A middle- and long-distance running evaluation method and system based on face and weight recognition

Abstract

The present invention relates to the technical field of middle- and long-distance running evaluation, and in particular to a middle- and long-distance running evaluation method and system based on face and body weight recognition, wherein the system comprises data acquisition, an algorithm host, a cloud server and a terminal module. In the data acquisition module, a camera group and a pressure sensing pad respectively collect image and pressure data; the algorithm host module processes data, identifies face and body features, tracks personnel, calculates scores and analyzes violations; the cloud server stores various types of data; and the terminal module is used for information entry and result display. The evaluation method includes steps such as data acquisition, identity recognition and tracking, score calculation, violation analysis, data integration storage and exception handling correction. The present invention solves many problems of existing RFID and AI face recognition evaluation methods, such as RFID cannot recognize biometric features and judge violations, and AI face recognition is interfered by the environment. Accurate identity recognition, full-process tracking, accurate score calculation, comprehensive violation detection and efficient data management are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of middle- and long-distance running evaluation, and in particular to a middle- and long-distance running evaluation method and system based on face and body weight recognition. Background Art

[0002] The existing middle and long-distance running evaluation methods are divided into two methods: RFID and AI face recognition, which have the following defects:

[0003] RFID method:

[0004] Technical limitations: RFID (Radio Frequency Identification) technology is mainly used in middle- and long-distance running tests to achieve timing and other functions by having test participants wear RFID tags; however, it cannot provide biometric identification, which means that it is impossible to accurately determine whether the person wearing the tag is the real test participant, which poses a risk of substitution. For example, in some large-scale middle- and long-distance running tests, if someone steals someone else's RFID tag and wears it, the system cannot identify the identity difference through the tag, making it possible for the test participant to succeed;

[0005] Poor experience and potential cheating: Participants are required to wear extra props, which not only brings extra burden to the participants and affects their running experience, but may also affect the test results due to problems such as improper wearing or damage of props. Moreover, RFID technology cannot judge whether a person has violated the rules or replay videos, and cannot effectively monitor illegal behaviors such as taking shortcuts, which gives cheaters an opportunity to take advantage and undermines the fairness and accuracy of the evaluation. For example, on a track with many curves, participants may bypass part of the track to shorten the running distance, and the RFID system cannot detect such cheating behavior.

[0006] AI face recognition method:

[0007] Environmental influence on recognition effect: AI face recognition technology is greatly affected by environmental factors such as lighting and posture in middle and long-distance running tests. Under strong light, the face may reflect light, making some facial features difficult to recognize. In terms of posture, when the participants run with their heads down or their heads turn or are blocked, it is difficult to extract facial features and the recognition accuracy decreases. For example, in outdoor middle and long-distance running tests, if the sun shines directly on the camera area, the facial images of some participants may appear too bright, affecting the accurate extraction of facial features by the face recognition algorithm.

[0008] Limited recognition area: There are problems with small recognition area and short distance. This means that participants can only be effectively recognized in a specific area close to the camera. If they deviate from this area or are too far away from the camera during running, the face recognition system may not work properly. In long-distance running tests, participants are distributed on a long track. This limitation of the recognition area may cause some people to be unable to be accurately recognized at key positions (such as the outside of a curve), thus affecting the accurate recording of the identity and running trajectory of the person during the entire evaluation process, and further affecting the accuracy of score calculation and violation monitoring.

[0009] Therefore, in order to solve the above problems, a middle- and long-distance running evaluation method and system based on face and weight recognition is proposed. Summary of the invention

[0010] The purpose of the present invention is to provide a method and system for evaluating middle- and long-distance running based on face and weight recognition, so as to solve the problems raised in the above-mentioned background technology.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A method for evaluating middle- and long-distance running based on face and weight recognition comprises the following steps:

[0013] Step S1, data collection:

[0014] Setting up a terminal module, which includes a terminal for assessors to input facial information, a terminal for candidates to register, and a terminal for displaying assessment result information. The terminal for assessors to input facial information and the terminal for candidates to register collect the facial information and registration information of the participants and synchronize them to the global character information database of the cloud server;

[0015] The image data is obtained by collecting images of the test persons through the camera groups arranged at the starting line, the finish line, the cornering, the cornering, and the cornering positions of the runway;

[0016] The pressure sensing pads laid at the circle points of the runway are used to collect the pressure characteristic points of the test participants when they pass by, and the pressure detection data is obtained;

[0017] Step S2: Identity recognition and tracking:

[0018] The collected image data is processed, and the face detection algorithm based on deep learning is used to detect the face and extract the features. The principal component analysis algorithm is used for dimensionality reduction. At the same time, the human posture recognition technology is used to extract the human posture features and construct the character identity feature model.

[0019] Maintain a real-time tracking person table for each camera, record the person ID, facial features, human image feature stream, and key point stream. After a person enters the camera screen, filter out targets with large offsets based on the key point stream information, and then use the human image features to match the remaining targets in the real-time tracking person table. If the similarity is greater than the preset threshold, the match is completed. Otherwise, the initial data of the person is added. If a person in the real-time tracking person table is not matched for more than 1 second, it is considered to have left the recognition area and moved out of the real-time tracking person table. During the tracking process, the relevant feature information of face and body weight recognition is updated in real time;

[0020] Step S3: Score calculation:

[0021] Combine the finish line camera image and the pressure sensor pad data to determine the time when the test person crosses the line. When the pressure value on the pressure sensor pad exceeds the preset threshold and lasts for a certain period of time, it is considered that the person is stepping on the pressure sensor pad. Combine this time to determine the time point when the person crosses the finish line , and calibrate the image frame time recorded by the finish line camera; record the start time , calculate running time ;

[0022] According to the standards of different test types, the running time is compared with the standards to get the evaluation results;

[0023] Step S4, violation analysis: After the running is finished, the process data is analyzed, the violation is judged, and the violation data is generated;

[0024] Step S5, data integration and storage: integrating the calculated evaluation results, violation behavior data, related identity recognition data, and running process data, and storing them in the cloud server for terminal access;

[0025] Step S6, abnormality processing and correction: If the pressure sensor data is found to be abnormal or the camera image acquisition is incomplete, the data is automatically marked and a manual check is prompted, and the timing error is automatically corrected according to a pre-set error correction algorithm.

[0026] As a preferred solution, when the face detection algorithm based on deep learning detects the face and extracts features, the face bounding box and key point prediction values ​​are calculated by building a convolutional neural network model. The face bounding box prediction formula is: , the face key point prediction formula is ,in, is the predicted face bounding box, is the sigmoid function, is the bounding box offset output by the network, is the predefined anchor box size, is the center coordinate of the anchor box, It is The first The predicted positions of key points, is the key point offset of the network output; the formula for the principal component analysis algorithm to perform dimensionality reduction is: ,in, is the original face image vector, is the average face vector, is a matrix consisting of the eigenvectors of the covariance matrix, Representation Matrix The transpose of is the feature vector after dimensionality reduction.

[0027] As a preferred solution, when extracting human posture features, human posture recognition technology calculates the joint confidence score by constructing a human posture convolutional neural network model. ,in, It is The confidence score of each joint point, It is The output value of the feature vector of each joint point after passing through the fully connected layer, is the total number of joint points, Indicates the The output value of the feature vector of the joint point after passing through the fully connected layer; the construction of the human body posture feature vector is based on the calculation of the angle and distance of the joint points, and the distance between the joint points Angle between the joint point line and the coordinate axis

[0028] , .

[0029] As a preferred solution, in human posture recognition technology, the improved Euclidean distance formula is used to calculate the optimal human feature similarity SIMim and the optimal gait feature similarity SIMstep ,in, and are the first and second eigenvectors of the two Dimensional component, is an adjustment parameter, and its value range is .

[0030] As a preferred solution, in the identity recognition and tracking step, when calculating the optimal feature similarity matching logic, the corresponding feature table of the participants is extracted from the global character information database. The feature table includes A two-dimensional matrix, where is the number of features, For feature dimensions, prepare MDim features to be matched, where: is the number of features, Dim is the feature dimension, and cosine similarity is used to compare two sets of features, taking the maximum similarity as the result. The cosine similarity formula is: ,in, Represents vectors and The components of n represent the vector and vector The number of dimensions is calculated by combining the feature vector extracted by face and body weight recognition technology during the feature comparison process.

[0031] As a preferred solution, in the step of calculating the score, each sensor on the pressure sensing mat records the pressure on the sole of the foot. Pressure value on the coordinate , forming three-dimensional data containing time series information. Each plane represents the spatial distribution of sole pressure at a specific moment. In the process of pressure feature recognition, a pressure distribution map is generated according to the pressure detection data, regional pressure differences are compared, and pressure change curves are drawn. The dynamic time warping comparison method is used to calculate the minimum distance and optimal path of two time series at different time points to measure similarity, and flexibly match different step cycle pressure change patterns. The formula of the dynamic time warping comparison method is: in, and For two time series, is the path weight, is the distance measure between two time points, calculated using Euclidean distance Among them, Dim is the feature dimension, which integrates the pressure feature with other features obtained by face and body weight recognition technology for analysis to assist identity recognition and score calculation.

[0032] As a preferred solution, in the data acquisition step, the camera uses the image pyramid technology to process images of different scales when collecting images to improve the detection ability of faces of different sizes. The size calculation formula of each layer of the image in the image pyramid is: in, It is layer image, is the original image, is the scaling factor, and l is the layer index of the image pyramid.

[0033] As a preferred solution, when analyzing violations, if any of the following two occurs, it will be considered a violation:

[0034] During the process, two cameras failed to monitor the person;

[0035] The camera's character path record when entering a turn shows that the person's path is significantly beyond the runway range in the inner lane direction.

[0036] A middle- and long-distance running evaluation system based on face and body weight recognition, wherein the middle- and long-distance running evaluation system is equipped with a middle- and long-distance running evaluation method based on face and body weight recognition.

[0037] It can be seen from the technical solutions provided by the present invention that the method and system for evaluating middle- and long-distance running based on face and body weight recognition provided by the present invention have the following beneficial effects:

[0038] 1. Solve the identity recognition problem and improve fairness:

[0039] Accurately identify identities and eliminate the phenomenon of taking exams for others: The existing RFID method cannot provide biometric identification, and AI face recognition is interfered by a variety of environmental factors. The present invention adopts face and body weight recognition technology, integrates face detection algorithm, principal component analysis algorithm dimensionality reduction processing and human posture recognition technology, and constructs a comprehensive and accurate character identity feature model. Through the comprehensive use of multimodal features, the identity of the test takers can still be accurately identified in complex environments. For example, in strong light and when the test takers have variable postures (such as lowering their heads, turning their heads, blocking, etc.), facial features and human posture features complement each other, effectively solving the difficulties of face recognition, ensuring that the identity of each test taker is authentic and reliable, fundamentally eliminating cheating behaviors such as taking exams for others, and greatly improving the fairness of the assessment.

[0040] Tracking personnel throughout the process to ensure complete trajectories: In order to solve the problem of small AI face recognition area and short distance, the present invention realizes seamless tracking of test participants throughout the process by rationally deploying multiple cameras at key positions of the runway and using camera topology information to design a dynamic update logic for camera recognition candidates. No matter where the test participants are on the runway, the system can accurately identify and record their trajectories, avoiding omissions of personnel or interruptions of trajectories due to restrictions on the recognition area, ensuring the continuity and accuracy of identity recognition throughout the evaluation process, and further maintaining the fairness and justice of the evaluation.

[0041] 2. Achieve comprehensive violation determination and maintain evaluation order:

[0042] Effectively monitor various violations: RFID cannot judge whether a person has violated a rule, and AI face recognition is susceptible to interference and may miss violations. The present invention collects image data through a camera group and pressure data through a pressure sensor pad, and can comprehensively monitor violations. For example, by recording and analyzing the key point flow of the human body through the camera character route at the corner, it can accurately determine whether there is a shortcut behavior that greatly exceeds the runway range in the inner lane direction. At the same time, by combing through the monitoring records of each camera, it can timely discover abnormal situations that are not monitored by two consecutive cameras, and effectively monitor violations such as running for others and leaving the field midway, which effectively maintains the order of the evaluation site.

[0043] Providing video evidence and enhancing the persuasiveness of judgment: The present invention adopts a multi-point camera control method to achieve video recording, monitoring and playback of the entire long-distance running process; when there is a suspicion of violation, detailed video evidence can be provided to facilitate management personnel to make accurate judgments and handle; compared with the RFID method that cannot provide video playback, the violation judgment of the present invention is more scientific and rigorous, reducing the misjudgment caused by insufficient evidence or subjective judgment, enhancing the persuasiveness of violation judgment, and ensuring the fairness and authority of the evaluation results;

[0044] 3. Optimize score calculation and improve evaluation accuracy:

[0045] Multi-source data fusion improves timing accuracy: RFID has limitations in timing, while the present invention combines the endpoint camera image and the pressure sensor pad data to determine the time when the test participants cross the line. The pressure sensor pad accurately senses the pressure change at the moment of stepping on the ground to provide a time node, and compares and calibrates it with the endpoint camera image frame time, effectively overcoming the errors that may exist in a single data source, such as time deviation caused by the camera frame rate or misjudgment of the pressure sensor pad, thereby more accurately calculating the running time and laying a solid foundation for obtaining accurate evaluation results.

[0046] Adapt to different test types and ensure fair results: The present invention can accurately calculate the running time and obtain the evaluation results according to the standards of different test types (such as 800 meters, 1000 meters, 3000 meters, marathon, etc.); whether it is a short-distance sprint or a long-distance endurance event, it can objectively evaluate the actual level of the test participants through accurate timing and reasonable standard comparison, avoiding the result deviation caused by inaccurate timing or inapplicable standards, ensuring the fairness and effectiveness of the results, and making the evaluation results more valuable for reference;

[0047] 4. Improve user experience and promote the development of evaluation technology:

[0048] No need to wear extra props, improving running experience: The RFID method requires participants to wear extra props, which affects the running experience; the present invention is based on face and body weight recognition technology, and does not require participants to wear any additional equipment, so that participants can perform middle and long-distance running tests more naturally and comfortably, reducing the inconvenience and interference caused by wearing props, and improving the overall running experience;

[0049] Innovative technology application promotes the progress of evaluation technology: The present invention comprehensively applies a variety of advanced technologies to solve many problems existing in the existing middle- and long-distance running evaluation methods, and provides an innovative and efficient solution for the field of middle- and long-distance running evaluation; its successful application will promote the continuous development and improvement of middle- and long-distance running evaluation technology, provide useful reference and reference for the subsequent research and improvement of related technologies, promote technological progress in the entire industry, and better meet the growing demand for middle- and long-distance running evaluation in activities such as sports education and competition selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of the steps of a middle- and long-distance running evaluation method based on face and body weight recognition according to the present invention;

[0051] Figure 2 It is a schematic diagram of the structure of the middle- and long-distance running evaluation system based on face and weight recognition in the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0054] like Figure 1-2 As shown, the embodiment of the present invention provides a middle- and long-distance running evaluation method based on face and body weight recognition, comprising the following steps:

[0055] Step S1, data collection:

[0056] Setting up a terminal module, which includes a terminal for assessors to input facial information, a terminal for candidates to register, and a terminal for displaying assessment result information. The terminal for assessors to input facial information and the terminal for candidates to register collect the facial information and registration information of the participants and synchronize them to the global character information database of the cloud server;

[0057] The image data is obtained by collecting images of the test persons through the camera groups arranged at the starting line, the finish line, the cornering, the cornering, and the cornering positions of the runway;

[0058] The pressure sensing pads laid at the circle points of the runway are used to collect the pressure characteristic points of the test participants when they pass by, and the pressure detection data is obtained;

[0059] Step S2: Identity recognition and tracking:

[0060] Step S2-1, process the collected image data, use the face detection algorithm based on deep learning to detect the face and extract features, use the principal component analysis algorithm to perform dimensionality reduction processing, and use the human posture recognition technology to extract human posture features to build a character identity feature model; specifically:

[0061] Step S2-1-1: Face detection and feature extraction based on deep learning:

[0062] Build a convolutional neural network model to calculate the face bounding box and key point prediction values;

[0063] The face bounding box prediction formula is: ,in, is the predicted face bounding box, is the sigmoid function, is the bounding box offset output by the network, is the predefined anchor box size, is the center coordinate of the anchor box;

[0064] The facial key point prediction formula is: ,in, It is The first The predicted positions of key points, is the offset of the key points output by the network;

[0065] Step S2-1-2: Dimensionality reduction processing using principal component analysis algorithm:

[0066] The principal component analysis algorithm is used for dimensionality reduction, and its formula is ,in, is the original face image vector, is the average face vector, is a matrix consisting of the eigenvectors of the covariance matrix, Representation Matrix The transpose of is the feature vector after dimensionality reduction;

[0067] Step S2-1-3: Human body posture recognition technology extracts human body posture features:

[0068] Construct a human posture convolutional neural network model to calculate the joint confidence score. The formula is:

[0069] ;

[0070] in, It is The confidence score of each joint point, It is The output value of the feature vector of each joint point after passing through the fully connected layer, is the total number of joint points, Indicates the The output value of the feature vector of the joint point after passing through the fully connected layer;

[0071] The human body posture feature vector is constructed based on the angle and distance calculation of the joint points. Angle between the joint point line and the coordinate axis

[0072] , ;

[0073] Step S2-1-4: Constructing a character identity feature model:

[0074] The facial features extracted above, the features after dimensionality reduction, and the human posture features are combined to construct a character identity feature model for subsequent identity recognition and tracking. For example, in the identity recognition and tracking step, a real-time tracking character table is maintained for each camera, recording information such as character ID, facial features, human image feature stream, key point stream, etc. The constructed character identity feature model is used to match and track characters. If the similarity is greater than the preset threshold, the match is completed, otherwise the character record initial data is added, etc. When calculating the optimal feature similarity matching logic, the corresponding feature table of the participants is extracted from the global character information database, combined with the feature vector extracted by the face and body weight recognition technology, and the two sets of features are compared pairwise using cosine similarity. The maximum similarity is taken as the result, and the application of the character identity feature model in the entire evaluation system is further improved;

[0075] Step S2-2, maintain a real-time tracking person table for each camera, record the person ID, facial features, human image feature stream, key point stream, after the person enters the camera screen, filter out the target with large offset according to the key point stream information, and then use the human image features to match the remaining targets in the real-time tracking person table. If the similarity is greater than the preset threshold, the match is completed, otherwise the initial data of the new person is recorded. If the person in the real-time tracking person table is not matched for more than 1 second, it is considered to have left the recognition area and moved out of the real-time tracking person table. During the tracking process, the relevant feature information of face and body weight recognition is updated in real time; Specifically:

[0076] Step S2-2-1, maintenance and initialization of real-time tracking character table:

[0077] Create and maintain a real-time tracking person table for each camera to record relevant information about people currently within the monitoring range of the camera;

[0078] The recorded information includes person ID (used to uniquely identify each test participant), facial features (extracted by the deep learning-based face detection algorithm mentioned above), human image feature stream (human image feature information continuously collected when the person is in the camera frame), and key point stream (time series of human key point information extracted by human posture recognition technology);

[0079] Step S2-2-2, processing flow after entering the camera screen:

[0080] Filtering targets with large offsets: When a person enters the camera screen, it is first processed based on the key point stream information. By analyzing the position changes of the key points of the human body in the key point stream, it is determined whether the person has a large position offset. If there is a large offset, it is considered as an abnormal target and filtered out without subsequent matching operations. This helps to exclude some abnormal targets caused by interference or false detection and improve tracking accuracy.

[0081] Human image feature matching: For the target after offset filtering, use its human image features to match with the remaining targets in the real-time tracking person table; calculate the similarity between the two, and the similarity calculation here may use a specific algorithm (such as the improved Euclidean distance formula mentioned in human posture recognition technology); if the calculated similarity is greater than the preset threshold, it is considered that the person is successfully matched with the existing target in the table, and the matching operation is completed; at this time, the relevant information of the person in the tracking table may be updated, such as the latest human image features, key point stream, etc.;

[0082] Add a new person record: If the similarity is not greater than the preset threshold, the person is determined to be an unknown person who has just entered the camera screen. At this time, a new person record is added to the real-time tracking person table, and the initial data of the person (such as the currently collected facial features, body image features, key point streams, etc.) is recorded to facilitate subsequent continuous tracking and identification.

[0083] Step S2-2-3: Determination and processing of a person leaving the recognition area:

[0084] Real-time monitoring and tracking of the matching status of each character in the character list; if a character is not matched for more than 1 second, that is, the character is not detected in the camera image for 1 second or the matching similarity with other targets is always lower than the threshold, it is considered that the character has left the recognition area of ​​the current camera;

[0085] Once it is determined that a person has left the recognition area, it will be removed from the real-time tracking person table to release system resources and update related data structures; at the same time, before the person leaves, its gait features will be extracted (as mentioned in the identity matching logic, when it is determined that a person in the tracking table has left the recognition area, the gait features will be extracted before removing the person target), and identity analysis will be performed. According to the previous identity matching logic (such as through facial feature matching, human feature similarity calculation, etc.), its identity information will be determined, and the camera records, human image features, gait features, key point streams and other related information will be submitted to the global person information database according to the rules or other corresponding processing will be performed;

[0086] Step S2-2-4, updating of feature information during tracking:

[0087] During the entire tracking process, the feature information related to face and body weight recognition is continuously updated; this includes the continuous collection of new face image features (if the face is visible in the camera image), body image features, and key point flow information; by updating these feature information in real time, the movement state, posture changes, and identity characteristics of the person can be tracked more accurately, improving the system's ability to identify and track people in complex environments, ensuring that all information about the participants can be accurately obtained during the middle and long-distance running evaluation process, and providing reliable data support for subsequent performance calculations, violation analysis, and other links; for example, in the identity matching logic, identity analysis and matching are performed based on the continuously updated feature information to ensure accurate identification of the person's identity and the integrity of the tracking record;

[0088] Step S3: Score calculation:

[0089] Step S3-1: Combine the camera image at the finish line and the pressure sensor pad data to determine the time when the participant crosses the line. When the pressure value on the pressure sensor pad exceeds the preset threshold and lasts for a certain period of time, it is considered that the person is stepping on the pressure sensor pad. Combine this time to determine the time when the person crosses the finish line. , and calibrate the image frame time recorded by the finish line camera; record the start time , calculate running time Specific:

[0090] Step S3-1-1, determine the end point crossing time:

[0091] Pressure sensor pad data judgment: Each sensor on the pressure sensor pad will record the surface of the sole of the foot Pressure value on the coordinate , forming three-dimensional data containing time series information, each plane represents the spatial distribution of sole pressure at a specific moment; when the pressure value at a certain point on the pressure sensor pad exceeds the preset threshold, and the pressure value lasts for a certain period of time (this duration can be set according to the actual situation to ensure that the person's foot is normally on the pressure sensor pad rather than an interference signal), it is considered that the person's foot is on the pressure sensor pad; at this time, the moment is recorded as the time point when the person's foot steps on the pressure sensor pad ;

[0092] Combined with camera image frame time calibration: At the same time, the endpoint camera records the image frame time of the test person crossing the line; the time point determined by the pressure sensor pad Compare and calibrate with the image frame time recorded by the endpoint camera; for example, if the camera frame rate is Frames / second, image frame time is (expressed in frame numbers, which need to be converted into actual time), then by comparing and , choose a more accurate time as the time when the participants cross the finish line If the difference between the two is within the allowable error range (the error range can be set in advance), the final ; If the difference is beyond the error range, further data analysis or manual verification is performed to determine the exact time of crossing the line;

[0093] Step S3-1-2, record the start time:

[0094] At the beginning of the test, the system records the time when the starting signal is given, that is, the time when the test participants start running. ; This time point can be accurately recorded while the system controls the on-site audio equipment to send the start signal, ensuring the accuracy of the start time;

[0095] Step S3-1-3, calculate running time:

[0096] Calculate running time , the formula is ; The time of crossing the finish line will be determined The starting time of the record Subtract the running time of the participants. ; This running time will be used as the key data for subsequent calculation of evaluation results, and will be used to compare with the standards of different test types to obtain the evaluation results of the test participants; for example, in different test types (such as 800 meters, 1000 meters, 3000 meters, marathon, etc.), the running time is accurately calculated in this way to evaluate the performance of the test participants in the corresponding items;

[0097] Step S3-2: According to different test type standards, the running time is compared with the standard to obtain the evaluation results; the specific steps are as follows:

[0098] Step S3-2-1, obtain the test type standard:

[0099] The system pre-stores standard data corresponding to different test types (such as 800 meters, 1000 meters, 3000 meters, marathon, etc.); these standard data include information such as the range of qualified scores for different genders and age groups (if subdivided) under the test type, and the standard for dividing grades of scores; for example, for an 800-meter run, it may be stipulated that the full score for boys in a certain age group is a certain number of seconds, the passing score is a certain number of seconds, and the grades corresponding to different score ranges (such as excellent, good, passing, failing, etc.);

[0100] Step S3-2-2, compare running time with the standard:

[0101] Compare the running time calculated by the participant in this assessment with the standard data for the corresponding test type; first determine the gender and age group (if applicable) of the participant, and then find the corresponding standard score range and level division;

[0102] For example, if the participant is a male and takes a 1000-meter run test, the running time is seconds; the system will search the standard data of male 1000-meter run to determine which performance range it is in, and thus determine its corresponding performance grade; if it is less than the time corresponding to the full score, it may be awarded an excellent grade; if it is between the passing score and the full score, it may be awarded a good or passing grade; if it is greater than the time corresponding to the passing score, it may be judged as failing;

[0103] Step S3-2-1: Determine the evaluation results:

[0104] According to the comparison results, the evaluation scores of the test participants are determined; the scores can be presented in the form of specific numbers (such as the difference between the running time in seconds and the standard score), grades (such as excellent, good, pass, fail, etc.) or a combination of the two; at the same time, the system will record the score and related comparison information for subsequent query and analysis; for example, after the score calculation is completed, the score data is associated with the identity information of the test participants and stored in the cloud server for the terminal to display the evaluation result information, so as to facilitate the candidates, examiners and other relevant personnel to view and understand the evaluation results; in the entire middle- and long-distance running evaluation process, this step ensures that the score of each test participant can be accurately and fairly obtained according to the corresponding test type standards, providing an objective basis for evaluating the long-distance running ability of the test participants;

[0105] Step S4, violation analysis: After the running is finished, the process data is analyzed, the violation is judged, and the violation data is generated; the specific steps are:

[0106] Step S4-1, process data collection and organization:

[0107] During the entire running process, the system continuously collects data through a variety of devices, including image data collected by cameras (covering image information at key locations such as the starting line, finish line, and bends), pressure detection data collected by pressure sensing pads, and person information recorded by each camera during the tracking process (such as person ID, facial features, body image feature stream, key point stream, etc.);

[0108] The collected data is collated and preprocessed so that it can be conveniently used for subsequent violation judgment analysis; for example, image data is arranged in chronological order, pressure detection data is formatted, and person tracking information is associated with corresponding time and location, etc.;

[0109] Step S4-2: Setting the basis for violation judgment:

[0110] The system pre-sets the basis for judging violations, which mainly include the following two situations:

[0111] Continuous camera failure to monitor the person: If a participant is not monitored by two cameras in a row during running, there may be violations; this means that the person's movement trajectory on the track is abnormal, and may have deviated from the normal running route or may be avoiding monitoring;

[0112] Abnormal route at cornering: If the route of the person recorded by the camera at cornering (through human key point flow analysis) shows that the person goes far beyond the track range in the inner lane, it will be considered as a violation. This may involve illegal behaviors such as taking shortcuts, affecting the fairness of the competition.

[0113] Step S4-3, Violation Judgment Analysis:

[0114] For the first type of violation judgment basis, the system checks the monitoring records of each camera to determine the appearance time and order of each participant in each camera screen; if it is found that a participant does not appear in the monitoring areas of two adjacent cameras (and objective reasons such as camera failure are excluded), the person is marked as possibly violating the rules;

[0115] For the judgment of violation when entering a turn, the system focuses on analyzing the human key point flow information in the image data collected by the camera at the turn; by calculating the relative position relationship between the human key points and the runway boundary, it is judged whether the person exceeds the specified runway range when entering the turn; for example, according to the joint point position determined by the human posture recognition technology, the distance between the joint point and the inner lane boundary of the runway is calculated. If the distance exceeds the preset allowable range, it is judged that the person has greatly exceeded the runway range in the inner lane direction, and thus it is determined to be a violation;

[0116] Step S4-4: Generate violation data:

[0117] Once a participant is determined to have violated the rules, the system will generate corresponding violation data; these data include the type of violation (such as failure to monitor by continuous cameras or abnormal route at the turn), the time when the violation occurred (roughly or specifically, determined based on the frequency of data collection and analysis accuracy), and the identity information of the violator (obtained by associating with previously collected facial and body weight recognition information).

[0118] The generated violation data is stored and marked for subsequent processing and query; for example, the violation data is associated with the evaluation performance data of the test participant, and the violation is displayed at the same time when the evaluation results are displayed; or the violation data is stored separately for management personnel to conduct statistical analysis, so as to supervise and improve the overall evaluation situation and ensure the fairness and standardization of subsequent middle- and long-distance running evaluation activities;

[0119] Step S5, data integration and storage: Integrate the calculated evaluation results, violation data, related identity recognition data, and running process data, and store them in the cloud server for terminal access; the specific steps are:

[0120] Step S5-1: Data integration:

[0121] Integration of evaluation results: Collect and organize the evaluation results of the test participants calculated in step S3 according to the standards of different test types; ensure the accuracy and completeness of the result data, including information such as running time and corresponding result level (if any);

[0122] Integration of violation data: Summarize the violation data generated in step S4, such as the violation type (continuous camera does not monitor the person or the route is abnormal when entering the curve, etc.), the time when the violation occurred, the identity information of the violators involved, etc.; associate these data with the corresponding test participants to facilitate subsequent query and analysis of the violation situation;

[0123] Identity data integration: Integrate all identity data previously obtained during data collection and identity recognition, such as the facial information of test participants, feature tables (including facial features, human image features, etc.) extracted from the global character information database, and personnel IDs; ensure that the identity information of each test participant accurately corresponds to other relevant data;

[0124] Integration of running process data: Collect image data of participants at various positions on the track (starting line, finish line, curves, etc.) captured by cameras (which may contain information such as the person's movement trajectory and posture changes) and pressure detection data collected by pressure sensor pads (such as pressure feature point data at lap points, reflecting the pace of the running process, etc.); sort and organize these data in chronological order and by participants to form a complete running process data record;

[0125] Step S5-2: Data storage and management:

[0126] Cloud server storage: upload and store all the integrated data to the cloud server; the cloud server provides large-capacity storage space and reliable data management functions to ensure data security and persistence; during the storage process, classify the data and establish a reasonable data structure, such as classifying by evaluation batch, participant ID and other dimensions to facilitate subsequent data retrieval and query;

[0127] Data indexing and tagging: In order to facilitate the terminal to quickly access data, a data indexing and tagging system is established; a unique identifier (such as ID) is set for the data of each participant, and different types of data (scores, violations, identities, process data, etc.) are marked, and key attributes of the data (such as data generation time, data source, etc.) are recorded; in this way, when the terminal needs to access a specific participant or a specific type of data, it can quickly locate and obtain the required data through indexing and tagging;

[0128] Step S5-3: Terminal access function support:

[0129] Interface provision: The cloud server provides data retrieval interfaces for terminal devices (such as terminals for assessors to input facial information, terminals for candidates to register, and terminals for displaying assessment result information); through these interfaces, terminal devices can send data requests to the cloud server, specifying conditions such as the type of data to be retrieved and the scope of test participants;

[0130] Data transmission and display: The cloud server retrieves and extracts the corresponding data from the stored data according to the request of the terminal, and then transmits the data back to the terminal device; after receiving the data, the terminal device displays the data according to its own functional requirements; for example, the terminal that displays the evaluation result information can present the test participants' scores, whether they have violated the rules, and key images of the running process (such as the start and sprinting scenes) to the user in an intuitive way, so that the user can view and analyze the evaluation results; in the entire middle- and long-distance running evaluation system, the data integration and storage steps ensure the effective management and utilization of all relevant data, and provide a solid data foundation for the query, statistical analysis and subsequent evaluation and improvement of the evaluation results;

[0131] Step S6, abnormality processing and correction: If the pressure sensor data is abnormal or the camera image acquisition is incomplete, the data is automatically marked and a manual check is prompted. The timing error is automatically corrected according to a pre-set error correction algorithm. The specific operations are as follows:

[0132] Step S6-1: Data anomaly detection:

[0133] Abnormal pressure sensor data detection: The system monitors the data collected by the pressure sensor pad in real time; each sensor on the pressure sensor pad records the pressure value on the plane coordinates of the sole of the foot, forming three-dimensional data containing time series information; by analyzing the change pattern and characteristics of these data, it is determined whether there is an abnormality; for example, check whether the pressure value fluctuates within a reasonable range (the upper and lower limits are set according to the empirical value of the sole pressure during normal running). If the pressure value of a certain sensor continues to exceed the normal range (too high or too low), or the pressure change curve shows abnormal fluctuations (such as sudden large jumps or interruptions), the pressure sensor data is determined to be abnormal;

[0134] Incomplete camera image acquisition detection: Check the integrity of the image data collected by the camera; this includes checking whether the image resolution meets the requirements, whether there are missing parts in the image (such as black screen or distorted screen in some areas), whether the image frame rate is stable, etc.; if the image is found to have the above problems, or if the complete image information cannot be obtained at key time points (such as the start, finish line, bends and other important locations), the camera image acquisition is determined to be incomplete;

[0135] Step S6-1, abnormal data marking and prompt manual verification:

[0136] Marking abnormal data: Once abnormal pressure sensor data or incomplete camera image acquisition is detected, the system will immediately mark the relevant data; the marking method can be to add a specific identification field in the data record to indicate the type of data abnormality (such as abnormal pressure sensor data: pressure value is too high, incomplete image acquisition: starting image is missing, etc.) and the time when the abnormality occurred (accurate to seconds or milliseconds, depending on the data record accuracy);

[0137] Prompt for manual verification: The system sends a prompt message to relevant staff (such as evaluation managers, technical maintenance personnel, etc.) to inform them of data anomalies and request manual verification; the prompt message can be conveyed in a variety of ways, such as popping up a warning box on the system management interface, sending SMS or email notifications, etc., to ensure that relevant personnel can know and handle the anomalies in a timely manner; manual verification personnel conduct a detailed inspection of the abnormal data based on the prompt message, determine whether the data anomaly is caused by equipment failure, environmental interference or other reasons, and take corresponding measures to repair or re-collect data (such as checking whether the pressure sensing pad sensor is damaged, adjusting the camera position or parameters, etc.);

[0138] Step S6-1, timing error correction:

[0139] Timing error analysis: The system analyzes factors that may cause timing errors, such as the difference between the crossing time determined by the pressure sensor pad and the image frame time recorded by the endpoint camera, data transmission delay, etc. By comparing the time information of different data sources, the size and direction of the timing error (i.e., whether the timing is too fast or too slow) are calculated;

[0140] Application of error correction algorithm: Automatically correct the timing error according to the pre-set error correction algorithm; the error correction algorithm may be based on statistical principles or empirical models, for example, weighted average processing of multiple measured time data, and weights are assigned according to the reliability of different data sources; or an error correction model is established based on known equipment performance parameters (such as the response time of the pressure sensor pad, the frame rate error of the camera, etc.), and the calculated timing error is applied to the actual running time calculation; for example, if it is found through analysis that the crossing time determined by the pressure sensor pad is seconds slower than the actual time, then when calculating the running time, the crossing time of the participant is subtracted by seconds to obtain a more accurate running time, thereby ensuring the accuracy of the evaluation results; in the entire middle- and long-distance running evaluation process, the exception handling and correction steps effectively ensure the data quality, improve the reliability of the system and the fairness of the evaluation results;

[0141] In this embodiment, in the identity recognition and tracking step, when calculating the optimal feature similarity matching logic, the corresponding feature table of the participants is extracted from the global character information database. The feature table includes A two-dimensional matrix, where is the number of features, For feature dimensions, prepare MDim features to be matched, where: is the number of features, Dim is the feature dimension, and cosine similarity is used to compare two sets of features, taking the maximum similarity as the result. The cosine similarity formula is: ,in, Represents vectors and The components of n represent the vector and vector The number of dimensions is calculated by combining the feature vector extracted by face and body weight recognition technology during the feature comparison process.

[0142] In this embodiment, in the step of calculating the performance, each sensor on the pressure sensing mat records the pressure on the sole of the foot. Pressure value on the coordinate , forming three-dimensional data containing time series information. Each plane represents the spatial distribution of sole pressure at a specific moment. In the process of pressure feature recognition, a pressure distribution map is generated according to the pressure detection data, regional pressure differences are compared, and pressure change curves are drawn. The dynamic time warping comparison method is used to calculate the minimum distance and optimal path of two time series at different time points to measure similarity, and flexibly match different step cycle pressure change patterns. The formula of the dynamic time warping comparison method is: ,in, and For two time series, is the path weight, is the distance measure between two time points, calculated using Euclidean distance Among them, Dim is the feature dimension, which integrates the pressure feature with other features obtained by face and body weight recognition technology for analysis to assist identity recognition and score calculation.

[0143] In this embodiment, in the data acquisition step, the camera uses the image pyramid technology to process images of different scales when acquiring images, so as to improve the detection capability of faces of different sizes. The size calculation formula of each layer of the image in the image pyramid is: ,in, It is layer image, is the original image, is the scaling factor, and l is the layer index of the image pyramid.

[0144] A middle-long distance running evaluation system based on face and body weight recognition, the middle-long distance running evaluation system is equipped with a middle-long distance running evaluation method based on face and body weight recognition; the middle-long distance running evaluation system includes a data acquisition module, an algorithm host module, a cloud server and a terminal module; specifically:

[0145] Data acquisition module:

[0146] Camera group: It is deployed at the starting line, finish line, cornering, turning and exiting positions of the runway to collect images of test participants and obtain image data. When collecting images, the camera can use the image pyramid technology to process images of different scales to improve the detection ability of faces of different sizes. The size calculation formula of each layer of the image is:

[0147] ;

[0148] in, It is layer image, is the original image, is the scaling factor, l is the layer index of the image pyramid;

[0149] Pressure sensing mat: It is laid at the circle point of the runway to collect the pressure characteristic points of the test participants when they pass by and obtain the pressure detection data; each sensor on the pressure sensing mat records the pressure on the sole plane Pressure value on the coordinate , forming three-dimensional data containing time series information, each plane represents the spatial distribution of sole pressure at a specific moment, these data can be used to assist in identity recognition and score calculation, such as by analyzing the pressure distribution map, regional pressure differences, pressure change curves and other features, using dynamic time warping comparison method for analysis;

[0150] Algorithm host module:

[0151] Receive image data and pressure detection data from the data acquisition module, process the image data, use the face detection algorithm based on deep learning to detect the face and extract features, use the principal component analysis algorithm to reduce the dimension, and use the human posture recognition technology to extract human posture features and build a character identity feature model;

[0152] Maintain a real-time tracking person table for each camera, record person ID, facial features, human image feature stream, key point stream and other information, perform person identification and tracking operations, including filtering targets with large offsets based on key point stream information, matching human image features with remaining targets in the real-time tracking person table, judging whether the person has left the identification area and updating relevant feature information, etc.; in the process of identification and tracking, multiple algorithm calculations are involved, such as using the improved Euclidean distance formula to calculate the optimal human feature similarity SIMim and the optimal gait feature similarity SIMstep (in, and are the first and second eigenvectors of the two Dimensional component, is an adjustment parameter, and its value range is ;When calculating the optimal feature similarity matching logic, extract the corresponding feature table of the participants from the global character information database, use cosine similarity to compare the two sets of features, and take the maximum similarity as the result. The cosine similarity formula is (in, Represents vectors and The components of n represent the vector and vector the number of dimensions of );

[0153] Combine the finish line camera image and the pressure sensor pad data to determine the time when the test participants cross the line, record the start time, calculate the running time, and get the evaluation results according to the standards of different test types; at the same time, after the run, analyze the process data for violation analysis to determine whether there are any violations (such as two consecutive cameras not monitoring the person during the process or the person's route record in the camera at the corner shows that the person's direction is far beyond the runway range), and generate violation data;

[0154] Cloud Server:

[0155] The facial information and registration information of the test takers collected by the terminal where the evaluators input their facial information and the terminal where the candidates register, respectively, are stored to form a global character information database;

[0156] Receive the evaluation results, violation data, related identity recognition data, running process data, etc. calculated by the algorithm host module, integrate and store these data for terminal access; during the data integration and storage process, ensure the classification management and efficient retrieval of data for subsequent query and statistical analysis;

[0157] Terminal module:

[0158] The terminal for evaluators to input facial information: used to collect the facial information of test participants and synchronize it to the global person information database of the cloud server to provide basic data for facial recognition and identity authentication;

[0159] Terminal for candidates to register: collects the registration information of candidates and synchronizes it to the cloud server so that the system can accurately identify the identity of the candidates and related examination information during the evaluation process;

[0160] Terminal for displaying assessment result information: obtains assessment results, violation data and other information from the cloud server and displays it to users (such as candidates, examiners, etc.) in an intuitive manner; users can view personal or overall assessment results through this terminal, including running time, score level, whether there is a violation and violation details (if any), so as to facilitate query and analysis of assessment results.

[0161] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating middle and long distance running based on face and weight recognition, characterized in that: The following steps are involved: Step S1, data collection: Setting up a terminal module, which includes a terminal for assessors to input facial information, a terminal for candidates to register, and a terminal for displaying assessment result information. The terminal for assessors to input facial information and the terminal for candidates to register collect facial information and registration information respectively and synchronize them to the global character information database of the cloud server; The image data is obtained by collecting images of the evaluators through camera groups arranged at the starting line, the finish line, the cornering, the cornering, and the cornering positions of the runway; The pressure sensing pads laid at the circle points of the runway are used to collect the pressure characteristic points when the evaluator passes by, and the pressure detection data is obtained; Step S2: Identity recognition and tracking: The collected image data is processed, and the face detection algorithm based on deep learning is used to detect the face and extract the features. The principal component analysis algorithm is used for dimensionality reduction. At the same time, the human posture recognition technology is used to extract the human posture features and construct the character identity feature model. Maintain a real-time tracking person table for each camera, record the person ID, facial features, human image feature stream, and key point stream. After the person enters the camera screen, filter out the target with a large offset based on the key point stream information, and then use the human image features to match the remaining targets in the real-time tracking person table. If the similarity is greater than the preset threshold, the match is completed. Otherwise, the initial data of the person is added. If the person in the real-time tracking person table is not matched for more than 1 second, it is considered to have left the recognition area and moved out of the real-time tracking person table. During the tracking process, the relevant feature information of face and body weight recognition is updated in real time; Step S3: Score calculation: Combine the finish line camera image and the pressure sensor pad data to determine the time when the person crosses the line. When the pressure value on the pressure sensor pad exceeds the preset threshold and lasts for a certain period of time, it is considered that the person is stepping on the pressure sensor pad. Combine this time to determine the time when the person crosses the finish line , and calibrate the image frame time recorded by the finish line camera; record the start time , calculate running time ; According to the standards of different test types, the running time is compared with the standards to get the evaluation results; Step S4, violation analysis: After the running is finished, the process data is analyzed, the violation is judged, and the violation data is generated; Step S5, data integration and storage: integrating the calculated evaluation results, violation behavior data, related identity recognition data, and running process data, and storing them in the cloud server for terminal access; Step S6, abnormality processing and correction: if the pressure sensor data is abnormal or the camera image acquisition is incomplete, the data is automatically marked and a manual check is prompted, and the timing error is automatically corrected according to a pre-set error correction algorithm; When the face detection algorithm based on deep learning detects faces and extracts features, a convolutional neural network model is constructed to calculate the face bounding box and key point prediction values. The face bounding box prediction formula is: , the face key point prediction formula is ,in, is the predicted face bounding box, is the sigmoid function, is the bounding box offset output by the network, is the predefined anchor box size, is the center coordinate of the anchor box, It is The first The predicted positions of key points, is the key point offset of the network output; the formula for the principal component analysis algorithm to perform dimensionality reduction is: ,in, is the original face image vector, is the average face vector, is a matrix consisting of the eigenvectors of the covariance matrix, Representation Matrix The transpose of is the feature vector after dimensionality reduction; In the identity recognition and tracking step, when calculating the optimal feature similarity matching logic, the corresponding feature table of the evaluator is extracted from the global character information database. The feature table includes A two-dimensional matrix, where is the number of features, For feature dimensions, prepare MDim features to be matched, where: is the number of features, Dim is the feature dimension, and cosine similarity is used to compare two sets of features, taking the maximum similarity as the result. The cosine similarity formula is: ,in, Represents vectors and The components of n represent the vector and vector The number of dimensions is calculated by combining the feature vector extracted by face and body weight recognition technology during feature comparison; In the score calculation step, each sensor on the pressure sensing mat records the pressure value on the plane coordinates of the sole of the foot, forming three-dimensional data containing time series information. Each plane represents the spatial distribution of sole pressure at a specific moment. In the process of pressure feature recognition, a pressure distribution map is generated based on the pressure detection data, regional pressure differences are compared, and pressure change curves are drawn. The dynamic time warping comparison method is used to measure the similarity by calculating the minimum distance and optimal path of two time series at different time points, flexibly matching different step cycle pressure change patterns, and integrating and analyzing the pressure features with other features obtained by face and body weight recognition technology to assist identity recognition and score calculation; In the data collection step, the camera uses image pyramid technology to process images of different scales when collecting images, improving the detection ability of faces of different sizes; When analyzing the above-mentioned violations, if any of the following two occurs, it will be considered a violation: During the process, two consecutive cameras failed to monitor the same evaluator; The camera's character route record when entering a turn shows that the person's route goes far beyond the runway in the inner lane.

2. A middle- and long-distance running evaluation system based on face and weight recognition, characterized in that: The middle- and long-distance running evaluation system is equipped with the middle- and long-distance running evaluation method based on face and body weight recognition as described in claim 1.

Images