Method for integrally evaluating class states of students by taking class as unit

Abstract

The invention relates to a method for integrally evaluating the class state of a student by taking a classroom as a unit, which belongs to the field of education quality management and comprises the following steps of: S1, acquiring a classroom video and segmenting the classroom video into a frame image sequence; s2, students in each frame of image are detected according to a human body detection algorithm, and human body areas are cut out; s3, outputting human skeleton points according to a posture detection algorithm, and generating a pure skeleton line segment image; s4, classifying the skeleton line segment images based on a pre-trained posture classification model, and evaluating the learning posture of each student; and S5, counting all student states in each frame of image, finally combining evaluation results of all frames, and calculating the student state of the whole class. According to the invention, through the training model, the learning state of the student can be objectively counted according to a posture detection algorithm, and the method has the characteristics of fast statistical speed, objective result, trainable model and adjustable standard.

Description

Technical Field

[0001] The present invention relates to the field of education quality management, and in particular to a method for comprehensively evaluating students' class status based on classrooms. Background Art

[0002] With the development of education informatization and intelligence, traditional classroom management methods are difficult to fully and real-time understand students' class status, and teachers can only rely on subjective experience to evaluate students' learning situation. This method not only has certain limitations, but also makes it difficult to accurately grasp the learning status of each student, especially in large classes, it is even more difficult to fully grasp students' attention, participation and understanding level.

[0003] In response to this situation, a method of evaluating students' classroom status as a whole based on the classroom was proposed, which is of great significance in improving teachers' education and teaching.

[0004] Attempts to detect human behavior have yielded many results: for the detection method of students' attendance status during lectures, the whole class is recorded by cameras, students' classroom behavior is detected, and then the overall status of students' attendance is evaluated. Summary of the invention

[0005] The present invention is proposed to alleviate or solve at least one aspect or at least one point of the above problems.

[0006] The present invention provides a method for evaluating students' class status in an overall manner based on a class, comprising the following steps:

[0007] S1: Collect classroom videos and segment them into frame image sequences;

[0008] S2: Detect the students in each frame of the image based on the human body detection algorithm and crop out the human body area;

[0009] S3: Output human skeleton points according to the posture detection algorithm to generate a pure skeleton line segment image;

[0010] S4: Based on the pre-trained posture classification model, the skeleton segment images are classified to evaluate each student’s learning posture;

[0011] S5: Count all student statuses in each frame image, and finally merge the evaluation results of all frames to calculate the student status of the entire class.

[0012] Preferably, the classroom video includes all students, the frame image sequence is divided into frame sequences of fixed length according to the time series, and the human body detection algorithm is a computer vision technology that aims to automatically identify and locate the presence of the human body from an image or video. The algorithm generates information containing the position and boundaries of the human body by analyzing the input image data.

[0013] Preferably, the human body detection algorithm achieves accurate detection and cropping of each student by applying a pre-trained multi-target tracking model, providing a reliable data basis for subsequent posture evaluation and behavior analysis. The multi-target tracking model is based on an efficient target detection algorithm, specifically used for multi-target tracking tasks, and uses an anchor-free method to predict target positions and categories, avoiding the complexity of anchor frame configuration, and utilizing a feature pyramid network to enhance the model's detection capabilities for targets of different scales.

[0014] Preferably, the efficient target detection algorithm includes:

[0015] Adopting an anchor-free detection paradigm, it does not rely on predefined anchor boxes for object localization.

[0016] Use backbone network;

[0017] The cross-stage residual stage is introduced to reduce the computational complexity by segmenting the feature map, allowing the model to extract more feature information in different paths;

[0018] Use a dynamic label allocation algorithm.

[0019] Preferably, the human detection algorithm inputs each frame of image into the multi-target tracking model for pedestrian detection, and the multi-target tracking model outputs the detection result by extracting features from the image.

[0020] Preferably, the human body detection algorithm uses the detection frame coordinates to crop a corresponding area from the original frame for each detected student.

[0021] Preferably, the posture detection algorithm outputs human skeleton points to generate a pure skeleton line segment image; the posture detection model consists of an input layer, a feature extraction network, a key point detection layer, a post-processing layer, and an output layer.

[0022] Preferably, the high-resolution human posture estimation algorithm includes:

[0023] A lightweight high-resolution network is used for feature extraction. By combining the permutation module of the shuffle network and the multi-branch structure of the high-resolution network, the conditional weighted channel unit is used to replace the traditional 1x1 convolution.

[0024] By analyzing the highest response area in the heat map, the coordinate position of each bone key point is found;

[0025] Adopt unbiased data encoding and decoding method;

[0026] Adopting information loss array augmentation technology, the generalization ability and robustness of the model are enhanced by randomly blocking part of the input data, simulating noise and information loss;

[0027] Use 16-bit floating point inference.

[0028] Preferably, the skeleton line segment generates a skeleton line segment graph representing the posture of the human body based on the identified skeleton point coordinates.

[0029] Preferably, the human body detection algorithm realizes image denoising and pure background generation. In the process of generating the skeleton line segment graph, unnecessary visual elements in the human body area are removed, and only key skeleton line segments are retained.

[0030] Beneficial effects:

[0031] The present invention records the entire class through a camera in the classroom, divides the classroom video into a frame image sequence, detects the students in each frame image according to a human body detection algorithm, outputs human skeleton points according to a posture detection algorithm, generates a pure skeleton line segment image, and classifies the skeleton line segment image based on a pre-trained posture classification model to evaluate the learning posture of each student; summarizes the evaluation results of all students to obtain the learning status of students in the entire class. The present invention can objectively count the learning status of students according to the posture detection algorithm through a training model, and has the characteristics of fast statistical speed, objective results, trainable model, and adjustable standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart for evaluating students’ class status based on the overall class performance;

[0033] Figure 2 Flowchart of the human body detection algorithm;

[0034] Figure 3 is the flow chart of the posture detection algorithm;

[0035] Figure 4 This is the architecture diagram of the posture detection model;

[0036] Figure 5 Flowchart for classifying skeleton line segment images based on the pose classification model. DETAILED DESCRIPTION

[0037] The following description of the embodiments of the present invention with reference to the accompanying drawings is intended to explain the overall inventive concept of the present invention, and should not be construed as a limitation of the present invention. In the present invention, the same reference numerals represent the same or similar components.

[0038] The features described herein can be implemented in different forms and should not be construed as being limited to the examples described herein. Rather, the examples described herein have been provided to illustrate only some of the many possible ways to implement the methods, devices, and / or systems described herein, which will be clear after understanding the disclosure of the present invention.

[0039] Although terms such as "first", "second", and "third" may be used herein to describe various members, components, regions, layers, or portions, these members, components, regions, layers, or portions should not be limited by these terms. Instead, these terms are only used to distinguish one member, component, region, layer, or portion from another member, component, region, layer, or portion.

[0040] In the specification, when an element (such as a layer, a region, or a substrate) is described as being “on”, “connected to”, or “coupled to” another element, the element may be directly “on”, “connected to”, or “coupled to” another element, or one or more other elements may be present therebetween. Conversely, when an element is described as being “directly on”, “directly connected to”, or “directly coupled to” another element, there may be no other elements present therebetween.

[0041] The terms used herein are only used to describe various examples and are not intended to limit the disclosure. Unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. The terms "comprise", "include" and "have" indicate the presence of the described features, quantities, operations, components, elements and / or combinations thereof, but do not exclude the presence or addition of one or more other features, quantities, operations, components, elements and / or combinations thereof.

[0042] In order to enable those skilled in the art to use the contents of the present invention, the following exemplary embodiments may be provided in combination with specific application scenarios, specific systems, parameters of devices and components, and specific connection methods. However, for those skilled in the art, these embodiments are only examples, and the general principles defined herein may be applied to other embodiments and application scenarios without departing from the spirit and scope of the present invention.

[0043] like Figure 1 As shown, an exemplary embodiment of the present invention includes the following steps:

[0044] S1: Collect classroom videos and segment them into frame image sequences.

[0045] The surveillance video of the whole class is obtained through the camera equipment installed in the classroom. The collected classroom video is divided into a frame sequence of fixed length according to the time series. Each frame image records the status of the student at the current moment, providing a data basis for subsequent processing. The specific method is as follows:

[0046] S11: The entire teaching process is recorded in real time using the high-definition video equipment installed in the classroom to generate a surveillance video containing all student activities. The collected video data is stored to ensure that the quality and resolution of the video file meet the requirements of subsequent analysis;

[0047] S12: Process the stored video files and set the frame rate of the video processing, for example, extract 30 frames per second, to ensure that sufficient information can be captured. The video is divided into a frame sequence of fixed length according to the time series, and the extracted frame images are stored in a dedicated database for fast retrieval and access.

[0048] S2: Detect the pupils in each frame of the image and crop the human body area according to the human body detection algorithm. The human body detection algorithm is a computer vision technology that aims to automatically identify and locate the presence of human bodies from images or videos. The algorithm generates information containing the location and boundaries of the human body by analyzing the input image data, such as Figure 2 As shown, the specific implementation process is as follows:

[0049] S21: Input each frame of the image and apply the human body detection algorithm to detect each student in the image. The human body detection algorithm uses a pre-trained multi-target tracking model to accurately detect and crop each student, providing a reliable data basis for subsequent posture evaluation and behavior analysis. The multi-target tracking model is based on an efficient target detection algorithm, specifically used for multi-target tracking tasks, using an anchor-free method to predict the target position and category, avoiding the complexity of anchor box configuration, and using a feature pyramid network to enhance the model's detection capabilities for targets of different scales.

[0050] The efficient target detection algorithm is an efficient target detection algorithm based on the Feipang deep learning framework. It is improved and enhanced on the basis of the target detection series model, aiming to provide higher detection accuracy and faster reasoning speed. The content of the efficient target detection algorithm is as follows:

[0051] (1) The anchor-free detection paradigm is adopted to locate the target without relying on predefined anchor boxes. This method reduces the need to tune the anchor box parameters, making the model more adaptable in different scenes and better able to handle targets of different sizes and shapes. Therefore, the model can automatically generate prediction boxes based on the image content.

[0052] (2) Use the backbone network to enhance the model's ability to extract features. The neck structure integrates the feature pyramid network to improve the detection ability of multi-scale targets. The backbone network is a pre-trained convolutional neural network that contains multiple convolutional layers, pooling layers, and activation functions to form a hierarchical structure to gradually extract more complex features. The neck structure fuses and transfers features of different scales to improve the model's ability to detect multi-scale targets.

[0053] (3) The cross-stage residual stage is introduced. By splitting the feature map to reduce the computational complexity, the model is allowed to extract more feature information in different paths, improving the accuracy and efficiency of feature extraction. The task head focuses on the processing of classification and regression tasks, and adopts a simplified structure to improve computational efficiency. The cross-stage residual stage is a network structure used to improve the computational efficiency and feature extraction accuracy of deep learning models. It divides the feature map into two parts, one of which is processed through residual connections, while the other is enhanced through additional feature extraction paths. The efficient task head is a part of the object detection model that focuses on the processing of classification and regression tasks.

[0054] (4) Use a dynamic label assignment algorithm. The dynamic label assignment algorithm is a strategy for training target detection models that aims to improve training efficiency and model generalization by evaluating the confidence and position accuracy of each prediction box in real time. Based on the real-time evaluation results, the prediction boxes with high confidence are matched with the true target and labels are dynamically assigned to the prediction boxes. The dynamic label assignment can be adjusted according to the performance of the model in training. It reduces the unevenness of labels and improves the generalization ability of the model.

[0055] (5) An efficient task head with speed and accuracy as the goal is proposed, which simplifies the alignment of the classification branch as a shortcut and replaces the alignment of the regression branch with a distribution focus loss layer. For classification and localization, zoom loss and distribution focus loss are selected and applied to the target detector to improve performance. Distribution focus loss is a loss function used in target detection, which aims to solve the problem of class imbalance and improve detection accuracy, and is used for key point localization in regression tasks. Zoom loss is another loss function for localization and classification tasks in target detection. It not only focuses on the accuracy of the classification task, but also focuses on the accuracy of localization. As shown in formula (1):

[0056]

[0057] Among them, α, β, and γ represent weight coefficients; loss GIoU represents the generalized intersection-combination loss; Represents the normalized target score. loss VFL Indicates zoom loss. loss DFLrepresents the distribution focal loss; N pos Represents the number of positive samples; i represents the number of indexed positive samples;

[0058] The feature pyramid network is a network architecture used for target detection and image segmentation tasks. It aims to enhance the model's ability to detect targets of different sizes through multi-scale feature fusion. It builds a pyramid structure to combine low-level features with high-level features, thereby enhancing the model's ability to recognize objects of different scales.

[0059] S22: Define the multi-target tracking model structure, use the predefined interface to import the model, and load the pre-trained weight file. In order to speed up the training or reasoning of the model, load the pre-trained weights. The weight file is trained with a large-scale data set, which can help the model converge faster or perform better during reasoning. Adjust the size of the input image to meet the requirements of the model.

[0060] S23: Input each frame of image to the multi-target tracking model, extract image features and output detection results, and filter out detection results with low confidence. The detection results include the bounding box coordinates of each pedestrian and the confidence score. The confidence score indicates the degree of confidence of the model in the detection result. Filter low-confidence detection results to reduce false detections.

[0061] S24: Using the detection frame coordinates, crop each student area in the original frame, and save the cropped student area image to a specified directory. The detection frame contains the coordinates of the bounding box, which can accurately locate the position of each student in the image, and save the cropped human body area image to a specified directory.

[0062] S3: Output human skeleton points based on the posture detection algorithm to generate a pure skeleton line segment image.

[0063] like Figure 3 As shown in the figure, the specific implementation process of the posture detection algorithm is as follows:

[0064] S31: Input each cropped image into a posture detection algorithm. The posture detection algorithm is a computer vision technology used to detect the posture of a human body from an image or video. It locates the key points of the human body (such as the head, shoulders, elbows, knees, ankles, etc.) through image preprocessing, feature extraction, heat map generation, and key point prediction, and generates a skeletal structure diagram representing the human posture.

[0065] S32: Load the posture detection model to ensure the accuracy and efficiency of the posture detection process. Figure 4As shown in the figure, the posture detection model consists of an input layer, a feature extraction network layer, a key point detection layer, a post-processing layer, and an output layer. It is based on a high-resolution human posture estimation algorithm. The high-resolution human posture estimation algorithm is used to identify and estimate the key points of the human body in real time. It is developed based on the Feijiang deep learning framework. The high-resolution human posture estimation algorithm is as follows:

[0066] (1) The input layer is responsible for receiving and preprocessing the input image for subsequent feature extraction network processing. The input image is resized to a fixed width, height, and number of channels. The pixel values ​​are normalized to a known range to improve the prediction stability of the model.

[0067] (2) A lightweight high-resolution network is used for feature extraction. By combining the permutation module of the shuffle network and the multi-branch structure of the high-resolution network, a conditional weighted channel unit is used to replace the traditional 1x1 convolution. This unit consists of two parts: a cross-resolution weighting function and a spatial weighting function. The conditional weighting part collects the resolution feature maps of other branches in the high-resolution network and determines the strength of the fusion based on the information content. This design further reduces the amount of calculation while ensuring accuracy.

[0068] (3) In the heat map generated by the model, the position of each skeletal key point is determined by finding the peak value in the heat map. The coordinate position of each skeletal key point is found by analyzing the highest response area in the heat map. In order to further improve the performance of the model under low-resolution heat maps, a dark channel module is integrated. This module dynamically adjusts the representation of key points so that the model can still accurately identify and locate human key points even in low-quality images.

[0069] (4) Adopting unbiased data encoding and decoding methods is to eliminate the bias introduced by data scale changes, image enhancement or other preprocessing steps during the data encoding and decoding process, ensure that the input and output of the model can maintain consistency, and thus improve the positioning accuracy of the model. Unbiased data encoding and decoding is a data processing method that aims to avoid introducing any systematic bias or distortion during the data encoding and decoding process.

[0070] (5) In order to improve the model's ability to locate key points, the array enhancement technology with information loss is used to enhance the generalization ability and robustness of the model by randomly blocking part of the input data, simulating noise and information loss. By simulating the situation of missing information, the model's robustness and adaptability to key points are enhanced, so that it can better handle various complex situations and environments in practical applications.

[0071] (6) In order to speed up the inference process, 16-bit floating-point inference is used. This method reduces the calculation accuracy of the model to 16-bit floating-point numbers, thereby reducing memory usage and computing burden. Through 16-bit floating-point inference, the model can not only achieve faster inference speed, but also improve the overall processing efficiency while ensuring accuracy. 16-bit floating-point inference is the process of using 16-bit floating-point numbers to perform model inference.

[0072] (7) The output layer is used to output the position of each key point and the confidence score of each key point to indicate the reliability of the detection results.

[0073] Formula (2) shows the calculation process:

[0074]

[0075] Where Y(i,j,k) represents the output value of the current pixel; X(i+m,j+n,k) represents a window position on the input image; W(m,n,k) represents the weight of the position and layer; X is the input feature map; W is the convolution kernel; k is the number of channels; H(x,y) represents the function value at the coordinate (x,y); x c ,y c represents the center position of Gaussian distribution; σ represents the standard deviation; exp represents the shape characteristics of Gaussian distribution; P real is the actual coordinate; P pixel is the pixel coordinate; S is the scale factor; 0.5 is a correction to compensate for quantization error; X FP16 Represents the converted 16-bit floating point number; X FP32 Represents 32-bit floating point format input data.

[0076] S33: Based on the key point coordinates obtained by posture detection, adjacent bone points are connected to generate bone segments to construct a complete bone structure diagram. In a specific implementation, a line segment structure reflecting the human body posture is generated by connecting adjacent bone points (such as head and shoulder, shoulder and elbow, elbow and wrist, etc.).

[0077] S34: Remove unnecessary visual elements in the image, retain only the human skeleton line segments, set the background to a pure monochrome (such as white or transparent), and highlight the human skeleton points and connecting line segments. In the process of generating the skeleton line segment graph, remove unnecessary visual elements in the human body area and retain only the key skeleton line segments. Set the background to a pure monochrome (such as white or transparent), highlight the human skeleton points and connecting line segments, ensure the clarity and simplicity of the skeleton line segment graph, and the output image is a skeleton line segment image with a pure background, clearly showing the human posture.

[0078] S4: Based on the pre-trained posture classification model, the skeleton segment images are classified to evaluate the learning posture of each student.

[0079] like Figure 5 As shown in the figure, based on the posture classification model, the process of classifying the skeleton line segment image is as follows:

[0080] S41: Use pre-defined posture classification standards and adopt deep learning posture classification model for training. The first step of this process is to collect a certain number of skeleton line segment images with manual annotations. These images are important datasets for training, validating and testing models to ensure that the model can accurately identify and classify different learning postures. Specifically:

[0081] In order to train and test the posture detection model, it is necessary to systematically collect posture-related data. Use a camera to record the learning process of students in a classroom or learning environment, and ensure that the video quality is good so that the posture can be clearly identified. This data will be used as a training set and a test set for model learning and evaluation; use the posture detection algorithm to obtain the skeleton point and line segment data of the video frame, and organize the posture data of the lowered head, tilted head, front, body tilted, raised head, and other reasons. Organize the collected data into a unified format.

[0082] S42: Use a lightweight network architecture to build a posture classification model to achieve higher reasoning speed and computational efficiency when classifying skeletal line segment images. In the process of classifying skeletal line segment images, adjust the number of classifications, feature extraction modules, learning rate adjustment strategy, and optimizer selection to further improve the classification ability of the model. Evaluate each student's learning posture based on the correctness of body posture and body movements.

[0083] The lightweight network architecture is a lightweight convolutional neural network architecture designed to improve model inference speed and computational efficiency. Specifically, it is a variant of the Shuffle Network V2 architecture, which means that the size of the network is reduced to 0.25 times the original. The lightweight network contains multiple stages, each of which contains multiple basic units. The basic unit usually includes a 1x1 convolution, a 3x3 depth convolution, and another 1x1 convolution. A channel shuffle operation is added after the 1x1 convolution, and no activation function is used after the 3x3 depth convolution. This version uses fewer parameters and computing resources while maintaining a certain classification accuracy. It uses point-by-point grouped convolution. Grouped convolution groups the feature maps of the input layer, and then uses different convolution kernels for convolution for each group, thereby reducing the amount of calculation. Channel shuffle is introduced, which allows information exchange between feature maps of different groups, thereby enhancing the feature extraction ability of the network.

[0084] S43: Select a posture classification model to classify skeleton point and line segment images. This model is particularly suitable for devices with limited resources due to its lightweight design, and can achieve high classification accuracy at a smaller model size. After building the model, efficient model training is achieved through loss function selection, optimizer selection, and learning rate adjustment strategy. To further improve the classification accuracy of the model, the feature extraction module can be adjusted during the training process. In the training cycle, the model gradually learns effective feature representations by adjusting weights, learning rates, and batch sizes. After training, the trained model is used to recognize and classify new posture images.

[0085] S44: In the real-time evaluation stage, each frame of the skeleton line segment image will be input into the pre-trained learning posture classification model. The model can automatically identify the student's current learning posture and output the corresponding classification result. Since each frame of the image may contain multiple students, it is necessary to classify and calculate the skeleton line segment image of each student one by one to obtain the class status of each student in the current frame.

[0086] S5: Count all student statuses in each frame image, and finally merge the evaluation results of all frames to calculate the student status of the entire class.

[0087] S51: For each frame of the image, the postures of the students are classified and the proportion of "lowering the head" is counted, and then compared with the types in the pre-collected standard posture library. The overlap between the current posture and the standard posture is determined by combining the Euclidean distance and time proportion statistics, and the class status of all students at the current moment can be obtained. This step summarizes the status information of each student to form the class status distribution at that point in time; specifically:

[0088] Euclidean distance is a commonly used similarity metric used to calculate the straight-line distance between two points (or vectors). In posture assessment, the position coordinates of the joints of the human body (such as head, shoulder, elbow, etc.) can be regarded as multidimensional vectors, and the similarity is measured by calculating the Euclidean distance between these joints in different postures, as shown in formula (3):

[0089]

[0090] d(P,Q) represents the distance between pose P and pose Q; n represents the total number of key points in the pose; x i represents the x coordinate of the i-th key point in posture P; y i represents the y coordinate of the i-th key point in posture P; x i ′ represents the x coordinate of the i-th key point in posture Q; y i ′ represents the y coordinate of the i-th key point in posture Q;

[0091] Time proportion statistics are used to evaluate students' behavior status throughout the class. By counting the similarity of students' postures in each time period, we can calculate the time proportion of students in different postures, and then get an overall posture analysis.

[0092] S52: Combine and calculate the evaluation results of all the frames during the whole class. This process includes accumulating the status results of all students in each frame, calculating the proportion of each status category, calculating the number of occurrences of each posture through the accumulated status data, and converting it into a proportion. Combining the proportions of each status, the overall class status of the students in the class is obtained.

[0093] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes may be made to the embodiments and combinations of elements may be made without departing from the principles and spirit of the invention, the scope of the invention being defined by the appended claims and their equivalents.

Claims

1. A method for evaluating students' class status in a classroom-based manner, characterized in that: The steps include: S1: Collect classroom videos and segment them into frame image sequences; S2: Detect the students in each frame of the image based on the human body detection algorithm and crop out the human body area; S3: Output human skeleton points according to the posture detection algorithm to generate a pure skeleton line segment image; S4: Based on the pre-trained posture classification model, the skeleton segment images are classified to evaluate each student’s learning posture; S5: Count all student statuses in each frame image, and finally merge the evaluation results of all frames to calculate the student status of the entire class.

2. The method for evaluating students' class status in a classroom-based manner according to claim 1, characterized in that: The classroom video contains all students, the frame image sequence is a frame sequence divided into fixed-length frames according to the time series, and the human body detection algorithm is a computer vision technology that aims to automatically identify and locate the presence of the human body from an image or video. The algorithm generates information containing the position and boundaries of the human body by analyzing the input image data.

3. The method for evaluating students' class status in a classroom-based manner according to claim 2, characterized in that: The human detection algorithm achieves accurate detection and cropping of each student by applying a pre-trained multi-target tracking model, providing a reliable data basis for subsequent posture evaluation and behavior analysis. The multi-target tracking model is based on an efficient target detection algorithm and is specifically used for multi-target tracking tasks. It uses an anchor-free method to predict target positions and categories, avoids the complexity of anchor frame configuration, and uses a feature pyramid network to enhance the model's detection capabilities for targets of different scales.

4. The method for evaluating students' class status in a classroom-based manner according to claim 3, characterized in that: The efficient target detection algorithm includes: Adopting an anchor-free detection paradigm, it does not rely on predefined anchor boxes for object localization. Use backbone network; The cross-stage residual stage is introduced to reduce the computational complexity by segmenting the feature map, allowing the model to extract more feature information in different paths; Use a dynamic label allocation algorithm.

5. The method for evaluating students' class status in a classroom-based manner according to claim 4, characterized in that: The human detection algorithm inputs each frame of image into the multi-target tracking model for pedestrian detection, and the multi-target tracking model outputs the detection result by extracting features from the image.

6. The method for evaluating students' class status in a classroom-based manner according to claim 5, characterized in that: The human detection algorithm uses the detection box coordinates to crop the corresponding area from the original frame for each detected student.

7. The method for evaluating students' class status in a classroom-based manner according to claim 6, characterized in that: The posture detection algorithm outputs human skeleton points and generates a pure skeleton line segment image; the posture detection model consists of an input layer, a feature extraction network, a key point detection layer, a post-processing layer, and an output layer.

8. The method for evaluating students' class status in a classroom-based manner according to claim 7, characterized in that: The high-resolution human posture estimation algorithm includes: A lightweight high-resolution network is used for feature extraction. By combining the permutation module of the shuffle network and the multi-branch structure of the high-resolution network, the conditional weighted channel unit is used to replace the traditional 1x1 convolution. By analyzing the highest response area in the heat map, the coordinate position of each bone key point is found; Adopt unbiased data encoding and decoding method; Adopting information loss array augmentation technology, the generalization ability and robustness of the model are enhanced by randomly blocking part of the input data, simulating noise and information loss; Use 16-bit floating point inference.

9. The method for evaluating students' class status in a classroom-based manner according to claim 8, characterized in that: The skeleton line segment generation is based on the identified skeleton point coordinates to represent the skeleton line segment graph of the human body posture.

10. The method for evaluating students' class status in a classroom-based manner according to claim 9, characterized in that: The human body detection algorithm realizes image denoising and pure background generation. In the process of generating the skeleton line segment graph, unnecessary visual elements in the human body area are removed, and only key skeleton segments are retained.

Images