Inquiry recognition auxiliary method and system based on yolov7

Abstract

The present application relates to a yolov7-based query recognition auxiliary method and system, which comprises the following steps: S1, data labeling processing is performed on the original image to obtain an anchor box, i.e., labeling information; S2, the labeled data is divided into pictures and labeling information, yolov7 model training is used, and the data is divided into training, testing and evaluation data; S3, in the yolov7 model training process, the batch value is set, and the training accuracy is determined by the number of training rounds; S4, the yolov7 model after training is tested, and it is determined whether the actual accuracy rate meets the requirements; S5, the yolov7 model meeting the requirements is converted into an onnx model; S6, the converted onnx model is used for inference and recognition; S7, according to the video frame extraction interval set according to the business scenario requirements, frame extraction is performed; S8, model inference is performed on the extracted pictures, the number of people in the business scenario at this time is confirmed, and it is determined whether to perform early warning processing.

Description

Technical Field

[0001] This invention relates to a query recognition assistance method and system based on YOLOv7. Background Technology

[0002] In the digital age, biometrics is becoming a field of great interest. With continuous technological advancements and expanding application scenarios, we need to develop this technology to meet social, economic, and security needs.

[0003] In certain scenarios, such as negotiations, scheme reviews, product testing meetings, interviews, visits, and interrogations, to prevent subjective bias, at least two people should participate in the inquiry to ensure the fairness and effectiveness of the inquiry process and prevent potential subjective influence, inadequate consideration, or other inappropriate behavior. This invention believes that it is urgent to use human body recognition technology to judge the compliance of the inquiry process by machine. This can not only promptly detect misjudgments, misjudgments, or non-compliance in the inquiry and communication process, but also save related labor costs.

[0004] Because people may move around and the camera angle may vary during the questioning and communication process, there may be situations where people overlap significantly or do not appear completely in the camera's view. This makes it difficult for the model to accurately identify the number of people during the questioning and communication process.

[0005] At the same time, the YOLOv7 model has poor generalization ability and requires a large number of scenarios for training, which puts a high demand on scenario collection and equipment.

[0006] Existing technologies such as YOLOv4 and YOLOv5 may have limitations in speed and accuracy, especially when dealing with large-scale datasets and complex scenarios. These limitations may be due to insufficiently optimized model structure, excessive computational complexity, or inaccurate label assignment strategies. While MediaPipe is fast on CPUs, it may not perform as well as YOLOv7 on GPUs. Summary of the Invention

[0007] The purpose of this invention is to provide a query recognition assistance method and system based on YOLOv7 by using predefined anchor boxes to aid training and recognizing human figures in images.

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] A query recognition assistance method based on YOLOv7 includes the following steps;

[0010] S1, Perform data annotation processing on the original image to obtain anchor boxes, i.e., annotation information;

[0011] S2, the labeled data is divided into images and label information, and the YOLOv7 model is used for training, testing and evaluation data.

[0012] S3, during the training of the YOLOv7 model, set the batch value and determine the training accuracy by the number of training epochs;

[0013] S4. Test the trained YOLOv7 model to see if the actual accuracy meets the requirements.

[0014] S5, convert the YOLOv7 model that meets the requirements into an ONNX model;

[0015] S6, use the transformed ONNX model for recognition and reasoning;

[0016] S7 allows you to set the video frame extraction interval based on business scenario requirements and perform frame extraction accordingly.

[0017] S8 performs ONNX model inference on the extracted frames to confirm the number of people in the current business scenario and determine whether to issue an alert.

[0018] Furthermore, step S1 includes the following steps;

[0019] S1.1, Image preprocessing, preliminary screening;

[0020] First, image data is collected in the defined business scenarios. All collected images are confidential data, and all image data processing is carried out in an environment without external network access. Then, after the image data is collected, preliminary screening is performed to remove image data with similarity or ambiguity exceeding a set threshold. The training image data needs to include all business scenarios, and each scenario should include different types of image data to meet the set quantity.

[0021] S1.2 After the initial screening, the image data is manually reviewed a second time to identify factors that may affect the accuracy of personnel identification. These factors are then added to the annotation process and the data is trained.

[0022] Factors affecting the accuracy of person recognition include: objects whose weight is greater than a set threshold that is similar to the human body in the image data; objects whose clothing color is greater than a set threshold that is similar to the human body in the image data; and objects that reflect the human body in the image data.

[0023] S1.3, Confirm the labels in the annotation process and set the annotation rules. Perform the annotation work on the image data according to the annotation rules. During annotation, first, determine the label content required for this annotation based on the set requirements. Then, annotate according to the annotation rules. Second, train the annotation data. Third, determine whether the annotation content needs to be adjusted. If so, return to the beginning of the annotation and re-annotate. Otherwise, export the annotated image and annotation file.

[0024] Furthermore, in step S2, the labeled data is classified, and the image data is divided into training set, test set, and validation set according to a set ratio;

[0025] The training set is used for model training, the test set is used to test the overall accuracy of the model trained on the training set, and the validation set is used to validate the model's performance.

[0026] The specific steps are as follows;

[0027] S2.1, Input the annotated image and annotation file from step S1.3;

[0028] S2.2, the labeled images and labeled files are divided into training set, test set, and validation set, which are used for training, testing, and evaluation, respectively;

[0029] S2.3, Based on training requirements, set up the YAML file in the model and confirm the data path and training parameter settings;

[0030] S2.4, Perform model training;

[0031] S2.5 First, evaluate the trained model; then, determine if the model meets the accuracy requirements; if it does, derive the model; if not, proceed to S2.6.

[0032] S2.6 First, test the model and determine the improvement needs based on the test images; then, determine whether to directly adjust the parameters and train; if so, return to S2.4; otherwise, execute S2.7.

[0033] S2.7 Determine whether the annotation needs to be adjusted; if so, modify the annotation content of the image accordingly; otherwise, increase the data volume of the image and add annotations.

[0034] Furthermore, in step S7, the camera is docked, and a set frame image is extracted within a set time. Frame extraction technology is used for frame extraction processing. The extracted static image is saved to a fixed position to become the original image of this process without secondary processing. All processed images are copies of the original images. The following steps are specifically executed.

[0035] S7.1, Creating the Environment;

[0036] S7.2, convert the trained model into ONNX format;

[0037] S7.3, Configure the model startup code according to the ONNX model;

[0038] S7.4, Determine if the model has started normally; if yes, the model deployment is successful; otherwise, proceed to S7.5;

[0039] S7.5, determine if the environment creation does not meet the requirements; if so, return to step S7.1; otherwise, proceed to S7.6;

[0040] S7.6 Determine if there is an error in the model conversion; if so, return to step S7.2; otherwise, return to step S7.3.

[0041] Furthermore, in step S8, the scene is set in a designated room; the scene is equipped with a close-up camera and a panoramic camera;

[0042] During the questioning process, firstly, the close-up camera is used for identification processing. The close-up camera is used to capture the person being questioned during the questioning process. Secondly, when the close-up camera identifies people in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera images extracted at the same time are identified. If the number of people appearing in the panoramic camera at this time is greater than the set number N, the questioning process is determined to be compliant. If the number of people identified is less than the set number N, the questioning process is determined to be non-compliant and defined as a failure.

[0043] Furthermore, in step S8, when someone enters or leaves the room, S frame-by-frame judgments are performed within a specified time according to the actual questioning and conversation requirements; when the number of violations reaches M, an early warning is issued, indicating that the questioning and conversation has violated regulations.

[0044] Furthermore, in step S8, the following steps are specifically performed;

[0045] S8.1 First, read the image and perform detection; then, determine whether the detection results indicate the presence of a set number of people.

[0046] S8.2 If not found, output the result; otherwise, proceed to step S8.3.

[0047] S8.3, Determine if it is higher than the threshold. If yes, output the result; otherwise, proceed to step S8.4.

[0048] S8.4, Delete the data and output the result.

[0049] Furthermore, during the YOLOv7 model training in step 2, firstly, based on the E-ELAN network, Planned re-parameterized convolution, and label allocation method, the label allocation of the lead and aux heads is controlled by the lead head according to the lead guided assigner method. The aux head learns the coarse label, and the lead head learns the fine label.

[0050] Then, YOLOv7 uses a box assignment method, assigning each ground truth label to no more than 3 grids. It then calculates the IoU (Interval of Value) between all assigned ground truth boxes and each predicted box. Next, it filters the first m = min(10, num(predict)) IoU values. Finally, it calculates a dynamic k-value based on the first m IoU values.

[0051] Dynamic _ k = ;

[0052] Since smiOTA considers both classification and regression scores, the cost of smiOTA is...

[0053] cost = Loss cls + 3Losss iou ;

[0054] Loss cls For classification loss, Losss iou For regression loss;

[0055] Based on the cost, obtain the indices of the Dynamic_k predicted labels with the lowest cost; the index refers to the predicted label that matches the true label.

[0056] Finally, when a predicted value matches multiple true values, the one with the lowest cost is selected as the match, and the others are removed.

[0057] A YOLOv7-based query recognition assistance system is used to perform the above-described method; the system includes a separate room in which close-up cameras and panoramic cameras are arranged;

[0058] The close-up camera and panoramic camera are electrically connected to the processor.

[0059] Compared with the prior art, the advantages of the present invention are: the model provided by the present invention can more accurately identify the human body and better distinguish objects that are close to the human body, and can better identify two human bodies with an overlap of less than 70%.

[0060] The specific advantages of this invention are: a. Improved detection speed and accuracy in real-time applications; b. Optimized model structure reduces computational complexity and improves the model's generalization ability; c. Training with an auxiliary head provides more image processing information, which helps in learning useful feature representations; d. Composite model scaling enables the model to adapt to different application requirements; e. The trainable BoF method improves model performance without increasing training costs. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the data annotation process of this invention.

[0062] Figure 2 This is a flowchart of the model training process implemented in this invention.

[0063] Figure 3 This is a flowchart illustrating the model deployment process of this invention.

[0064] Figure 4 This is a flowchart of the image recognition process implemented in this invention.

[0065] Figure 5 This is a business process diagram illustrating the implementation of this invention. Detailed Implementation

[0066] The technical solution adopted in this invention is:

[0067] S1, Perform data annotation processing on the original image to obtain anchor boxes, i.e., annotation information;

[0068] S2, the labeled data is divided into two parts: images and labeling information. The YOLOv7 model is used for training, testing and evaluation data.

[0069] S3, During training, the batch value can be set, and the training accuracy can be determined by the number of training rounds;

[0070] S4. Test the trained model to see if the actual accuracy meets the requirements.

[0071] S5 converts the model that meets the requirements into an ONNX model;

[0072] S6, use the transformed ONNX model for recognition and reasoning;

[0073] S7 allows you to set the video frame extraction interval based on business scenario requirements and perform frame extraction accordingly.

[0074] S8 performs model inference on the extracted frames to confirm the number of people in the current business scenario and determine whether to issue an early warning.

[0075] For image annotation, step S1 also includes a preceding step S1.1, image preprocessing, for preliminary screening;

[0076] First, due to the specific business scenarios used in this invention, image acquisition is required within these defined scenarios. All acquired images are confidential data and cannot be disseminated; all data processing is conducted without an external network. Second, after data collection, images with high similarity or those that are blurry require preliminary filtering to avoid affecting the overall accuracy of the trained model. Third, to ensure the generalization ability of the trained model, the training data needs to encompass all business scenarios, with approximately 100-200 data points of different types for each scenario.

[0077] The core steps of image annotation are as follows: S1.2, after initial screening, the images are manually reviewed to identify factors that affect the accuracy of human recognition. These factors are then incorporated into the annotation process and the image is trained to distinguish it from humans. Factors affecting the accuracy of human recognition include: objects whose weight similarity to a person in the image data exceeds a set threshold, such as chairs, sofas, computers, mice, keyboards, and lights; objects whose color similarity to the clothing of a person in the image data exceeds a set threshold, such as clothes and coat hangers; and objects that reflect the person in the image data, such as televisions and mirrors.

[0078] S1.3, Based on the above, confirm the labels for the annotation process and set annotation rules to prevent different annotation situations from affecting the accuracy of recognition. Annotate the relevant image data according to the annotation rules. The data volume needs to reach approximately 5000 images across different scenarios; naturally, more data will result in higher accuracy. During annotation, first, determine the required label content based on the set requirements; then, annotate according to the annotation rules; second, train the annotated data; third, determine if the annotation content needs adjustment. If so, return to the beginning of the annotation process and re-annotate; otherwise, export the annotated images and annotation files. In computer science, labels are a general concept used to describe the meaning of a piece of code or data.

[0079] For model training, in step S2, the labeled data is classified. In this invention, the image data is divided into training set, test set, and validation set according to an 8:1:1 ratio.

[0080] The training set is used for model training, the test set is used to test the overall accuracy of the model trained on the training set, and the validation set is used to validate the model's performance. The specific steps are as follows:

[0081] S2.1, Input the annotated image and annotation file from step S1.3;

[0082] S2.2, the labeled images and labeled files are divided into training set, test set, and validation set, which are used for training, testing, and evaluation, respectively;

[0083] S2.3, Based on training requirements, set up the YAML file in the model and confirm the data path and training parameter settings; the YAML file is a highly readable format used to express data serialization.

[0084] S2.4, Perform model training;

[0085] S2.5 First, evaluate the trained model; then, determine if the model meets the accuracy requirements; if it does, derive the model; if not, proceed to S2.6.

[0086] S2.6 First, test the model and determine the improvement needs based on the test images; then, determine whether to directly adjust the parameters and train; if so, return to S2.4; otherwise, execute S2.7.

[0087] S2.7, Determine if the annotations need to be adjusted; if so, modify the annotations for the image accordingly; otherwise, increase the image data size and add annotations.

[0088] For video stream frame extraction, in step S7, the camera is connected, and 5 frames are extracted per minute. This invention uses Java for frame extraction technology. The extracted static images are saved to a fixed location as the original images for this process. The original images are not processed again throughout the entire development process. All processed images are copies of the original images to prevent the original images from being modified and affecting subsequent processes. The specific steps are as follows;

[0089] S7.1, Creating the Environment;

[0090] S7.2, convert the training model from pt format to onnx format;

[0091] S7.3, Configure the model startup code according to the ONNX model;

[0092] S7.4, Determine if the model has started normally; if yes, the model deployment is successful; otherwise, proceed to S7.5;

[0093] S7.5, determine if the environment creation does not meet the requirements; if so, return to step S7.1; otherwise, proceed to S7.6;

[0094] S7.6, Determine if there is an error in the model conversion; if so, return to step S7.2; otherwise, return to step S7.3.

[0095] Regarding the early warning processing, in step S8, the main auxiliary scenarios of this invention are interview processes, accounting processes, negotiation processes, or multiple people questioning a single person, equipped with a close-up camera and a panoramic camera. During the questioning and communication, the two cameras need to be processed. First, the close-up camera is processed for identification. The close-up camera is used to capture the person being questioned during the questioning process. Second, when the close-up camera identifies people in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera images extracted at the same time are identified. If the number of people appearing in the panoramic camera is greater than a set number N, for example, N=3, then the questioning and communication process is judged to be compliant. If the number of people identified is less than the set number N, then the questioning and communication process is judged to be in violation and defined as a failure.

[0096] However, given that special circumstances may arise during the entire inquiry and conversation process, such as personnel needing to temporarily leave to sign documents or use the restroom, this invention incorporates further processing. Based on the actual inquiry and conversation requirements, S frame-by-frame judgments are performed within a specified time, for example, S=5. When the number of violations reaches M, an early warning is issued, clearly indicating that the inquiry and conversation has entered a violation state.

[0097] The specific steps are as follows;

[0098] S8.1 First, read the image and perform detection; then, determine whether the detection results show that someone is present.

[0099] S8.2 If not found, output the result; otherwise, proceed to step S8.3.

[0100] S8.3, Determine if it is higher than the threshold. If yes, output the result; otherwise, proceed to step S8.4.

[0101] S8.4, Delete the data and output the result.

[0102] In step 2, when training the YOLOv7 model, firstly, based on the E-ELAN network, Planned re-parameterized convolution, and label allocation method, YOLOv7 proposes the lead-guided assigner method, which controls the label allocation of the lead and aux heads through the lead head, the aux head learns the coarse label, and the lead head learns the fine label.

[0103] Previous auxiliary loss heads typically used a decoupled approach, assigning labels separately to each. However, this made it difficult to align the loss of the aux head with that of the lead head. Therefore, the lead head guided label assigner was proposed. Based on this, a method was proposed whereby the aux head learns coarse labels and the lead head learns fine labels. This allows the aux head to learn more labels and achieve better recall, while the lead head focuses more on accuracy.

[0104] Then, YOLOv7 uses a box assignment method, assigning each ground truth label to no more than 3 grids. The IoU (Interval of Value) is calculated between all assigned ground truth boxes and each predicted box; next, the first m = min(10, num (predict)) IoU values ​​are filtered. Finally, a dynamic k-value is calculated based on the first m IoU values, i.e.

[0105] Dynamic _ k = ;

[0106] Since smiOTA considers both classification and regression scores, the cost of smiOTA is...

[0107] cost = Loss cls + 3Losss iou ;

[0108] Loss cls For classification loss, Losss iou For regression loss;

[0109] Based on the cost, obtain the indices of the Dynamic_k predicted labels with the lowest cost. Here, the index refers to the predicted label that matches the true label.

[0110] Finally, we need to handle the case where the predicted value matches multiple true values. Here, we also select the one with the lowest cost as the match and remove the others.

[0111] Notes on technical terminology:

[0112] Yolov7: The seventh version of the YOLO algorithm, which improves upon the original version, increasing detection speed and accuracy.

[0113] Deep Learning: A subfield of machine learning that is based on the learning of artificial neural networks, particularly layers with multiple nonlinear transformations (i.e., deep neural networks).

[0114] Person identification: Identifying individuals in images or videos using computer vision technology, typically involving techniques such as facial recognition and gait analysis.

[0115] Data preprocessing: A series of processing steps performed on data before model training, including normalization, scaling, and augmentation, to improve model performance.

[0116] Image scaling: Changing the size of an image to fit the requirements of the model input.

[0117] Normalization: Scaling the data proportionally to fit it into a small, specific range, typically [0, 1] or [-1, 1], to speed up training and improve model stability.

[0118] Data augmentation: Artificially increasing the size and diversity of a dataset through methods such as rotation, scaling, and cropping to improve the generalization ability of a model.

[0119] Annotation: Labeling images in a dataset to indicate the category and location of objects contained in the images is a necessary step in training models in supervised learning.

[0120] Model Training: Using labeled datasets to train deep learning models so that they can learn patterns and features in the data.

[0121] Bounding Box: A rectangular box used to represent the location of a target in an image, usually defined by coordinates (x, y) (top left corner) and width and height.

[0122] Class Probability: The confidence level of a model in predicting that a target belongs to a certain class, usually output by the softmax function.

[0123] Generalization ability: The ability of a model to make predictions on unseen data, which is an important metric for measuring model performance.

[0124] Graphical User Interface (GUI): An interface that allows users to interact with electronic devices, typically including elements such as windows, icons, and buttons.

[0125] Data Interface: A mechanism that allows different software or systems to exchange data, such as an API or database connection.

[0126] Frame Extraction: The process of extracting image frames at a specific time point from a video stream.

[0127] The present invention has been described in detail for the purpose of making the disclosure clearer, and the prior art will not be listed in detail.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. It is obvious to those skilled in the art that multiple technical solutions of the present invention can be combined. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All technical contents not described in detail in the present invention are well-known technologies.

Claims

1. A query recognition auxiliary method based on YOLOv7, characterized in that: Includes the following steps; S1, Perform data annotation processing on the original image to obtain anchor boxes, i.e., annotation information; S2, the labeled data is divided into images and label information, and the YOLOv7 model is used for training, testing and evaluation data. S3, during the training of the YOLOv7 model, set the batch value and determine the training accuracy by the number of training epochs; S4. Test the trained YOLOv7 model to see if the actual accuracy meets the requirements. S5, convert the YOLOv7 model that meets the requirements into an ONNX model; S6, use the transformed ONNX model for reasoning and recognition; S7 allows you to set the video frame extraction interval based on business scenario requirements and perform frame extraction accordingly. S8, perform ONNX model inference on the extracted frames to confirm the number of people in the current business scenario and determine whether to issue an early warning. When training the YOLOv7 model in step 2, firstly, based on the E-ELAN network, Planned re-parameterized convolution, and label allocation method, the label allocation of the lead and aux heads is controlled by the lead head according to the lead guided assigner method. The aux head learns the coarse label, and the lead head learns the fine label. Then, YOLOv7 uses a box assignment method, assigning each ground truth label to no more than 3 grids. It then calculates the IoU (Interval of Value) between all assigned ground truth boxes and each predicted box. Next, it filters the first m = min(10, num(predict)) IoU values. Finally, it calculates a dynamic k-value based on the first m IoU values. Dynamic _ k = ; Since smiOTA considers both classification and regression scores, the cost of smiOTA is... cost = Loss cls + 3Losss iou ; Loss cls For classification loss, Losss iou For regression loss; Based on the cost, obtain the indices of the Dynamic_k predicted labels with the lowest cost; the index refers to the predicted label that matches the true label. Finally, when a predicted value matches multiple true values, the one with the lowest cost is selected as the match, and the others are removed.

2. The query recognition assistance method based on YOLOv7 according to claim 1, characterized in that: Step S1 includes the following steps; S1.1, Image preprocessing, preliminary screening; First, image data is collected in the set business scenario. All collected images are confidential data, and all image data processing is carried out in the absence of an external network environment. Then, after the image data is collected, a preliminary screening process is performed to remove image data with similarity greater than a set threshold or ambiguity greater than a set threshold; the training image data needs to include all business scenarios, and each scenario includes different types of image data to meet the set quantity; S1.2 After the initial screening, the image data is manually reviewed a second time to identify factors that may affect the accuracy of personnel identification. These factors are then added to the annotation process and the data is trained. Factors affecting the accuracy of person recognition include: objects whose weight is greater than a set threshold compared to the human body in the image data, objects whose clothing color is greater than a set threshold compared to the human body in the image data, and objects that reflect the human body in the image data. S1.3, Confirm the labels in the annotation process and set the annotation rules. Perform image data annotation work according to the annotation rules. During annotation, first, determine the label content required for this annotation based on the set requirements. Then, annotate according to the annotation rules. Second, train the annotated data. Third, determine whether the annotation content needs to be adjusted. If so, return to the beginning of annotation and re-annotate. Otherwise, export the annotated image and annotation file.

3. The query recognition assistance method based on YOLOv7 according to claim 2, characterized in that: In step S2, the labeled data is classified, and the image data is divided into training set, test set, and validation set according to the set ratio; The training set is used for model training, the test set is used to test the overall accuracy of the model trained on the training set, and the validation set is used to validate the model's performance. The specific steps are as follows; S2.1, Input the annotated image and annotation file from step S1.3; S2.2, the labeled images and labeled files are divided into training set, test set, and validation set, which are used for training, testing, and evaluation, respectively; S2.3, Based on training requirements, set up the YAML file in the model and confirm the data path and training parameter settings; S2.4, Perform model training; S2.5 First, evaluate the trained model; then, determine if the model meets the accuracy requirements; if it does, derive the model; if not, proceed to S2.

6. S2.6 First, test the model and determine the improvement needs based on the test images; then, determine whether to directly adjust the parameters and train; if so, return to S2.4; otherwise, execute S2.

7. S2.7 Determine whether the annotation needs to be adjusted; if so, modify the annotation content of the image accordingly; otherwise, increase the data volume of the image and add annotations.

4. The query recognition assistance method based on YOLOv7 according to claim 3, characterized in that: In step S7, the camera is docked, and a set frame image is extracted within a set time. Frame extraction technology is used to perform frame extraction processing. The extracted static image is saved to a fixed position as the original image of this process and is not processed again. All processed images are copies of the original images. The following steps are executed in detail. S7.1, Creating the Environment; S7.2, convert the trained model into ONNX format; S7.3, Configure the model startup code according to the ONNX model; S7.4, Determine if the model starts normally; If so, the model deployment was successful; Otherwise, execute S7.5; S7.5, Determine if the environment creation does not meet the requirements; if so, return to step S7.1; Otherwise, execute S7.6; S7.6 Determine if there is an error in the model conversion; if so, return to step S7.2; otherwise, return to step S7.

3.

5. The query recognition assistance method based on YOLOv7 according to claim 4, characterized in that: In step S8, the scene is set in a designated room; the scene is equipped with a close-up camera and a panoramic camera; During the questioning process, firstly, the close-up camera is used for identification processing. The close-up camera is used to capture the person being questioned during the questioning process. Secondly, when the close-up camera identifies people in the scene, it is determined that the room has entered the questioning state. At this time, the panoramic camera images extracted at the same time are identified. If the number of people appearing in the panoramic camera is greater than the set number N, the questioning process is considered compliant. If the number of people identified is less than the set number N, the questioning process is considered non-compliant and is defined as a failure.

6. The query recognition assistance method based on YOLOv7 according to claim 5, characterized in that: In step S8, when someone enters or leaves the room, S frame-by-frame checks are performed within a specified time according to the actual questioning and conversation requirements. When the number of violations reaches M, an early warning is issued, indicating that a violation has occurred during the questioning and conversation.

7. The query recognition assistance method based on YOLOv7 according to claim 1, characterized in that: In step S8, the following steps are specifically performed; S8.1 First, read the image and perform detection; then, determine whether the detection results indicate the presence of a set number of people. S8.2, if not, output the result; Otherwise, proceed to step S8.3; S8.3, determine whether it is higher than the threshold, and if so, output the result; Otherwise, proceed to step S8.4; S8.4, Delete the data and output the result.

8. A query recognition assistance system based on YOLOv7, characterized in that: For performing the method according to any one of claims 1-7; the system includes a separate room in which a close-up camera and a panoramic camera are arranged; The close-up camera and panoramic camera are electrically connected to the processor.

Images