An intelligent and safe pick-up and drop-off system for children

Abstract

The present invention discloses an intelligent and safe pick-up and drop-off system for infants, comprising on-site equipment, a cloud platform and a mobile terminal. The on-site equipment comprises: an entry-to-school portrait collection device, an exit-to-school portrait collection device, an entry-to-school dynamic face recognition host, an exit-to-school dynamic face recognition host, an analysis host and a POE switch; the cloud platform comprises a read-in database, a standard database, an APP platform, and a Web portrait input direct delivery; the entry-to-school dynamic face recognition device and the exit-to-school dynamic face recognition device perform face recognition and transmit the comparison record to the analysis host, the analysis host performs analysis on the real-time entry-to-school and exit-to-school comparison records, performs logical judgment, and transmits the analysis result to the cloud platform, and after the cloud platform obtains the data, the APP platform pushes the information.

Description

Technical Field

[0001] The present invention relates to a safety system based on computer-aided biometric identification, and more particularly to an intelligent and safe child transportation system. Background Art

[0002] Kindergarten and primary school education is the starting point of early childhood education in my country. It is the earliest educational platform for basic education. The age of students is basically around 3 to 12 years old. Children of this age have begun to become sensible and their curiosity has gradually increased. Parents also give them more freedom. In this case, the issue of student safety has also surfaced, and schools and parents have to pay great attention to it. At the same time, social security issues are also prominent. In 2018, children were killed at the school gates of Shanghai World Foreign Language Primary School, Chongqing Banan District Kindergarten, and Beijing Xicheng District Xuanshi No. 1 Primary School without the protection of teachers and parents.

[0003] With the rapid development of computer technology, the automatic recognition technology of biometric features has been widely researched and developed. Face recognition is based on facial image identification, which aims to enable computers to identify identities through facial images. It is an intelligent system based on high technologies such as image understanding, pattern recognition, and computer vision. Compared with other human biometric feature recognition technologies, it is direct, friendly, and convenient, and is the most natural, direct, and convenient means.

[0004] Face recognition technology has broad application prospects, especially in non-contact environments and without disturbing the person being detected. The superiority of face recognition technology far exceeds other biological detection methods. In general, human facial features are very stable and reliable, and easy to carry. Face image is also the primary feature used to distinguish people. Because of this, human images are widely used in security department investigations, security protection, and access control. Secondly, the collection of face images is very convenient, especially the image collection method based on standard video. At present, most security systems use standard cameras within the visible light range, which fully meet the requirements for the collection of face images. The collection of face images is non-contact and does not require active cooperation from people, which maximizes the response speed of the system.

[0005] This system is an intelligent system developed based on the cutting-edge technology of dynamic face recognition. According to the actual situation of the school, a set of face recognition security intelligent pick-up and drop-off solutions based on student safety are tailored to the school. It uses artificial intelligence AI big data analysis, combined with face recognition algorithms, to intelligently determine the identity of the people entering, and then combines with the school security room staff to create a fast, benign, efficient and safe intelligent attendance solution. Summary of the invention

[0006] The present invention is a smart and safe pick-up and drop-off system for young children, including on-site equipment, a cloud platform and a mobile terminal. The on-site equipment includes: portrait collection equipment for entering school, portrait collection equipment for leaving school, dynamic face recognition equipment for entering school, dynamic face recognition equipment for leaving school, an analysis host and a POE switch; the cloud platform includes a read-in database, a specification database (which backs up the read-in database), a data analysis module, an APP platform (APP portrait template collection and distribution, APP instant information push, APP advertising release), and Web portrait input direct distribution; the dynamic face recognition equipment for entering school and the dynamic face recognition equipment for leaving school transmit the comparison records to the analysis host after performing face recognition, and the analysis host analyzes the comparison records of entering and leaving school in real time, and transmits the analysis results to the cloud platform. After the cloud platform obtains the data, the APP platform pushes the information. The portrait collection equipment for entering school and the portrait collection equipment for leaving school are provided with a real-time portrait collection system, and the real-time portrait collection system includes a same-person tracking and recognition module:

[0007] The same person tracking and identification module adopts n:N face recognition technology, and compares the feature values ​​formed by the portraits collected in real time with all the portraits collected within a certain time range before. The feature value refers to when n faces are captured in the lens, a temporary feature value is formed in the cache, and an n:N interlaced comparison is performed with the N feature values ​​already retained in the original cache. If the comparison result is greater than the set threshold, it is considered to be the same person and the new feature value is not retained in the cache. Otherwise, the new feature value is retained in the cache and the feature value in the cache becomes N+1.

[0008] The portrait collection device for entering the school and the portrait collection device for leaving the school are provided with a real-time portrait collection system, and the real-time portrait collection system includes an image collection and screening module:

[0009] By setting the head ratio parameters and image quality parameters, if the ratio of the head in the lens is less than the set parameters, no acquisition will be performed; if the pixel value of both eyes is calculated and is lower than the set value, it does not meet the requirements of the image quality parameters and no acquisition will be performed, thereby realizing the screening of image acquisition.

[0010] The portrait collection device for entering the school and the portrait collection device for leaving the school are provided with a real-time portrait collection system, and the real-time portrait collection system also includes a portrait quality optimization module for the same person:

[0011] That is, during the period of time from the time when the person appears in front of the portrait acquisition device, the eyes and nose wings of the captured face are used as the coordinate system and projected into the positive coordinate system. Through vector calculation, the one with the smallest deviation is the most frontal portrait frame. From all the captured portrait frames, the most frontal frame with the most pixels for both eyes is automatically selected as the comparison frame.

[0012] The dynamic face recognition host is provided with a face comparison module, which includes: a 2D restoration 3D modeling module, also called 2D+ modeling, that is, a face 3D restoration feature value extraction process. First, the 3D face data is divided into 2D RGB data + depth data, and then the 2D face modeling method is used to process the 2D RGB data, and then the depth data is processed, and the 3D face image organ template is expanded to calculate the associated three-dimensional geometric vectors, that is, to generate 3D restoration feature values.

[0013] Compared with 2D face modeling, the data of 2D+ face modeling is relatively complete and three-dimensional, and the generated facial feature values ​​are 3D restored feature values, so the recognition angle is wide.

[0014] After 2D+ modeling, the head portrait captured in real time in the video can be deflected at a certain angle, ±25° left and right, ±15° up and down, and ±10° rotated, and the accuracy of liveness detection is greatly improved.

[0015] At the same time, due to the difficulty in collecting 3D face photos, 2D+ modeling that restores 3D through 2D has practical significance.

[0016] An interference restoration correction module, the interference restoration correction module comprising an illumination restoration module, a physiological age restoration module, a race restoration module, an expression restoration module, a posture restoration module and a cover restoration module;

[0017] Comparison engine module, including: neural network design module, loss function design module, deep learning module, feature comparison module;

[0018] Neural Network Design Module:

[0019] The neural network model for feature extraction is a neural network architecture based on imagenet. The recognition accuracy and parameter quantities of these neural network models on imagenet are shown in the following table:

[0020]

[0021] You can choose one of the neural network models as the underlying architecture as needed.

[0022] Loss function design module:

[0023] After the neural network model is selected, the next thing to do is to design a loss function to serve as a standard for adjusting the neural network weight parameters during training.

[0024] Commonly used loss functions are: softmax, sphereface, cosface, arcface.

[0025] A good loss function can basically determine the quality of the feature extraction model. Different loss functions can adjust the weight parameters of the position of the comparison points of the selected face, the weight of the comparison points, the number of comparison points, the distribution ratio of the comparison area, and the number of digits of precision of floating-point operations.

[0026] The system can automatically calculate the number of pixels between the two eyes according to the quality of the collected image, and automatically select a corresponding one of the above four loss functions according to the number of pixels between the two eyes.

[0027] Deep Learning Modules:

[0028] Deep learning is divided into two steps: training and inference

[0029] Training: Prepare the data and train the designed neural network model. A large number of data sets can come in handy at this time. The larger the data set, the better the value of the trained neural network weight parameters, the higher the generalization ability, and the better the robustness.

[0030] Reasoning: Input face data, calculate through neural network, and output feature data. The accuracy of the output feature data is determined by the previous neural network design, loss function design, and training.

[0031] Feature matching module:

[0032] After completing the feature extraction step, the next step is to calculate the similarity of the features. Features are structured data, whether one-dimensional, two-dimensional or N-dimensional, so you only need to calculate the similarity between the features of the image to be identified and the features in the library, and then based on the similarity, you can determine whether they are the same person and realize face recognition. The feature similarity calculation is to calculate the similarity of two sets of data. The methods are:

[0033] 1. Cosine similarity: the difference in vector direction.

[0034]

[0035] 2. Euclidean distance: calculate the spatial distance between two points.

[0036]

[0037] Cosine similarity considers the spatial angle, while Euclidean distance considers more the distance between two straight lines in space. Of course, in addition to these two, there are also Manhattan distance, Minkowski distance, Pearson correlation coefficient, Spearman rank correlation coefficient, etc. for similarity correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the working principle diagram of this patent.

[0039] Figure 2 This is the system diagram of this patent.

[0040] Figure 3 It is a schematic diagram of three-dimensional modeling.

[0041] Figure 4 This is the flow chart of this patent.

[0042] Figure 5 This is the functional diagram of this patented system. DETAILED DESCRIPTION

[0043] The specific implementation is described below in conjunction with the accompanying drawings:

[0044] 1. If Figure 1 , the process of this patent is:

[0045] First, the dynamic face recognition host performs face recognition and uploads the matching records to the analysis host. Then, the analysis host analyzes the real-time matching records of entering and leaving the school and reports the analysis results to the cloud platform. After the cloud platform obtains the data, the APP pushes it to the teachers and parents corresponding to the student account.

[0046] School Enrollment: During normal school hours, students enter the school through the school gate camera (entry), and at the same time, a headshot image is generated and sent to the class teacher and parents APP to inform the parents that their children have entered the school. After arriving at the class, the teacher confirms again and signs back, and the student account will receive a push message again, reminding the student's parents that the child has arrived in the classroom.

[0047] School dismissal: Leaving school during normal school dismissal hours, the school gate camera (out) is used to record and capture the student and parent portraits, and the student and parent information is matched. If the match is successful, the child leaving the school and the matching parent portrait are produced, and the APP pushes it to the class teacher and parents, informing them that the child has been picked up by the parent and which parent picked him up. If the match is not successful, it is automatically pushed to the security room PC, prompting an alarm, and the child is left alone without anyone to pick him up.

[0048] Leaving school during abnormal time periods: The school gate camera (exit) is used to record and capture the headshots of the student and the parent, and the information of whether the student's parents have asked for leave to leave school is matched. If no match is successful, it will be automatically pushed to the PC in the security room to prompt an alarm. If a student leaves school alone during abnormal time periods, it will be automatically pushed to the PC in the security room to prompt an alarm.

[0049] In order to avoid blocking each other, when a large number of people enter the school, everyone is required to stand at the entrance identification line, stay for about 1-2 seconds, self-confirm, and see their names displayed on the electronic screen before they are allowed to enter the school; similarly, when a large number of people leave the school, everyone is required to stand at the exit identification line, stay for about 1-2 seconds, self-confirm, and see their names displayed on the electronic screen before they are allowed to leave the school.

[0050] 2. Figure 2 This system can adopt a distributed architecture. Each school is equipped with an IP camera, a dynamic face recognition host (one camera corresponds to one host), a local analysis host, and a POE switch. The analysis results will be uploaded to the virtual server on the cloud platform in real time. The cloud platform virtual server adopts the B / S architecture + APP mode. The administrator can manage the relevant data of each school through the web page. The cloud platform pushes the results uploaded by different schools to the parents and class teachers corresponding to each student account through the APP.

[0051] According to different needs, you can use:

[0052] 1. Multi-channel video acquisition:

[0053] High-definition POE camera: 5 million pixels, wide-angle lens, manual focus (not automatic), installation height 2.2-3.0 meters, focus depth of field about 5.0 meters, effective field width (monitoring width) <3 meters. During installation, the horizontal axis should be kept horizontal and cannot be tilted, the downward inclination of the vertical axis is less than 100, and the number of cameras is determined according to the width of the school gate.

[0054] Dynamic face recognition host: J1900, 4G memory, 64G SSD, each recognition host is deployed with one dynamic face recognition software. The number of recognition hosts is determined according to the number of cameras, one recognition host for each camera.

[0055] Analysis host (VMS): i5, 16G memory, SSD 128G, quantity: 1 unit, deployment: matching logic program, template management software, identification record management software, communication and equipment management software, equipment Mysql database software, 1 set each.

[0056] POE switch: 8 POE network ports, 8 ordinary network ports, 1 WAN port.

[0057] Electronic display screen: At least one 40-inch display screen is required for each entrance and exit. The number of electronic display screens can also be determined based on the width of the door and the number of cameras.

[0058] 2. Two-way video acquisition: If the school only needs two cameras (one in and one out)

[0059] 1. Camera configuration is the same as 1 above.

[0060] 2. Configure a high-performance host (i7, 16G memory, SSD 128G) to replace the hosts in parts 2 and 3 above, which can simultaneously meet the computing performance of two-way dynamic face recognition and analysis.

[0061] 3. No need.

[0062] 4. The POE switch is changed to: 2 POE network ports, 2 ordinary network ports, and 1 WAN port.

[0063] 5. Electronic display screen: 1 in, 1 out, 40 inches, 2 in total.

[0064] 3. Real-time portrait acquisition software

[0065] 1. Real-time and effective portrait collection:

[0066] 1.1、Fan tracking identification:

[0067] 1.1.1. Human trajectory tracking: Based on an algorithm that combines motion and human shape templates, the real-time collected human portraits are identified and tracked at the front end.

[0068] 1.1.2. Deduplication of same-person faces by face recognition: Due to occlusion or leaving the camera range and then re-entering, the trajectory of the same-person face cannot be tracked. N:N face recognition technology is used to compare the instantly captured portrait with the feature values ​​(stored in the cache) formed by all the portraits captured in the previous 10 minutes (the length of time can be set according to the cache size. The larger the cache, the more feature values ​​of the portrait can be retained and the longer the time). The same-person face recognition automatically deduplicates the portrait.

[0069] The innovation is: when n faces are captured in the camera, temporary feature values ​​are formed in the cache, and an n:N interlaced comparison is performed with the N feature values ​​already retained in the original cache. If the comparison result is greater than the set threshold, it is considered to be the same person and the new feature value is not retained in the cache. Otherwise, the new feature value is retained in the cache and the feature value in the cache becomes N+1.

[0070] 1.2. The software can set the ratio parameters of human heads and image quality parameters. If the ratio of human heads in the lens is less than the set parameters, no collection will be done. If the image quality parameters are not met, no collection will be done, because even if the collected portraits cannot meet the pixel requirements for comparison. The settings of these two software parameters can filter and collect portraits, and can effectively cooperate with cameras with different pixels to effectively collect portraits.

[0071] 1.3. Quality optimization: For tracking of the same person, the person will be automatically selected from all captured portrait frames, with the most positive frame and the one with the most eye pixels, within 3 seconds before the person appears in the camera lens, or before the person leaves (when the person appears in the camera lens for less than 3 seconds). Because the positive coordinate system of the portrait can be calibrated and set by the software, the eyes and nose wings of the face captured by the camera are the coordinate system, which is projected to the positive coordinate system. Through vector calculation, the frame with the smallest deviation is the most positive portrait frame.

[0072] 2. The multi-algorithm fusion engine uses the above algorithms to perform mixed operations to screen the same person in the fastest and most effective way. If it is confirmed to be the same person, the portrait collected again will no longer be transmitted to the background. This greatly saves the background CPU computing capacity, transmission bandwidth, and storage hardware resources.

[0073] 3. Capture portrait photos and panoramic photos simultaneously:

[0074] Synchronously combined with the video, the captured portrait is marked with a timestamp and location stamp; double-click the portrait photo to pop up a panoramic photo and trace the scene.

[0075] 4. About face matching software:

[0076] 1. If Figure 3 , about 2D portrait collection, 3D restoration modeling. Wide recognition angle (3D modeling allows the head portrait in the video to rotate at a certain angle, left and right ±25°, up and down ±15°, and rotate ±10°).

[0077] 2. Interference restoration correction algorithm: lighting restoration, physiological age restoration, race restoration, expression restoration, posture restoration, and cover restoration.

[0078] 3. Hybrid matching engine: The eigenvalue matching algorithm is combined with the deep learning algorithm. The two algorithms are integrated. According to the obtained image quality, the effective focus of the two algorithms is automatically selected, and the location, number and weight distribution ratio of the matching points that are conducive to accurate matching are selected, which effectively solves the problem of being unable to model and match due to low image clarity. Therefore, the recognition rate is high and the matching speed is fast.

[0079] 5. Matching logic module (on the analysis host)

[0080] 1. Database association: Each child has a unique ID, and their parents are associated with their children in the database, including photos and basic information;

[0081] 2. When a child appears in the camera and is successfully identified, the parent associated with the child's unique identifier is matched among all successfully matched portraits within 15 seconds before and after the successful identification (this time can be set) to find the parent, and analyze the matched records to find out the records of the student entering and leaving the school with the parents.

[0082] 6. Template management and recognition record management software (on the analysis host)

[0083] 1. Face template management: You can select the terminal to upload and issue the template, realize the template upload and issue function between different terminals; realize the unified add, delete and modify functions. 2. Recognition record reception statistics management: face recognition results / report statistics, support export function, export to Excel table.

[0084] 3. Query reported records, local templates, and local snapshot records.

[0085] 7. Communication and equipment management software (on the analysis host)

[0086] 1. Upload the analyzed data to the cloud server in real time, mark the data that has been uploaded successfully or failed, and consider how to resume the data upload and re-upload it at a scheduled time when the data upload fails or the network is interrupted to ensure that the records are not lost.

[0087] 2. Monitor the working status of cameras, switches, dynamic face recognition hosts and the machine in real time, and upload them to the cloud platform so that the cloud platform knows the working status of each device and realizes real-time monitoring of the operating status of each device in the cloud.

[0088] 3. Equipment management: realize the addition, deletion, modification and query of each subsystem, permission setting, user management, alarm function management and communication interface management.

[0089] 4. Regularly upload heartbeat messages to actively handshake with the server. The server can also send or modify the face template by analyzing the socket interface of the host.

[0090] 5. Reserve an external data interface so that other systems can obtain data with specified conditions through the interface.

[0091] 8. Cloud Server Deployment

[0092] 1. Cloud server is a virtual server that can be expanded according to actual needs.

[0093] 2. The cloud server mainly implements two functional modules: one is the management and access of web pages, and the other is to support access by mobile APP.

[0094] 3. Cloud server software deployment

[0095] 3.1 Database system:

[0096] The data read into the database is some original ecological data. First of all, the data needs to be layered and extracted to the ODS layer. After processing, a standard data layer is formed and some data conversion is done. After the data is standardized, a standard database is formed. The data of the two databases can be used as backup for each other. Then, data screening and statistics are performed. Finally, the data is extracted and displayed using the data automatic report generation tool to form various visual reports.

[0097] 3.1.1. Receive and save the entry and exit records from different analysis hosts (different schools);

[0098] 3.1.2、Monitor the operating status of each subsystem, generate logs and alarm for abnormal situations;

[0099] 3.1.3. Subsystem management ensures that data uploaded from different subsystems do not conflict.

[0100] 3.1.4 User management: except for the super administrator, other administrators can only create users with a lower level than themselves. Administrators can only manage leaf node users created by them.

[0101] 3.1.5. Permission management: the permissions assigned by a parent user to a child user can only be the permissions owned by the parent user.

[0102] 3.1.6 Various statistical reports

[0103] 3.2、Data automatic generation report system software:

[0104] The report system mainly consists of two parts: report design template and report parsing template. The report design module can design various functions in reports such as tables, graphics, parameters, controls, filling, printing, and exporting. It is an integrated platform for report application development, debugging, and deployment; the report parsing module is mainly used to read and parse the templates made by the design module, and convert the templates into HTML pages, so that users can view, modify, and print data through the browser. The report parsing module parses the Servlet form of the report in the web environment, and users interact with it through the browser.

[0105] The reporting system is built with a three-tier structure that has distinct hierarchies. It connects to different data sources through a variety of connection methods. All report production work is completed in the design module and is ultimately presented to users through the analysis module.

[0106] 3.2.1. Implement report management, user management, and authority management;

[0107] 3.2.2. Component-based design, responsive layout, adaptive screen size and resolution;

[0108] 3.2.3. Directly connect to the database through JDBC, associate multiple data sources, bind data once, switch between multiple charts, and combine charts as desired;

[0109] 3.2.4. By dragging the series, you can easily classify and accumulate analysis, making the display more direct;

[0110] 3.2.5. Can integrate Baidu / Google and other GIS maps to achieve map zooming, location positioning and drilling;

[0111] 3.2.6, 18 kinds of controls, making human-computer data interaction simpler;

[0112] 3.2.7. Data display - interactive analysis - multi-dimensional analysis

[0113] 3.2.8. Support concurrent processing of a large number of users and real-time data analysis

[0114] 3.3 Data query and filtering software

[0115] 3.3.1、Data display - interactive analysis - controls - parameter linkage, conditional interaction, layer-by-layer constraints, precise positioning, and rapid screening;

[0116] 3.3.2. The user inputs query conditions through the interface to control the content and format of the report display.

[0117] 3.3.3. In addition to the parameters input by the user, some parameters required in the report are determined by the system environment, such as the user name, role, current date and time of the current logged-in user, etc. These can be obtained through design or configuration for flexible data analysis.

[0118] 3.3.4. The parameters in the report can be divided into template parameters and global parameters according to their scope of use. Global parameters: all templates under the current project application can be used; template parameters: only the current template can be used. According to their different usage methods, they can be divided into template parameters and global parameters. Global parameters are to directly retrieve data that meets the conditions. Global parameters are to retrieve all data and then set filtering conditions to obtain the data of the corresponding parameters. Therefore, if the amount of data is large, using global parameters can improve the access efficiency of the report.

[0119] 9. Mobile APP software

[0120] 1. APP account establishment: adopt hierarchical control permissions, that is, each user can only manage the sub-users or grandchildren users created by him, and the permissions that can be issued are only the permissions owned by the current user. The student user is created by default when the system is initialized. This is an account for parents.

[0121] 2. Collection of photos from the database through the APP: The student’s facial template information can be provided by the school or collected through the teacher-side APP. Parents need to log in to the student’s account, take a photo on the APP and enter relevant information, submit a parent association application to the system, and the application will be reviewed and approved by the system administrator and sent to the corresponding school database.

[0122] 3. Functional authority control function: The pages or functions that can be operated by different users logging into the system need to be divided according to the authority level of the current logged-in user.

[0123] 4. APP information release:

[0124] 4.1. Pick-up and drop-off information push

[0125] 4.2. Advertising Push

Claims

1. A smart and safe pick-up and drop-off system for children, comprising a field device, a cloud platform and a mobile terminal, wherein the field device comprises: The portrait collection equipment for entering the school, the portrait collection equipment for leaving the school, the dynamic face recognition host for entering the school, the dynamic face recognition host for leaving the school, the analysis host and the POE switch; the cloud platform includes a read-in database, a specification database, an APP platform, and a Web portrait input directly issued; the dynamic face recognition host for entering the school and the dynamic face recognition host for leaving the school transmit the comparison records to the analysis host after face recognition, and the analysis host analyzes the real-time comparison records of entering and leaving the school, makes logical judgments and matches, and transmits the analysis results to the cloud platform. After the cloud platform obtains the data, the APP platform pushes the information; The portrait collection device for entering the school and the portrait collection device for leaving the school are provided with a real-time portrait collection system, and the real-time portrait collection system includes a same-person tracking and recognition module: The n:N face recognition technology is used to compare the feature values ​​formed by the portraits collected in real time with those formed by all the portraits collected within a certain time range before. The feature value refers to when n faces are captured in the camera, a temporary feature value is formed in the cache, and an n:N interlaced comparison is performed with the N feature values ​​already reserved in the original cache. If the comparison result is greater than the set threshold, it is considered to be the same person and the new feature value is not retained in the cache. Otherwise, the new feature value is retained in the cache and the feature value in the cache becomes N+1. The portrait collection device for entering the school and the portrait collection device for leaving the school are provided with a real-time portrait collection system, and the real-time portrait collection system includes an image collection and screening module: By setting the head ratio parameters and image quality parameters, if the head ratio in the lens is less than the set parameters, the image will not be captured; Calculate the pixel value of both eyes. If it is lower than the set value, it does not meet the requirements of the image quality parameters and will not be collected, thus realizing the screening of image collection.

2. The intelligent and safe child transportation system according to claim 1, characterized in that: The portrait collection device for entering the school and the portrait collection device for leaving the school are provided with a real-time portrait collection system, and the real-time portrait collection system includes a portrait quality optimization module for the same person: That is, during the period of time from the time when the person appears in front of the portrait acquisition device, the eyes and nose wings of the captured face are used as the coordinate system and projected into the positive coordinate system. Through vector calculation, the one with the smallest deviation is the most frontal portrait frame. From all the captured portrait frames, the most frontal frame with the most pixels for both eyes is automatically selected as the comparison frame.

3. The intelligent and safe pick-up and drop-off system for children according to claim 1 is characterized in that: the dynamic face recognition host for entering and leaving the school is provided with a face comparison module, and the face comparison module comprises: 3.

1. 2D restoration 3D modeling module, the 2D restoration 3D modeling module first divides the 3D face data into 2D RGB data plus depth data, then uses the 2D face modeling method to process the 2D RGB data, then processes the depth data, expands it according to the 3D face organ template, and calculates the related three-dimensional geometric vectors; 3.

2. Interference restoration correction module, which includes an illumination restoration module, a physiological age restoration module, a race restoration module, an expression restoration module, a posture restoration module and a cover restoration module; 3.

3. Comparison engine module, including: neural network design module, loss function design module, deep learning module, feature comparison module; 3.3.

1. The neural network design module refers to a neural network model for feature extraction, which is a neural network architecture based on imagenet; 3.3.

2. After the loss function design module selects the neural network model, the standard for adjusting the neural network weight parameters during training; 3.3.

3. The deep learning module is input with face data and outputs feature data. The accuracy of the output feature data is determined by the neural network design, loss function design, and training of the deep learning module. 3.3.

4. The feature comparison module calculates the similarity between the features of the image to be identified and the features in the library, and then determines whether they are the same person based on the similarity, thereby realizing face recognition.

Images