A method and system for predicting the severity of harm caused by dangerous human behavior based on video information

Abstract

This invention discloses a method and system for predicting the severity of injury from dangerous human behavior based on video information. The method includes: acquiring video information of dangerous behavior and force-time curves generated by force sensors; using a FuzzyPose model to identify human skeletal points in the video information and extracting kinematic parameters of the skeletal points; selecting the kinematic parameters of the skeletal points as input to the dataset; using the parameters in the force-time curves and the number of penetration layers caused by the dangerous behavior to a protective material sample as dangerous behavior injury parameters, and using the dangerous behavior injury parameters corresponding to the input of the dataset as the output of the dataset; constructing a V2D model, training the V2D model based on the input and output of the dataset; and inputting real-time dangerous behavior into the V2D model for prediction to obtain the predicted severity of injury from dangerous human behavior.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and artificial intelligence, specifically to a method and system for predicting the degree of harm caused by dangerous human behavior based on video information. Background Technology

[0002] With the research and development of protective materials, especially stab-resistant materials, stab-resistant materials such as aramid fiber reinforced plastic (AFRP) have been widely used in various protective equipment, such as stab-resistant vests and protective clothing. These materials provide effective puncture resistance through the layered structure of high-strength fibers, protecting the wearer from sharp objects to a certain extent. However, existing stab-resistant materials mainly rely on their physical properties, and their protective capabilities are designed and evaluated primarily based on standardized experimental tests, such as static puncture tests or dynamic impact tests. This method of evaluating protective capabilities based on static and laboratory conditions cannot fully reflect the performance of stab-resistant materials in actual use, especially in complex dynamic stab attacks.

[0003] Existing research on protective materials lacks in-depth analysis of the relationship between stabbing behavior and protective performance. Especially in real-world scenarios, the perpetrator's movement characteristics, such as stabbing speed and the manner of force application, significantly impact the penetration effectiveness of stab-resistant materials. Current research methods cannot effectively capture and utilize these movement characteristics, resulting in protective material designs that fail to fully consider the complexity and diversity of actual stabbing behavior.

[0004] Furthermore, with the development of artificial intelligence technology, motion capture and analysis technologies based on computer vision and deep learning have gradually matured. This provides new possibilities for real-time analysis and prediction of complex movements and their impact on protective materials. However, currently in the field of protective materials, there is no research or application based on video information and motion capture technology to predict protective performance.

[0005] Therefore, there is an urgent need for a new analytical method and tool that can dynamically predict the protective effect of protective materials in actual use by combining real-world stabbing behavior, thereby providing a scientific basis for the design and optimization of protective equipment. This will open up new directions for the intelligent and personalized design of protective materials and provide technical support for improving the effectiveness and reliability of protective equipment in actual use. Summary of the Invention

[0006] To address the above technical problems, this invention provides a method and system for predicting the degree of harm caused by dangerous human behavior based on video information, the method comprising:

[0007] Step S1: Collect video information of dangerous behavior and force change curves over time generated by force sensors;

[0008] Step S2: Use the FuzzyPose model to identify human skeleton points in the video information and extract the kinematic parameters of the skeleton points;

[0009] Step S3: Select the kinematic parameters of the skeletal points as input to the dataset;

[0010] Step S4: Take the parameters in the force-time curve and the number of penetration layers caused by the dangerous behavior to the protective material sample as the dangerous behavior damage parameters, and take the dangerous behavior damage parameters corresponding to the input of the dataset as the output of the dataset.

[0011] Step S5: Construct a V2D model and train the V2D model based on the input and output of the dataset;

[0012] Step S6: Input the real-time dangerous behavior into the V2D model for prediction to obtain the prediction result of the degree of harm caused by the dangerous behavior.

[0013] Optionally, in step S2, four kinematic parameters are extracted from the skeletal points: four kinematic parameters are selected from the kinematic parameters of the skeletal points, namely: the x-direction displacement change ΔX, the y-direction displacement change ΔY, and the maximum instantaneous velocity V in the x-direction extracted from the kinematic trajectory. x_max The maximum instantaneous velocity V in the y-direction y_max and the maximum instantaneous resultant velocity V max .

[0014] Optionally, in step S2, the FuzzyPose model specifically includes:

[0015] The matrix consists of four layers: a standard convolutional layer with a 3×3 kernel, seven depth-separable convolutional layers connected in sequence, a bottleneck layer with a spread factor of 6 and a 3×3 kernel, and an output layer with 8×8×128 channels.

[0016] The fully connected layer adds two convolutional layers with 1×1 kernels and one fully connected layer to the base layer;

[0017] The output of the fully connected layer is input into a 3×3 deconvolution layer and a 1×1 heatmap convolution layer.

[0018] Optionally, in step S5, the V2D model specifically includes:

[0019] The V2D model consists of one input layer, three hidden layers, and one output layer.

[0020] The input layer consists of the kinematic parameters of key skeletal points;

[0021] The number of neurons in the three hidden layers are 256, 128 and 64, respectively. The activation function used is LeakyR eLU. A Dropout layer is added after each hidden layer to prevent overfitting.

[0022] The output layer has 4 neurons and uses a linear activation function to output continuous values ​​in the regression task.

[0023] This invention also discloses a system for predicting the degree of harm caused by dangerous human behavior based on video information. The system includes: a video acquisition module, a feature extraction module, an input selection module, a harm parameter determination module, a model construction module, and a protection prediction module.

[0024] The video acquisition module is used to acquire video information of dangerous behaviors and force-time curves generated by force sensors.

[0025] The feature extraction module is used to identify human skeleton points in the video information and extract the kinematic parameters of the skeleton points using the FuzzyPose model.

[0026] The input selection module is used to select the kinematic parameters of the skeletal points as input to the dataset;

[0027] The damage parameter determination module is used to take the parameters in the force-time curve and the number of penetration layers caused by the dangerous behavior to the protective material sample as the dangerous behavior damage parameters, and take the dangerous behavior damage parameters corresponding to the input of the dataset as the output of the dataset.

[0028] The model building module is used to build a V2D model and train the V2D model based on the input and output of the dataset.

[0029] The protection prediction module is used to input real-time dangerous behaviors into the V2D model for prediction, and obtain the prediction result of the degree of harm caused by dangerous human behaviors.

[0030] Optionally, in the feature extraction module, four kinematic parameters of the extracted skeletal points are selected: four kinematic parameters are selected from the kinematic parameters of the skeletal points, namely: the change in displacement in the x-direction ΔX, the change in displacement in the y-direction ΔY, and the maximum instantaneous velocity V in the x-direction extracted from the kinematic trajectory. x_max The maximum instantaneous velocity V in the y-direction y_max and the maximum instantaneous resultant velocity V max .

[0031] Optionally, the FuzzyPose model specifically includes:

[0032] The matrix consists of four layers: a standard convolutional layer with a 3×3 kernel, seven depth-separable convolutional layers connected in sequence, a bottleneck layer with a spread factor of 6 and a 3×3 kernel, and an output layer with 8×8×128 channels.

[0033] The fully connected layer adds two convolutional layers with 1×1 kernels and one fully connected layer to the base layer;

[0034] The output of the fully connected layer is input into a 3×3 deconvolution layer and a 1×1 heatmap convolution layer.

[0035] Optionally, the V2D model specifically includes:

[0036] The V2D model consists of one input layer, three hidden layers, and one output layer.

[0037] The input layer consists of the kinematic parameters of key skeletal points;

[0038] The number of neurons in the three hidden layers are 256, 128 and 64, respectively. The activation function used is LeakyR eLU. A Dropout layer is added after each hidden layer to prevent overfitting.

[0039] The output layer has 4 neurons and uses a linear activation function to output continuous values ​​in the regression task.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention collects video information of human stabbing behavior and the curve of the force exerted by the hand on the knife during the stabbing process over time; it uses a FuzzyPose model to identify 33 human skeletal points in the video information and extracts the kinematic parameters of these points; based on feature engineering, it selects the kinematic parameters of 6 key skeletal points as input to the dataset; it uses the parameters obtained from the force-over-time curve and the number of penetration layers caused by the stab wound as puncture injury parameters, and uses the puncture injury parameters corresponding to the input of the dataset as the output of the dataset; it constructs a machine learning regression model (V2D model), trains the V2D model using the input and output of the dataset, and the V2D model enables the prediction of puncture injury parameters based on the kinematic parameters obtained from the video information. This invention enables the prediction of human injury to protective materials based on video information of dangerous behavior, which not only expands the understanding and analysis methods of human stabbing behavior, but also provides technical support for the research of personalized protective equipment.

[0042] This invention enables the prediction of puncture injury parameters based on video information of human stabbing behavior. This invention not only expands the methods for understanding and analyzing stabbing behavior but also provides a scientific basis and technical support for the development of personalized protective equipment and the formulation of protective standards. Attached Figure Description

[0043] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the method for predicting the degree of harm caused by dangerous human behavior based on video information, according to an embodiment of the present invention.

[0045] Figure 2 This is a functional diagram of the V2D model according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the structure for predicting the degree of harm caused by dangerous human behavior based on video information, according to an embodiment of the present invention.

[0047] Figure 4 This is a comparison chart of the prediction results and actual results of the V2D model on the test set data in an embodiment of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example 1

[0050] A method for predicting the severity of harm caused by dangerous human behavior based on video information, such as Figure 1 As shown, the method includes:

[0051] Step S1: Collect video information of dangerous behavior and force change curves generated by force sensors over time.

[0052] In this embodiment, the dangerous behavior is a stabbing act. When collecting video information about the dangerous behavior, a stabbing test was first conducted. The stabbing test followed the principles of the 2013 Declaration of Helsinki and was approved by the local ethics committee. All volunteers read and signed a detailed informed consent form before the test to ensure they understood the purpose, procedure, and potential risks of the experiment. All volunteers were instructed to hold the knife held in the clamp and stab the aramid fiber reinforced composite sample in front of the backing as horizontally as possible. Normal force application habits and movements were maintained during each test. The distance between the knife tip and the sample surface was maintained at 5cm, 10cm, 15cm, 20cm, 25cm, and 30cm during the test. Each volunteer repeated the stabbing test 10 times at each distance.

[0053] A human motion video capture system was designed to collect video information. It uses... RealSense TM The D435 camera is a powerful and versatile camera suitable for various applications. It features a high-resolution RGB webcam, producing high-quality color images. Furthermore, its frame rate reaches 30 frames per second, enabling it to capture changes in human motion. To use the D435 camera, first connect it to your computer via USB 3.1. Then download and install the necessary software. RealSense TM The SDK ensures that all camera functions can be used normally. RealSense TM Viewer software allows you to view and adjust camera settings, including resolution, frame rate, and other parameters.

[0054] During the human stabbing test, a motion capture system fixed to the right side of the person... RealSense TM The D435 camera captured complete video of the puncture procedure. The duration of the complete video capture is 1 second. The video frame rate is 30 FPS, and the resolution is 1920×1080.

[0055] A puncture injury parameter acquisition system was constructed to collect and store pressure signals in real time during stab tests on humans. First, a clamp with a force sensor was designed to hold the dagger. The dagger used conformed to the NIJS Standard-0115.00 standard. The handle was designed as a cylindrical shape, 125mm long and 20mm in diameter, sized for most adult hands, and was fitted with a rubber sleeve for anti-slip during experiments. The dagger clamping part was designed with a 90mm long, 3mm thick connecting groove and two threaded positioning holes to ensure consistent clamping position and exposed length of the dagger in each test. The handle and dagger clamping part were CNC machined from aluminum alloy-7075 with a minimum surface roughness of Ra3.2. A DYMH-103 diaphragm sensor was threadedly fixed between the handle and clamping part. This sensor has a range of 200N and features high accuracy, good stability, and a high dynamic response frequency. The force sensor signal was transmitted in real time to a handheld force gauge via a data cable. This force gauge can acquire and transmit the force sensor signal to a computer in real time for display and storage. In addition, the backing used for the puncture test consists of four layers of 5.8mm thick neoprene foam, one layer of 31mm thick polyethylene closed-cell foam, and two layers of 6.4mm thick natural rubber. The aramid fiber reinforced composite sample is secured to the backing with two 10cm wide elastic straps with Velcro fasteners.

[0056] Ten volunteers (six men and four women, aged 22-30) were invited to participate in the live biopsy test; their details are shown in Table 1. The test strictly followed the principles of the 2013 Declaration of Helsinki and was approved by the local ethics committee. After approval was granted, all volunteers read and signed a detailed informed consent form to ensure they understood the purpose, procedures, and potential risks of the experiment.

[0057] Table 1

[0058] 1 male 29 180 78 2 male 25 178 70 3 male 24 178 70 4 male 24 175 70 5 male 23 180 75 6 male 23 175 77 7 female 26 158 56 8 female 25 160 47 9 female 24 172 60 10 female 23 165 68

[0059] To ensure the reliability of the test results, all volunteers were instructed to hold the blade held in the clamp and pierce the aramid fiber reinforced composite sample in front of the backing as horizontally as possible. They maintained their normal force application habits and movements during each test. The distance between the blade tip and the sample surface was kept at 5cm, 10cm, 15cm, 20cm, 25cm, and 30cm. Each volunteer repeated the puncture test 10 times at each distance. Therefore, 60 sets of data were collected from each volunteer, for a total of 600 sets of data to meet the needs of subsequent training of the machine learning model. During each test, the membrane sensor in the puncture injury parameter acquisition module collected the impact force as a function of time, and the experimenters recorded the number of sample layers penetrated. The human motion video acquisition module was fixed on the right side... RealSense TM The D435 camera captured complete video of the puncture procedure. The duration of the complete video capture is 1 second. The video frame rate is 30 FPS, and the resolution is 1920×1080.

[0060] Step S2: Use the FuzzyPose model to identify human skeleton points in the video information and extract the kinematic parameters of the skeleton points. The FuzzyPose model includes: a matrix, a key point localization module, and a heatmap generation module; the matrix includes: a standard convolutional layer with a "3×3" kernel as the first layer, seven sequentially connected depthwise separable convolutional layers in the middle, a bottleneck layer with a channel expansion factor of 6 and a kernel size of "3×3" as the next layer, and finally an 8×8×128 output layer; the key point localization module includes: two convolutional layers with a "1×1" kernel and one fully connected layer added on the basis of the matrix; the heatmap generation module includes: one deconvolutional layer with a "3×3" kernel and one heatmap convolutional layer with a "1×1" kernel. The human skeletal points identified and the extracted kinematic parameters in this embodiment are as follows: The 33 identified human skeletal points include: nose, left inner corner of eye, left eye, left outer corner of eye, right inner corner of eye, right eye, right outer corner of eye, left ear, right ear, left side of mouth, right side of mouth, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left thumb, right thumb, left index finger, right index finger, left little finger, right little finger, left hip, right hip, left knee, right knee, left ankle, right ankle, left heel, right heel, left toe, and right toe; The 4 extracted kinematic parameters include: the change in displacement in the x-direction ΔX, the change in displacement in the y-direction ΔY, and the maximum instantaneous velocity V in the x-direction extracted from the kinematic trajectory. x_max The maximum instantaneous velocity V in the y-direction y_max and the maximum instantaneous resultant velocity V maxBased on feature engineering, the kinematic parameters of key skeletal points were selected as input to the dataset, including: the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V max These are the kinematic parameters of key skeletal points.

[0061] FuzzyPose uses a lightweight convolutional neural network, MobileNetV2, as its backbone for extracting image features. MobileNetV2 employs depthwise separable convolution and inverse residual blocks, enabling it to provide high feature extraction capabilities while maintaining computational efficiency. Its main parameters are shown in Table 2. In Table 2, t represents the expansion factor in the inverse residual block. In each inverse residual block, the number of input channels is multiplied by the expansion factor before depthwise separable convolution is performed. This expansion factor improves feature representation while maintaining computational efficiency. n represents the number of times the network layer is repeated for a specific configuration. FuzzyPose's keypoint localization module predicts the precise locations of human keypoints from the feature maps extracted from the backbone, outputting the coordinates and confidence score of each keypoint. The confidence score indicates the reliability of the model's prediction of the skeleton point's location, typically represented by a value between 0 and 1. The heatmap generation module is responsible for representing the probability distribution of each keypoint in the image. The probability distribution is used to evaluate whether each keypoint was predicted.

[0062] Table 2

[0063] <![CDATA[224 2 ×3]]> conv2d - 32 1 2 <![CDATA[112 2 ×32]]> bottleneck 1 16 1 1 <![CDATA[112 2 ×16]]> bottleneck 6 24 2 2 <![CDATA[56 2 ×24]]> bottleneck 6 32 3 2 <![CDATA[28 2 ×32]]> bottleneck 6 64 4 2 <![CDATA[14 2 ×64]]> bottleneck 6 96 3 1 <![CDATA[14 2 ×96]]> bottleneck 6 160 3 2 <![CDATA[7 2 ×160]]> bottleneck 6 320 1 1 <![CDATA[7 2 ×320]]> conv2d1×1 - 1280 1 1 <![CDATA[7 2 ×1280]]> Avgpool7×7 - - 1 - 1×1×1280 conv2d1×1 - k -

[0064] FuzzyPose detected 33 key skeletal points covering the major joints and feature points of the human body, including the head, torso, limbs, and hands. The positional information of the major joints is crucial for pose estimation and motion recognition. The additional detection of key points on the hands and feet adds further detail, allowing the model to be applied to more complex scenarios. Therefore, this model was applied to motion capture and motion analysis of a knife-stabbing test in this study.

[0065] The FuzzyPose model was used to capture skeletal points in the human stabbing test process. The original FuzzyPose model exhibited significant bias in skeletal point recognition. Furthermore, it showed marked errors in predicting incompletely visible skeletal points based on visible ones. This indicates that the pre-trained FuzzyPose model is unsuitable for directly applying to the human stabbing action dataset constructed in this study. Therefore, transfer learning was employed to train the FuzzyPose model. Based on the pre-trained FuzzyPose model, fine-tuning was performed using human stabbing action data. The network structure, except for the three layers after the keypoint localization module, was frozen. The learning rate was adjusted to 0.001, and the FuzzyPose model was retrained on the human stabbing action dataset using the Adam optimizer. The model correctly predicted human movements and achieved relatively accurate skeletal point recognition. Therefore, the FuzzyPose model after transfer learning is suitable for motion capture in human stabbing actions.

[0066] Step S3: Select the kinematic parameters of the skeletal points as input to the dataset. The process of selecting the kinematic parameters of key skeletal points is as follows:

[0067] The coordinates of skeletal points and their confidence scores for each frame detected by the FuzzyPose model were recorded. Firstly, since the FuzzyPose model assumes the human head is always visible during training, the confidence scores of facial skeletal points are close to 1, rendering them unreliable. Secondly, because the camera primarily captures the upper body and right-hand piercing action, information on skeletal points on the left and lower body is incomplete. Therefore, key skeletal points on the right side of the body were selected, including the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb.

[0068] Analysis of the movement trajectory reveals that even when volunteers were instructed to perform a horizontal puncture during the test, the puncture blade still exhibits a certain degree of vertical movement during the puncture procedure. Therefore, it can be observed that during the abrupt change in the x-axis coordinate, the y-axis coordinate also changes to some extent. Correspondingly, V... x When there are significant changes, V y It will also change. Therefore, V is related to V in most cases. x Maintain consistency, but also with V x There are instances of deviation. Therefore, to comprehensively consider various puncture scenarios, the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the six key bone points during the puncture action are considered. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V maxThe extracted data served as kinematic parameters for human puncture behavior. Furthermore, the extraction of kinematic parameters at different test distances revealed that the puncture test distance does indeed affect human motor behavior. The puncture speed gradually increases with distance, and the puncture force on the AFRP also changes. Therefore, kinematic data for puncture tests at six different distances ranging from 5cm to 30cm were acquired.

[0069] The saved video was input into the FuzzyPose model after transfer learning for detection. The coordinates of the skeletal points and their confidence scores for each detected frame were recorded. Firstly, the BlazePose model assumed during training that the human head was always visible; therefore, the confidence scores of facial skeletal points were close to 1 and thus lost their reference value. The camera primarily captured the upper body of the person performing a right-hand piercing action; information on the skeletal points on the left side and lower body was incomplete. Therefore, the focus was primarily on key skeletal points on the right side of the body, including the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb.

[0070] The kinematic trajectories of the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb during puncture were reconstructed, and kinematic parameters were extracted. Comprehensive analysis of the motion trajectories revealed that even though volunteers were instructed to perform horizontal punctures during the test, the puncture blade still exhibited a certain degree of vertical movement during the puncture. Therefore, in the coordinate changes, a change in the y-axis coordinate was observed during the abrupt change in the x-axis coordinate. Correspondingly, V... x When there are significant changes, V y It will also change. Therefore, V is related to V in most cases. x Maintain consistency, but also with V x There are instances of deviation. Therefore, to comprehensively consider various puncture scenarios, the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the six key bone points during the puncture action are considered. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V max The data were extracted as kinematic parameters in human puncture behavior. In addition, the puncture speed gradually increases with distance, and the puncture force on the AFRP also changes. Therefore, kinematic data from puncture tests at six different distances from 5cm to 30cm were acquired.

[0071] Step S4: The parameters in the force-time curve and the number of penetration layers caused by the hazardous behavior to the protective material sample are used as hazardous behavior injury parameters. The hazardous behavior injury parameters corresponding to the input of the dataset are used as the output of the dataset. The parameters obtained from the force-time curve and the number of penetration layers caused by the stab wound are used as puncture injury parameters. The puncture injury parameters corresponding to the input of the dataset are used as the output of the dataset.

[0072] The puncture injury parameters include: the maximum force F obtained by processing the curve of the force exerted by the hand on the knife over time. max The maximum rate of change of force F' max The integral of the force J, and the number of layers N penetrated by the sample recorded after the puncture test. In this embodiment, the selection process for the puncture injury parameters is as follows:

[0073] The impact force data collected by the diaphragm sensor during each puncture test was processed. After fitting the original Ft curve, three key features were extracted: the maximum impact force, the maximum rate of change of impact force, and the integral of the force application process (impulse). The maximum impact force F... max This represents the maximum force applied to the cutting tool during a human puncture, and also reflects the maximum impact load experienced by the sample during puncture resistance. The maximum value of the rate of change of impact force, F' max The impulse J reflects the maximum velocity the scalpel can apply during a puncture, helping to evaluate the protective performance of a sample at different puncture velocities. The force integral, or impulse J, reflects the cumulative force throughout the puncture process, representing the total energy transferred to the sample. A larger impulse indicates greater energy the sample must withstand. Therefore, impulse is related to the energy absorption capacity of the protective material and can assess its protective effect under long-term use and repeated impacts. In addition, the number of penetration layers recorded after each puncture test is also used as a puncture injury parameter. The number of penetration layers is a quantitative indicator that directly measures the puncture resistance of a sample and is crucial for evaluating the effectiveness of the material in practical applications.

[0074] The impact force data collected by the diaphragm sensor during each puncture test was processed. After fitting the original curve, three key features were extracted: the maximum impact force, the maximum rate of change of impact force, and the integral of the force application process (impulse). The maximum impact force F... max This represents the maximum force applied to the cutting tool during a human puncture, and also reflects the maximum impact load experienced by the sample during puncture resistance. The maximum value of the rate of change of impact force, F' maxThe impulse J reflects the maximum velocity the scalpel can apply during a puncture, helping to evaluate the protective performance of a sample at different puncture velocities. The force integral, or impulse J, reflects the cumulative force throughout the puncture process, representing the total energy transferred to the sample. A larger impulse indicates greater energy the sample must withstand. Therefore, impulse is related to the energy absorption capacity of the protective material and can assess its protective effect under long-term use and repeated impacts. In addition, the number of penetration layers recorded after each puncture test is also used as a puncture injury parameter. The number of penetration layers is a quantitative indicator that directly measures the puncture resistance of a sample and is crucial for evaluating the effectiveness of the material in practical applications.

[0075] Step S5: Construct a V2D model and train it based on the input and output of the dataset. Step S6: Input real-time dangerous behaviors into the V2D model for prediction to obtain a prediction result of the degree of harm caused by the dangerous behavior.

[0076] A V2D model is constructed and trained using the input and output of the dataset. The V2D model enables the prediction of puncture injury parameters based on kinematic parameters obtained from video information.

[0077] The V2D model consists of: one input layer, three hidden layers, and one output layer;

[0078] The input layer consists of the kinematic parameters of key skeleton points;

[0079] The number of neurons in the three hidden layers are 256, 128, and 64, respectively. The activation function used is LeakyReLU. A Dropout layer is added after each hidden layer to prevent the model from overfitting.

[0080] The output layer has 4 neurons, consistent with the number of puncture damage parameters. The output layer uses a linear activation function to ensure continuous output values ​​in the regression task.

[0081] The functions of V2D models include Figure 2 As shown.

[0082] Specifically, in this embodiment, the training details of the V2D model are as follows:

[0083] The training cycle of the model is set to 2000 to ensure that the model can fully learn the relationship between the kinematic parameters of the input key skeletal points and the output puncture injury parameters.

[0084] The dataset was divided into an 80% training set and a 20% test set.

[0085] The V2D model uses a custom loss function, the formula of which is:

[0086]

[0087] Where L is the loss value, N is the number of samples, i is the sample index, and ω i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. i Let δ be the true value of the i-th sample, and δ be an adjustable threshold parameter.

[0088] In addition, R 2 To evaluate the predictive performance of the model.

[0089] After determining the kinematic parameters of key skeletal points and corresponding puncture injury parameters in a stabbing incident, a V2D model for puncture injury prediction, from "video" to "injury," will be established. The self-developed V2D model has three hidden layers. Each hidden layer introduces non-linear computation through the LeakyReLU activation function, and a Dropout layer is used to prevent overfitting. The overall model configuration is determined by balancing dataset complexity and computational efficiency. The V2D model is trained for 2000 epochs. The kinematic parameters of key skeletal points extracted from the stabbing incident video information are used as input, and the puncture injury parameters are used as output. The dataset is divided into 80% for training and 20% for testing. The V2D model achieves F... max , F' max The predictive performance of the four puncture injury parameters J, N, and N is shown in Table 3. Figure 4 This section compares the predictions of the V2D model with the actual results on the test set. It can be observed that most of the deviations from the actual values ​​are around the center line, indicating that the V2D model can effectively predict the actual puncture injury parameters.

[0090] Table 3

[0091]

[0092] Example 2

[0093] A system for predicting the severity of harm caused by dangerous human behavior based on video information, such as Figure 3 As shown, the system includes: a video acquisition module, a feature extraction module, an input selection module, a damage parameter determination module, a model building module, and a protection prediction module;

[0094] The video acquisition module is used to acquire video information of dangerous behaviors and the force exerted by the hazardous material on the protective material sample during the dangerous behavior as a function of time.

[0095] In this embodiment, the dangerous behavior is a stabbing act. When collecting video information about the dangerous behavior, a stabbing test was first conducted. The stabbing test followed the principles of the 2013 Declaration of Helsinki and was approved by the local ethics committee. All volunteers read and signed a detailed informed consent form before the test to ensure they understood the purpose, procedure, and potential risks of the experiment. All volunteers were instructed to hold the knife held in the clamp and stab the aramid fiber reinforced composite sample in front of the backing as horizontally as possible. Normal force application habits and movements were maintained during each test. The distance between the knife tip and the sample surface was maintained at 5cm, 10cm, 15cm, 20cm, 25cm, and 30cm during the test. Each volunteer repeated the stabbing test 10 times at each distance.

[0096] A human motion video capture system was designed to collect video information. It uses... RealSense TM The D435 camera is a powerful and versatile camera suitable for various applications. It features a high-resolution RGB webcam, producing high-quality color images. Furthermore, its frame rate reaches 30 frames per second, enabling it to capture changes in human motion. To use the D435 camera, first connect it to your computer via USB 3.1. Then download and install the necessary software. RealSense TM The SDK ensures that all camera functions can be used normally. RealSense TM Viewer software allows you to view and adjust camera settings, including resolution, frame rate, and other parameters.

[0097] During the human stabbing test, a motion capture system fixed to the right side of the person... RealSense TM The D435 camera captured complete video of the puncture procedure. The duration of the complete video capture is 1 second. The video frame rate is 30 FPS, and the resolution is 1920×1080.

[0098] A puncture injury parameter acquisition system was constructed to collect and store pressure signals in real time during stab tests on humans. First, a clamp with a force sensor was designed to hold the dagger. The dagger used conformed to the NIJS Standard-0115.00 standard. The handle was designed as a cylindrical shape, 125mm long and 20mm in diameter, sized for most adult hands, and was fitted with a rubber sleeve for anti-slip during experiments. The dagger clamping part was designed with a 90mm long, 3mm thick connecting groove and two threaded positioning holes to ensure consistent clamping position and exposed length of the dagger in each test. The handle and dagger clamping part were CNC machined from aluminum alloy-7075 with a minimum surface roughness of Ra3.2. A DYMH-103 diaphragm sensor was threadedly fixed between the handle and clamping part. This sensor has a range of 200N and features high accuracy, good stability, and a high dynamic response frequency. The force sensor signal was transmitted in real time to a handheld force gauge via a data cable. This force gauge can acquire and transmit the force sensor signal to a computer in real time for display and storage. In addition, the backing used for the puncture test consists of four layers of 5.8mm thick neoprene foam, one layer of 31mm thick polyethylene closed-cell foam, and two layers of 6.4mm thick natural rubber. The aramid fiber reinforced composite sample is secured to the backing with two 10cm wide elastic straps with Velcro fasteners.

[0099] Ten volunteers (six men and four women, aged 22-30) were invited to participate in a real biopsy. The test strictly followed the principles of the 2013 Declaration of Helsinki and was approved by the local ethics committee. After approval was granted, all volunteers read and signed a detailed informed consent form to ensure they understood the purpose, procedures, and potential risks of the experiment.

[0100] To ensure the reliability of the test results, all volunteers were instructed to hold the blade held in the clamp and pierce the aramid fiber reinforced composite sample in front of the backing as horizontally as possible. They maintained their normal force application habits and movements during each test. The distance between the blade tip and the sample surface was kept at 5cm, 10cm, 15cm, 20cm, 25cm, and 30cm. Each volunteer repeated the puncture test 10 times at each distance. Therefore, 60 sets of data were collected from each volunteer, for a total of 600 sets of data to meet the needs of subsequent training of the machine learning model. During each test, the membrane sensor in the puncture injury parameter acquisition module collected the impact force as a function of time, and the experimenters recorded the number of sample layers penetrated. The human motion video acquisition module was fixed on the right side... RealSense TMThe D435 camera captured complete video of the puncture procedure. The duration of the complete video capture is 1 second. The video frame rate is 30 FPS, and the resolution is 1920×1080.

[0101] The feature extraction module is used to identify human skeletal points in the video information using the FuzzyPose model and extract the kinematic parameters of the skeletal points. The input selection module is used to select the kinematic parameters of the skeletal points as input to the dataset.

[0102] The FuzzyPose model includes: a matrix, a keypoint localization module, and a heatmap generation module. The matrix consists of: a standard convolutional layer with a 3×3 kernel as the first layer, seven sequentially connected depthwise separable convolutional layers in the middle, a bottleneck layer with a channel expansion factor of 6 and a kernel size of 3×3, and finally an 8×8×128 output layer. The keypoint localization module adds two convolutional layers with a 1×1 kernel and one fully connected layer to the matrix. The heatmap generation module includes one deconvolutional layer with a 3×3 kernel and one heatmap convolutional layer with a 1×1 kernel. The human skeletal points identified and the extracted kinematic parameters in this embodiment are as follows: The 33 identified human skeletal points include: nose, left inner corner of eye, left eye, left outer corner of eye, right inner corner of eye, right eye, right outer corner of eye, left ear, right ear, left side of mouth, right side of mouth, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left thumb, right thumb, left index finger, right index finger, left little finger, right little finger, left hip, right hip, left knee, right knee, left ankle, right ankle, left heel, right heel, left toe, and right toe; The 4 extracted kinematic parameters include: the change in displacement in the x-direction ΔX, the change in displacement in the y-direction ΔY, and the maximum instantaneous velocity V in the x-direction extracted from the kinematic trajectory. x_max The maximum instantaneous velocity V in the y-direction y_max and the maximum instantaneous resultant velocity V max Based on feature engineering, the kinematic parameters of key skeletal points were selected as input to the dataset, including: the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V max These are the kinematic parameters of key skeletal points.

[0103] FuzzyPose uses a lightweight convolutional neural network, MobileNetV2, as its backbone for extracting image features. MobileNetV2 employs depthwise separable convolution and inverse residual blocks, enabling it to provide high feature extraction capabilities while maintaining computational efficiency. Its main parameters are shown in Table 2. In Table 4, t represents the expansion factor in the inverse residual block. In each inverse residual block, the number of input channels is multiplied by the expansion factor before depthwise separable convolution. This expansion factor improves feature representation while maintaining computational efficiency. n represents the number of times the network layer is repeated for a specific configuration. FuzzyPose's keypoint localization module predicts the precise locations of human keypoints from the feature maps extracted from the backbone, outputting the coordinates and confidence score of each keypoint. The confidence score indicates the reliability of the model's prediction of the bone point's location, typically represented by a value between 0 and 1. The heatmap generation module represents the probability distribution of each keypoint in the image. The probability distribution is used to evaluate whether each keypoint was predicted.

[0104] Table 4

[0105] <![CDATA[224 2 ×3]]> conv2d - 32 1 2 <![CDATA[112 2 ×32]]> bottleneck 1 16 1 1 <![CDATA[112 2 ×16]]> bottleneck 6 24 2 2 <![CDATA[56 2 ×24]]> bottleneck 6 32 3 2 <![CDATA[28 2 ×32]]> bottleneck 6 64 4 2 <![CDATA[14 2 ×64]]> bottleneck 6 96 3 1 <![CDATA[14 2 ×96]]> bottleneck 6 160 3 2 <![CDATA[7 2 ×160]]> bottleneck 6 320 1 1 <![CDATA[7 2 ×320]]> conv2d1×1 - 1280 1 1 <![CDATA[7 2 ×1280]]> Avgpool7×7 - - 1 - 1×1×1280 conv2d1×1 - k -

[0106] FuzzyPose detected 33 key skeletal points covering the major joints and feature points of the human body, including the head, torso, limbs, and hands. The positional information of the major joints is crucial for pose estimation and motion recognition. The additional detection of key points on the hands and feet adds further detail, allowing the model to be applied to more complex scenarios. Therefore, this model was applied to motion capture and motion analysis of a knife-stabbing test in this study.

[0107] The FuzzyPose model was used to capture skeletal points in the human stabbing test process. The pre-trained FuzzyPose model exhibited significant bias in skeletal point recognition. Furthermore, it showed marked errors in predicting incompletely visible skeletal points based on visible ones. This indicates that the pre-trained FuzzyPose model is unsuitable for directly applying to the human stabbing action dataset constructed in this study. Therefore, transfer learning was employed to train the FuzzyPose model. Based on the pre-trained FuzzyPose model, fine-tuning was performed using human stabbing action data. The network structure, except for the three layers after the keypoint localization module, was frozen. The learning rate was adjusted to 0.001, and the FuzzyPose model was retrained on the human stabbing action dataset using the Adam optimizer. The model correctly predicted human movements and achieved relatively accurate skeletal point recognition. Therefore, the FuzzyPose model after transfer learning is suitable for capturing motion in human stabbing actions.

[0108] The damage parameter determination module is used to take the parameters in the force-time curve and the number of penetration layers caused by the dangerous behavior to the protective material sample as the dangerous behavior damage parameters, and take the dangerous behavior damage parameters corresponding to the input of the dataset as the output of the dataset.

[0109] The coordinates of skeletal points and their confidence scores for each frame detected by the FuzzyPose model were recorded. Firstly, since the FuzzyPose model assumes the human head is always visible during training, the confidence scores of facial skeletal points are close to 1, rendering them unreliable. Secondly, because the camera primarily captures the upper body and right-hand piercing action, information on skeletal points on the left and lower body is incomplete. Therefore, key skeletal points on the right side of the body were selected, including the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb.

[0110] Analysis of the movement trajectory reveals that even when volunteers were instructed to perform a horizontal puncture during the test, the puncture blade still exhibits a certain degree of vertical movement during the puncture procedure. Therefore, it can be observed that during the abrupt change in the x-axis coordinate, the y-axis coordinate also changes to some extent. Correspondingly, V... x When there are significant changes, V y It will also change. Therefore, V is related to V in most cases. x Maintain consistency, but also with V x There are instances of deviation. Therefore, to comprehensively consider various puncture scenarios, the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the six key bone points during the puncture action are considered. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V max The extracted data served as kinematic parameters for human puncture behavior. Furthermore, the extraction of kinematic parameters at different test distances revealed that the puncture test distance does indeed affect human motor behavior. The puncture speed gradually increases with distance, and the puncture force on the AFRP also changes. Therefore, kinematic data for puncture tests at six different distances ranging from 5cm to 30cm were acquired.

[0111] The saved video was input into the FuzzyPose model after transfer learning for detection. The coordinates of the skeletal points and their confidence scores for each detected frame were recorded. Firstly, the BlazePose model assumed during training that the human head was always visible; therefore, the confidence scores of facial skeletal points were close to 1 and thus lost their reference value. The camera primarily captured the upper body of the person performing a right-hand piercing action; information on the skeletal points on the left side and lower body was incomplete. Therefore, the focus was primarily on key skeletal points on the right side of the body, including the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb.

[0112] The kinematic trajectories of the right shoulder, right elbow, right wrist, right little finger, right index finger, and right thumb during puncture were reconstructed, and kinematic parameters were extracted. Comprehensive analysis of the motion trajectories revealed that even though volunteers were instructed to perform horizontal punctures during the test, the puncture blade still exhibited a certain degree of vertical movement during the puncture. Therefore, in the coordinate changes, a change in the y-axis coordinate was observed during the abrupt change in the x-axis coordinate. Correspondingly, V... x When there are significant changes, V y It will also change. Therefore, V is related to V in most cases. x Maintain consistency, but also with V x There are instances of deviation. Therefore, to comprehensively consider various puncture scenarios, the x-direction displacement change ΔX, y-direction displacement change ΔY, and maximum x-direction velocity V of the six key bone points during the puncture action are considered. x_max Maximum velocity V in the y direction y_max and the maximum resultant velocity V max The data were extracted as kinematic parameters in human puncture behavior. In addition, the puncture speed gradually increases with distance, and the puncture force on the AFRP also changes. Therefore, kinematic data from puncture tests at six different distances from 5cm to 30cm were acquired.

[0113] The model building module is used to build a V2D model and train the V2D model based on the input and output of the dataset.

[0114] The parameters obtained from the force-time curve and the number of penetration layers caused by the stab wound are used as puncture damage parameters. The puncture damage parameters corresponding to the input of the dataset are used as the output of the dataset.

[0115] The puncture injury parameters include: the maximum force F obtained by processing the curve of the force exerted by the hand on the knife over time. max The maximum rate of change of force F' maxThe integral of the force J, and the number of layers N penetrated by the sample recorded after the puncture test. In this embodiment, the selection process for the puncture injury parameters is as follows:

[0116] The impact force data collected by the diaphragm sensor during each puncture test was processed. After fitting the original Ft curve, three key features were extracted: the maximum impact force, the maximum rate of change of impact force, and the integral of the force application process (impulse). The maximum impact force F... max This represents the maximum force applied to the cutting tool during a human puncture, and also reflects the maximum impact load experienced by the sample during puncture resistance. The maximum value of the rate of change of impact force, F' max The impulse J reflects the maximum velocity the scalpel can apply during a puncture, helping to evaluate the protective performance of a sample at different puncture velocities. The force integral, or impulse J, reflects the cumulative force throughout the puncture process, representing the total energy transferred to the sample. A larger impulse indicates greater energy the sample must withstand. Therefore, impulse is related to the energy absorption capacity of the protective material and can assess its protective effect under long-term use and repeated impacts. In addition, the number of penetration layers recorded after each puncture test is also used as a puncture injury parameter. The number of penetration layers is a quantitative indicator that directly measures the puncture resistance of a sample and is crucial for evaluating the effectiveness of the material in practical applications.

[0117] The impact force data collected by the diaphragm sensor during each puncture test was processed. After fitting the original curve, three key features were extracted: the maximum impact force, the maximum rate of change of impact force, and the integral of the force application process (impulse). The maximum impact force F... max This represents the maximum force applied to the cutting tool during a human puncture, and also reflects the maximum impact load experienced by the sample during puncture resistance. The maximum value of the rate of change of impact force, F' max The impulse J reflects the maximum velocity the scalpel can apply during a puncture, helping to evaluate the protective performance of a sample at different puncture velocities. The force integral, or impulse J, reflects the cumulative force throughout the puncture process, representing the total energy transferred to the sample. A larger impulse indicates greater energy the sample must withstand. Therefore, impulse is related to the energy absorption capacity of the protective material and can assess its protective effect under long-term use and repeated impacts. In addition, the number of penetration layers recorded after each puncture test is also used as a puncture injury parameter. The number of penetration layers is a quantitative indicator that directly measures the puncture resistance of a sample and is crucial for evaluating the effectiveness of the material in practical applications.

[0118] The protection prediction module is used to input real-time dangerous behaviors into the V2D model for prediction, and obtain the prediction result of the degree of harm caused by dangerous human behaviors.

[0119] A V2D model is constructed and trained using the input and output of the dataset. The V2D model enables the prediction of puncture injury parameters based on kinematic parameters obtained from video information.

[0120] The V2D model consists of: one input layer, three hidden layers, and one output layer;

[0121] The input layer consists of the kinematic parameters of key skeleton points;

[0122] The number of neurons in the three hidden layers are 256, 128, and 64, respectively. The activation function used is LeakyReLU. A Dropout layer is added after each hidden layer to prevent the model from overfitting.

[0123] The output layer has 4 neurons, consistent with the number of puncture damage parameters. The output layer uses a linear activation function to ensure continuous output values ​​in the regression task.

[0124] The functions of V2D models include Figure 2 As shown.

[0125] Specifically, in this embodiment, the training details of the V2D model are as follows:

[0126] The training cycle of the model is set to 2000 to ensure that the model can fully learn the relationship between the kinematic parameters of the input key skeletal points and the output puncture injury parameters.

[0127] The dataset was divided into an 80% training set and a 20% test set.

[0128] The V2D model uses a custom loss function, the formula of which is:

[0129]

[0130] Where L is the loss value, N is the number of samples, i is the sample index, and ω i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. i Let δ be the true value of the i-th sample, and δ be an adjustable threshold parameter.

[0131] In addition, R 2 To evaluate the predictive performance of the model.

[0132] After determining the kinematic parameters of key skeletal points and corresponding puncture injury parameters in a stabbing incident, a V2D model for puncture injury prediction, from "video" to "damage," will be established. The self-built V2D model has three hidden layers. Each hidden layer introduces non-linear computation through the LeakyReLU activation function, and a Dropout layer is used to prevent overfitting. The overall model configuration is determined by balancing dataset complexity and computational efficiency. The V2D model has 2000 training cycles. The kinematic parameters of key skeletal points extracted from the stabbing incident video information are used as input, and the puncture injury parameters are used as output. The dataset is divided into 80% for training and 20% for testing. Figure 4 This section compares the predictions of the V2D model with the actual results on the test set. It can be observed that most of the deviations from the actual values ​​are around the center line, indicating that the V2D model can effectively predict the actual puncture injury parameters.

[0133] Example 3

[0134] This application also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method or module.

[0135] Example 4

[0136] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described method.

[0137] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the degree of harm caused by dangerous human behavior based on video information, characterized in that, The method includes: Step S1: Collect video information of dangerous behavior and force change curves over time generated by force sensors; Step S2: Use the FuzzyPose model to identify human skeleton points in the video information and extract the kinematic parameters of the skeleton points; Step S3: Select the kinematic parameters of the skeletal points as input to the dataset; Step S4: Use the parameters in the force-time curve and the number of penetration layers caused by the hazardous behavior to the protective material sample as hazardous behavior injury parameters, and use the hazardous behavior injury parameters corresponding to the input of the dataset as the output of the dataset; the hazardous behavior injury parameters include: the maximum value of the force obtained from the force-time curve of the hand holding the knife. Maximum rate of change of force Integral of force and the number of layers penetrated by the sample after the stab test. ; Step S5: Construct a V2D model and train the V2D model based on the input and output of the dataset; Step S6: Input the real-time dangerous behavior into the V2D model for prediction to obtain the prediction result of the degree of harm caused by the dangerous behavior; In step S2, the FuzzyPose model specifically includes: The matrix consists of four layers, namely: a convolution kernel of 3... A standard convolutional layer of 3, 7 depthwise separable convolutional layers connected sequentially, with a spread factor of 6 and a kernel size of 3. The bottleneck layer is 3 and the output layer has 8×8×128 channels; The fully connected layer adds two convolutional kernels to the base layer, making it 1. One convolutional layer and one fully connected layer; The output of the fully connected layer is input into a 3-layer deconvolution kernel. A 3-layer deconvolutional layer and a 1-layer convolutional kernel are 1 1 heatmap convolutional layer; The saved video was input into the FuzzyPose model after transfer learning for detection. The coordinates of the skeletal points and their confidence scores for each detected frame were recorded. The kinematic trajectories of the right shoulder, right elbow, right wrist, right little finger, right index finger and right thumb during the puncture process were reconstructed and the kinematic parameters were extracted. In step S5, the V2D model specifically includes: The V2D model consists of one input layer, three hidden layers, and one output layer. The input layer consists of the kinematic parameters of key skeletal points; The number of neurons in the three hidden layers are 256, 128 and 64 respectively. The activation function is LeakyReLU. A Dropout layer is added after each hidden layer to prevent overfitting. The output layer has 4 neurons and uses a linear activation function to output continuous values ​​in the regression task.

2. The method for predicting the degree of harm caused by dangerous human behavior based on video information according to claim 1, characterized in that, In step S2, five kinematic parameters of the extracted skeletal points are selected. These five kinematic parameters are: the change in displacement in the x-direction extracted from the kinematic trajectory. y-direction displacement change Maximum instantaneous velocity in the x-direction Maximum instantaneous velocity in the y-direction and the maximum instantaneous resultant velocity .

3. A system for predicting the degree of harm caused by dangerous human behavior based on video information, wherein the system is used to implement the method for predicting the degree of harm caused by dangerous human behavior based on video information as described in any one of claims 1-2, characterized in that, The system includes: a video acquisition module, a feature extraction module, an input selection module, a damage parameter determination module, a model building module, and a protection prediction module; The video acquisition module is used to acquire video information of dangerous behaviors and force-time curves generated by force sensors. The feature extraction module is used to identify human skeleton points in the video information and extract the kinematic parameters of the skeleton points using the FuzzyPose model. The input selection module is used to select the kinematic parameters of the skeletal points as input to the dataset; The damage parameter determination module is used to take the parameters in the force-time curve and the number of penetration layers caused by the dangerous behavior to the protective material sample as the dangerous behavior damage parameters, and take the dangerous behavior damage parameters corresponding to the input of the dataset as the output of the dataset. The model building module is used to build a V2D model and train the V2D model based on the input and output of the dataset. The protection prediction module is used to input real-time dangerous behaviors into the V2D model for prediction, and obtain the prediction result of the degree of harm caused by dangerous human behaviors.

4. The system for predicting the degree of harm caused by dangerous human behavior based on video information according to claim 3, characterized in that, In the feature extraction module, five kinematic parameters of the extracted skeletal points are selected. These five kinematic parameters are: the change in displacement in the x-direction extracted from the kinematic trajectory. y-direction displacement change Maximum instantaneous velocity in the x-direction Maximum instantaneous velocity in the y-direction and the maximum instantaneous resultant velocity .

Images