 美女黄网站色视频| 久久久a久久爽久久v久久| 免费av毛片视频| 别揉我奶头 嗯啊视频| 精品国产三级普通话版| 色尼玛亚洲综合影院| 亚洲美女视频黄频| 乱码一卡2卡4卡精品| 久久久久久伊人网av| 亚洲第一电影网av| 久久久久久久久久黄片| 亚洲三级黄色毛片| 色哟哟·www| 99久久精品一区二区三区| 精品久久久久久久久亚洲| 亚洲欧美清纯卡通| 午夜日韩欧美国产| 色噜噜av男人的天堂激情| 国产不卡一卡二| 搡老岳熟女国产| 亚洲va在线va天堂va国产| 国产高清三级在线| 色噜噜av男人的天堂激情| 极品教师在线视频| 3wmmmm亚洲av在线观看| 国产精品永久免费网站| 色综合亚洲欧美另类图片| 久久久久精品国产欧美久久久| 免费av毛片视频| 成人亚洲精品av一区二区| 寂寞人妻少妇视频99o| 一夜夜www| 亚洲久久久久久中文字幕| 久久久久久九九精品二区国产| 久久午夜福利片| 欧美日韩国产亚洲二区| 婷婷色综合大香蕉| 国产欧美日韩一区二区精品| 天天一区二区日本电影三级| 精品乱码久久久久久99久播| 成人性生交大片免费视频hd| 日日摸夜夜添夜夜添av毛片| 免费不卡的大黄色大毛片视频在线观看 | 夜夜看夜夜爽夜夜摸| 狂野欧美激情性xxxx在线观看| 老女人水多毛片| av天堂在线播放| 免费观看的影片在线观看| 99久久九九国产精品国产免费| 国产精品精品国产色婷婷| 亚洲国产精品国产精品| 欧美日韩在线观看h| 国产淫片久久久久久久久| av天堂中文字幕网| 久久国内精品自在自线图片| 色哟哟哟哟哟哟| 人妻少妇偷人精品九色| 自拍偷自拍亚洲精品老妇| 免费看日本二区| 久久精品久久久久久噜噜老黄 | 久久久久性生活片| 亚洲成人精品中文字幕电影| 欧美另类亚洲清纯唯美| 丰满人妻一区二区三区视频av| 国产色婷婷99| 综合色av麻豆| 在线国产一区二区在线| 精品一区二区免费观看| 三级毛片av免费| 亚洲中文字幕一区二区三区有码在线看| 免费观看人在逋| 赤兔流量卡办理| 久久久久久久久久黄片| 中文字幕免费在线视频6| 国产精品福利在线免费观看| 亚洲国产色片| 日韩一本色道免费dvd| 成年免费大片在线观看| 久久亚洲精品不卡| 亚洲在线观看片| 久久久精品94久久精品| 日本熟妇午夜| 一边摸一边抽搐一进一小说| 国产高清视频在线播放一区| 色视频www国产| 五月伊人婷婷丁香| 国产精品永久免费网站| 99在线人妻在线中文字幕| 午夜日韩欧美国产| 日日摸夜夜添夜夜爱| 99视频精品全部免费 在线| 99在线视频只有这里精品首页| 此物有八面人人有两片| 蜜桃亚洲精品一区二区三区| 久久久成人免费电影| 精品久久久久久久久久免费视频| 亚洲精品国产成人久久av| 欧美绝顶高潮抽搐喷水| 99riav亚洲国产免费| 精品一区二区三区人妻视频| 免费在线观看成人毛片| 久久亚洲精品不卡| 联通29元200g的流量卡| 少妇裸体淫交视频免费看高清| 看十八女毛片水多多多| 国产欧美日韩一区二区精品| 国产午夜精品久久久久久一区二区三区 | 国产精品久久视频播放| 久久精品国产亚洲av涩爱 | 国产黄色视频一区二区在线观看 | 久久草成人影院| 无遮挡黄片免费观看| 亚洲aⅴ乱码一区二区在线播放| 国产毛片a区久久久久| 男女视频在线观看网站免费| 网址你懂的国产日韩在线| 两个人视频免费观看高清| 亚洲最大成人中文| 欧美性猛交╳xxx乱大交人| 日韩欧美在线乱码| 深爱激情五月婷婷| 特级一级黄色大片| 日韩成人av中文字幕在线观看 | 看非洲黑人一级黄片| 晚上一个人看的免费电影| 99国产极品粉嫩在线观看| 三级毛片av免费| 人人妻,人人澡人人爽秒播| 日韩欧美精品免费久久| 插逼视频在线观看| 男插女下体视频免费在线播放| 小说图片视频综合网站| 午夜福利在线观看吧| av天堂中文字幕网| 久久久久国内视频| 精品一区二区三区视频在线观看免费| 一区二区三区免费毛片| 免费av不卡在线播放| 天天躁日日操中文字幕| h日本视频在线播放| 午夜福利在线观看吧| 婷婷精品国产亚洲av| 毛片一级片免费看久久久久| 精品福利观看| 在线观看午夜福利视频| 国产视频一区二区在线看| 俺也久久电影网| 欧美三级亚洲精品| 一本久久中文字幕| 身体一侧抽搐| 日本一二三区视频观看| 亚洲国产高清在线一区二区三| 亚洲精品国产av成人精品 | 国产精华一区二区三区| 成人亚洲欧美一区二区av| 天天躁日日操中文字幕| 少妇高潮的动态图| 嫩草影院入口| 你懂的网址亚洲精品在线观看 | 亚洲一区二区三区色噜噜| 日韩亚洲欧美综合| 高清午夜精品一区二区三区 | 久久精品夜色国产| 国产三级中文精品| 最近2019中文字幕mv第一页| av天堂在线播放| 韩国av在线不卡| 国产三级在线视频| 夜夜夜夜夜久久久久| ponron亚洲| 在线免费十八禁| 99久久久亚洲精品蜜臀av| 亚洲av免费高清在线观看| 我的老师免费观看完整版| 热99re8久久精品国产| 亚洲欧美成人综合另类久久久 | 最近视频中文字幕2019在线8| 99久久成人亚洲精品观看| 自拍偷自拍亚洲精品老妇| 久久人人爽人人片av| 免费看a级黄色片| 精品人妻一区二区三区麻豆 | 中文字幕久久专区| 两个人视频免费观看高清| 我的老师免费观看完整版| 晚上一个人看的免费电影| 亚洲av成人av| 天天一区二区日本电影三级| 亚洲国产色片| 男女下面进入的视频免费午夜| 国产欧美日韩精品一区二区| 少妇猛男粗大的猛烈进出视频 | 欧美性猛交╳xxx乱大交人| 插逼视频在线观看| 女人被狂操c到高潮| 深夜精品福利| 久久久久国产精品人妻aⅴ院| 麻豆成人午夜福利视频| 精品国产三级普通话版| 人人妻人人看人人澡| 成人无遮挡网站| 人人妻,人人澡人人爽秒播| 中文在线观看免费www的网站| 亚洲性夜色夜夜综合| 国产精品久久久久久久电影| 国产69精品久久久久777片| 国产精品免费一区二区三区在线| 特级一级黄色大片| 亚洲国产色片| 男女做爰动态图高潮gif福利片| 日韩欧美在线乱码| 午夜福利视频1000在线观看| 亚洲成人中文字幕在线播放| 久久人妻av系列| 国产精品三级大全| 日韩在线高清观看一区二区三区| 美女xxoo啪啪120秒动态图| 欧美激情国产日韩精品一区| 午夜亚洲福利在线播放| 国产精品乱码一区二三区的特点| 一级毛片久久久久久久久女| 亚洲经典国产精华液单| av卡一久久| 不卡一级毛片| 国产中年淑女户外野战色| 97人妻精品一区二区三区麻豆| 亚洲美女搞黄在线观看 | 特大巨黑吊av在线直播| 久久这里只有精品中国| 最近最新中文字幕大全电影3| 麻豆国产97在线/欧美| 人人妻人人看人人澡| 一个人免费在线观看电影| 国产白丝娇喘喷水9色精品| 一级毛片我不卡| 成人特级黄色片久久久久久久| 午夜精品在线福利| 91午夜精品亚洲一区二区三区| 国产视频内射| 欧美日韩精品成人综合77777| 成人亚洲精品av一区二区| 久久久成人免费电影| 可以在线观看毛片的网站| 一级毛片久久久久久久久女| 小说图片视频综合网站| 国产真实乱freesex| 最好的美女福利视频网| 男人狂女人下面高潮的视频| 国产成人91sexporn| 真人做人爱边吃奶动态| 久久6这里有精品| 成年女人看的毛片在线观看| 一级黄片播放器| 五月伊人婷婷丁香| 欧美国产日韩亚洲一区| 国产高清视频在线观看网站| 国产精品电影一区二区三区| 成年女人永久免费观看视频| 国产亚洲精品av在线| 午夜亚洲福利在线播放| 日韩精品中文字幕看吧| 美女内射精品一级片tv| 免费观看在线日韩| 在线免费观看不下载黄p国产|