Online car-hailing driver face key point positioning method and device based on deep learning

Abstract

The invention relates to the field of deep learning, and discloses a method and a device for positioning face key points of an online car-hailing driver based on deep learning, which are used for solving the problem that face key points cannot be accurately positioned when an emergency occurs in face key point positioning. Randomly selecting a large number of real online car-hailing verification photos as a data set, performing feature engineering processing on the data set, repeating the steps to obtain a data set, performing pre-neural network model design, evaluating evaluation indexes of the pre-neural network model according to data in the data set, and adjusting model parameters to obtain a final neural network model. According to the method, real-time online car-hailing verification photos are collected, feature engineering processing is carried out on the real-time online car-hailing verification photos, the real-time online car-hailing verification photos subjected to feature engineering processing are input into the final neural network model, circle selection of the face position in the photo of each driver is obtained, and the accuracy of face key point positioning is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and more specifically to a method and device for locating key points of a face of an online car-hailing driver based on deep learning. Background Art

[0002] Deep learning is a method of pattern recognition and decision-making that simulates the working principles of the human nervous system. It is built on the basis of artificial neural networks and uses a multi-level neuron structure, with each layer performing feature extraction and representation learning on the input data. This hierarchical structure enables deep learning to automatically learn and discover abstract features and patterns from large amounts of data. Deep learning has achieved great success in computer vision, natural language processing, speech recognition and other fields. It can be trained with large-scale data sets to achieve highly accurate classification, prediction and generation tasks. The core algorithm of deep learning is the back-propagation algorithm, which adjusts network parameters by minimizing the error between the predicted results and the actual results, thereby improving the performance of the model.

[0003] Facial key point localization is a research task in the field of computer vision, which aims to automatically detect and locate the key points in face images. These key points usually include facial features with important semantic information such as eyes, eyebrows, nose, mouth, etc. In the technology of facial key point localization, a common method is to use deep learning models such as convolutional neural networks (CNN) or their variants. These models learn the location patterns and feature representations of facial key points by training large-scale annotated data sets. Once the model is trained, it can predict new face images and output the exact location of each key point in the image.

[0004] Existing facial key point positioning and detection methods may fail to accurately locate facial key points when unexpected situations occur, such as changes in lighting, occlusion, and changes in expression, thereby reducing the accuracy of face verification.

[0005] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and device for locating key points of the face of an online car-hailing driver based on deep learning to solve the problems existing in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for locating key points of a face of an online car-hailing driver based on deep learning, comprising the following steps:

[0009] Step 1: Randomly select a large number of real online car-hailing verification photos from the online database as the data set;

[0010] Step 2: Perform feature engineering on the data set and divide the data set into training set, test set, and validation set.

[0011] Step 3: Repeat steps 1 and 2 to generate multiple sets of data, recorded as data sets;

[0012] Step 4: Pre-neural network model design;

[0013] Step 5: Read the validation set data of each data set in the data group, and input the validation set data into the pre-neural network model obtained in step 4;

[0014] Step 6: Repeat steps 4 and 5, adjust the model parameters by evaluating the evaluation index of the pre-neural network model, and obtain the final neural network model. The evaluation index of the model is accuracy, and its calculation formula is Among them, ZQ represents accuracy, TP represents true positive examples, TN represents true negative examples, FP represents false positive examples, and FN represents false negative examples;

[0015] Step 7: Collect real-time online car-hailing verification photos and perform feature engineering on the real-time online car-hailing verification photos;

[0016] Step 8: Input the real-time online car-hailing verification photos after feature engineering into the final neural network model to obtain the circled face position in each driver's photo.

[0017] Preferably, the step of performing feature engineering on the data set is:

[0018] Preprocessing the data set, wherein the preprocessing includes rotation, scaling, translation and mirroring;

[0019] The image samples in the data set are annotated using Python's image annotation tool to annotate the key points of the face, which include the position points of the eyes, nose, and corners of the mouth. A rectangular frame is obtained by connecting the position points of the eyes and corners of the mouth.

[0020] The annotated image samples are saved to obtain corresponding annotation information, wherein the annotation information includes the annotation category and the coordinate position of the intersection point of the two diagonal lines of the rectangular frame.

[0021] Preferably, the steps of designing the pre-neural network model are:

[0022] Load the training set and test set in the data group and generate a data iterator for step-by-step calling of the iterative model;

[0023] Design an initial neural network model and divide the image samples into n rectangular boxes on average as predetermined anchor boxes;

[0024] Use the smooth L1 loss algorithm to calculate the loss between the rectangular box and the predetermined anchor box and design the loss function;

[0025] The initial neural network model is trained using the training set until the loss function converges to obtain a pre-neural network model.

[0026] Preferably, the steps of designing the initial neural network model are:

[0027] Use convolutional neural network as the backbone network, and use ResNet50 to perform preliminary feature extraction on the image to obtain C3, C4, and C5 feature layers.

[0028] Use feature pyramid network, connect from top to bottom and side to side, sum element by element, convolution fusion, and output P3, P4, P5 effective feature layers;

[0029] The SSH structure is used, and a 5×5 convolution stack is utilized to obtain an effective output layer, wherein the output layer includes SSH1, SSH2, and SSH3.

[0030] Preferably, the steps of using ResNet50 to perform preliminary feature extraction on the image for the backbone network to obtain the C3, C4, and C5 feature layers are:

[0031] Get the pre-trained ResNet50 model from the pre-trained model library;

[0032] The ResNet50 model is truncated, retaining the first few convolutional blocks of ResNet50 and discarding the last fully connected layer;

[0033] Input the image into the truncated ResNet50 model for forward propagation;

[0034] Extract the outputs of the C3, C4, and C5 feature layers from the results of the forward propagation.

[0035] Preferably, the smooth L1 loss algorithm is used to calculate the loss between the rectangular box and the predetermined anchor box, and the loss function design step is:

[0036] The predicted bounding box is The ground-truth bounding box is in and Represent the center coordinates of the bounding box, and Represented as the width and height of the bounding box respectively;

[0037] The loss between bounding boxes is calculated using the smooth L1 loss function, which is calculated as follows: Where i represents the coordinate latitude of the bounding box, SmoothL1 is the smooth L1 loss function, and its expression is

[0038] Preferably, the step of adjusting the model parameters by evaluating the evaluation index of the pre-neural network model to obtain the final neural network model is to compare the accuracy with a preset threshold. If the accuracy is greater than the preset threshold, the current model is output and recorded as the final neural network model. If the accuracy is less than the preset threshold, return to step four to continue optimizing the model until the accuracy is greater than the preset threshold.

[0039] Preferably, a device for locating key points of a face of an online car-hailing driver based on deep learning comprises:

[0040] A face image acquisition module is used to acquire face images through a camera device and transmit the acquired face images to a feature engineering processing module;

[0041] A feature engineering processing module is used to receive the face image transmitted by the face image acquisition module, perform feature engineering processing on the image, and divide the image after feature engineering processing into a training set, a test set and a validation set. The feature engineering processing includes data set preprocessing and image sample annotation, and the divided image is transmitted to the neural network model construction module;

[0042] A neural network model building module is used to receive the image transmitted by the feature engineering processing module, build a final neural network model, and transmit the final neural network model to the key point positioning module;

[0043] The key point positioning module is used to receive the final neural network model transmitted by the neural network model building module, and use the final neural network model to locate the key points of the real-time photo.

[0044] Technical effects and advantages of the present invention:

[0045] Technical effects:

[0046] A large number of real online car-hailing verification photos are randomly selected as the data set, and feature engineering is performed on the data set. The above steps are repeated to obtain the data group, and a pre-neural network model is designed. The evaluation indicators of the pre-neural network model are evaluated according to the data in the data group, and the model parameters are adjusted to obtain the final neural network model. Real-time online car-hailing verification photos are collected, and feature engineering is performed on the real-time online car-hailing verification photos. The real-time online car-hailing verification photos after feature engineering are input into the final neural network model to obtain the circled face position in the photo of each driver.

[0047] advantage:

[0048] 1. Deep learning has a powerful feature extraction capability. Through the training of neural networks, deep learning can automatically learn richer and more abstract feature representations from raw data to better capture the details and features of the face. Compared with traditional methods, deep learning does not require manual feature design, but can automatically extract high-level features from data, avoiding the tediousness and subjectivity of feature design. This feature makes deep learning perform well in face recognition tasks;

[0049] 2. Through training, deep learning models can learn deeper internal laws, making them more sensitive to abnormal situations and emergencies. Compared with traditional methods, traditional methods usually rely on pre-defined rules or manually designed feature engineering, which limits their ability to deal with unknown situations. Deep learning models automatically extract features and representations from data without the need to define rules in advance, and have a stronger ability to adapt to unknown situations. Therefore, deep learning performs well in dealing with outliers and emergencies;

[0050] 3. The model demonstrates good generalization ability and can adapt to various scenarios and changes in data distribution. At the same time, by increasing the network structure or introducing more training data, the performance of the model can be further improved, showing excellent scalability and extensibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the overall flow chart of the present invention.

[0052] Figure 2 It is a structural diagram of the device of the present invention. DETAILED DESCRIPTION

[0053] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are only examples. The method and device for locating facial key points of online car-hailing drivers based on deep learning involved in the present invention are not limited to the various structures recorded in the following embodiments. All other implementations obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0054] The present invention provides a method for locating key points of a face of an online car-hailing driver based on deep learning, comprising the following steps:

[0055] Step 1: Randomly select a large number of real online car-hailing verification photos from the online database as the data set;

[0056] Step 2: Perform feature engineering on the data set and divide the data set into training set, test set, and validation set.

[0057] Step 3: Repeat steps 1 and 2 to generate multiple sets of data, recorded as data sets;

[0058] Step 4: Pre-design neural network model for data;

[0059] Step 5: Read the validation set data of each data set in the data group, and input the validation set data into the pre-neural network model obtained in step 4;

[0060] Step 6: Repeat steps 4 and 5 to generate the final neural network model by evaluating the evaluation index of the model. The evaluation index of the model is accuracy, which is the ratio of the number of correctly classified samples to the total number of samples. It can measure the accuracy of the overall classification of the model. The calculation formula is: Where ZQ represents accuracy, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative. The accuracy is one of the commonly used indicators for evaluating the performance of classification models, indicating the proportion of the number of samples correctly predicted by the model to the total number of samples;

[0061] Step 7: Collect real-time online car-hailing verification photos and perform feature engineering on the real-time online car-hailing verification photos;

[0062] Step 8: Input the real-time online car-hailing verification photos after feature engineering into the model to obtain the circled face position in each driver's photo.

[0063] By locating the key points of the driver's face, it can be used to identify and verify the driver's identity. This is very important for online ride-hailing service providers, as it ensures that only certified drivers can accept orders and improves the safety of passengers. By analyzing the key points of the driver's face, the driver's expression and emotional state can be inferred. This is beneficial for online ride-hailing platforms, as they can adjust service strategies based on the driver's emotional state and improve passenger satisfaction and safety. By detecting the posture of the driver's face, the driver's fatigue, distraction or discomfort can be detected in time, thereby reminding the driver to take a break or adjust his posture in time to reduce the risk of traffic accidents.

[0064] In this embodiment, it should be specifically explained that the steps of performing feature engineering on the data are:

[0065] Preprocessing the data set, wherein the preprocessing includes rotation: rotating the image at different angles to increase the diversity of the data;

[0066] Scaling: resizing images proportionally so that training data has different scales;

[0067] Translation: translate the image horizontally and vertically to simulate faces in different positions;

[0068] Mirror: Flip the image horizontally to expand the data set;

[0069] The image samples in the data set are annotated using Python's image annotation tool to annotate the key points of the face, which include the position points of two eyes, a nose, and two corners of the mouth. The position points of the eyes and the corners of the mouth are connected to obtain a rectangular frame;

[0070] The annotated image samples are saved to obtain corresponding annotation information, wherein the annotation information includes the annotation category and the coordinate position of the intersection point of the two diagonal lines of the rectangular frame.

[0071] Facial data is usually high-dimensional. Feature engineering can help us extract the most relevant and valuable features from the raw data to facilitate subsequent model training and prediction; feature engineering can help us reduce the dimension of the data, reduce the complexity and redundancy of the data, thereby improving the training speed and prediction efficiency of the model; through feature engineering processing, we can improve the model's ability to express data, improve the model's performance and generalization ability, thereby improving the model's prediction accuracy and stability; feature engineering can help us deal with missing values ​​and outliers in the data, thereby improving the robustness and stability of the model and reducing the impact of noise and interference on the model.

[0072] In this embodiment, it should be specifically explained that the steps of pre-designing a neural network model for the data are:

[0073] Loading training sets and test sets, generating data iterators for step-by-step calling of iterative models, wherein loading training sets and test sets refers to reading data from a storage medium (such as a hard disk) into memory for subsequent processing; generating data iterators refers to creating a data flow mechanism so that the model can access data batch by batch; step-by-step calling of iterative models refers to obtaining data from the data iterator batch by batch during model training or evaluation, and inputting the data into the model for forward propagation and backward propagation;

[0074] Design an initial neural network model and divide the image samples into n rectangular boxes on average as predetermined anchor boxes for predicting key points of images at different scales and aspect ratios;

[0075] The loss between the rectangular box and the predetermined anchor box is calculated using the smooth L1 loss algorithm, and the loss function is designed. The smooth L1 loss algorithm is a loss function for regression tasks, which is usually used in target detection, especially when calculating the loss between the predicted bounding box and the true bounding box. Compared with the mean square error (MSE) loss, the smooth L1 loss has a smooth curve with a lower slope at the boundary, which makes it more robust to outliers;

[0076] The initial neural network model is trained using the training set until the loss function converges to obtain a pre-neural network model.

[0077] In machine learning, especially when training neural networks, the convergence of the loss function refers to the process in which the model's prediction results gradually approach the true label and the loss value gradually decreases as the training progresses. In other words, the convergence of the loss function means that the model gradually learns the characteristics of the data during the training process and is able to make better predictions on the data.

[0078] When the loss function reaches a relatively stable value and fluctuates around this value, we can say that the loss function has converged. This means that the model has learned the pattern of the data and performs well for the samples in the training dataset. At this point, the model can usually generalize well on unseen data, that is, it can also perform well on the test dataset.

[0079] The convergence of the loss function does not necessarily mean that the model has fully learned all the characteristics of the data, because problems such as overfitting may affect the generalization ability of the model. Therefore, in practical applications, we usually verify whether the model converges by monitoring the performance of the model on the validation set or test set.

[0080] The steps of calling the iterative model step by step are:

[0081] Use a data iterator (such as PyTorch's 'DataLoader') to load a batch of data from the dataset. The data iterator is a tool for loading and processing data batch by batch, which is particularly suitable for large-scale datasets in deep learning and machine learning. The data iterator can load data from the dataset on demand, perform preprocessing and enhancement, and provide it to the model in small batches, avoiding the waste of resources and inefficiency caused by loading the entire dataset into memory at one time;

[0082] The loaded data is input into the model, and forward propagation calculation is performed to obtain the prediction result. The mean square error loss function is used to calculate the difference between the model's prediction result and the actual label. The calculation formula is: Where N is the number of samples, y iRepresented as the actual value of the i-th sample, It is expressed as the predicted value of the i-th sample. The mean square error is a commonly used loss function, which is usually used in regression tasks. It is used to measure the average value of the square difference between the model prediction value and the true value. The smaller the mean square error loss, the closer the model prediction result is to the true value.

[0083] Calculate the gradient of the loss function with respect to the model parameters, and use the optimizer to update the model parameters based on the calculated gradient.

[0084] In this embodiment, it should be specifically explained that the steps of designing the initial neural network model are:

[0085] A convolutional neural network (CNN) is used as the backbone network. ResNet50 is used for the backbone network to perform preliminary feature extraction on the image and obtain C3, C4, and C5 feature layers. The ResNet50 is a classic convolutional neural network architecture with strong feature extraction capabilities. When we use ResNet50 as the backbone network, by forward propagating the input image, we can obtain some specific feature layers in certain layers in the middle of the network (usually the last few convolution blocks). In ResNet50, these feature layers are usually called C3, C4, and C5. The C3 feature layer refers to the output feature map of the third convolution block in the backbone network. This feature layer captures lower-level semantic information in the image, such as edges, textures, etc. The C4 feature layer refers to the output feature map of the fourth convolution block in the backbone network. This feature layer captures higher-level semantic information in the image, such as object parts and shapes, etc. The C5 feature layer refers to the output feature map of the last convolution block in the backbone network. This feature layer is usually considered to be the feature layer with the most semantic information and can be used for tasks such as object detection and semantic segmentation.

[0086] Use Feature Pyramid Network (FPN), connect from top to bottom and sideways, sum each element, perform convolution fusion, and output P3, P4, P5 effective feature layers. The Feature Pyramid Network (FPN) is a deep neural network structure for extracting image features, which is usually used for tasks such as target detection and semantic segmentation. The FPN structure can detect targets at different scales and has good semantic information retention capabilities.

[0087] Using the SSH structure, a 5×5 convolution stack is used to obtain an effective output layer. The output layer includes SSH1, SSH2, and SSH3. The purpose is to increase the receptive field. The SSH structure is a network structure for face key point detection tasks. It is usually used to detect key points on the face, such as eyes, nose, mouth, etc. The SSH structure includes three parallel convolution structures, each of which uses a 5x5 convolution kernel for feature extraction. The three parallel structures are called SSH1, SSH2, and SSH3 respectively.

[0088] The top-down approach starts from the high-level features of the backbone network (such as the C5 feature layer in ResNet), and gradually reduces the resolution of the feature map through downsampling operations to obtain a multi-scale feature map;

[0089] The lateral connection is to extract features from the high-level features of the backbone network (such as the C3, C4, and C5 feature layers in ResNet) and adjust these features to the same size as the low-level features through convolution operations;

[0090] The element-by-element summation is to add the top-down feature map and the corresponding adjusted feature map element-by-element to obtain a fused feature map;

[0091] The convolution fusion is to perform a convolution operation on the fused feature map to further fuse the features and generate the final P3, P4, and P5 effective feature layers.

[0092] In this embodiment, it should be specifically explained that the steps of using ResNet50 to perform preliminary feature extraction on the image and obtain the C3, C4, and C5 feature layers are as follows:

[0093] Get a pre-trained ResNet50 model, usually by downloading and loading it from a repository containing pre-trained models;

[0094] The ResNet50 model consists of several convolutional blocks, and we need to truncate the model to get the intermediate layer outputs of interest. Specifically, we need to keep the first few convolutional blocks of ResNet50 and discard the last fully connected layer;

[0095] Input the image into the truncated ResNet50 model for forward propagation. Through the forward propagation of the network, the output of the image in the intermediate layer of interest can be obtained, which is the C3, C4, and C5 feature layers;

[0096] The outputs of the C3, C4, and C5 feature layers are extracted from the results of the forward propagation. These feature layer outputs usually have different spatial resolutions and semantic levels, and can be used for subsequent tasks such as object detection and semantic segmentation.

[0097] In this embodiment, it should be specifically explained that the smooth L1 loss algorithm is used to calculate the loss between the rectangular box and the predetermined anchor box, and the steps of designing the loss function are:

[0098] Set the predicted bounding box The ground-truth bounding box is in and Represent the center coordinates of the bounding box, and Represented as the width and height of the bounding box respectively;

[0099] The loss between bounding boxes is calculated using the smooth L1 loss function, which is calculated as follows: Where i represents the coordinate latitude of the bounding box, SmoothL1 is the smooth L1 loss function, and its expression is

[0100] In this embodiment, it should be specifically explained that the step of generating the final neural network model by evaluating the evaluation index of the evaluation model is to compare the accuracy with the preset threshold. If the accuracy is greater than the preset threshold, the current model is output and recorded as the final neural network model. If the accuracy is less than the preset threshold, return to step four to continue optimizing the model until the accuracy is greater than the preset threshold.

[0101] In this embodiment, it should be specifically explained that Figure 2 As shown, a device for locating key points of a face of an online car-hailing driver based on deep learning, the device comprising:

[0102] A face image acquisition module is used to acquire face images through a camera device and transmit the acquired face images to a feature engineering processing module;

[0103] A feature engineering processing module is used to receive the face image transmitted by the face image acquisition module, perform feature engineering processing on the image, and divide the image after feature engineering processing into a training set, a test set and a validation set. The feature engineering processing includes data set preprocessing and image sample annotation, and the divided image is transmitted to the neural network model construction module;

[0104] A neural network model building module is used to receive the image transmitted by the feature engineering processing module, build a final neural network model, and transmit the final neural network model to the key point positioning module;

[0105] The key point positioning module is used to receive the final neural network model transmitted by the neural network model building module, and use the final neural network model to locate the key points of the real-time photo.

[0106] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0107] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for locating key points of a ride-hailing driver's face based on deep learning, characterized in that: The following steps are involved: Step 1: Randomly select a large number of real online car-hailing verification photos from the online database as the data set; Step 2: Perform feature engineering on the data set and divide the data set into training set, test set, and validation set. Step 3: Repeat steps 1 and 2 to generate multiple sets of data, recorded as data sets; Step 4: Pre-neural network model design; Step 5: Read the validation set data of each data set in the data group, and input the validation set data into the pre-neural network model obtained in step 4; Step 6: Repeat steps 4 and 5, adjust the model parameters by evaluating the evaluation index of the pre-neural network model, and obtain the final neural network model. The evaluation index of the model is accuracy, and its calculation formula is Among them, ZQ represents accuracy, TP represents true positive examples, TN represents true negative examples, FP represents false positive examples, and FN represents false negative examples; Step 7: Collect real-time online car-hailing verification photos and perform feature engineering on the real-time online car-hailing verification photos; Step 8: Input the real-time online car-hailing verification photos after feature engineering into the final neural network model to obtain the circled face position in each driver's photo.

2. According to claim 1, a method for locating key points of the face of an online car-hailing driver based on deep learning is characterized by: The steps of feature engineering the data set are as follows: Preprocessing the data set, wherein the preprocessing includes rotation, scaling, translation and mirroring; The image samples in the data set are annotated using Python's image annotation tool to annotate the key points of the face, which include the position points of the eyes, nose, and corners of the mouth. A rectangular frame is obtained by connecting the position points of the eyes and corners of the mouth. The annotated image samples are saved to obtain corresponding annotation information, wherein the annotation information includes the annotation category and the coordinate position of the intersection point of the two diagonal lines of the rectangular frame.

3. The method for locating key points of a driver's face of an online car-hailing vehicle based on deep learning according to claim 1, characterized in that: The steps of designing the pre-neural network model are as follows: Load the training set and test set in the data group and generate a data iterator for step-by-step calling of the iterative model; Design an initial neural network model and divide the image samples into n rectangular boxes on average as predetermined anchor boxes; Use the smooth L1 loss algorithm to calculate the loss between the rectangular box and the predetermined anchor box and design the loss function; The initial neural network model is trained using the training set until the loss function converges to obtain a pre-neural network model.

4. The method for locating key points of a driver's face of an online car-hailing vehicle based on deep learning according to claim 3, characterized in that: The steps of designing the initial neural network model are: Use convolutional neural network as the backbone network, and use ResNet50 to perform preliminary feature extraction on the image to obtain C3, C4, and C5 feature layers. Use feature pyramid network, connect from top to bottom and side to side, sum element by element, convolution fusion, and output P3, P4, P5 effective feature layers; The SSH structure is used, and a 5×5 convolution stack is utilized to obtain an effective output layer, wherein the output layer includes SSH1, SSH2, and SSH3.

5. The method for locating key points of the face of an online car-hailing driver based on deep learning according to claim 4, characterized in that: The steps for performing preliminary feature extraction on the image using ResNet50 in the backbone network to obtain the C3, C4, and C5 feature layers are as follows: Get the pre-trained ResNet50 model from the pre-trained model library; The ResNet50 model is truncated, retaining the first few convolutional blocks of ResNet50 and discarding the last fully connected layer; Input the image into the truncated ResNet50 model for forward propagation; Extract the outputs of the C3, C4, and C5 feature layers from the results of the forward propagation.

6. The method for locating key points of a driver's face of an online car-hailing vehicle based on deep learning according to claim 3, characterized in that: The smooth L1 loss algorithm is used to calculate the loss between the rectangular box and the predetermined anchor box, and the steps of designing the loss function are: The predicted bounding box is The ground-truth bounding box is in and Represent the center coordinates of the bounding box, and Represented as the width and height of the bounding box respectively; The loss between bounding boxes is calculated using the smooth L1 loss function, which is calculated as follows: Where i represents the coordinate latitude of the bounding box, SmoothL1 is the smooth L1 loss function, and its expression is 7. The method for locating key points of a driver's face of an online car-hailing vehicle based on deep learning according to claim 1, characterized in that: The step of adjusting the model parameters by evaluating the evaluation index of the pre-neural network model to obtain the final neural network model is to compare the accuracy with the preset threshold. If the accuracy is greater than the preset threshold, the current model is output and recorded as the final neural network model. If the accuracy is less than the preset threshold, return to step 4 to continue optimizing the model until the accuracy is greater than the preset threshold.

8. A device for locating key points of a ride-hailing driver's face based on deep learning, characterized in that: The device comprises: A face image acquisition module is used to acquire face images through a camera device and transmit the acquired face images to a feature engineering processing module; A feature engineering processing module is used to receive the face image transmitted by the face image acquisition module, perform feature engineering processing on the image, and divide the image after feature engineering processing into a training set, a test set and a validation set. The feature engineering processing includes data set preprocessing and image sample annotation, and the divided image is transmitted to the neural network model construction module; A neural network model building module is used to receive the image transmitted by the feature engineering processing module, build a final neural network model, and transmit the final neural network model to the key point positioning module; The key point positioning module is used to receive the final neural network model transmitted by the neural network model building module, and use the final neural network model to locate the key points of the real-time photo.

Images