Intelligent falling injury rapid identification and grading prediction method and system

Abstract

The application belongs to the field of artificial intelligence and medical technology, and relates to a method and system for intelligent fall injury rapid identification and grading prediction. The method comprises the following steps: acquiring fall site data, positioning data and historical fall injury case data; data integration; calling a GIS map; extracting fall data; calculating visual risk parameters, geographic risk parameters and historical risk parameters; obtaining a comprehensive fall injury risk through a multi-feature fusion decision algorithm; determining an injury grade and calculating the probability of each injury type corresponding to each injury grade; constructing a gradient boosting decision tree model, training the gradient boosting decision tree model, and obtaining an injury prediction model; and calculating the injury grade of the current falling personnel and the probability of each injury type corresponding to each injury grade. The application solves the problems of slow fall injury evaluation speed, low precision and data fragmentation in the prior art, realizes second-level fall injury identification and grading prediction through data collaboration, and improves the evaluation accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence and medical technology, and specifically relates to an intelligent fall injury rapid identification and grading prediction method and system. BACKGROUND

[0002] A large number of fall injury accidents occur every year. At present, the fall injury scene assessment is still mainly artificial assessment, that is, relying on the description of eyewitnesses or the artificial inspection after the arrival of first aid personnel. Most cases have a delay in first aid response, that is, missing the clinically recognized golden ten-minute rescue period. According to the statistics of the emergency center, from the accident alarm to the arrival of professional medical personnel at the scene to complete the basic injury judgment (such as the point of force and the suspected damage site), the entire assessment process takes much longer than the time window of the golden rescue period.

[0003] Fall injuries (such as spinal cord injuries and brain injuries) have a strong concealment, and the initial judgment on the scene is easily affected by environmental and experience factors. The existing technology uses artificial assessment and wearable device monitoring to judge the force point and high-risk damage site and other fall core parameters, which has a high misjudgment rate. Misjudgment (such as misjudging spinal cord injury as soft tissue injury) often leads to secondary injury during transportation and first aid, further aggravating the patient's condition.

[0004] In the existing technical system, medical data (such as patient's past medical history and allergy history), geographic information data (such as the hardness of the accident scene terrain and the distribution of surrounding first aid resources), and the historical case database cannot be real-time fused and called due to protocol incompatibility and data format heterogeneity. For example, wearable devices can only provide positioning and fall alarm, cannot associate the influence of the hardness of the scene terrain (such as concrete or grass) on the injury, and cannot retrieve the treatment plan of similar historical cases, resulting in a lack of comprehensive support for the assessment results.

[0005] The existing AI medical auxiliary tools (such as AI models for injury assessment) can only analyze a single type of data (such as X-ray and CT images in hospitals), and cannot process dynamic visual information (such as human body posture and environmental obstacles in accident scene photos and short videos) on the scene. The scene visual information is the core basis for judging the fall height, force point and collision force, for example, the fall height can be calculated from the step height in the photo, and the spine stress condition can be judged from the human body landing posture in the video. The existing technology lacks the ability to analyze visual information, resulting in a single assessment dimension and difficulty in improving accuracy.

[0006] In summary, the existing fall injury assessment technology has insurmountable defects in real-time, accuracy and collaboration, and an intelligent system that can integrate scene visual information, real-time fusion of multi-source data and end-to-cloud collaborative rapid calculation is urgently needed. SUMMARY

[0007] In order to solve the above technical problems, the application provides a method and system for intelligent fall injury rapid identification and grading prediction.

[0008] In the first aspect, the application provides a method for intelligent fall injury rapid identification and grading prediction, comprising:

[0009] Obtaining fall site data, positioning data and historical fall injury case data of the falling personnel; the fall site data includes fall site image data and fall injury condition data; the fall injury condition data includes injury condition grades;

[0010] Constructing a GIS map database and a historical injury condition database, and integrating data to extract the material hardness coefficient of the falling point; the historical injury condition database is used to store the historical fall injury case data;

[0011] Calling the GIS map, positioning the user location information and the hospital location information according to the positioning data, and determining the fall injury diagnosis and treatment hospital with the shortest driving distance;

[0012] Extracting the falling data from the fall site image data, including: identifying the falling height according to the fall site image data; identifying the body node of the falling personnel through semantic segmentation, generating a heat map by combining contour density analysis, calculating the collision probability of each body node based on a collision probability model, and determining the force point parameters of the falling personnel; determining the falling posture of the falling personnel through continuous frame analysis; identifying the obstacle type of the falling point and setting the obstacle label;

[0013] According to the falling height, the force point parameters, the falling posture and the obstacle label, the energy impact index is calculated, the visual risk parameter is calculated according to the energy impact index, and the geographical risk parameter is calculated according to the material hardness coefficient of the falling point; according to the similarity of the current fall injury case data and the historical fall injury case data, the historical risk parameter is calculated according to the similarity value;

[0014] The visual risk parameter, the geographical risk parameter and the historical risk parameter are weighted by a multi-feature fusion decision algorithm to obtain the fall injury comprehensive risk;

[0015] According to the fall injury comprehensive risk, the injury condition grade is determined, and the probability of each injury type corresponding to each injury condition grade is calculated;

[0016] A gradient boosting decision tree model is constructed, the falling height, the force point parameters, the falling posture, the obstacle label and the historical fall injury case data are taken as inputs, the injury condition grade and the probability of each injury type corresponding to each injury condition grade are taken as outputs, the gradient boosting decision tree model is trained, and an injury prediction model is obtained;

[0017] According to the falling height of the current falling person, the force point parameter, the falling posture, the obstacle label and the historical falling injury case data, the injury prediction model is used to calculate the injury grade of the current falling person and the probability of each injury type corresponding to each injury grade.

[0018] In a second aspect, the application provides a smart falling injury rapid identification and grading prediction system, comprising an acquisition unit, a construction unit, a calling unit, an extraction unit, a parameter calculation unit, a feature fusion unit, a data processing unit, a model construction unit and an output unit.

[0019] The acquisition unit is used to acquire falling site data, positioning data and historical falling injury case data of a falling person; the falling site data includes falling site image data and falling injury data; the falling injury data includes an injury grade.

[0020] The construction unit is used to construct a GIS map database and a historical injury database, and to integrate data and extract the material hardness coefficient of the falling point; the historical injury database is used to store historical falling injury case data.

[0021] The calling unit is used to call a GIS map, locate user location information and hospital location information according to the positioning data, and determine the falling injury diagnosis and treatment hospital with the shortest travel distance.

[0022] The extraction unit is used to extract falling data from the falling site image data, including: identifying the falling height according to the falling site image data; identifying the body nodes of the falling person through semantic segmentation, generating a heat map by combining contour density analysis, calculating the collision probability of each body node based on a collision probability model, and determining the force point parameter of the falling person; determining the falling posture of the falling person through continuous frame analysis; identifying the obstacle type of the falling point and setting the obstacle label.

[0023] The parameter calculation unit is used to calculate the energy impact index according to the falling height, the force point parameter, the falling posture and the obstacle label, calculate the visual risk parameter according to the energy impact index, calculate the geographical risk parameter according to the material hardness coefficient of the falling point, and calculate the historical risk parameter according to the similarity between the current falling injury case data and the historical falling injury case data.

[0024] The feature fusion unit is used to weight the visual risk parameter, the geographical risk parameter and the historical risk parameter through a multi-feature fusion decision algorithm to obtain the falling injury comprehensive risk.

[0025] The data processing unit is used to determine the injury grade according to the falling injury comprehensive risk and calculate the probability of each injury type corresponding to each injury grade.

[0026] The model construction unit is configured to construct a gradient boosting decision tree model, take the falling height, the landing point parameter, the falling posture, the obstacle label and the historical falling injury case data as inputs, take the injury grade and the probability of each injury type corresponding to each injury grade as outputs, train the gradient boosting decision tree model, and obtain an injury prediction model.

[0027] The output unit is configured to calculate the injury grade and the probability of each injury type corresponding to each injury grade of the current falling personnel by using the injury prediction model according to the falling height, the landing point parameter, the falling posture, the obstacle label and the historical falling injury case data of the current falling personnel.

[0028] Based on the above technical solutions, the application can be further improved as follows.

[0029] Further, the falling height is identified according to the falling scene image data, including: if there is a reference object with a known height, a reference object scale transformation method is used to calculate the falling height based on the reference object pixel height and the reference object actual height; the falling height is the ratio of the reference object pixel height to the reference object actual height multiplied by the pixel height of the falling starting point; when there is no reference object, a three-dimensional posture projection algorithm is used to calculate the falling height, assuming that the falling height is h, the body type coefficient library characteristic value is C, the pixel length of the spine is L, the camera angle is θ, and the falling height is represented as: h=C*L*sinθ C is a body type coefficient library characteristic value, L is a pixel length of the spine, θ is a camera angle, and .

[0030] Further, the body node of the falling personnel is identified through semantic segmentation, including:

[0031] Each pixel in the falling scene image data is classified, and each pixel is labeled with a category to which the pixel belongs, and a pixel-level category mask is output;

[0032] The body node name is defined, and the pixel coordinates of all body nodes are labeled;

[0033] The semantic segmentation subnetwork is constructed, the body mask is output by using the semantic segmentation subnetwork, and the body image is obtained by performing pixel-level multiplication on the mask and the original image;

[0034] The body node detection subnetwork is constructed, the intermediate layer features of the semantic segmentation subnetwork are input into the body node detection subnetwork in the body mask obtained through semantic segmentation, and the coordinates of all body nodes are located;

[0035] The semantic segmentation subnetwork is pre-trained, the parameters of the semantic segmentation subnetwork are frozen, and the body node detection subnetwork is trained;

[0036] The semantic segmentation subnetwork and the body node detection subnetwork are jointly trained, and all network parameters are fine-tuned to adapt the features of the semantic segmentation subnetwork and the body node detection subnetwork.

[0037] Further, the heat map is generated in combination with the contour density analysis, and the collision probability of each body node is calculated based on a collision probability model, including: setting the collision probability as , the pixel density as , the coefficient of the pixel density as , the shadow deformation degree as , the coefficient of the shadow deformation degree as , the texture damage degree as , and the coefficient of the texture damage degree as , the collision probability model is represented as: .

[0038] Further, the force point parameter of the falling person is determined, including: determining the body node with the maximum probability value in the result of the collision probability model, and taking the label value of the body node as the force point parameter.

[0039] Further, the falling posture of the falling person is determined through continuous frame analysis, including:

[0040] Single-frame posture estimation is performed through a lightweight posture estimation model, and a fifteen-point skeleton of each frame is detected;

[0041] The hip speed and the change in the spine angle are calculated in a loop to obtain dynamic parameters;

[0042] The cumulative stress of the spine is calculated, and the risk level is output by comparing a preset threshold;

[0043] The hip trajectory and the spine angle change curve are drawn to obtain the falling posture of the falling person.

[0044] Further, the Euclidean distance algorithm is used to calculate the similarity value of the current falling injury case data and the historical falling injury case data, and when the Euclidean distance of the retrieved historical falling injury case data and the current falling event is less than or equal to a set threshold, the similar case weighted data is calculated according to the Euclidean distance, and the similar case weighted data is used as the historical risk parameter.

[0045] Further, the injury type includes a brain impact index, a spine dynamics load, an organ damage coefficient, and an old-age bone fragility factor;

[0046] The brain impact index is set as , the head landing probability is , and the falling height is , the brain impact index is represented as:

[0047] ;

[0048] Let the spinal dynamics load be , the spinal bending angle be , the inter-frame momentum difference be , then: ;

[0049] Let the organ injury coefficient be , the abdominal landing mark be , the body mass index be , and the ground stiffness coefficient be , then: ;

[0050] Let the old bone fragility factor be , the age be , and the pelvic landing probability be , then: .

[0051] Further, let the comprehensive risk of falling injury be , the visual risk be , the geographical risk be , and the historical risk be , is the coefficient of visual risk, is the coefficient of geographical risk, is the coefficient of historical risk, is the image quality parameter, is the cloud coverage, is the freshness of historical data, the visual risk is the energy impact index, the geographical risk is the material hardness coefficient, and the historical risk is the difference between 1 and the weighted data of similar cases, then the comprehensive risk of falling injury is represented as:

[0052] ;

[0053] Wherein:

[0054] ;

[0055] ;

[0056] .

[0057] The beneficial effects of the present application are: the present application solves the problems of slow evaluation speed, low precision and data fragmentation in the prior art, through mobile visual acquisition, cloud multi-source data integration and intelligent terminal and cloud collaborative mode, realizing second-level falling injury recognition and grading prediction, improving the evaluation accuracy, and providing efficient and accurate technical support for first aid decision-making. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the intelligent fall injury rapid identification and classification prediction method provided in Embodiment 1 of the present invention;

[0059] Figure 2 This is a schematic diagram of the intelligent fall injury rapid identification and classification prediction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0061] Example 1

[0062] As an example, see the attached document. Figure 1 As shown, to solve the above-mentioned technical problems, this embodiment provides an intelligent method for rapid identification and classification prediction of fall injuries, including:

[0063] Acquire fall scene data, location data, and historical fall injury case data for fallen personnel; fall scene data includes fall scene image data and fall injury data; fall injury data includes injury severity level;

[0064] Construct a GIS (Geographic Information System) map database and a historical injury database, integrate the data, and extract the material hardness coefficient of the fall points; the historical injury database is used to store historical fall injury case data;

[0065] By calling up the GIS map and locating the user's location information and the hospital's location information based on the location data, the hospital with the shortest driving distance for treating fall injuries can be determined.

[0066] The process involves extracting fall data from fall scene image data, including: identifying the fall height based on the fall scene image data; identifying the body nodes of the fallen person through semantic segmentation, generating a heat map by combining contour density analysis, calculating the collision probability of each body node based on a collision probability model, and determining the force point parameters of the fallen person; determining the fall posture of the fallen person through continuous frame analysis; identifying the obstacle type at the fall point and setting obstacle labels.

[0067] According to the falling height, the force point parameter, the falling posture and the obstacle label, an energy impact index is calculated, a visual risk parameter is calculated according to the energy impact index, and a geographical risk parameter is calculated according to the material hardness coefficient of the falling point; according to the similarity of the current falling injury case data and the historical falling injury case data, a historical risk parameter is calculated according to the similarity value;

[0068] The visual risk parameter, the geographical risk parameter and the historical risk parameter are weighted by a multi-feature fusion decision algorithm to obtain a falling injury comprehensive risk;

[0069] According to the falling injury comprehensive risk, the injury grade is determined, and the probability of each injury type corresponding to each injury grade is calculated;

[0070] A gradient boosting decision tree model is constructed, the falling height, the force point parameter, the falling posture, the obstacle label and the historical falling injury case data are taken as inputs, the injury grade and the probability of each injury type corresponding to each injury grade are taken as outputs, the gradient boosting decision tree model is trained, and an injury prediction model is obtained;

[0071] According to the falling height, the force point parameter, the falling posture, the obstacle label and the historical falling injury case data of the current falling personnel, the injury prediction model is used to calculate the injury grade of the current falling personnel and the probability of each injury type corresponding to each injury grade.

[0072] In actual application process, through the application program of mobile terminal carrying Android operating system, image and video collection upload and GPS positioning data upload are carried out, so as to obtain visual data and geographical position information of falling scene. The collected visual data is subjected to light weight pretreatment (such as compression to within 100KB), only the key area information such as human body and reference object is reserved, and the data transmission bandwidth occupation is reduced; when it is detected that the cloud network is unobstructed (delay≤50ms), the pretreated visual data and GPS positioning information are automatically uploaded.

[0073] GIS map database access: store the terrain data in global range, including the hardness coefficient of common terrain such as cement ground, asphalt road, grassland and soil (such as: the hardness coefficient of cement ground is 3.5, the hardness coefficient of asphalt road is 3.2, and the hardness coefficient of grassland is 1.5), the terrain type and the corresponding hardness coefficient of the falling scene can be matched in real time according to the GPS positioning information uploaded by the mobile terminal.

[0074] By accessing GIS map database and historical injury database on the server side to store medical data and historical falling injury case data, the integration of multi-source geographical and medical data is realized. The medical data includes hospital outpatient information and hospital location information, according to the user location information and the hospital location information, the falling injury diagnosis and treatment hospital which meets the falling injury diagnosis and treatment conditions and has the shortest driving distance is determined.

[0075] In practical application, a historical falling injury case database with a sample size greater than 1000 cases is constructed for subsequent injury prediction model training and comprehensive risk calculation. Each historical falling injury case contains fields such as falling height, landing point, posture, terrain type, and injury grade.

[0076] Optionally, the falling height is identified according to the falling scene image data, including: if there is a known height reference object, a reference object scale transformation method is used to calculate the falling height based on the reference object pixel height and the reference object actual height; the falling height is the ratio of the reference object pixel height to the reference object actual height multiplied by the pixel height of the falling starting point; when there is no reference object, a three-dimensional posture projection algorithm is used to calculate the falling height, assuming is the falling height, is the body shape coefficient library characteristic value, is the pixel length of the spine, is the camera angle of elevation, then the falling height is represented as: .

[0077] Optionally, the body nodes of the falling person are identified through semantic segmentation, including:

[0078] Each pixel in the falling scene image data is classified, and each pixel is labeled with the class to which it belongs, and a pixel-level class mask is output;

[0079] Define the body node name and label the pixel coordinates of all body nodes;

[0080] Construct a semantic segmentation subnetwork, use the semantic segmentation subnetwork to output a body mask, and perform pixel-level multiplication between the mask and the original image to obtain a body image;

[0081] Construct a body node detection subnetwork, input the intermediate layer features of the semantic segmentation subnetwork into the body node detection subnetwork within the body mask obtained by semantic segmentation, and locate the coordinates of all body nodes;

[0082] Pre-train the semantic segmentation subnetwork, freeze the parameters of the semantic segmentation subnetwork, and train the body node detection subnetwork;

[0083] Jointly train the semantic segmentation subnetwork and the body node detection subnetwork, fine-tune all network parameters, and adapt the features of the semantic segmentation subnetwork and the body node detection subnetwork.

[0084] Optionally, a heat map is generated by combining contour density analysis, and the collision probability of each body node is calculated based on a collision probability model, including: assuming that the collision probability is , the pixel density is , the coefficient of the pixel density is , and the shadow deformation degree is , the coefficient of the shadow deformation degree is , the texture damage degree is , the coefficient of the texture damage degree is , the collision probability model is represented as: .

[0085] When calculating the collision probability of each region, the region with the highest probability is taken as the main force point, and the region with the second highest probability is taken as the secondary force point.

[0086] Optionally, the force point parameter of the falling person is determined, including: determining the body node with the maximum probability value in the result of the collision probability model, and taking the label value of the body node as the force point parameter.

[0087] Optionally, the falling posture of the falling person is determined through continuous frame analysis, including:

[0088] Single-frame posture estimation is performed through a lightweight posture estimation model, and a fifteen-point skeleton is detected for each frame;

[0089] The hip speed and spine angle change are calculated in a loop to obtain dynamic parameters;

[0090] The cumulative stress of the spine is calculated, and the risk level is output by comparing with a preset threshold;

[0091] The hip trajectory and spine angle change curve are drawn to obtain the falling posture of the falling person.

[0092] A fifteen-point simplified skeleton model (key points: head, cervical vertebra, lumbar vertebra, left / right shoulder, left / right hip, left / right knee, left / right ankle, left / right wrist, and left / right elbow) is used, the inter-frame hip speed and spine angle change are calculated through 30fps video frame analysis, and the cumulative stress of the spine is calculated based on integration.

[0093] Optionally, the Euclidean distance algorithm is used to calculate the similarity value of the current falling injury case data and the historical falling injury case data, when the Euclidean distance of the retrieved historical falling injury case data and the current falling event is less than or equal to a set threshold, the similar case weighted data is calculated according to the Euclidean distance, and the similar case weighted data is taken as the historical risk parameter.

[0094] The Euclidean distance algorithm is used to retrieve similar cases, for example: when the Euclidean distance of the historical falling injury case and the falling injury case of the current event is less than or equal to 0.35, it is determined as a similar case.

[0095] When constructing the gradient boosting decision tree model, the falling height, force point parameter, falling posture, obstacle label and historical falling injury case data are taken as input, and the injury grade and the probability corresponding to each injury type are taken as output, the gradient boosting decision tree model is trained to obtain an injury prediction model;

[0096] The input data of the gradient boosting decision tree model includes the height, force point parameters (for example, the head force point parameter is 1, the torso force point parameter is 2, and the leg force point parameter is 3) output by the visual engine, the posture code (for example, the posture code corresponding to the bending is 1, the posture code corresponding to the standing is 2, and the posture code corresponding to the lying is 3), the obstacle label (for example, the obstacle label corresponding to the hard ground is 1, the obstacle label corresponding to the metal is 2, and the obstacle label corresponding to the sharp object is 3), and the material hardness coefficient of the falling point and the similar case weighting data integrated and output by the server side.

[0097] Optionally, the injury type includes a brain impact index, a spine dynamics load, an organ damage coefficient, and an old-age bone fragility factor.

[0098] Let the brain impact index be , the head landing probability be , and the falling height be , the brain impact index is expressed as: ;

[0099] Let the spine dynamics load be , the spine bending angle be , and the inter-frame momentum difference be , then: ;

[0100] Let the organ damage coefficient be , the abdominal landing identification be , the body mass index be , and the ground stiffness coefficient be , then: ;

[0101] Let the old-age bone fragility factor be , the age be , and the pelvic landing probability be , then: .

[0102] Optionally, let the falling injury comprehensive risk be , the visual risk be , the geographical risk be , and the historical risk be , is the coefficient of the visual risk, is the coefficient of the geographical risk, is the coefficient of the historical risk, is the image quality parameter, is the cloud coverage, is the historical data freshness, the visual risk is the energy impact index, and the geographical risk is the material hardness coefficient, historical risk is the difference between 1 and the weighted data of similar cases, and the comprehensive risk of falling injury is represented as:

[0103] ;

[0104] wherein:

[0105] ;

[0106] ;

[0107] .

[0108] The implementation process of the intelligent falling injury rapid identification and grading prediction system of the application is described in detail in combination with specific application scenarios as follows:

[0109] (1) Implementation scenario setting, a user falls at the stairs of a community, a person accompanying the user uses a mobile phone to open a terminal application program, takes a photo (including stairs as a reference object, the actual height of the stairs is 15 cm) and a 10-second short video (30 fps, 1080P resolution) of the scene, and uploads them to the cloud; the user is 55 years old, BMI = 22, and the APP obtains the longitude and latitude of the scene through the mobile phone GPS.

[0110] (2) Mobile terminal processing process: the mobile terminal application program calls the mobile terminal camera to take photos and short videos, automatically compresses the photos to 80 KB, and compresses the videos to 100 KB (only the human body and the stairs area are retained); calls the GPS module to obtain positioning information (error 8 meters), detects that the cloud network delay is 30 ms (in the ready state), uploads the compressed visual data and positioning information to the cloud; runs the YOLO_Nano model to detect the human body, extracts the coordinates of the head, trunk, hip and other key parts, and the processing time is 0.8 seconds;

[0111] (3) Server-side data integration process, the data integration submodule matches the GIS map according to the positioning information, determines that the falling scene is a cement ground, and the corresponding hardness coefficient = 3.5; searches the historical injury database, filters out the cases of "falling height 5-7m, head landing, cement ground", calculates the Euclidean distance, wherein the Euclidean distance of 1 case (falling height 6.2m, head landing, cement ground, injury grade 8, probability of brain injury 90%) is 0.28, which is determined as a similar case; the historical risk parameter is 1-0.28 = 0.72;

[0112] Height recognition: The cloud GAN model enhances the uploaded photo, and the ResNet152 model measures the step pixel height = 30 pixels, the falling point (head) pixel height = 600 pixels, and substitutes into the reference scale transformation formula: estimated height = (30 pixels / 15 cm) x 600 pixels = 600 cm = 6 m, error 9.5% (actual falling height 6.6 m);

[0113] Force point positioning: Identify the head and torso area through human semantic segmentation, calculate the pixel density of the head area = 0.9, shadow deformation = 0.8, texture damage = 0.7, and substitute into the collision probability model: the output of the collision probability model is P(collision), P(collision) = 0.6 x 0.9 + 0.25 x 0.8 + 0.15 x 0.7 = 0.845, the collision probability of the torso area = 0.32, and determine the main force point as the head and the secondary force point as the torso.

[0114] Fall posture analysis: Identify key points using a fifteen-point simplified skeleton model, analyze 10-second video (300 frames) through mmDetect model, calculate the average value of inter-frame hip speed = 0.5 m / s, and the average value of spine angle change = 15°, set trapz as numerical integration operation, and calculate the cumulative stress of the spine based on integration = trapz(0.5 x sin15° x 300) = trapz(0.5 x 0.2588 x 300) = trapz(38.82) = 38.82 (stress unit).

[0115] Injury prediction model processing process:

[0116] (1) Input parameter integration: height = 6 m, force point code = 1 (head), posture code = 1 (bending), obstacle label = 1 (hard ground), terrain hardness coefficient = 3.5, similar case weighted data (injury level 8 points x 0.72, probability of brain injury 90% x 0.72);

[0117] (2) Feature calculation:

[0118] Cranial impact index = 0.845 x 6^(1.5) = 0.845 x 14.696 ≈ 12.42;

[0119] Spine dynamics load = ∑(15° x 0.5 m / s x 300 frames) = 2250 (load unit);

[0120] Visceral injury coefficient = 0 (non-abdominal landing) x 22 x 3.5 = 0;

[0121] Old age bone fragility factor = (55-60) x 0.3 x 0 (non-pelvic landing) = 0 (age less than 60, take 0);

[0122] (3) Multi-source fusion decision: Assuming the image quality parameter is 0.9, the cloud coverage is 20%, and the historical data freshness is 8, then:

[0123] ;

[0124] ;

[0125] ;

[0126] Overall Risks: ;

[0127] (4) Output results: Based on the GBDT model, the output injury level is 8 points (10 points for the most critical), the probability of craniocerebral injury is 92%, the probability of spinal injury is 38%, the probability of pelvic injury is 12%, the probability of massive bleeding is 16%, and the processing time is 0.4 seconds.

[0128] The overall time taken for this implementation was as follows: 0.8 seconds for mobile processing, 0.3 seconds for data transmission, 0.5 seconds for cloud data integration, 2.2 seconds for visual engine processing, and 0.4 seconds for injury prediction model processing, totaling 4.2 seconds. Subsequent medical diagnosis showed that the user suffered a traumatic brain contusion (the probability of injury was consistent with the system's prediction) and a minor compression fracture of the spine (the probability was slightly higher than the predicted 38%, with the error stemming from individual differences in bone density). The overall assessment accuracy was 92.5%, verifying the effectiveness of the method of this invention.

[0129] This invention solves the problems of slow speed, low accuracy and fragmented data in the existing technology of fall injury assessment. By using mobile visual acquisition, cloud multi-source data integration and smart terminal and cloud collaboration, it can achieve fall injury identification and classification prediction in seconds, improve the accuracy of assessment and provide efficient and accurate technical support for emergency decision-making.

[0130] Example 2

[0131] Based on the same principle as the method shown in Embodiment 1 of the present invention, as illustrated in the appendix. Figure 2 As shown, the embodiments of the present invention also provide an intelligent fall injury rapid identification and classification prediction system, including an acquisition unit, a construction unit, a calling unit, an extraction unit, a parameter calculation unit, a feature fusion unit, a data processing unit, a model construction unit, and an output unit;

[0132] The acquisition unit is used to acquire fall scene data, location data, and historical fall injury case data of fallen personnel; the fall scene data includes fall scene image data and fall injury data; the fall injury data includes the injury level;

[0133] A construction unit is configured to construct a GIS map database and a historical injury database, integrate data, and extract a material hardness coefficient of a falling point.

[0134] A calling unit is configured to call a GIS map, locate user position information and hospital position information according to positioning data, and determine a falling injury treatment hospital with the shortest driving distance.

[0135] An extraction unit is configured to extract falling data from falling site image data, including: identifying a falling height according to the falling site image data; identifying body node of a falling person through semantic segmentation, generating a heat map by combining contour density analysis, calculating collision probability of each body node based on a collision probability model, and determining force point parameters of the falling person; determining a falling posture of the falling person through continuous frame analysis; and identifying an obstacle type of a falling point and setting an obstacle label.

[0136] A parameter calculation unit is configured to calculate an energy impact index according to the falling height, the force point parameters, the falling posture, and the obstacle label, calculate a visual risk parameter according to the energy impact index, and calculate a geographical risk parameter according to the material hardness coefficient of the falling point; calculate a historical risk parameter according to a similarity between current falling injury case data and historical falling injury case data, and according to a similarity value.

[0137] A feature fusion unit is configured to perform weighting on the visual risk parameter, the geographical risk parameter, and the historical risk parameter through a multi-feature fusion decision algorithm to obtain a falling injury comprehensive risk.

[0138] A data processing unit is configured to determine an injury grade according to the falling injury comprehensive risk, and calculate a probability of each injury type corresponding to each injury grade.

[0139] A model construction unit is configured to construct a gradient boosting decision tree model, take the falling height, the force point parameters, the falling posture, the obstacle label, and the historical falling injury case data as inputs, take the injury grade and the probability of each injury type corresponding to each injury grade as outputs, train the gradient boosting decision tree model, and obtain an injury prediction model.

[0140] An output unit is configured to calculate an injury grade of a current falling person and a probability of each injury type corresponding to each injury grade of the current falling person by using the injury prediction model according to the falling height, the force point parameters, the falling posture, the obstacle label, and the historical falling injury case data of the current falling person.

[0141] Optionally, the falling height is identified according to the falling scene image data, including: if there is a reference object with a known height, a reference object scale transformation method is used to calculate the falling height based on the reference object pixel height and the reference object actual height; the falling height is the ratio of the reference object pixel height to the reference object actual height multiplied by the pixel height of the falling starting point; when there is no reference object, a three-dimensional posture projection algorithm is used to calculate the falling height, assuming that the falling height is H, the camera angle is a, the body shape coefficient is k, the pixel length of the spine is L, and the pixel length of the body is L', then the falling height is represented as: H = k * L * sin(a) . . . . .

[0142] Optionally, the body node of the falling person is identified through semantic segmentation, including:

[0143] Each pixel in the falling scene image data is classified, and the category to which each pixel belongs is labeled, and a pixel-level category mask is output;

[0144] The body node name is defined, and the pixel coordinates of all body nodes are labeled;

[0145] The semantic segmentation subnetwork is constructed, the body mask is output by using the semantic segmentation subnetwork, and the body image is obtained by performing pixel-level multiplication on the mask and the original image;

[0146] The body node detection subnetwork is constructed, the intermediate layer features of the semantic segmentation subnetwork are input into the body node detection subnetwork within the body mask obtained by semantic segmentation, and the coordinates of all body nodes are located;

[0147] The semantic segmentation subnetwork is pre-trained, the parameters of the semantic segmentation subnetwork are frozen, and the body node detection subnetwork is trained;

[0148] The semantic segmentation subnetwork and the body node detection subnetwork are jointly trained, all network parameters are fine-tuned, and the features of the semantic segmentation subnetwork and the body node detection subnetwork are adapted.

[0149] Optionally, a heat map is generated by combining contour density analysis, and the collision probability of each body node is calculated based on a collision probability model, including: assuming that the collision probability is P, the pixel density is D, the coefficient of the pixel density is a, the shadow deformation degree is S, the coefficient of the shadow deformation degree is b, the texture damage degree is T, and the coefficient of the texture damage degree is c, then the collision probability model is represented as: P = D * a + S * b + T * c . . . . . . .

[0150] Optionally, the landing point parameter of the falling person is determined, including: determining a body node with the maximum probability value in the result of the collision probability model, and taking a label value of the body node as the landing point parameter.

[0151] Optionally, the falling posture of the falling person is determined through continuous frame analysis, including:

[0152] Single-frame posture estimation is performed through a lightweight posture estimation model, and a fifteen-point skeleton of each frame is detected;

[0153] The hip speed and the change in the spine angle are calculated in a loop to obtain dynamic parameters;

[0154] The cumulative stress of the spine is calculated, and a risk level is output by comparing a preset threshold value;

[0155] The hip trajectory and the change curve of the spine angle are drawn to obtain the falling posture of the falling person.

[0156] Optionally, the Euclidean distance algorithm is used to calculate the similarity value of the current falling injury case data and the historical falling injury case data, when the Euclidean distance of the retrieved historical falling injury case data and the current falling event is less than or equal to a set threshold value, the similar case weighted data is calculated according to the Euclidean distance, and the similar case weighted data is taken as the historical risk parameter.

[0157] Optionally, the injury type includes a brain impact index, a spine dynamics load, an organ damage coefficient, and an old person bone fragility factor;

[0158] The brain impact index is set as , the head landing probability is , and the falling height is , and the brain impact index is represented as: ;

[0159] The spine dynamics load is set as , the spine bending angle is , and the inter-frame momentum difference is , and: ;

[0160] The organ damage coefficient is set as , the abdominal landing identification is , the body mass index is , and the ground stiffness coefficient is , and: ;

[0161] The old person bone fragility factor is set as , the age is , and the pelvic landing probability is , and: .

[0162] Optionally, the comprehensive risk of falling injury is , the visual risk is , the geographical risk is , the historical risk is , is the coefficient of the visual risk, is the coefficient of the geographical risk, is the coefficient of the historical risk, is the image quality parameter, is the cloud coverage, is the historical data freshness, the visual risk is the energy impact index, the geographical risk is the material hardness coefficient, the historical risk is the difference between 1 and the weighted data of similar cases, and the comprehensive risk of falling injury is represented as:

[0163] ;

[0164] wherein:

[0165] ;

[0166] ;

[0167] .

[0168] The above only is the preferred embodiment of the present application, and is not used for limiting the present application, for the person skilled in the art, the present application can have various changes and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for intelligent fall injury rapid identification and grading prediction, characterized in that, The method comprises the following steps: Obtain the falling site data, positioning data and historical falling injury data of the falling personnel; the falling site data comprises falling site image data and falling injury data; the falling injury data comprises injury grade; Construct a GIS map database and a historical injury database, integrate the data, and extract the material hardness coefficient of the falling point; the historical injury database is used to store the historical falling injury data; Call the GIS map, locate the user location information and hospital location information according to the positioning data, and determine the falling injury diagnosis and treatment hospital with the shortest driving distance; Extract the falling data from the falling site image data, including: identifying the falling height according to the falling site image data; identifying the body node of the falling personnel through semantic segmentation, generating a heat map by combining contour density analysis, calculating the collision probability of each body node based on a collision probability model, and determining the force point parameters of the falling personnel; determining the falling posture of the falling personnel through continuous frame analysis; identifying the obstacle type of the falling point and setting the obstacle label; Calculate the energy impact index according to the falling height, force point parameters, falling posture and obstacle label, calculate the visual risk parameter according to the energy impact index, and calculate the geographical risk parameter according to the material hardness coefficient of the falling point; calculate the historical risk parameter according to the similarity of the current falling injury data and the historical falling injury data; Weight the visual risk parameter, geographical risk parameter and historical risk parameter through a multi-feature fusion decision algorithm to obtain the falling injury comprehensive risk; Determine the injury grade according to the falling injury comprehensive risk, and calculate the probability of each injury type corresponding to each injury grade; Construct a gradient boosting decision tree model, take the falling height, force point parameters, falling posture, obstacle label and historical falling injury data as inputs, take the injury grade and the probability of each injury type corresponding to each injury grade as outputs, train the gradient boosting decision tree model, and obtain an injury prediction model; Calculate the injury grade and the probability of each injury type corresponding to each injury grade of the current falling personnel by using the injury prediction model according to the falling height, force point parameters, falling posture, obstacle label and historical falling injury data of the current falling personnel.

2. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, According to the falling scene image data, the falling height is identified, including: if there is a known height reference object, a reference object scale transformation method is used, and the falling height is calculated based on the reference object pixel height and the reference object actual height; the falling height is the ratio of the reference object pixel height and the reference object actual height multiplied by the pixel height of the falling starting point; when there is no reference object, a three-dimensional posture projection algorithm is used to calculate the falling height, assuming is the falling height, is the body shape coefficient library characteristic value, is the spine pixel length, is the camera angle of elevation, and the falling height is expressed as: .

3. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, Identify the body node of the falling personnel through semantic segmentation, including: Classify each pixel in the falling site image data, label the category of each pixel, and output the category mask at the pixel level; Define the name of the body node, and label the pixel coordinates of all body nodes; Construct a semantic segmentation subnetwork, output the body mask using the semantic segmentation subnetwork, and perform pixel-level multiplication on the mask and the original image to obtain a body image; Construct a body node detection subnetwork, input the intermediate layer features of the semantic segmentation subnetwork into the body node detection subnetwork within the body mask obtained by semantic segmentation, and locate the coordinates of all body nodes; Pre-train the semantic segmentation subnetwork, freeze the parameters of the semantic segmentation subnetwork, and train the body node detection subnetwork; Jointly train the semantic segmentation subnetwork and the body node detection subnetwork, fine-tune all network parameters, and adapt the features of the semantic segmentation subnetwork and the body node detection subnetwork.

4. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, The heat map is generated in combination with the contour density analysis, and the collision probability of each body node is calculated based on a collision probability model, including: setting the collision probability as , the pixel density as , the coefficient of the pixel density as , the shadow deformation degree as , the coefficient of the shadow deformation degree as , the texture damage degree as , and the coefficient of the texture damage degree as , so that the collision probability model is represented as: .

5. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, The force point parameter of the falling person is determined, including: determining the body node with the maximum probability value in the result of the collision probability model, and taking the label value of the body node as the force point parameter.

6. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, The falling posture of the falling person is determined through continuous frame analysis, including: Single-frame posture estimation is performed through a lightweight posture estimation model, and a fifteen-point skeleton of each frame is detected; The hip speed and the change of the spine angle are calculated in a loop to obtain dynamic parameters; The cumulative stress of the spine is calculated, and the risk level is output by comparing with a preset threshold; The hip trajectory and the change curve of the spine angle are drawn to obtain the falling posture of the falling person.

7. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, The similarity value of the current falling injury case data and the historical falling injury case data is calculated by using the Euclidean distance algorithm. When the Euclidean distance of the retrieved historical falling injury case data and the current falling event is less than or equal to a set threshold, the similar case weighted data is calculated according to the Euclidean distance, and the similar case weighted data is taken as the historical risk parameter.

8. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, The injury types include the brain impact index, the spine dynamics load, the organ damage coefficient and the bone fragility factor of the elderly; Let the head impact index be , the head landing probability be , and the falling height be , then the head impact index is expressed as: ; Let the spinal dynamic load be , the spinal bending angle be , and the inter-frame momentum difference be , then: ; The organ damage coefficient is , the abdominal landing mark is , the body mass index is , and the ground stiffness coefficient is , then: ; Let the bone fragility factor of the elderly be , the age be , and the probability of pelvis landing be , then: .

9. The intelligent fall injury rapid identification and grading prediction method according to claim 1, characterized in that, Let the fall injury composite risk be , the visual risk be , the geographical risk be , the historical risk be , a coefficient of the visual risk, a coefficient of the geographical risk, a coefficient of the historical risk, an image quality parameter, a cloud cover, a historical data freshness, the visual risk an energy impact index, the geographical risk a material hardness coefficient, the historical risk a difference between 1 and the similar case weighted data, then the fall injury composite risk is represented as: ； Among them: ； ； 。 10. An intelligent fall injury rapid identification and grading prediction system, characterized in that, It includes an acquisition unit, a construction unit, a calling unit, an extraction unit, a parameter calculation unit, a feature fusion unit, a data processing unit, a model construction unit and an output unit; The acquisition unit is used for acquiring falling site data, positioning data and historical falling injury case data of the falling person; the falling site data includes falling site image data and falling injury data; the falling injury data includes injury level; The construction unit is used for constructing GIS map database and historical injury database, and integrating data, and extracting material hardness coefficient of falling point; the historical injury database is used for storing historical falling injury case data; The calling unit is used for calling GIS map, positioning user location information and hospital location information according to positioning data, and determining the falling injury diagnosis and treatment hospital with the shortest driving distance; The extraction unit is used for extracting falling data from the falling site image data, including: identifying the falling height according to the falling site image data; identifying the body node of the falling person through semantic segmentation, generating a heat map based on contour density analysis, calculating the collision probability of each body node based on the collision probability model, and determining the force point parameter of the falling person; determining the falling posture of the falling person through continuous frame analysis; identifying the obstacle type of the falling point and setting the obstacle label; The parameter calculation unit is used for calculating the energy impact index according to the falling height, the force point parameter, the falling posture and the obstacle label, calculating the visual risk parameter according to the energy impact index, and calculating the geographical risk parameter according to the material hardness coefficient of the falling point; calculating the historical risk parameter according to the similarity of the current falling injury case data and the historical falling injury case data, and calculating the historical risk parameter according to the similarity value; The feature fusion unit is used for weighting the visual risk parameter, the geographical risk parameter and the historical risk parameter through a multi-feature fusion decision algorithm to obtain the comprehensive risk of falling injury; The data processing unit is used for determining the injury level according to the comprehensive risk of falling injury, and calculating the probability of each injury type corresponding to each injury level. The model construction unit is configured to construct a gradient boosting decision tree model, take the falling height, the force point parameter, the falling posture, the obstacle label, and the historical falling injury case data as inputs, take the injury grade and the probability of each injury type corresponding to each injury grade as outputs, train the gradient boosting decision tree model, and obtain an injury prediction model. The output unit is configured to calculate the injury grade and the probability of each injury type corresponding to each injury grade of the current falling personnel by using the injury prediction model according to the falling height, the force point parameter, the falling posture, the obstacle label, and the historical falling injury case data of the current falling personnel.

Images