Deep learning-based air-riding student shape and temperament scoring method and deep learning-based air-riding student shape and temperament scoring system

Abstract

The invention discloses a deep learning-based air-riding student shape and temperament scoring method and system, and the method comprises the steps: collecting a training video through a multi-view motion capture camera, extracting 25 skeleton key points through a YOLO v8 algorithm, and generating a three-dimensional posture data dot matrix in combination with camera parameters; performing data smoothing processing by using a time sequence convolutional network, and comparing the data with an expert standard action template; analyzing data through a space-time diagram convolution attention network, extracting spatial features by using a GCN layer, learning time sequence features by using bidirectional LSTM, and dynamically distributing weights by using a cross attention mechanism; performing intelligent scoring based on the fine-tuned Qwen-7B large model, and outputting a three-dimensional scoring report; a virtual image and a thermodynamic diagram are generated through a digital twinning technology, and training deviation is fed back in real time; according to the invention, the long-term dynamic tracking evaluation of the body and posture is realized, the limitation of traditional subjective evaluation is broken through, and the evaluation scientificity and traceability are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent scoring technology, and in particular relates to a method and system for scoring the physical appearance and temperament of flight attendant students based on deep learning. Background Art

[0002] With the rapid development of the aviation industry in recent years, airlines have placed higher demands on the professional qualities and appearance of flight attendants. Physical appearance and demeanor, as a crucial component of a flight attendant's professional image, have become a key criterion in airline recruitment and assessment. To meet industry demands, vocational colleges specializing in flight attendant majors generally incorporate physical training courses into their core training programs. Through systematic training, students' posture, demeanor, and overall demeanor are enhanced to meet airline hiring standards.

[0003] Currently, physical training courses for flight attendants at vocational colleges primarily rely on manual observation to assess students' physical appearance and demeanor. Specifically, instructors or industry mentors visually observe students' standing, sitting, walking, and overall demeanor, then assess their performance based on their personal experience.

[0004] However, these manual evaluation methods have significant limitations: the results are susceptible to subjective factors. Due to the lack of a unified quantitative standard, different teachers' scores can vary significantly, making it difficult to ensure objectivity and consistency. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method for grading the body and temperament of flight attendants based on deep learning, which can evaluate the body and posture of flight attendants in a long-term and dynamic manner, and improve the objectivity of the assessment; on the other hand, to provide a system for grading the body and temperament of flight attendants based on deep learning.

[0006] Technical solution: The method for scoring the physical and temperament of flight attendant students of the present invention comprises the following steps:

[0007] (1) Using multi-view high-definition motion capture cameras to collect flight attendant students’ physical training video streams, and using facial recognition technology to associate student identity information, ensuring the integrity of data collection and identity accuracy, and providing a high-quality video stream data source for subsequent analysis;

[0008] (2) Based on the YOLO v8 target detection algorithm, the 25 skeletal key points of the human body in the video are extracted into a two-dimensional coordinate sequence. Combined with the camera space parameters, multi-view triangulation positioning is performed to generate a three-dimensional posture data matrix, overcoming the single-view occlusion problem and restoring the dynamic action details in the real space, laying a data foundation for body analysis.

[0009] (3) Use the time convolution network to smooth the three-dimensional posture data point array, eliminate the action jitter noise, improve the data continuity, and compare with the pre-constructed expert standard action template library, quantify the deviation of the student action from the industry standard, and realize the preliminary posture evaluation;

[0010] (4) Adopt the space-time graph convolution attention network to analyze the three-dimensional posture data, wherein: the GCN layer is used to extract the spatial topological relationship of the nodes, the bidirectional LSTM unit is used to learn the action time sequence characteristics, and the cross attention mechanism is used to dynamically allocate weights to the eight core shape indexes, so as to realize the intelligent weighted fusion of multi-dimensional shape characteristics;

[0011] (5) Fine-tune the Qwen-7B large model, and based on the fine-tuned Qwen-7B large model, intelligent scoring is performed, the airline recruitment standard knowledge graph and the teacher scoring case are fused, and a three-dimensional scoring report containing sub-item scores, improvement suggestions and airline matching index is output, so as to improve the professional depth of the evaluation;

[0012] (6) Generate a student virtual image and a posture heat map through digital twinning technology, intuitively present the action deviation, strengthen the student cognition, real-time feedback the training deviation and update the personalized growth file, and realize the data-driven personalized teaching optimization.

[0013] Preferably, the generating the three-dimensional posture data point array in step 2 comprises:

[0014] The YOLO v8 model with an input resolution of 1280×1280 is adopted to extract 25 civil aviation shape evaluation key points in real time at 32 FPS on the Jetson platform;

[0015] The dynamic weight decay mechanism is adopted to process the student shielding problem in the training process;

[0016] Based on the calibration parameters of the multi-view camera, a triangulation matrix is constructed by using the camera baseline distance, installation angle and focal length;

[0017] The three-dimensional coordinates are calculated by the least square method to generate the three-dimensional point array data of the student shape posture.

[0018] The high-resolution YOLO v8 model is used to realize real-time key point detection on the embedded platform, ensuring the accuracy and efficiency of the civil aviation shape evaluation; the dynamic weight decay mechanism effectively alleviates the shielding problem, improving the robustness of key point extraction; combined with the multi-view calibration parameters to construct a triangulation matrix, and using the least square method to optimize the three-dimensional coordinate calculation, finally generating high-precision and high-stability three-dimensional posture point array data, providing a reliable space motion representation basis for subsequent shape analysis.

[0019] Preferably, the construction of the expert standard action template library in step 3 comprises:

[0020] 12 standard motion data of the body teacher and the airline cabin crew are collected by the motion capture system;

[0021] A professional characteristic motion database including the range of bowing, gait cycle and gesture trajectory is established.

[0022] By collecting standardized motion data of professional teachers and airline cabin crew, an expert template library covering 12 core motions (such as bowing and gait) is constructed, and key professional characteristic parameters (such as range, cycle, trajectory, etc.) are quantified, providing an authoritative benchmark for subsequent student motion comparison, and realizing standardized evaluation and precise deviation detection of body training.

[0023] Preferably, step 4 comprises:

[0024] A graph convolution network (GCN) layer is used to construct a human joint topology graph, extract spatial correlation features between adjacent joints, and output a joint feature vector;

[0025] The time series data of consecutive frames are processed by a bidirectional LSTM unit to capture the phase change rule of the motion, wherein the hidden layer dimension of the forward LSTM and the backward LSTM is 64.

[0026] An attention cross-layer is designed to dynamically allocate weights to 8 core body indicators, including a cervical spine angle weight coefficient of 0.3 and a shoulder line levelness weight coefficient of 0.25, and a hard attention selection is realized by a Softmax function, and finally an 128-dimensional composite feature vector is output.

[0027] The spatial topology relationship of human joints is mined by a graph convolution network, the action time evolution rule is accurately modeled by a bidirectional LSTM, and the core body indicators are dynamically focused by a cross-attention mechanism, and finally an 128-dimensional high-representation composite feature vector is output, realizing fine analysis of body motion in the space-time dimension and adaptive weighted fusion of key indicators.

[0028] Preferably, the fine-tuning of the Qwen-7B large model in step 5 comprises:

[0029] The preset recruitment standards of 12 airlines are converted into a structured knowledge graph, and the knowledge graph is injected into the Qwen-7B large model by low-rank adaptive technology;

[0030] The scoring case data of several experienced body instructors of civil aviation airports and airlines are collected, and a near-end strategy optimization algorithm is used to train the Qwen-7B large model for preference alignment;

[0031] A sliding time window mechanism is designed to compare the current body data of the student with the average data of the past 30 days, and a Z-score standardization method is used to calculate the progress index.

[0032] Generate 200 typical error posture data using deep convolutional generative adversarial networks, including a pelvic tilt angle exceeding 10 degrees and an arm swing asymmetry exceeding 15 degrees;

[0033] The output includes a three-dimensional score report containing six dimensions of percentage scores, TOP3 improvement suggestions ranked by promotion potential, and an airline matching index ranging from 0 to 1.

[0034] By injecting the airline recruitment standard knowledge graph into the Qwen-7B model using low-rank adaptive technology, and combining with the preference alignment training of experienced instructor scoring cases, the industry adaptability of the model is significantly improved; the time window mechanism and Z-score standardization are introduced to dynamically track the training progress, supplemented by 200 typical error posture data generated by the adversarial generation to enhance the generalization ability, finally output the intelligent report covering multi-dimensional score, priority improvement suggestion and quantitative matching index, realizing the deep and accurate matching of physical assessment and airline demand.

[0035] Preferably, the real-time feedback training deviation and updating the personalized growth profile of step 6 comprises:

[0036] Push real-time posture correction instructions to the teaching large screen through the WebSocket communication protocol, synchronously display the three-dimensional skeleton wireframe, and send training reports containing text prompts and two-dimensional posture analysis graphs to iOS and Android mobile terminals;

[0037] Based on the Unity three-dimensional engine, a student virtual image is constructed, and a red-yellow-green three-color gradient mapping technology is used to identify the posture deviation area;

[0038] A knowledge graph with student ID as the unique index is constructed, and student physical action data and historical scoring records are stored in association, supporting personalized training programs according to time dimension.

[0039] Through WebSocket, multi-terminal real-time synchronization of three-dimensional skeleton wireframe and correction instructions is realized, combined with the virtual image of Unity engine and three-color heat map to intuitively present the deviation; at the same time, a knowledge graph with student ID as the core is constructed, integrating historical data and scoring records, realizing visual tracking of the training process and dynamic optimization of personalized programs, forming a closed-loop intelligent teaching feedback system.

[0040] The air hostess student physical temperament scoring system of the application comprises:

[0041] The physical training teaching monitoring module is configured with a multi-view high-definition motion capture camera array and a face recognition unit, used for collecting and associating student training videos and identity information;

[0042] The body data collection module is deployed with a YOLO v8 target detection algorithm and a three-dimensional reconstruction unit, and is used to extract 25 skeleton key points from a video stream and generate a three-dimensional posture data point array;

[0043] The data preprocessing module integrates a time sequence convolution network and an expert action comparison unit, and is used for smoothing the three-dimensional posture data point array and comparing it with an expert standard action template library;

[0044] The body data analysis module includes a spatio-temporal graph convolution attention network, which is used to extract the spatial topology relationship of the nodes through the GCN layer, learn the action time sequence features through the bidirectional LSTM unit, and dynamically allocate weights to the eight core body indicators through the cross-attention mechanism.

[0045] The intelligent scoring module is built-in with a fine-tuned Qwen-7B large model and a three-dimensional report generator, which is used for intelligent scoring and outputs a three-dimensional scoring report containing sub-item scores, improvement suggestions and airline matching index.

[0046] The feedback module is used to generate a student virtual image and a posture heat map through digital twinning technology, to real-time feedback training deviation and to update the personalized growth profile.

[0047] Preferably, the body training teaching monitoring module includes 4 groups of 1280 million pixel high-definition cameras uniformly deployed on the top of the training room, with a frame rate not less than 60 FPS; and a face recognition API service connected with the school administrative system, supporting automatic binding of student ID numbers.

[0048] Through the 4 groups of high-frame-rate 1280 million pixel overhead cameras, no dead angle motion capture is achieved, and the student ID number is automatically associated with the training data through the face recognition API connected with the administrative system, ensuring high-definition collection and accurate identity matching during the body training process, and providing high-quality and traceable raw data support for subsequent intelligent scoring.

[0049] Preferably, the body data collection module includes a YOLO v8 model with an input resolution of 1280x1280; a 32FPS real-time inference unit realized by a Jetson platform; and a three-dimensional coordinate calculation unit based on a triangulation matrix.

[0050] The high-resolution YOLO v8 model combined with the 32FPS real-time inference capability of the Jetson platform realizes efficient and accurate skeleton key point detection, and quickly calculates the three-dimensional coordinates through the triangulation matrix, providing a low-latency and high-precision three-dimensional posture data source for body evaluation, ensuring the real-time and reliability of subsequent analysis.

[0051] Preferably, the feedback module comprises: a real-time instruction transmission channel implemented by a WebSocket protocol; a posture heat map generation unit with a red-yellow-green three-color gradient mapping; and a two-dimensional image report system compatible with iOS / Android mobile terminals.

[0052] The low-delay instruction transmission channel is established through the WebSocket protocol, the posture deviation is intuitively presented in combination with the three-color heat map, and the two-dimensional report system is compatible with multiple terminals, so that real-time visual feedback and cross-platform seamless interaction of training data are realized, and the immediacy of body correction and the accuracy of teaching guidance are significantly improved.

[0053] Advantages: Compared with the prior art, the present application has the following advantages: 1. Through multi-view motion capture and three-dimensional posture reconstruction technology, the body shape of the flight attendant student can be tracked and evaluated for a long time, continuously and dynamically, which breaks through the subjective limitations of traditional manual observation, generates an objective score report based on quantitative data, and significantly improves the scientific nature and traceability of body training evaluation; 2. Deeply integrate AI technology and flight attendant professional standards, convert airline recruitment requirements into a quantifiable intelligent scoring system, realize the transformation of traditional oral teaching mode to digital and standardized teaching; 3. Through expert database data comparison and digital twin feedback technology, real-time recognition of student posture deviation and generation of targeted improvement plan form a closed-loop training system of "evaluation-feedback-correction", effectively improving teaching efficiency and training quality; 4. Apply spatio-temporal graph convolution network and large model technology to body training field, provide a demonstration case for cross-field integration of artificial intelligence in vocational education, and promote the landing innovation of AI technology in traditional skill training. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The present application is a method flowchart;

[0055] Figure 2 The present application is a body and temperament intelligent scoring system design flowchart;

[0056] Figure 3 The present application is a body and temperament intelligent scoring system architecture schematic diagram. DETAILED DESCRIPTION

[0057] The technical solutions of the present application will be further described below with reference to the accompanying drawings.

[0058] As Figure 1As shown, the flight attendant student physical temperament scoring method provided by the present application first combines the advantages of CNN-LSTM deep learning neural network in storing time-related information, builds a flight attendant student physical temperament intelligent scoring system based on deep learning and expert feedback, collects flight attendant student three-dimensional posture data points array combining the YOLO v8 target detection algorithm of the high-definition motion capture camera, inputs the real-time dynamic physical data of the flight attendant student as the basic data, takes the professional quality standards of the civil aviation industry and the airlines for the physical temperament of the flight attendants as the standard, intelligently scores the physical temperament of the flight attendant student in real time and dynamically, and gives personalized training guidance to the relatively weak part. Therefore, it is beneficial to dynamically track the physical temperament of the flight attendants, realize objective and comprehensive evaluation. Not only can it provide more scientific and accurate teaching feedback for the teachers, but also increase the employment competitiveness of the flight attendant students in future employment.

[0059] Specifically, the flight attendant student physical temperament scoring method provided by the present application comprises the following steps:

[0060] First, install a high-definition motion capture camera in the physical training room, which can track the action posture of the person in real time, use the device to form a flight attendant student physical posture collection port, collect training videos of the flight attendant student in the classroom process from different angles (such as front, side and oblique upper), and the collected video stream is transmitted to the teaching large screen in the physical training room for synchronous display and the workstation in the background for subsequent data processing; at the same time, through the face recognition system, each student participating in the training is accurately matched according to the student ID, the change of the student in the whole physical training process is recorded, and the information security of the student is also ensured to avoid data leakage.

[0061] According to the video data obtained by the teaching monitoring, the YOLO v8 target detection algorithm is used to reconstruct the three-dimensional posture of the student in the physical training process: first, the human body skeleton key points are extracted, and the planar two-dimensional coordinate sequence mainly composed of 25 main joint points is obtained; second, the spatial triangulation is performed combining the parameters of the high-definition motion capture camera, and the three-dimensional posture data point array of the student is generated; then the posture sequence is smoothed through the time convolution network (TCN), and the noise interference generated by the lens jitter, the overlapping of the students in the training, and the wrinkles of the training clothes is eliminated; at the same time, the professional characteristic actions (such as bowing amplitude, gait cycle, gesture trajectory) of the physical teachers and the flight attendants are parameterized and calibrated as the expert library in the physical training process, which provides data support for subsequent analysis.

[0062] The collected student three-dimensional posture data point array data is captured through a multi-layer convolutional neural network to capture the limb coordination relationship in the spatial dimension; at the same time, the LSTM unit is used to learn the action evolution law in the time dimension; in addition, the cross attention mechanism is introduced to assign dynamic weights to the air hostess etiquette professional quality assessment indicators (such as cervical angle, shoulder line levelness, and center of gravity transfer trajectory), and a composite feature matrix containing static beauty parameters (spinal physiological curvature, head-body ratio, etc.) and dynamic temperament parameters (action fluency, rhythm and rhythm) is generated.

[0063] According to the recruitment standards of well-known civil aviation enterprises and airlines for flight attendants at home and abroad, combined with the experience and knowledge of flight attendant experts, a general and comprehensive system directly reflecting the core professional quality and physical index system of flight attendants is obtained by using large model fine-tuning technology, such as using Qwen large model to fuse and summarize the standards, while a self-adaptive tracking mechanism is designed to realize progress quantification by comparing historical data of students, and an adversarial generative network (GAN) is introduced to simulate typical error posture, enhance the fault tolerance and robustness of the scoring system. Finally, a three-dimensional scoring report containing sub-item scores, improvement suggestions and professional potential is output.

[0064] A human-computer interaction feedback system is developed to build a student physical training quality growth file, and posture problems and correction instructions occurring in each training course are fed back to the training large screen, workstation and student mobile phone in real time; the stage feedback adopts digital twin technology to generate a three-dimensional virtual image of the student himself, and the posture deviation heat map is directly displayed through color mapping, and a personalized teaching plan is customized for each student to maximize the requirements of airlines on the physical and temperament of students.

[0065] As shown in Figure 2 and 3 , it is an air hostess student physical and temperament intelligent scoring system based on deep learning and expert feedback.

[0066] (1) Physical training teaching monitoring module

[0067] Four groups of high-definition motion cameras are evenly arranged on the top of the physical training room of the school, forming a spatial coordinate coverage network, and the camera calibration parameters are transmitted to the workstation for storage; in addition, the cameras are connected with the workstation through the network, realizing real-time transmission of student physical data to the teaching large screen and the background workstation.

[0068] The face recognition module is based on the student information library in the school's teaching management system, and the face recognition API is connected with the student information library to bind the student information captured by the camera. The whole student physical training data transmission adopts encryption method to ensure the safety of student information, and only the authorized teacher end can decrypt the historical data.

[0069] (2) Student physical data acquisition module

[0070] YOLO v8 model is deployed on a workstation with an input resolution of 1280x1280, achieving 32FPS real-time inference on the Jetson platform. Three-dimensional pose reconstruction is implemented in three steps:

[0071] Key point extraction: Extract 25 key points (including C7 vertebrae, acromion, anterior superior iliac spine, etc. for civil aviation body assessment) and introduce a dynamic weight decay mechanism to effectively solve the student occlusion problem;

[0072] Spatial positioning: Based on the camera's calibration parameters, use the camera's baseline distance, installation angle, and focal length parameters to construct a triangulation matrix to achieve multi-view spatial triangulation. Finally, calculate the three-dimensional coordinates using the least squares method to generate a three-dimensional point array data of the student's body posture;

[0073] Temporal filtering: Design a TCN-ResNet hybrid model for motion smoothing, with a convolution kernel size of 7x1 to capture 128 frames of temporal features and effectively suppress coordinate jitter "noise" caused by clothing wrinkles.

[0074] Expert library construction: Use motion capture systems to collect body standard actions of body teachers and airline cabin crew instructors to establish a template library containing 12 parameters such as body posture, bowing angle, and gait cycle.

[0075] (3) Student body data analysis module

[0076] Construct a spatio-temporal graph convolutional attention network (ST-GAT) that includes:

[0077] Spatial modeling: Use GCN layers to extract the topological relationship and feature vector of the key points;

[0078] Temporal modeling: Bidirectional LSTM units process temporal data, effectively capturing phase changes in actions;

[0079] Attention mechanism: Design a cross-attention mechanism to dynamically allocate attention to 8 core indicators such as cervical angle (weight 0.3) and shoulder line levelness (0.25). Implement hard attention selection through Softmax to generate a 128-dimensional complex feature vector.

[0080] (4) Intelligent scoring module

[0081] Based on the Qwen-7B model, build a scoring engine and implement three-stage fine-tuning:

[0082] Knowledge injection: Convert the recruitment standards of 12 airlines such as China Southern Airlines and China Eastern Airlines into a structured knowledge graph and inject it into the large model through LoRA technology;

[0083] Preference learning: Collect 30 scoring cases from experienced body shape instructors of civil aviation airports and airlines, design industry expert feedback mechanism, and use PP fine-tuning Qwen-7B large model for preference alignment training.

[0084] Adaptive calibration: Design a sliding window mechanism, compare the current data of the student with the 30-day average, and use Z-score standardization for progress calculation.

[0085] In the adversarial training phase, 200 typical error postures (such as pelvic anteversion angle > 10°, arm swing asymmetry > 15%) are generated using DC-GAN.

[0086] Output three-dimensional scoring report: including 6-dimensional score (percentage), TOP3 improvement items (sorted by potential improvement), and airline matching index (range 0-1).

[0087] (5) Feedback module

[0088] Develop a cross-platform feedback system based on Unity 3D, including:

[0089] Real-time feedback: Push posture correction instructions through WebSocket, display real-time skeleton wireframe on the teaching big screen, and receive text prompts and two-dimensional image reports on the student's mobile phone (iOS / Android).

[0090] Digital twin: Build a virtual avatar based on the Unity engine, and map the deviation area through color gradient (red->yellow->green to represent error level), and generate a posture heat map.

[0091] Growth profile: Build a "student-action-score" knowledge graph, and use student ID as the unique index to realize personalized solution query.

[0092] Based on deep learning and expert feedback, the system breaks through the technical bottleneck of traditional body shape scoring system "two-dimensional and subjective" through the deep integration of three-dimensional kinematics modeling and expert knowledge graph. The implementation data shows that the system significantly improves certain indicators such as posture recognition accuracy and scoring consistency compared to existing technologies, effectively shortens the student's body training period, improves the effectiveness of classroom teaching, and has significant technical foresight and industrial application value.

Claims

1. A method for scoring the physical and temperament of flight attendant students based on deep learning, characterized by: The following steps are involved: (1) Capture flight attendant students’ physical training video streams through multi-view high-definition motion capture cameras, and use face recognition technology to associate student identity information; (2) Based on the YOLO v8 target detection algorithm, the 25 skeletal key points of the human body in the video are extracted into a two-dimensional coordinate sequence, and the multi-view triangulation positioning is performed in combination with the camera spatial parameters to generate a three-dimensional posture data matrix; (3) Use a temporal convolutional network to smooth the 3D posture data points and compare them with a pre-built expert standard action template library; (4) A spatiotemporal graph convolutional attention network is used to analyze 3D posture data, where: the spatial topological relationship of joints is extracted through the GCN layer, the action timing features are learned through the bidirectional LSTM unit, and the weights of the eight core body indicators are dynamically assigned through the cross-attention mechanism; (5) Fine-tune the Qwen-7B model, perform intelligent scoring based on the fine-tuned Qwen-7B model, integrate the airline recruitment standard knowledge graph and instructor scoring cases, and output a three-dimensional scoring report containing sub-item scores, improvement suggestions, and airline matching index; (6) Generate students’ virtual images and posture heat maps through digital twin technology, provide real-time feedback on training deviations and update personalized growth files.

2. The method for scoring the physical and temperament of flight attendant students according to claim 1, characterized in that: The step 2 of generating the three-dimensional posture data lattice includes: Using the YOLO v8 model with an input resolution of 1280×1280, 25 key points for civil aviation shape assessment were extracted in real time at 32FPS on the Jetson platform. A dynamic weight decay mechanism is used to handle student occlusion during training; Based on the calibration parameters of the multi-view camera, the triangulation matrix is ​​constructed using the camera baseline distance, installation angle and focal length; The three-dimensional coordinates are calculated by the least square method to generate three-dimensional dot matrix data of the student's body posture.

3. The method for scoring the physical and temperament of flight attendant students according to claim 1, characterized in that: The construction of the expert standard action template library described in step 3 includes: The motion capture system is used to collect 12 standard movement data of physical teachers and airline flight attendant instructors; Establish a database of occupational characteristic movements including bowing amplitude, gait cycle, and gesture trajectory.

4. The method for scoring the physical and temperament of flight attendant students according to claim 1, characterized in that: Step 4 includes: The graph convolutional network (GCN) layer is used to construct the topological map of human joints, extract the spatial correlation features between adjacent joints, and output the joint feature vector; The bidirectional LSTM unit processes the sequential data of consecutive frames to capture the phase variation of the action. The hidden layer dimensions of both the forward LSTM and the backward LSTM are 64-dimensional. A cross-attention layer is designed to dynamically assign weights to eight core body indicators, including a cervical spine angle weight coefficient of 0.3 and a shoulder line horizontality weight coefficient of 0.

25. Hard attention selection is implemented through the Softmax function, and finally a 128-dimensional composite feature vector is output.

5. The method for scoring the physical and temperament of flight attendant students according to claim 1, characterized in that: The fine-tuning of the Qwen-7B large model described in step 5 includes: The preset recruitment standards of 12 airlines were converted into a structured knowledge graph, and the knowledge graph was injected into the Qwen-7B large model through low-rank adaptive technology; We collected scoring case data from several senior physical instructors at civil aviation airports and airlines, and used a proximal strategy optimization algorithm to perform preference alignment training on the Qwen-7B large model. Design a sliding time window mechanism to compare students' current physical data with the historical 30-day average data, and use the Z-score standardization method to calculate the progress index; Use a deep convolutional generative adversarial network to generate 200 types of typical incorrect posture data, including incorrect postures with pelvic tilt angles exceeding 10 degrees and arm swing asymmetry exceeding 15 degrees; The output includes a 100-point score for each of the six dimensions, the top three improvement suggestions ranked by potential for improvement, and a three-dimensional score report with an airline matching index ranging from 0 to 1.

6. The method for scoring the physical and temperament of flight attendant students according to claim 1, characterized in that: The real-time feedback of training deviation and updating of personalized growth profile described in step 6 include: Push real-time posture correction instructions to the teaching screen via the WebSocket communication protocol, synchronously display the 3D skeleton wireframe diagram, and send training reports containing text prompts and 2D posture analysis diagrams to iOS and Android mobile terminals; The student's virtual image was constructed based on the Unity 3D engine, and the red, yellow, and green gradient mapping technology was used to identify the posture deviation areas; Build a knowledge graph with student ID as the unique index, associate and store student body movement data and historical scoring records, and support querying personalized training plans by time dimension.

7. A deep learning-based body shape and temperament scoring system for flight attendant students, characterized by: include: The physical training teaching monitoring module is equipped with a multi-view HD motion capture camera array and a facial recognition unit to collect and associate student training videos with their identity information; The body data acquisition module deploys the YOLO v8 target detection algorithm and 3D reconstruction unit to extract 25 skeletal key points from the video stream and generate 3D posture data points; The data preprocessing module integrates a temporal convolutional network and an expert motion comparison unit to smooth the 3D posture data points and compare them with the expert standard motion template library; The body data analysis module includes a spatiotemporal graph convolutional attention network, which is used to extract the spatial topological relationship of joints through the GCN layer, learn the temporal characteristics of actions through bidirectional LSTM units, and dynamically assign weights to eight core body indicators through a cross-attention mechanism; The intelligent scoring module, which includes a fine-tuned Qwen-7B large model and a 3D report generator, is used for intelligent scoring and outputs a 3D scoring report containing sub-item scores, improvement suggestions, and airline matching index. The feedback module is used to generate student virtual images and posture heat maps through digital twin technology, provide real-time feedback on training deviations, and update personalized growth files.

8. The flight attendant student body and temperament scoring system according to claim 7 is characterized in that: The body training teaching monitoring module includes: 4 groups of 12.8 million pixel high-definition cameras evenly deployed on the top of the training room, with a frame rate of not less than 60FPS; a face recognition API service connected to the school's academic affairs system, which supports automatic binding of student numbers.

9. The flight attendant student body shape and temperament scoring system according to claim 7, characterized in that: The body data acquisition module includes: a YOLO v8 model with an input resolution of 1280×1280; a 32FPS real-time inference unit implemented on the Jetson platform; and a three-dimensional coordinate calculation unit based on a triangulation matrix.

10. The flight attendant student body shape and temperament scoring system according to claim 7, characterized in that: The feedback module includes: a real-time instruction transmission channel implemented by the WebSocket protocol; a posture heat map generation unit with red, yellow and green gradient mapping; and a two-dimensional image reporting system compatible with iOS / Android mobile terminals.

Images