• <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 ?

    Human Faces Detection and Tracking for Crowd Management in Hajj and Umrah

    2022-08-23 02:22:52RiadAlharbeyAmeenBanjarYahiaSaidMohamedAtriAbdulrahmanAlshdadiandMohamedAbid
    Computers Materials&Continua 2022年6期

    Riad Alharbey,Ameen Banjar,Yahia Said,Mohamed Atri,Abdulrahman Alshdadi and Mohamed Abid

    1Department of Information Systems and Technology,College of Computer Science and Engineering,University of Jeddah,Jeddah,Saudi Arabia

    2Electrical Engineering Department,College of Engineering,Northern Border University,Arar,Saudi Arabia

    3Laboratory of Electronics and Microelectronics(LR99ES30),Faculty of Sciences of Monastir,University of Monastir,Tunisia

    4College of Computer Sciences,King Khalid University,Abha,Saudi Arabia

    5CES Laboratory,ENIS,University of Sfax,Tunisia

    Abstract: Hajj and Umrah are two main religious duties for Muslims.To help faithfuls to perform their religious duties comfortably in overcrowded areas, a crowd management system is a must to control the entering and exiting for each place.Since the number of people is very high, an intelligent crowd management system can be developed to reduce human effort and accelerate the management process.In this work, we propose a crowd management process based on detecting,tracking,and counting human faces using Artificial Intelligence techniques.Human detection and counting will be performed to calculate the number of existing visitors and face detection and tracking will be used to identify all the humans for security purposes.The proposed crowd management system is composed form three main parts which are: (1) detecting human faces, (2) assigning each detected face with a numerical identifier, (3) storing the identity of each face in a database for further identification and tracking.The main contribution of this work focuses on the detection and tracking model which is based on an improved object detection model.The improved Yolo v4 was used for face detection and tracking.It has been very effective in detecting small objects in highresolution images.The novelty contained in this method was the integration of the adaptive attention mechanism to improve the performance of the model for the desired task.Channel wise attention mechanism was applied to the output layers while both channel wise and spatial attention was integrated in the building blocks.The main idea from the adaptive attention mechanisms is to make the model focus more on the target and ignore false positive proposals.We demonstrated the efficiency of the proposed method through expensive experimentation on a publicly available dataset.The wider faces dataset was used for the train and the evaluation of the proposed detection and tracking model.The proposed model has achieved good results with 91.2% of mAP and a processing speed of 18 FPS on the Nvidia GTX 960 GPU.

    Keywords:Crowd management;Hajj and Umrah;face detection;object tracking;convolutional neural networks(CNN);adaptive attention mechanisms

    1 Introduction

    For Muslims,doing regional duties such as Hajj and Umrah is critical.But such duties are known by their extremely crowded spaces due to the big number of people.Hajj and Umrah are considered one of the largest human gatherings, where millions of people gather in a specific time and place,whether in the Two Holy Mosques(Makkah and Madinah),in addition to millions of pilgrims in the holy sites.The Kingdom works to maintain the safety and security of pilgrims and protect them from the dangers of crowding and gathering during the annual Hajj and Umrah seasons, in line with the Kingdom’s vision 2030 to harness the potentials and capabilities to serve the guests of Rahman.

    Hajj and Umrah are a major focus of Vision 2030 as the Kingdom seeks to increase the number of Umrah performers to 30 million by 2030.To help the faithful to perform their religious duties comfortably in overcrowded areas,a crowd management system is a must to control the entering and exiting for each place.Since the number of people is very high, an intelligent crowd management system can be developed to reduce human effort and accelerate the management process.A crowd management process is based on detecting, tracking, and counting human faces using Artificial Intelligence techniques.Human faces detection and counting will be performed to calculate the number of existing visitors and face recognition will be used to identify all the humans for security and health purposes.

    First and foremost,effective crowd management helps to ensure the safety of those at Hajj and Umrah,from the guests to the security staff,and the workers.When the Hajj takes place in Makkah,everyone in the venue should be able to perform their duties without worrying about their safety.The consequences of a poorly managed crowd can be disastrous, people can be injured and lives can be lost.Effective crowd management can help minimize the risk of overcrowding occurrences.

    The main importance of the crowd management system comes out when manipulating the crowd and ensuring comfortable movement and a safe environment.In effect,controlling the entering and exiting of authorized people can enable the manipulation of the crowd flow through an automatic system.Also, the crowd management system allows detecting dangerous situation because of overcrowded areas and ensure a quick intervention of the special teams to fix the problem.Automatic crowd management helps to reduce human effort and accelerate the process.Watching many surveillance systems is a hard task that needs a focus to detect dangerous situations.

    In Hajj and Umrah situation,people are continuously moving.So,there is a possibility to count a person twice or more and that may cause a problem.To overcome this problem, we propose to add a face recognition framework to the crowd management system to assign each person with an ID and track it.This method will allow controlling the entering of only authorized people to avoid overcrowded areas and to eliminate the multiple counts of the same person to get precise statistics.Besides, focusing on faces detection allows enhancing the counting process since it is impossible to detect the entire body of the person in an overcrowded area.

    The proposed crowd management system is mainly based on image processing tasks and data storage tasks.For image processing,the collected images are analyzed and human faces are detected and identified.Then each face’s identifier (ID) is stored in a database for further recognition and eliminating the possibility of multiple counting.The recent advances in image processing techniques[1]have boosted the state-of-the-art to a new level for many tasks such as image recognition[2],scene recognition[3],object detection[4],traffic signs detection[5],face identification[6],image inpainting[7], and medical image retrieval [8].The success of those techniques comes from the use of very deep neural networks [9] with the ability to learn directly from input data and through a prediction methodology that mimics the biological brain.

    In this work, we propose to use the Yolo v4 object detection framework [10] as our baseline and applied many improvements.Starting with the backbone,we proposed a better model with high accuracy and good processing speed.Then, we added adaptive attention mechanisms to make the model focus on the target to detect.For backbone, we applied the Cross-Stage-Partial-connections on the ResNeXt model.The ResneXt model [9] is a very powerful model that achieved state-ofthe-art image recognition on the ImageNet dataset with a top-1 error of 21.2%.Cross-Stage-Partialconnections(CSP)[11]was proposed to reduce the computation cost of deep neural networks without decreasing the accuracy.It was very effective and allow a reduction of 20%of the computations.The CSP guarantee to achieve real-time processing using very deep neural networks on low-performance computers equipped with low-end graphics processing units(GPU).

    The Yolo v4 has collected many techniques at the same framework.It started by proposing novel data augmentation techniques and designed better loss functions.Then, spatial pyramid pooling[12] and path aggregation [13] were applied at the detection stage.Those additions have improved the detection, accuracy of the Yolo v4 compared to the older Yolo version and state-of-the-art object detection frameworks.Combining many improvement techniques allows achieving better performances.In this work, we propose to add an adaptive attention mechanism to the Yolo v4 at the features extraction and detection stages to make more focus on the detected target and ignore false positive predictions.Besides,we propose a new set of anchors that fit for face detection at long distances.The proposed model was trained and evaluated on the wider face dataset[14].

    The rest of the paper is organized as follows:section 2 will present an overview of related works with a discussion on the limitation of existing works.The proposed approach will be presented and detailed in section 3.In section 4, we present the experiment and report the achieved results while presenting a deep discussion on the efficiency of the proposed approach.Conclusions and future works will be presented in section 5.

    2 Related Works

    Crowd management systems are very important systems for controlling overcrowded spaces.Hajj and Umrah are the best situations to test the crowd management system because of the huge number of existing people and the complexity of the environment.

    Generally, a crowd management system is based on detecting, tracking, and counting existing humans in a defined space.Many works have been proposed in the context of human detection for a variety of applications.

    Ayachi et al.[15]proposed a pedestrian detection system for advanced driver assistance systems.The proposed system was based on lightweight separable convolution blocks,which are designed for possible embedded implementation.The detection was performed through a linear regression method to accelerate the processing time.In [16] a pedestrian detection system was proposed based on the Yolo v2 model[17]with a lightweight backbone.The SqueezeNet model[18]was used as a backbone to achieve better results and low model size.

    For crowd management,humans must be detected and counted to control the flow.Lamba et al.[19]presented a survey on the most recent advances for crowd management and monitoring.In[20]a crowd management system was proposed to predict overcrowded urban areas when natural disasters happen such as an earthquake, typhoon, and national festivals.In such a situation the behavior of humans may differ from daily situations.The proposed system was based on an autoencoder with custom convolutional long-short term memory(LSTM)layers[21].The proposed approach was used to detect crowd flow and can predict the flow density through the same model using a Multitask Convolutional LSTM Encoder-Decoder network.The achieved results were very impressive when the proposed approach was evaluated in different events such as the earthquake,New Year’s Day,and the Tokyo marathon.

    Das et al.[22]proposed a Convolutional Neural Network with a special attention mechanism for crowd counting to be used for crowd management.The proposed approach aims to separate standing people and sitting people and classify them into different categories.First, a base Convolutional Neural Network was used to predict the number of people in different categories.Second, a global density map was generated based on crowd counting.Finally, a refinement process was performed through a weighted linear regression layer to split the density map into standing and sitting density maps.A custom dataset was built for evaluating the proposed approach and an extensive experiment was performed to prove the efficiency of the method.Reported results show that the proposed approach achieved an accuracy of 86.49% and a mean absolute error of 4.80 and 4.15 for standing and sitting categories respectively which outperform existing state-of-the-art methods.

    Seema et al.[23] propose to combine traffic analyses and crowd management for smart cities surveillance.The proposed system was based on the single-shot multi-box detection model (SSD)[24].The SSD was used to count the objects of interest in the video.Based on counting vehicles and pedestrians, the traffic flow was managed.The vehicle’s density was used to manage traffic signals.In addition, a license plate detection system was integrated to detect vehicles violating any traffic regulations.The surveillance videos were analyzed to predict the crowd statistics for managing the crowd in emergency cases safely.

    A crowd counting and density estimation was proposed in [25] for crowd management.The proposed approach was based on the combination of a Convolutional Neural Network with two output signals,the first used for crowd counting and the second is used for density map estimation.For better prediction,only human heads were considered in the count.The proposed approach was evaluated on a custom-made dataset with 107 images containing 45000 annotated humans.The images have a range of between 58 and 2200 humans in each image.The achieved results show the efficiency of the proposed approach.

    3 Proposed Approach

    First, will present an overview on the proposed crowd management system and define its components.Second, we will move on to introduce the proposed backbone and the architecture of the building blocks.Also,we will present the applied technique to reduce the computations.Third,we will present the design of the Yolo v4 and its main parts.Finally,we will present the proposed adaptive attention mechanisms to improve the performances of the Yolo v4 for human face detection.

    The proposed approach for crowd management in Hajj and Umrah was combined from an offline process and an online process.In the offline process, we train an object detection model for human face detection.In the online process,each detected face will be assigned with a numerical ID and stored for further identification and tracking.The pipeline of the proposed approach is illustrated in Fig.1.

    Figure 1:Proposed pipeline for human face detection and identification for crowd management system

    In the first step,RGB cameras were used to collect data in a specific area.In the second step,a face detection model was used to detect human faces by processing the data provided by the cameras.In the third step,the detected face is cropped for further processing.In the fourth step,each cropped face is compared to the stored data and identified.If the cropped face was already identified and stored in the database,it is assigned with its old identifier.But if the face was not already identified,it will be assigned with a new identifier(ID).In the final step,the cropped face and its ID are stored in a database for further identification and tracking.This procedure eliminates the need for the identification of all humans present in an overcrowded space manually which reduces the human effort and facilitates the management of the crowds.

    The challenging part of the proposed pipeline is face detection.Due to the challenging conditions such as the small size of the target,occlusion,geometric deformation,and so on,it is very necessary to build a robust face detection system that can overcome those challenges.So,we propose to use a powerful object detection framework with a very deep Convolutional Neural Network as a backbone.Also,we applied many techniques to enhance the performance and the processing speed.

    The Yolo v4 was used as an object detection framework.It was designed to achieve real-time processing while getting high detection accuracy.The proposed improvement by Yolo v4 has enhanced the accuracy by 10%and the processing speed by 12%compared to the Yolo v3[26].Besides,the model training becomes easier and requires less computation,and can be performed on a single GPU.

    In Yolo v4, they start by enhancing the backbone by applying a bag of freebies and a bag of specials.In this work, we proposed to replace the original darknet model with a deeper and more accurate model.For that,we proposed the use of the ResNeXt model[27].The model was an extension of the ResNet model.The main innovation was to use a set of transformers with the same topology inside the same block(Fig.2).

    Figure 2:Building block of the ResNeXt model compared to ResNet model

    The size of the transformers was called cardinality.Empirical experimentations on the ImageNet dataset[28]have proved that expanding the cardinality is more efficient than increasing the network depth or width to get better accuracy.In ResNeXt, an aggregation transformation was proposed to consider each neuron and network which is called network-in-neuron.The main idea was to project an input into a low dimension output and transform it.Then all the transformed outputs are aggregated to a single output.The use of a set of low dimension transformations made the network more flexible.Two main rules were considered to ensure that the computation complexity of the blocks is roughly the same:(1)in case of generating spatial maps of the same size,then the same hyper-parameters were shared in the block,and(2)for downsampling a spatial map by a factor of two,the width of the block must be doubled.Those rules were very important for maintaining the computation complexity and eliminated the design of each block separately.So,they design one block and repeated it until building the network while just changing the hyper-parameters at each level.The ResNeXt-101 has ranked in second place in the ImageNet large scale visual recognition challenge(ILSVRC)2016[28]with a top-1 error of 21.2%

    As the ResNeXt-101 has a high computation complexity,we proposed to apply the Cross-Stage-Partial-connections(CSP)[11]to reduce the computation complexity without damaging the accuracy.The CSP has proved that 20% of the computation can be reduced without any loss in accuracy.The Cross-Stage-Partial-connections were inspired by the DenseNet[29]that comes with the idea of connecting all previous layers to the actual layer and propagate the bottom features to the top of the network cheaper and more effectively.

    For CSP architecture,there are two main concepts.Fig.3 illustrated the application of the CSP on the ResNeXt blocks.The first is the partial dense block and the second is the partial transition layer.For the partial dense block,the output of the previous block is divided into two parts where a part goes through the actual block and the other goes directly to the partial transition layer.The main purposes of this concept are the following: (i) doubling the gradient path and eliminating gradient vanishing.(ii)balancing the computation of each block bypassing only the half channels through the block while the other half passes to the end directly.(iii)reducing memory usage by saving half of the memory of part that passes without computations.For the partial transition layer, a hierarchical feature fusion mechanism was designed to prevent different layers from learning the same gradient information by splitting the gradient flow.Due to balancing the computation and reducing the memory usage,CSP allows reducing the computation while guarding or enhancing the accuracy by enhancing the gradient pass.

    Figure 3:Applying the cross-stage-partial-connections on the ResNeXt model

    To design the face detection system,we propose to integrate the proposed backbone to the Yolo v4 object detection framework[10].The Yolo v4 has introduced the combination of almost all innovation techniques such as SPPNet[12],path aggregation[13],FPN[30],attention mechanism[31],and novel data augmentation techniques.This combination has widely enhanced the accuracy and the processing speed while reducing the computation overhead.

    The Yolo v4 is composed of 4 main stages which are:the input,the backbone,the neck,and the prediction.Fig.4 presents the architecture of the Yolo v4.The input is the image or the video that will be processed to detect existing objects.The backbone is a deep neural network used to extract relevant features for further processing.The neck is used for region proposals that may contain the target object.The prediction is the stage of score assignment for each detected object.

    Figure 4:Yolo v4 architecture

    The optimization techniques applied to the Yolo v4 were divided into two categories.All techniques applied to get better accuracy without increasing the inference speed were called bags of freebies.All techniques that enhance the accuracy and influence the inference speed were called bag of specials.

    Generally, an object detection model is trained offline which allows to development of more efficient training techniques that result in achieving better accuracy without damaging the inference speed.Data augmentation techniques were the most used strategies to enhance accuracy.In effect,the main purpose of data augmentation techniques is to increase data variability to meet real-world conditions.Thus,improved the generalization power of the model and make it more robust when tested with new data.For real scene images,geometric and photometric distortions are one of the challenges to handle.So, applying a data augmentation technique that mimics those challenges was a very effective solution.Random scaling,cropping,translation,rotation,and flipping were the most used techniques to deal with geometric distortion while adjusting the contrast,hue,saturation,and noise were effectively deployed for dealing with photometric distortion.In Yolo v4,more data augmentation techniques were proposed mixing images by multiplying and superimposing with different coefficient ratios,and then adjusting the label with these superimposed ratios.Also,a CutMix technique[32]was used by implementing the ground truth of a random image into another image.Furthermore,a style transfer generative adversarial model (GAN) was used as a data augmentation technique.This was useful to reduce the texture bias learned by the detection model.

    Usually,the training data is randomly collected and there is a problem of data imbalance between classes.The focal loss [33] was proposed to handle this problem by making the model focus on hard samples.There are more problems related to data collection and labeling such as expressing the relationship of the degree of association between different classes.To overcome this issue,a label smoothing[34]was proposed which convert hard labels to soft label for the training process.

    Training a neural network model is based on optimizing a loss function using gradient descent algorithms.So,the loss function is a critical component for the performance of the model.Generally,cross-entropy and its variant are used for classification problems,and mean square error is used for regression problems which are used to predict the parameters of the bounding box.For a direct estimation of the bounding box parameters, each parameter must be treated as an independent variable.But such a method does not consider the integrity of the target object.To solve the problem,the intersection over union(IoU)was used as a loss function[35].The newest variant CIoU loss[36]was proposed to consider the overlapping area,the distance between the center points of the ground truth and the predicted bounding boxes,and the aspect ratio.

    For the bag of special,many techniques were proposed to enhance the receptive field and increase the capability of features integration.The SPP was integrated into Yolo v4 to enlarge the receptive field.Since the SPP was originally designed to generate a vector in output and this cannot be applied for dense prediction using convolution layers, it was modified by concatenating different outputs to a tensor and used as input to the next layer.Besides, an attention model was deployed to enhance the accuracy.The spatial attention module(SAM)[37]was modified for use in Yolo v4 by replacing the pooling operation followed by convolution with only a convolution layer.The attention model results in small additional computation and does not affect the inference speed.Another important focus in the design of the neural network model is the activation function.Since the first success of the convolutional neural network for the computer vision task, the rectified linear unit (ReLU) was the first activation function that solves the problem of gradient vanishing.Then,many activation functions were proposed to enhance the performance such as the PReLU and Leaky ReLU,Scaled Exponential Linear Unit(SELU),swish and Mish.The Mish activation function was adopted by Yolo v4 since it achieved state-of-the-art performances and it is a continuously differentiable activation function.

    For the final prediction process, the non-maximum suppression (NMS) technique is applied to select the best-fit bounding box from a set of bounding boxes that predicts the same object.The original NMS does not consider the object context.So,for Yolo v4 the DIoU NMS[36]was used to consider the difference between center points as context information.

    The Yolo v4 was designed for general object detection and does not work well for the detection of tiny human faces.So,we propose to add an adaptive attention mechanism to enhance the focus of the model on human faces and allow their detection in high-resolution images.Generally, channel-wise attention is solving the problem of what to focus on and spatial attention is used to solve the problem of where to focus.So,we propose to combine both attention mechanisms for better performances.

    Traditional attention mechanisms are designed through pooling layers across the channel dimensions.However, we designed an adaptive spatial attention mechanism through a fully convolutional layer.As pooling layers are parameterless, using convolutional layers instead enhances the learning capability of the model without any additional computations.Fig.5 presents the proposed spatial attention mechanisms.It was designed in a way to generate an attention map that recalibrates the features from different spatial locations.The 1×1 convolution layer was used to squeeze the features across the channel dimension.Besides, it prevents the influence of the backpropagation on the backbone directly.The 3×3 convolution is used to excite a local area reaction to enhance the efficacity.Since the adaptive spatial attention mechanism is composed of 1×1 and 3×3 convolution,the relative position and the receptive fields of the spatial attention map are similar to the output of the backbone.So, the pixels of the spatial attention network weigh the pixels of output feature maps at the same location.The adaptive spatial attention mechanism was integrated into the backbone in a plug-in manner.

    Figure 5:Adaptive spatial attention mechanism.(t is the compression ratio)

    The adaptive channel-wise attention was designed by a squeeze and excitation structure followed by a domain attention network.Fig.6 illustrates the proposed adaptive channel-wise attention mechanism.Mainly, different pooling layers are important for the attention mechanism.The main idea behind this is for an input feature map,applying a global average pooling allows the identification of the object extend, and applying a global max pooling tends to identify the location of the object which are two main features for the detection task.Using global max pooling instead of max pooling is more useful for the detection of small objects.

    Figure 6:Adaptive channel-wise attention mechanism.(s and r are the compression ratios)

    Considering the mentioned above,many works have designed a channel-wise attention mechanism based on the combination of global max pooling and global average pooling.Then,weigh both paths equally.But in reality, objects have different scales and aspect ratios and an equal weight may work well for some objects but the bad result will be achieved for others.In this work,we deal with human faces at different aspect ratios which we will focus on.The adaptive channel-wise attention mechanism will be designed to handle the difference of aspect ratios to achieve the ultimate results.As mentioned earlier,we added a domain attention network to the pooling structure which is the main novelty of the proposed attention mechanism compared to existing ones.The proposed domain attention network was designed concerning three main rules.First, it must be fully data-driven where intermediate features and outputs can be adapted to the input.Second, the network must be powerful to weigh raw vectors.

    Finally,the network must be lightweight to avoid many additional computations and reduce the overall complexity.The domain attention network is composed form three fully connected layers and a hidden layer.The output of the network is weight tensor sensitive to the target domain.This vector is used to recalibrate the raw channel-wise generated by the previous pooling structure.The adaptive channel-wise attention mechanism was integrated into the detection stage and the backbone at the ResNeXt blocks where low semantic features can be detected.

    The integration of the proposed adaptive attention mechanisms was performed in a way to maintain the backbone structure to take advantage of the pre-trained weight and we make several changes on the detection stage.The adaptive spatial attention mechanism was integrated into the ResNeXt building blocks(ResX).Also,the adaptive channel-wise attention mechanism was integrated into the ResX blocks after applying the adaptive spatial attention mechanism.In the detection stage,only the adaptive channel-wise attention mechanism was applied.Since the top layers contain rich sematic features and less positional information, it was important to implement the channelwise attention but the spatial attention has no impact.Fig.7 present the design of the proposed implementation of the adaptive attention mechanisms on the Yolo v4.For clearance, the presented structure of the Yolo v4 was simplified and only important parts were illustrated.

    Figure 7:The proposed implementation of adaptive attention mechanisms in the Yolo v4

    The proposed adaptive attention mechanisms were designed to be implemented in a plug-in manner.Due to the lack of positional features in the top layers and the small size of feature maps,channel wise attention mechanism has been integrated.Subsequently,spatial and channel-wise attention mechanisms have been integrated in the bottom building blocks of the backbone because of the lack of semantic features at those layers.This configuration enabled a quick initialization of the model using pre-trained weight which accelerate the training process and guarantee high performances.

    Considering the mentioned analyses,the design of the proposed model was based on the Yolo v4 with ResNeXt model as backbone with additional adaptive attention mechanisms.As shown in Fig.7,the channel-wise attention is implemented in the ACBL block and the ARESX building block while the spatial attention mechanism is only implemented in the ARESX building block.

    4 Experiments and Results

    4.1 Dataset

    For training and evaluation, the wider face dataset [14] was used.The data was collected from the internet using search engines such as Google and Bing.Then,the data were manually filtered and annotated.The dataset contains 32203 images with a total of 393703 annotated faces.The dataset is very challenging due to the diversity of the collected data and the capturing conditions such as occlusions,pose variation,and geometric deformation.Most of the images in the dataset were collected from events in overcrowded areas.That makes it very useful for the studied task and will help to achieve better performance.The dataset was randomly divided into training,validation and testing sets where 40%of the data was used for training,10%for validation and 50%for testing.Pascal VOC evaluation metric was adopted for the evaluation of performances.

    4.2 Implementation Details

    All the experiments were carried out on a desktop running the Ubuntu 20.04 LTS equipped with an Intel i7 CPU,32 GB of RAM,an Nvidia GTX 960 GPU.TensorFlow Deep Learning framework was used for the development of the proposed model with support of CUDA acceleration and cuDNN library.The OpenCV library was used for images manipulation and display.

    Model training was performed using the Adam optimizer which is a gradient descent variant that optimizes the learning rate alongside the parameters and accelerates the convergence process.The model was trained for 40 epochs with an initial learning rate of 0.001.The size of the input images was fixed to 320×320 for both training and testing to achieve high performance and to respect real-time constraints.The batch size was fixed to 4 due to the limited memory of the used GPU.Backbone was initialized using the pre-trained weights on the ImageNet dataset.The model has trained alternatively by training the detection stage and freezing the backbone then training the complete model.The compression ratios were fixed as follow:r=s=16 and t=32.The training was performed for 110k iterations and lasted for two days.An early stop condition was established if the loss is not reduced for 10 K iterations.

    4.3 Evaluation and Comparison

    The performance of the model was evaluated based on different metrics such as mean Average Precision (mAP), processing speed (FPS), and floating-point operations (FLOPS).The proposed model was evaluated using the standard parameters of Yolo v4.The loss optimization curves are presented in Fig.8.The loss was reduced from 65.34 to 1.05 which proved the efficiency of the learning algorithm based on Adam optimizer for searching the minimum.Besides,the use of the pre-trained weight has been very important for reaching the convergence.The proposed model has achieved an mAP of 92.1%while running with a speed of 18 FPS.The achieved results outperform most state-of the-art models tested on the wider faces dataset.

    Figure 8:Different loss curves of the proposed model compared to the original Yolo v4.Bleu for the Yolo v4 and red for the proposed model

    To further improve the efficiency of the proposed method,we compared against the state-of-theart works on the same dataset.Tab.1 presents a comparison against the most recent works on the used dataset.The proposed method widely achieved the existing works with a big margin in terms of both precision and speed.Even methods with good precision struggles from slow processing and fast methods achieve low precision.However,the proposed method has achieved a good trade-off between speed and precision.

    4.4 Ablation Study

    An ablation study was conducted to evaluate the effectiveness of the adaptive attention mechanisms.To show the impact of the proposed improvement,we evaluated the performance of the original Yolo v4 on the same dataset.Tab.2 presents the achieved results for the original Yolo v4 and the improved version with adaptive attention mechanisms.The original Yolo v4 has fewer GFLOPS that is due to the additional attention mechanism implemented in the improved version.The proposed version has better precision and a similar processing speed to the original one.As shown in Tab.2,the proposed adaptive attention mechanisms have a positive impact on precision and do not explode the computation complexity.

    Table 1: Comparison against most recent works on the wider faces dataset

    Table 2: Achieved results compared to the original Yolo v4

    4.5 Implementation Demo

    The proposed face detection model was integrated into a crowd management system based on detecting and counting human faces to estimate crowd density and facilitate their management.A demo of human faces detection in hajj is presented in Fig.9.The proposed system has proved its efficiency for detecting tiny faces at a complex background and degraded conditions such as occlusion and deformation.The generalization power of the detection model was very high since it was not trained on Hajj and Umrah images.Using the adaptive attention mechanisms has a great impact on the overall performance.

    Figure 9:Demo of the human faces detection in Hajj

    5 Conclusion

    Due to the importance of Hajj and Umrah for Muslims, it is very critical to do their duties in comfortable situations.Crowd management systems are a good solution to manage the crowd to avoid dangerous situations.In this paper, we proposed a crowd management system based on detecting,tracking,and counting human faces.It is more efficient to detect human faces instead of detecting the whole body in a crowded area due to challenging conditions such as occlusion and deformation.The proposed face detection method was based on the Yolo v4 object detection framework with a ResNeXt backbone and additional adaptive attention mechanisms.Extensive experimentation has proved the efficiency of the proposed adaptive attention mechanism.We proposed two kinds of attention to taking advantage of all the features.The adaptive spatial attention was used to solve the problem of object position and the adaptive channel-wise attention was used to resolve the problem of what object to focus on.Compared to many existing works,the proposed method achieved a good balance between precision and speed.

    Funding Statement:This work was funded by the University of Jeddah,Jeddah,Saudi Arabia,under Grant No.(UJ-21-ICL-4).The authors,therefore,acknowledge with thanks the University of Jeddah technical and financial support.

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

    久久久久久九九精品二区国产| 99在线人妻在线中文字幕| 俄罗斯特黄特色一大片| 无人区码免费观看不卡| 日韩欧美国产在线观看| 国产精品一区二区三区四区免费观看 | 国产成人av教育| 国产高清videossex| 特级一级黄色大片| 中文字幕av成人在线电影| 久久久久久人人人人人| 91在线观看av| 国产亚洲精品av在线| 国产高清三级在线| 美女 人体艺术 gogo| bbb黄色大片| 国产中年淑女户外野战色| 三级毛片av免费| 色视频www国产| 中出人妻视频一区二区| 少妇的丰满在线观看| 亚洲成a人片在线一区二区| 国内精品久久久久久久电影| 欧美色欧美亚洲另类二区| 99久国产av精品| 欧美成人a在线观看| 亚洲中文日韩欧美视频| 久99久视频精品免费| 可以在线观看的亚洲视频| 女警被强在线播放| 成年免费大片在线观看| 日本三级黄在线观看| 国产久久久一区二区三区| 久久精品国产自在天天线| 国产成人福利小说| 欧美成人性av电影在线观看| 丝袜美腿在线中文| 一本久久中文字幕| 亚洲av免费在线观看| 成年女人永久免费观看视频| 亚洲狠狠婷婷综合久久图片| 免费人成视频x8x8入口观看| 欧美日韩乱码在线| 在线观看免费午夜福利视频| 欧洲精品卡2卡3卡4卡5卡区| 国产综合懂色| 中文字幕高清在线视频| 欧美日韩瑟瑟在线播放| 国产精品亚洲一级av第二区| 此物有八面人人有两片| 久久亚洲精品不卡| 国产一区二区在线av高清观看| 天美传媒精品一区二区| 桃红色精品国产亚洲av| 国产中年淑女户外野战色| 99热这里只有精品一区| 人妻丰满熟妇av一区二区三区| 国产探花在线观看一区二区| 成人永久免费在线观看视频| 在线观看免费视频日本深夜| 国产乱人视频| 一级a爱片免费观看的视频| 国产精品免费一区二区三区在线| 国产亚洲欧美在线一区二区| 丝袜美腿在线中文| 别揉我奶头~嗯~啊~动态视频| www国产在线视频色| 中文字幕人妻熟人妻熟丝袜美 | 日本五十路高清| 黄色女人牲交| 国产精品三级大全| 亚洲黑人精品在线| 日韩有码中文字幕| av女优亚洲男人天堂| 日韩欧美免费精品| 色综合站精品国产| 欧美一区二区国产精品久久精品| 成人三级黄色视频| 亚洲美女视频黄频| 国产高清视频在线播放一区| 网址你懂的国产日韩在线| 日本成人三级电影网站| 国产爱豆传媒在线观看| 精品久久久久久,| 国产精品av视频在线免费观看| 香蕉丝袜av| 亚洲人与动物交配视频| 亚洲成人精品中文字幕电影| 国内揄拍国产精品人妻在线| 99久久综合精品五月天人人| 国产av麻豆久久久久久久| 不卡一级毛片| 久久精品国产清高在天天线| 欧美日韩中文字幕国产精品一区二区三区| 18禁裸乳无遮挡免费网站照片| 亚洲欧美精品综合久久99| 天堂av国产一区二区熟女人妻| 久久精品国产99精品国产亚洲性色| 一进一出抽搐动态| 一级毛片女人18水好多| 国语自产精品视频在线第100页| 国内少妇人妻偷人精品xxx网站| 国产av在哪里看| 色噜噜av男人的天堂激情| 国产精品一及| 丝袜美腿在线中文| 欧美中文日本在线观看视频| 小蜜桃在线观看免费完整版高清| 搞女人的毛片| 亚洲中文字幕一区二区三区有码在线看| 国产探花极品一区二区| 老司机在亚洲福利影院| 亚洲 国产 在线| 国产伦精品一区二区三区视频9 | 成人特级av手机在线观看| 色播亚洲综合网| 国产成人av教育| 国产又黄又爽又无遮挡在线| 欧美绝顶高潮抽搐喷水| 精品国产超薄肉色丝袜足j| 搡老岳熟女国产| 久久精品综合一区二区三区| 丝袜美腿在线中文| 欧美成狂野欧美在线观看| 中文字幕精品亚洲无线码一区| 亚洲男人的天堂狠狠| 99热精品在线国产| 国产久久久一区二区三区| 欧美另类亚洲清纯唯美| 国产av一区在线观看免费| 午夜免费男女啪啪视频观看 | 99久久久亚洲精品蜜臀av| 国产精品av视频在线免费观看| 中文字幕熟女人妻在线| 99riav亚洲国产免费| 国产97色在线日韩免费| 一级毛片女人18水好多| 一个人看的www免费观看视频| 欧美又色又爽又黄视频| 色吧在线观看| 男女床上黄色一级片免费看| 国产97色在线日韩免费| bbb黄色大片| 在线观看免费午夜福利视频| 老师上课跳d突然被开到最大视频 久久午夜综合久久蜜桃 | 久久99热这里只有精品18| 国产三级黄色录像| 亚洲精品久久国产高清桃花| 欧美丝袜亚洲另类 | 欧美一区二区国产精品久久精品| 嫁个100分男人电影在线观看| 欧美成人免费av一区二区三区| 国产亚洲精品综合一区在线观看| 日本与韩国留学比较| 18禁美女被吸乳视频| 亚洲avbb在线观看| 日本黄色视频三级网站网址| 九九热线精品视视频播放| 狂野欧美激情性xxxx| 亚洲av五月六月丁香网| 长腿黑丝高跟| 国内精品久久久久久久电影| 欧美区成人在线视频| 国产私拍福利视频在线观看| 桃红色精品国产亚洲av| 国产伦在线观看视频一区| 3wmmmm亚洲av在线观看| 国产成人影院久久av| 国产伦精品一区二区三区四那| 午夜福利18| 免费搜索国产男女视频| 中文字幕高清在线视频| 九九久久精品国产亚洲av麻豆| 日韩亚洲欧美综合| 香蕉久久夜色| 日韩欧美三级三区| 亚洲一区高清亚洲精品| 亚洲电影在线观看av| 成人特级av手机在线观看| 精品午夜福利视频在线观看一区| 国产精品综合久久久久久久免费| 51午夜福利影视在线观看| 午夜久久久久精精品| 亚洲成人久久性| 变态另类丝袜制服| 男女之事视频高清在线观看| 1000部很黄的大片| 亚洲国产高清在线一区二区三| 国产精品久久久久久久电影 | 久久久久亚洲av毛片大全| 国产精品,欧美在线| 欧美日韩乱码在线| 国产伦人伦偷精品视频| 国产淫片久久久久久久久 | 噜噜噜噜噜久久久久久91| 欧美乱妇无乱码| 最近最新中文字幕大全电影3| 老熟妇乱子伦视频在线观看| 人人妻人人澡欧美一区二区| 天堂√8在线中文| 午夜激情福利司机影院| 亚洲人成网站在线播放欧美日韩| 俄罗斯特黄特色一大片| 亚洲av成人精品一区久久| 神马国产精品三级电影在线观看| 日韩高清综合在线| 久久精品国产综合久久久| 热99在线观看视频| 一本一本综合久久| 午夜福利成人在线免费观看| 精品久久久久久久久久免费视频| 免费一级毛片在线播放高清视频| 久久香蕉精品热| 草草在线视频免费看| 51午夜福利影视在线观看| 欧美黄色淫秽网站| 好看av亚洲va欧美ⅴa在| 国产爱豆传媒在线观看| 天堂√8在线中文| 精品欧美国产一区二区三| 黑人欧美特级aaaaaa片| 免费av不卡在线播放| 久久婷婷人人爽人人干人人爱| 精品免费久久久久久久清纯| 一级黄片播放器| 国产精品野战在线观看| www日本在线高清视频| xxxwww97欧美| 男女那种视频在线观看| 一个人看视频在线观看www免费 | 国产精品日韩av在线免费观看| 俄罗斯特黄特色一大片| 免费在线观看成人毛片| 午夜视频国产福利| 亚洲成人久久性| 国产欧美日韩精品亚洲av| 亚洲成av人片免费观看| 亚洲在线自拍视频| 亚洲国产高清在线一区二区三| 一进一出好大好爽视频| 在线观看午夜福利视频| 色尼玛亚洲综合影院| 欧美日韩综合久久久久久 | 国产精品嫩草影院av在线观看 | 国产伦精品一区二区三区视频9 | 每晚都被弄得嗷嗷叫到高潮| 舔av片在线| 色播亚洲综合网| 久久人人精品亚洲av| 在线免费观看的www视频| 淫秽高清视频在线观看| 国产成人av教育| 亚洲国产日韩欧美精品在线观看 | 久久精品国产亚洲av涩爱 | 日韩欧美国产在线观看| 在线十欧美十亚洲十日本专区| 女人高潮潮喷娇喘18禁视频| 天堂动漫精品| 狂野欧美白嫩少妇大欣赏| 日日夜夜操网爽| 一进一出抽搐gif免费好疼| 国产高清视频在线观看网站| 亚洲精品久久国产高清桃花| 亚洲avbb在线观看| 久久国产精品影院| 国产精品香港三级国产av潘金莲| 精品免费久久久久久久清纯| 99久国产av精品| h日本视频在线播放| 欧美午夜高清在线| 中国美女看黄片| 国产精品一区二区免费欧美| 老汉色∧v一级毛片| 欧美性猛交黑人性爽| 色精品久久人妻99蜜桃| 欧美黄色淫秽网站| 蜜桃久久精品国产亚洲av| 日韩大尺度精品在线看网址| 午夜视频国产福利| 成人永久免费在线观看视频| 国产精品电影一区二区三区| 少妇高潮的动态图| 欧美+亚洲+日韩+国产| eeuss影院久久| 国产私拍福利视频在线观看| 男女视频在线观看网站免费| 无人区码免费观看不卡| 国内久久婷婷六月综合欲色啪| 国产精品永久免费网站| 老司机午夜福利在线观看视频| 无人区码免费观看不卡| 老司机午夜福利在线观看视频| 日韩欧美国产在线观看| 欧美乱色亚洲激情| 网址你懂的国产日韩在线| 亚洲av成人不卡在线观看播放网| 欧美一区二区国产精品久久精品| av片东京热男人的天堂| 国产精品99久久久久久久久| 3wmmmm亚洲av在线观看| 免费看十八禁软件| 免费人成在线观看视频色| 国产男靠女视频免费网站| 最近最新中文字幕大全免费视频| 精品久久久久久久久久久久久| 免费在线观看影片大全网站| 久久精品综合一区二区三区| www国产在线视频色| 在线观看66精品国产| 久久中文看片网| 美女 人体艺术 gogo| 亚洲va日本ⅴa欧美va伊人久久| 欧美午夜高清在线| 欧美日韩一级在线毛片| 欧美日本视频| 亚洲片人在线观看| 久9热在线精品视频| 麻豆一二三区av精品| 久久婷婷人人爽人人干人人爱| 很黄的视频免费| 亚洲中文字幕日韩| 超碰av人人做人人爽久久 | 亚洲精品日韩av片在线观看 | 国产精品 欧美亚洲| 1000部很黄的大片| 天天添夜夜摸| 久久久久精品国产欧美久久久| 悠悠久久av| 少妇裸体淫交视频免费看高清| 99热这里只有精品一区| 两个人的视频大全免费| 欧美一区二区亚洲| 麻豆国产av国片精品| 成人欧美大片| 91字幕亚洲| 成人无遮挡网站| 欧美高清成人免费视频www| 久久国产乱子伦精品免费另类| 午夜精品在线福利| www日本在线高清视频| 国产一区二区激情短视频| 午夜福利成人在线免费观看| 1000部很黄的大片| 超碰av人人做人人爽久久 | 99久久久亚洲精品蜜臀av| 亚洲av一区综合| 国产在线精品亚洲第一网站| 最好的美女福利视频网| 成人午夜高清在线视频| 啦啦啦韩国在线观看视频| 啦啦啦免费观看视频1| 18禁裸乳无遮挡免费网站照片| 搡老熟女国产l中国老女人| 精华霜和精华液先用哪个| 午夜免费激情av| 99久久精品一区二区三区| 两个人的视频大全免费| 精品免费久久久久久久清纯| 亚洲天堂国产精品一区在线| 神马国产精品三级电影在线观看| 久久亚洲真实| 亚洲国产欧美网| 色综合亚洲欧美另类图片| 欧美日韩福利视频一区二区| 中文字幕av在线有码专区| 亚洲片人在线观看| 午夜福利视频1000在线观看| 亚洲午夜理论影院| 国产精品久久久人人做人人爽| 亚洲成人久久爱视频| 少妇高潮的动态图| 九九热线精品视视频播放| 亚洲天堂国产精品一区在线| 女人高潮潮喷娇喘18禁视频| a在线观看视频网站| 人人妻人人看人人澡| 国产一级毛片七仙女欲春2| 国产三级在线视频| 亚洲av熟女| 制服人妻中文乱码| 97碰自拍视频| 亚洲av成人av| 欧美色欧美亚洲另类二区| 亚洲一区二区三区不卡视频| 色综合亚洲欧美另类图片| 成人av在线播放网站| 岛国在线观看网站| 色综合站精品国产| 老师上课跳d突然被开到最大视频 久久午夜综合久久蜜桃 | www.www免费av| 美女高潮的动态| 亚洲人成电影免费在线| 久久久精品大字幕| 我的老师免费观看完整版| 亚洲av一区综合| 日韩大尺度精品在线看网址| 男人舔女人下体高潮全视频| 在线免费观看不下载黄p国产 | 亚洲五月天丁香| av福利片在线观看| 最后的刺客免费高清国语| 又紧又爽又黄一区二区| 久久久色成人| 97碰自拍视频| www国产在线视频色| 3wmmmm亚洲av在线观看| 99久久成人亚洲精品观看| 日韩欧美一区二区三区在线观看| 精品一区二区三区av网在线观看| 国产亚洲欧美在线一区二区| 日日摸夜夜添夜夜添小说| 国产精品久久久久久精品电影| 亚洲欧美激情综合另类| 国产午夜福利久久久久久| 18美女黄网站色大片免费观看| 国产老妇女一区| 久久九九热精品免费| 国产一级毛片七仙女欲春2| av女优亚洲男人天堂| 嫩草影视91久久| 99久久精品国产亚洲精品| 神马国产精品三级电影在线观看| 嫁个100分男人电影在线观看| 国产aⅴ精品一区二区三区波| 久久国产精品人妻蜜桃| 日韩免费av在线播放| 91久久精品国产一区二区成人 | 成人永久免费在线观看视频| 一二三四社区在线视频社区8| 欧美极品一区二区三区四区| 老汉色∧v一级毛片| 国产高清有码在线观看视频| 亚洲av中文字字幕乱码综合| 国产欧美日韩一区二区三| 国产69精品久久久久777片| 在线观看免费午夜福利视频| 岛国在线免费视频观看| 级片在线观看| 中文字幕av成人在线电影| 啪啪无遮挡十八禁网站| 欧美+日韩+精品| 一夜夜www| h日本视频在线播放| 日本一二三区视频观看| 日本黄大片高清| 在线观看免费午夜福利视频| 日韩欧美精品v在线| 亚洲avbb在线观看| 88av欧美| 久久精品国产综合久久久| 一个人免费在线观看电影| 精品福利观看| 黄片大片在线免费观看| 每晚都被弄得嗷嗷叫到高潮| 欧美3d第一页| 精品国产三级普通话版| 啦啦啦观看免费观看视频高清| 国产精品三级大全| 午夜激情欧美在线| 99国产极品粉嫩在线观看| 欧美3d第一页| 最新中文字幕久久久久| 黄色丝袜av网址大全| 国产精品久久久久久精品电影| 两个人看的免费小视频| 成人特级av手机在线观看| 国产探花极品一区二区| 俺也久久电影网| 日韩免费av在线播放| 欧美黄色淫秽网站| 一进一出好大好爽视频| 亚洲av中文字字幕乱码综合| 最近最新中文字幕大全免费视频| 亚洲电影在线观看av| 午夜激情欧美在线| 国产不卡一卡二| 亚洲久久久久久中文字幕| 国产免费av片在线观看野外av| 免费观看精品视频网站| 淫秽高清视频在线观看| 亚洲色图av天堂| 嫩草影院入口| 日韩欧美 国产精品| 美女高潮的动态| 久久久成人免费电影| 99精品在免费线老司机午夜| 叶爱在线成人免费视频播放| 一级黄色大片毛片| 国产成人aa在线观看| 搡老岳熟女国产| 成人18禁在线播放| 久久精品综合一区二区三区| 少妇熟女aⅴ在线视频| 午夜免费激情av| 欧美一区二区精品小视频在线| e午夜精品久久久久久久| 久久精品国产亚洲av香蕉五月| 尤物成人国产欧美一区二区三区| 欧美午夜高清在线| 午夜日韩欧美国产| 亚洲五月天丁香| 国产精品综合久久久久久久免费| 国产单亲对白刺激| 亚洲av美国av| 9191精品国产免费久久| 啪啪无遮挡十八禁网站| 每晚都被弄得嗷嗷叫到高潮| 亚洲精品456在线播放app | 黄片大片在线免费观看| 欧美丝袜亚洲另类 | 亚洲18禁久久av| 国产精品 欧美亚洲| 最好的美女福利视频网| 国产成人aa在线观看| 麻豆久久精品国产亚洲av| 最后的刺客免费高清国语| 国产免费男女视频| 88av欧美| xxx96com| www.色视频.com| 在线看三级毛片| 伊人久久大香线蕉亚洲五| 国产真实乱freesex| 亚洲色图av天堂| 天美传媒精品一区二区| 18+在线观看网站| 在线免费观看的www视频| 欧美极品一区二区三区四区| 亚洲 国产 在线| 青草久久国产| 在线观看免费视频日本深夜| 91在线精品国自产拍蜜月 | 色吧在线观看| 国产久久久一区二区三区| 精品国产美女av久久久久小说| 中国美女看黄片| 成人鲁丝片一二三区免费| 久久久久久九九精品二区国产| 男插女下体视频免费在线播放| 在线a可以看的网站| 亚洲黑人精品在线| 伊人久久精品亚洲午夜| 女人十人毛片免费观看3o分钟| 在线十欧美十亚洲十日本专区| 欧美成人性av电影在线观看| 又黄又爽又免费观看的视频| 免费搜索国产男女视频| 亚洲色图av天堂| 久久6这里有精品| 不卡一级毛片| 老司机福利观看| 九色成人免费人妻av| 婷婷亚洲欧美| 亚洲人成电影免费在线| 草草在线视频免费看| 午夜免费男女啪啪视频观看 | 内射极品少妇av片p| 日本一二三区视频观看| 国产成人系列免费观看| 免费在线观看日本一区| 成人午夜高清在线视频| 在线观看av片永久免费下载| 亚洲国产中文字幕在线视频| 综合色av麻豆| 亚洲中文字幕日韩| 亚洲精品456在线播放app | 国产伦精品一区二区三区视频9 | 国产欧美日韩精品亚洲av| 久久精品国产亚洲av涩爱 | 亚洲在线观看片| 精品久久久久久久毛片微露脸| 99视频精品全部免费 在线| 亚洲人成电影免费在线| 国产精品久久久人人做人人爽| 韩国av一区二区三区四区| 禁无遮挡网站| 啦啦啦免费观看视频1| 99久久精品一区二区三区| 午夜福利在线观看吧| 人人妻人人澡欧美一区二区| 国产成人a区在线观看| 国产伦人伦偷精品视频| 久久久久久人人人人人| 亚洲国产中文字幕在线视频| 久久久精品欧美日韩精品| 亚洲 欧美 日韩 在线 免费| 国产精品久久久久久久久免 | 亚洲七黄色美女视频| 99久久久亚洲精品蜜臀av| 亚洲国产精品久久男人天堂| 丰满乱子伦码专区| 在线天堂最新版资源| 国产97色在线日韩免费| 久久婷婷人人爽人人干人人爱| 国内精品久久久久精免费| 精品久久久久久久毛片微露脸| www日本在线高清视频| a在线观看视频网站| 婷婷精品国产亚洲av在线| 国产又黄又爽又无遮挡在线| 51午夜福利影视在线观看| 大型黄色视频在线免费观看| 中文在线观看免费www的网站| 日本a在线网址| 久久精品人妻少妇| xxx96com| 免费人成视频x8x8入口观看| 国产av在哪里看| 一级毛片女人18水好多| 国内毛片毛片毛片毛片毛片| 成人国产综合亚洲| 午夜精品在线福利| 久久人人精品亚洲av| 日日夜夜操网爽| 国产成人影院久久av|