Highly infectious disease screening system based on ai vision analysis and behavior recognition technology

Abstract

The present application relates to the technical field of computer vision and public health epidemic prevention, solves the problems of single screening technology means, high missed diagnosis rate caused by disconnection of symptom recognition and behavior monitoring, and insufficient privacy protection. The system constructs a disease feature knowledge base, synchronously acquires visible light images, infrared thermal images and environmental parameters through a multi-modal data acquisition module; an AI vision analysis module uses an improved U-Net network to segment the eye and rash areas, combines with Wood lamp enhanced features, and uses a 3D pose estimation algorithm to quantify joint protective posture and gait abnormalities; a behavior recognition module captures action timing features, and a risk assessment and early warning module matches the features with the knowledge base to generate a structured report. The present application realizes the synchronous non-contact screening of specific signs and high-risk behaviors of various highly contagious diseases, significantly improves the recognition accuracy and prevention and control efficiency, and takes into account data privacy and security.

Description

Technical Field

[0001] This invention relates to the field of computer vision and public health and epidemic prevention, specifically to a screening system for highly infectious diseases based on AI visual analysis and behavior recognition technology. Background Technology

[0002] Highly infectious diseases are characterized by rapid transmission, wide impact, and high severity, especially in densely populated places such as airports, train stations, and hospitals. Once an infection case occurs, it can easily lead to a cluster outbreak, posing a serious threat to public health and safety. With the acceleration of globalization and increased frequency of population movement, the prevention and control of highly infectious diseases such as influenza, COVID-19, chikungunya, and Crimean-Congo hemorrhagic fever are becoming increasingly severe. Building an efficient, accurate, and contactless screening system to achieve early warning and rapid response is a key link in improving epidemic prevention and control capabilities. In recent years, artificial intelligence and computer vision technologies have made significant progress in the field of medical image analysis, providing new technical pathways for intelligent screening of highly infectious diseases. However, how to achieve deep fusion of multimodal data and accurate identification in complex scenarios remains a pressing technical challenge.

[0003] Current technologies primarily rely on manual screening and single-indicator detection devices. Traditional manual screening typically involves healthcare workers using handheld infrared thermometers at public entrances to check the temperature of passersby, observe for obvious symptoms such as cough and fatigue, and inquire about their travel history. While simple to implement, this method has significant limitations: First, manual screening is inefficient, easily causing congestion in high-traffic areas, and prolonged work can lead to staff fatigue and increase the risk of missed diagnoses. Second, single-indicator detection has blind spots; infrared thermometers only obtain surface temperature data and cannot correlate it with other specific symptoms such as rash, conjunctival congestion, and joint pain, lacking effective means of identifying latent infections or patients without obvious fever symptoms. Third, manual observation is highly subjective; non-professionals often struggle to accurately diagnose mild neurological abnormalities (such as drowsiness or apathy) or musculoskeletal abnormalities (such as limping or protective joint postures), leading to missed or misdiagnosed cases. Finally, manual screening lacks data retention and traceability capabilities, failing to provide accurate data support for subsequent epidemiological investigations.

[0004] Existing technologies have attempted to introduce automated screening techniques based on computer vision, but these suffer from fragmented functionality and poor coordination. Some existing technologies use facial recognition algorithms combined with thermal imaging for fever screening, capturing facial images and measuring body temperature via cameras, triggering an alarm when the temperature exceeds a threshold. However, these technologies often focus on identifying single symptoms and lack the ability to monitor high-risk behaviors (such as crowding, improper mask-wearing, and close-range conversations), which are crucial factors in the spread of highly infectious diseases. Other existing technologies attempt to monitor crowd gatherings using behavioral recognition algorithms, but often employ general target detection models without specific optimization for disease-related behavioral patterns (such as limping in chikungunya patients or abnormal hemostasis in hemorrhagic fever patients). This leads to a significant drop in accuracy in complex lighting, occlusion, and crowded environments. Furthermore, existing technologies generally lack multimodal data fusion mechanisms, failing to effectively correlate heterogeneous data such as visible light images, infrared thermal imaging, and environmental parameters, resulting in a disconnect between symptom recognition and behavioral monitoring, making it difficult to form a complete chain of screening evidence.

[0005] There are significant shortcomings in data privacy protection and system generalization capabilities. During the screening process, the system needs to collect sensitive visual data such as facial features and behavioral trajectories. Existing technologies often lack robust data anonymization and encryption mechanisms, which can easily lead to the leakage of personal information and pose privacy and security risks. Some technologies excessively collect data in pursuit of recognition accuracy, increasing the pressure on data transmission and storage and reducing the system's real-time responsiveness. At the same time, existing AI models are mostly trained on data from specific scenarios or specific diseases, resulting in insufficient generalization capabilities and difficulty in quickly adapting to the screening needs of emerging infectious diseases. For example, models trained for novel coronavirus infection cannot be directly applied to identify the joint pain characteristics of Chikungunya fever or the retinitis-related behaviors of Riften Valley fever, resulting in a lack of flexibility in responding to emerging infectious diseases.

[0006] In summary, existing screening technologies for highly infectious diseases suffer from problems such as low efficiency and high rate of missed diagnoses in manual screening, obvious limitations of single testing methods, fragmented application of AI technology with poor coordination, and insufficient balance between privacy protection and screening efficiency. Summary of the Invention

[0007] The purpose of this invention is to provide a screening system for highly infectious diseases based on AI visual analysis and behavior recognition technology. This integrated screening system combines AI visual analysis and behavior recognition technology to simultaneously monitor specific symptoms such as fever, rash, and neurological abnormalities with high-risk behaviors such as gathering and close contact, thereby improving the accuracy and efficiency of screening while ensuring data privacy and security. This addresses the technical bottlenecks in existing technologies, such as the disconnect between symptom recognition and behavior monitoring, single-scenario adaptation, and insufficient coverage of multiple diseases.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A screening system for highly infectious diseases based on AI visual analysis and behavior recognition technology includes: The multimodal data acquisition module is used to simultaneously acquire visible light images, infrared thermal imaging data, and environmental parameter data of the target. A disease feature knowledge base is used to store the mapping relationship between specific clinical manifestations and behavioral patterns of various highly infectious diseases. The specific clinical manifestations include joint pain features of Chikungunya fever, retinitis-related behavioral features of Rift Valley fever, and facial flushing features of Crimean-Congo hemorrhagic fever. The AI ​​visual analysis module preprocesses the visible light image, uses an improved U-Net network to perform pixel-level segmentation of the eye and rash areas, and combines Wood's lamp auxiliary light source to enhance the morphological and color features of the rash; it uses a 3D pose estimation algorithm to extract key points of the whole skeleton and quantitatively analyzes joint protective posture and limping features. The behavior recognition module is used to analyze the temporal characteristics of the target's actions based on continuous frame images, and to identify behaviors related to systemic poisoning symptoms, gastrointestinal symptoms, and abnormal nervous system behaviors. The risk assessment and early warning module is used to input the visual features extracted by the AI ​​visual analysis module and the behavioral features identified by the behavior recognition module into the disease feature knowledge base for matching, calculate the risk level and generate a structured screening report.

[0009] Furthermore, the AI ​​visual analysis module includes a facial thermal distribution analysis unit. The facial thermal distribution analysis unit constructs a facial thermal distribution map using infrared thermal imaging data, calculates the temperature gradient of different areas of the face, identifies a fever pattern of high fever in the trunk accompanied by cold extremities, and a specific facial flushing pattern. The specific facial flushing pattern includes the flushing features of a drunken appearance corresponding to Crimean-Congo hemorrhagic fever.

[0010] Furthermore, the 3D pose estimation algorithm uses the OpenPose algorithm to extract 33 key points of the whole body skeleton. The quantitative analysis of joint protective posture includes: calculating the joint angle change rate, limb movement trajectory curvature and gait symmetry parameters. When a specific joint range of motion is detected to be lower than a preset threshold and there is compensatory movement, it is determined to be a joint protective posture.

[0011] Furthermore, the behavior recognition module includes a respiratory symptom recognition unit, which uses a dual-stream convolutional neural network to extract spatial and temporal features, fuses and analyzes the frequency of chest undulations, the timing of mouth opening actions, and neck vibration features in continuous frames, distinguishes between sudden cough, persistent cough, and violent sneezing behaviors, and identifies occlusion compensation features in the context of wearing a mask.

[0012] Furthermore, the system also includes a multimodal data fusion module, which is used to align the timestamps and spatial coordinates of the visible light image and infrared thermal imaging data, introduce environmental parameter data as a weighting factor into the feature fusion layer, and construct a two-dimensional feature vector of symptoms and behavior. The environmental parameter data includes light intensity, population density, and temperature and humidity data.

[0013] Furthermore, the risk assessment and early warning module refers to constructing a risk assessment model based on the analytic hierarchy process, quantifying symptom indicators, behavioral indicators, and environmental indicators into risk scores, setting graded early warning thresholds, and automatically generating a risk tracing report containing suspected disease types, abnormal signs, and movement trajectories when the risk score reaches a high-risk threshold.

[0014] Furthermore, the system also includes a privacy protection module, which performs blurring and desensitization processing on the faces of people in visible light images during the data acquisition stage; adopts SSL / TLS encryption protocol during the data transmission stage; adopts AES-256 encryption algorithm during the data storage stage, and sets a data storage period, automatically desensitizing and deleting non-critical data that exceeds the period.

[0015] Furthermore, the improved U-Net network introduces an attention mechanism module at the encoder end to enhance the perception of areas with blurred rash boundaries and subtle features of conjunctival congestion. The Wood's lamp auxiliary light source excites the fluorescence response of the rash area with ultraviolet light with wavelengths from 320 nm to 400 nm.

[0016] Furthermore, the disease feature knowledge base supports online iterative updates of the model, receives clinical feature data of emerging infectious diseases, adjusts the model parameters of the AI ​​visual analysis module through transfer learning, and updates the mapping relationships in the disease feature knowledge base to adapt to the screening needs of emerging infectious diseases.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) Because the present invention adopts a multimodal fusion screening architecture driven by a disease feature knowledge base, it deeply correlates visible light images, infrared thermal imaging data and behavioral pattern data, thus solving the problem of disconnect between symptom identification and behavior monitoring in the prior art. Compared with the prior art that only relies on single temperature measurement or isolated behavior analysis, the present invention can simultaneously identify specific signs such as fever, rash, and joint pain, as well as high-risk behaviors such as gathering and close contact, and constructs a screening evidence chain of "symptoms + behavior", which significantly improves the identification accuracy of a variety of highly infectious diseases such as Chikungunya fever, Rift Valley fever, and Crimean-Congo hemorrhagic fever.

[0018] (2) By setting up a 3D pose estimation and quantification analysis mechanism based on the OpenPose algorithm, the system accurately captures abnormal behaviors of the musculoskeletal system such as joint protective posture and limping. Subjective pain behavior is transformed into objective quantitative parameters such as joint angle change rate and gait symmetry, overcoming the shortcomings of strong subjectivity and difficulty in quantification in manual observation. This enables the system to effectively screen for diseases such as Chikungunya fever, which are characterized by joint pain, and fills the gap in existing screening technologies for musculoskeletal symptoms.

[0019] (3) Because the present invention adopts an improved U-Net network combined with a Wood's lamp auxiliary light source in the AI ​​visual analysis module, it significantly enhances the ability to recognize subtle signs such as rash morphology and conjunctival congestion. By exciting the fluorescence reaction with ultraviolet light with a wavelength of 320nm to 400nm and using the attention mechanism to enhance the feature perception of blurred boundary areas, the technical problem of the difficulty in early detection of small rashes and bleeding points under conventional visible light is solved, and high-sensitivity detection of early signs of hemorrhagic fever diseases is achieved.

[0020] (4) By configuring a multimodal data fusion module and a dynamic environmental parameter weighting mechanism, the robustness of the system in complex scenarios is improved. The technical principle is to incorporate environmental factors such as light intensity and personnel density into the recognition model and dynamically adjust the feature fusion weights, effectively eliminating the impact of environmental interference such as strong light, backlight, and personnel occlusion on recognition accuracy, so that the system can operate stably in high-risk places with complex lighting and dense crowds, such as airports and train stations.

[0021] (5) Because this invention constructs a full-link privacy protection system, it employs fuzzy desensitization, SSL / TLS encryption, AES-256 encryption, and periodic automatic deletion technologies in each stage of data collection, transmission, and storage, achieving a balance between screening efficiency and privacy security. Compared with existing technologies that neglect data security or cause privacy leaks due to excessive collection, this invention maximizes the protection of personal identity information while ensuring the integrity of the screening function, thus meeting the ethical and legal requirements of public health and epidemic prevention.

[0022] (6) By setting up a disease feature knowledge base that supports online iterative updates, the system is given the ability to quickly adapt to emerging infectious diseases. When a new infectious disease emerges, only new clinical feature data needs to be input, and the system can adjust the model parameters through transfer learning without replacing hardware or redeveloping the software system, which greatly shortens the response cycle of epidemic prevention and control and reduces the operation and maintenance cost of the prevention and control system. Attached Figure Description

[0023] Figure 1 This is a block diagram of the overall structure of the highly infectious disease screening system based on AI visual analysis and behavior recognition technology of the present invention.

[0024] Figure 2 This is a schematic diagram of the internal structure of the AI ​​visual analysis module.

[0025] Figure 3 This is a flowchart illustrating the workflow of the risk assessment and early warning module.

[0026] The diagram is labeled as follows: 10-Multimodal data acquisition module; 11-Visible light image acquisition unit; 12-Infrared thermal imaging acquisition unit; 13-Environmental parameter acquisition unit; 14-Wood's lamp auxiliary light source; 20-Disease feature knowledge base; 30-AI visual analysis module; 31-Facial thermal distribution analysis unit; 32-Rash and eye feature segmentation unit; 33-Skeletal key point extraction unit; 34-Joint posture quantification unit; 40-Behavior recognition module; 41-Respiratory system symptom recognition unit; 42-Temporal feature analysis unit; 50-Multimodal data fusion module; 51-Spatiotemporal alignment unit; 52-Feature fusion unit; 60-Risk assessment and early warning module; 61-Risk score calculation unit; 62-Graded early warning unit; 63-Source tracing report generation unit; 70-Privacy protection module; 71-Facial desensitization unit; 72-Data encryption unit; 80-Model iteration update module. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0028] like Figure 1As shown, the highly infectious disease screening system based on AI visual analysis and behavior recognition technology provided in this embodiment includes a multimodal data acquisition module 10, a disease feature knowledge base 20, an AI visual analysis module 30, a behavior recognition module 40, a multimodal data fusion module 50, a risk assessment and early warning module 60, a privacy protection module 70, and a model iteration and update module 80. The multimodal data acquisition module 10, as the system's front-end perception layer, is deployed at the screening site (such as an airport entrance or hospital lobby) and is responsible for simultaneously acquiring visible light images, infrared thermal imaging data, and environmental parameter data of the target, providing a multi-source heterogeneous data foundation for subsequent analysis. The disease feature knowledge base 20 is stored on the system's back-end server and pre-sets mapping relationships between specific clinical manifestations and behavioral patterns of various highly infectious diseases such as Chikungunya fever, Rift Valley fever, and Crimean-Congo hemorrhagic fever. The AI ​​visual analysis module 30 and the behavior recognition module 40 run in parallel on an edge computing unit or cloud server, respectively responsible for deep feature extraction of physiological signs and behavioral patterns. The multimodal data fusion module 50 is responsible for spatiotemporal alignment and feature fusion of heterogeneous data. The risk assessment and early warning module 60 performs risk calculation and early warning output based on the fused features. The privacy protection module 70 ensures personal information security throughout the entire data lifecycle. First, the multimodal data acquisition module 10 performs non-contact data acquisition. After being de-identified by the privacy protection module 70, the data is transmitted to the AI ​​visual analysis module 30 and the behavior recognition module 40 for parallel feature extraction. Then, the multimodal data fusion module 50 generates a comprehensive feature vector. Finally, the risk assessment and early warning module 60 combines the data with the disease feature knowledge base 20 to perform matching and discrimination and output the screening results.

[0029] The multimodal data acquisition module 10 includes a visible light image acquisition unit 11, an infrared thermal imaging acquisition unit 12, an environmental parameter acquisition unit 13, and a Wood's lamp auxiliary light source 14. The visible light image acquisition unit 11 uses an industrial-grade high-definition network camera with a resolution of at least 1080P and a frame rate of at least 30fps. It supports wide dynamic range (WDR) and low-light imaging functions, and the lens focal length can be selected from 4mm to 12mm depending on the monitoring distance. It is used to acquire visible light image data of facial expressions, limb movements, and skin surface features (such as rashes and bleeding points). The infrared thermal imaging acquisition unit 12 uses an uncooled vanadium oxide microbolometer focal plane detector with a thermal sensitivity (NETD) better than 50mK, a temperature measurement accuracy of ±0.3℃, and a temperature measurement range of -20℃ to 150℃. It is used to acquire thermal distribution images of the face and body surface of personnel, achieving non-contact body temperature monitoring through infrared thermal imaging technology. The environmental parameter acquisition unit 13 includes temperature and humidity sensors and light sensors, used to monitor the ambient temperature, humidity, and light intensity at the screening site in real time. The environmental parameter data will serve as a compensation factor for the subsequent algorithm model, eliminating the interference of environmental changes on recognition accuracy. The Wood's lamp auxiliary light source 14 emits ultraviolet light with wavelengths from 320nm to 400nm, primarily used to assist in the identification of skin lesions such as rashes and pigmentation abnormalities. Under specific wavelengths of ultraviolet light, certain fungal infections or metabolic skin diseases will emit specific fluorescence, thereby enhancing the AI ​​visual analysis module 30's ability to extract subtle rash features. Each acquisition unit is connected to the edge computing unit via gigabit Ethernet or Wi-Fi 6 wireless networks, achieving millisecond-level low-latency data transmission.

[0030] The disease feature knowledge base 20 is built on a relational database management system (such as PostgreSQL) and uses ontology modeling technology to provide a structured description of the clinical features of highly infectious diseases. The knowledge base comprises three core sub-bases: a physiological signs knowledge base, a behavioral pattern knowledge base, and a disease mapping rule base. The physiological signs knowledge base records detailed visual feature parameters of signs such as fever, rash, conjunctival congestion, and facial flushing. For example, the "drunken appearance" feature of Crimean-Congo hemorrhagic fever (CCHF) is described as a combination of facial flushing, conjunctival congestion, and eyelid edema; the joint pain feature of Chikungunya fever is described as swelling and limited movement of specific joints (such as the wrist and ankle). The behavioral pattern knowledge base records temporal features of behaviors such as fatigue, vomiting, limping, and unsteady gait. For example, early symptoms of encephalitis in Nipah virus disease patients include olfactory impairment and unsteady holding. The disease mapping rule base stores the association weights between signs, behaviors, and disease types, supporting rapid matching via SQL queries or a rule engine. The knowledge base supports online updates via the model iteration update module 80. When a new infectious disease emerges, medical staff can enter new clinical feature data through the management backend, and the system will automatically adjust the model parameters and update the knowledge base content through transfer learning.

[0031] The AI ​​visual analysis module 30 includes a facial thermal distribution analysis unit 31, a rash and eye feature segmentation unit 32, a skeletal key point extraction unit 33, and a joint posture quantization unit 34. The facial thermal distribution analysis unit 31 receives thermal image data transmitted by the infrared thermal imaging acquisition unit 12. First, it uses a median filtering algorithm to remove thermal noise. Then, it performs image segmentation based on the Otsu method to extract facial regions, calculates the average and maximum temperatures of different facial regions (forehead, cheeks, nose tip, and chin), and constructs a facial thermal distribution vector. This unit is specifically designed with an abnormal thermal pattern recognition algorithm, capable of identifying the centrifugal temperature difference pattern of "high fever in the trunk accompanied by cold extremities" and specific facial flushing patterns. When the facial thermal distribution vector matches the "drunken appearance" thermal feature template of Crimean-Congo hemorrhagic fever, it is automatically labeled as a high-risk feature.

[0032] The rash and eye feature segmentation unit 32 employs an improved U-Net fully convolutional neural network for image semantic segmentation. This network introduces an attention mechanism module at the encoder end, enhancing the model's ability to perceive subtle features such as blurred rash boundaries and conjunctival congestion through weighted coefficients, while suppressing background noise interference. For the input visible light image, this unit first performs illumination correction preprocessing to eliminate the influence of strong light or backlight environments, and then segments the eye region and suspected rash regions. For the eye region, conjunctival color histogram features are extracted, and the mean and variance of the red, green, and blue (RGB) channels are calculated to determine whether conjunctival congestion or hemorrhage exists. For the rash region, combined with fluorescence images acquired by the Wood's lamp-assisted light source 14, the morphology (macules, papules, vesicles), color, and distribution range of the rash are analyzed, and the rash classification results are output.

[0033] The skeletal keypoint extraction unit 33 constructs a 3D pose estimation model based on the OpenPose algorithm. It extracts image features through a convolutional neural network and outputs the two-dimensional coordinates and confidence scores of 33 skeletal keypoints throughout the human body through multi-stage inference. Then, combined with camera calibration parameters and geometric constraints, it solves the three-dimensional spatial coordinates of the keypoints. The joint pose quantization unit 34 receives the three-dimensional coordinate sequence of skeletal keypoints and calculates the angle, angular velocity, and range of motion (ROM) between adjacent bone vectors. For the joint pain characteristics of diseases such as chikungunya fever, this unit designs a "joint-protective posture" judgment logic: when the angle of motion of a certain joint (such as the right ankle joint) is detected to be continuously less than a preset threshold (such as 15%) of the contralateral joint during movement, and accompanied by a prolonged gait cycle and a shortened stance phase time, it is judged as a joint-protective posture, indicating that there may be motor dysfunction caused by joint pain.

[0034] The behavior recognition module 40 includes a respiratory symptom recognition unit 41 and a temporal feature analysis unit 42. The respiratory symptom recognition unit 41 employs a dual-stream convolutional neural network architecture, where the spatial flow network processes single-frame images to extract facial expressions and lip features, and the temporal flow network processes continuous-frame optical flow images to capture dynamic changes in movement. This unit can recognize three typical respiratory symptom behaviors: sudden cough, persistent cough, and violent sneezing. For mask-wearing scenarios, this unit introduces an occlusion feature compensation mechanism. By analyzing eye muscle contractions (such as blinking), brow creases, and neck muscle vibration features, it infers coughing behavior under the mask's obstruction, effectively solving the recognition blind spot problem caused by mask occlusion. The temporal feature analysis unit 42 constructs a behavior sequence classifier based on a Long Short-Term Memory (LSTM) network. It performs long-term temporal analysis on the person's behavioral trajectory, dwell time, and interaction frequency, identifying behaviors related to systemic poisoning symptoms (such as prolonged sitting and lethargy), behaviors related to gastrointestinal symptoms (such as frequent toilet visits and abdominal clutching), and abnormal nervous system behaviors (such as unsteady gait and unstable holding).

[0035] The multimodal data fusion module 50 includes a spatiotemporal alignment unit 51 and a feature fusion unit 52. The spatiotemporal alignment unit 51 is responsible for time synchronization and spatial registration of data from different modalities. Since there may be time differences and viewing angle differences between visible light images and infrared thermal imaging, this unit adopts an image registration algorithm based on mutual information. By maximizing the mutual information value of the two images, it finds the optimal spatial transformation matrix to achieve pixel-level precise alignment. At the same time, it uses hardware trigger signals or software timestamp synchronization mechanisms to ensure that the time error of multiple data streams is controlled within 10ms. The feature fusion unit 52 adopts a feature-level fusion strategy, which concatenates the vital sign feature vectors (such as body temperature, rash confidence, and joint angle deviation) output by the AI ​​visual analysis module 30 with the behavioral feature vectors (such as cough frequency and gait abnormality index) output by the behavior recognition module 40. It also introduces environmental weight factors (such as the correction coefficient of ambient temperature for body temperature measurement and the influence of light intensity on the confidence of visible light recognition) output by the environmental parameter acquisition unit 13. Through a fully connected layer, feature dimensionality reduction and nonlinear mapping are performed to generate a comprehensive feature vector containing "symptoms + behavior + environment" information.

[0036] The risk assessment and early warning module 60 includes a risk score calculation unit 61, a graded early warning unit 62, and a source tracing report generation unit 63. The risk score calculation unit 61 constructs a risk assessment model based on the Analytic Hierarchy Process (AHP), assigning different weight values ​​to symptom indicators (fever, rash, cough, etc.), behavioral indicators (gathering, not wearing masks, limping, etc.), and environmental indicators (personnel density, ventilation). For example, the weight of fever symptom is set to 0.4, the weight of joint-protective posture is set to 0.2, and the weight of gathering behavior is set to 0.1. A comprehensive risk score is calculated through weighted summation, with a score range of 0 to 100. The graded early warning unit 62 sets three early warning thresholds: low risk (0-30 points), medium risk (31-70 points), and high risk (71-100 points). High-risk individuals will trigger a red warning, and medium-risk individuals will trigger a yellow warning. The source tracing report generation unit 63 automatically retrieves the travel records, captured images, movement trajectories, and close contact information of high-risk personnel to generate a structured risk source tracing report. The report includes suspected disease type (such as "suspected chikungunya fever"), abnormal signs (such as "limited movement of the left ankle joint"), risk score details, and treatment suggestions, and is pushed to the terminal of the prevention and control personnel through an encrypted channel.

[0037] The privacy protection module 70 includes a face desensitization unit 71 and a data encryption unit 72. During the data acquisition phase, the face desensitization unit 71 employs a face anonymization algorithm based on Generative Adversarial Networks (GANs) to blur and desensitize or replace features on the faces of individuals in visible light images. While retaining necessary feature information for symptom recognition (such as facial flushing and conjunctival congestion), it removes identity recognition features (such as iris texture and facial contour details), ensuring that the original image cannot be used to reconstruct an individual's identity. During the data transmission phase, the data encryption unit 72 uses the SSL / TLS 1.3 protocol to establish an encrypted communication link, preventing data from being eavesdropped on or tampered with during transmission. During the data storage phase, it uses the AES-256 encryption algorithm to encrypt and store both structured and unstructured data. Furthermore, the system implements a strict data lifecycle management strategy, with non-critical data storage lasting 30 days. Data exceeding this period is automatically deleted after desensitization, reducing the risk of privacy leaks at the source.

[0038] The model iteration and update module 80 provides the system with online learning and model optimization capabilities. This module receives feedback data from medical staff on the screening results (such as false positive corrections and missed positive entries), and stores the feedback data as labeled samples in the training database. When the number of samples accumulates to a certain amount (e.g., every 100 new samples) or reaches a preset period (e.g., weekly), the model fine-tuning process is triggered. Model fine-tuning uses transfer learning technology, keeping the basic layer parameters of the original network unchanged, and only retraining the top-level classifier. This allows the model to quickly adapt to the characteristics of emerging infectious diseases or environmental changes in specific scenarios without destroying the original knowledge structure, thus achieving continuous evolution of the system.

[0039] During operation, firstly, when personnel enter the screening area, the visible light image acquisition unit 11 and the infrared thermal imaging acquisition unit 12 of the multimodal data acquisition module 10 are simultaneously triggered to acquire visible light images and thermal imaging data of the personnel. If the ambient light is insufficient, the Wood's lamp auxiliary light source 14 automatically turns on to supplement the light. The environmental parameter acquisition unit 13 records the on-site temperature and humidity in real time. Next, the raw data undergoes real-time desensitization processing by the facial desensitization unit 71 of the privacy protection module 70 to remove identity features and retain only disease-related features. Subsequently, the desensitized data is transmitted to the AI ​​visual analysis module 30 and the behavior recognition module 40. The AI ​​visual analysis module 30 executes three branches of processing in parallel: the facial thermal distribution analysis unit 31 extracts body temperature and thermal pattern features; the rash and eye feature segmentation unit 32 segments and identifies rash and conjunctival congestion features; and the skeletal key point extraction unit 33 and the joint posture quantification unit 34 analyze human posture and joint movement features. At the same time, the behavior recognition module 40 performs temporal recognition of behaviors such as coughing, sneezing, and limping. Then, the multimodal data fusion module 50 performs spatiotemporal alignment and weighted fusion of the aforementioned multi-path features to generate a comprehensive feature vector. Finally, the risk assessment and early warning module 60 inputs the comprehensive feature vector into the disease feature knowledge base 20 for matching. The risk score calculation unit 61 calculates the risk score, and the graded early warning unit 62 triggers the corresponding early warning based on the score level. If the risk is determined to be high, the source tracing report generation unit 63 automatically generates a report and pushes it to the prevention and control terminal, and the model iteration and update module 80 records the screening data for subsequent model optimization.

[0040] This embodiment is applied to the screening of highly infectious diseases at the entrance of an international airport terminal. A multimodal data acquisition module 10 is deployed at the entrance gate. The visible light image acquisition unit 11 uses a Hikvision DS-2CD2T87FWD-L network camera with a resolution of 4 megapixels, a frame rate of 50fps, and a lens focal length of 8mm. The infrared thermal imaging acquisition unit 12 uses a Hikvision thermal imager with a resolution of 640×512 and a temperature measurement accuracy of ±0.3℃. The Wood's lamp auxiliary light source 14 uses a 365nm wavelength ultraviolet LED array with a power of 10W. The edge computing unit uses the NVIDIA Jetson AGX Xavier development kit. The system parameters are configured as follows: the fever recognition threshold is set to 37.3℃; the joint protective posture judgment threshold is set to a difference in joint angles greater than 15% between the two sides; the cough behavior recognition time window is set to 5 seconds; and the high-risk warning threshold is set to 70 minutes. In an actual screening, a passenger passed through the gate. The visible light image acquisition unit 11 captured that the passenger's right foot had a significantly shortened contact time with the ground while walking, exhibiting a limping gait. The skeletal key point extraction unit 33 extracted key points of the passenger's entire skeleton, and the joint posture quantification unit 34 calculated that the dorsiflexion angle of the right ankle joint was only 5 degrees, far below the normal value (approximately 20 degrees), and the proportion of the right lower limb's support phase time was only 35% (the normal value is approximately 60%), classifying it as a "right ankle joint protective posture." Simultaneously, the facial thermal distribution analysis unit 31 detected a facial temperature of 36.8℃, which did not reach the fever threshold. The rash and eye feature segmentation unit 32 did not detect any rash or conjunctival congestion. The multimodal data fusion module 50 fused the "right ankle joint protective posture" feature (weight 0.3) with the normal body temperature feature (weight -0.1), calculating a risk score of 45 points, triggering a yellow alert. The system prompted "Suspected joint pain symptoms, re-examination recommended." Subsequent re-examination by medical staff confirmed that the passenger was an early-stage Chikungunya fever patient. This embodiment verifies the system's ability to screen for non-febrile highly infectious diseases.

[0041] This embodiment is applied to the fever clinic triage station of a specialized infectious disease hospital. The system configuration is basically the same as in Embodiment 1, except that a high-precision Wood's lamp auxiliary light source 14 (wavelength 320nm) is added, and the weight configuration of the disease feature knowledge base 20 is adjusted to improve the sensitivity to hemorrhagic fever diseases. A patient comes for treatment. The infrared thermal imaging acquisition unit 12 detects that the patient's highest facial temperature is 38.5℃, and the facial thermal distribution analysis unit 31 analyzes and finds that the patient's face presents a typical "drunken appearance" thermal pattern (abnormally high temperature in the flushed area of ​​the face, and blurred boundaries). At the same time, the rash and eye feature segmentation unit 32, under the illumination of the Wood's lamp auxiliary light source 14, detects scattered petechiae and ecchymoses on the patient's chest skin, and the fluorescence reaction is positive. The AI ​​visual analysis module 30 integrates the above features and combines them with the Crimean-Congo hemorrhagic fever (CCHF) feature template in the disease feature knowledge base 20, with a matching degree of up to 92%. The risk assessment and early warning module 60 immediately triggers a red warning and generates a report prompting "Highly suspected CCHF, please isolate immediately". This embodiment verifies the system's ability to accurately identify specific signs of hemorrhagic fever.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A screening system for highly infectious diseases based on AI visual analysis and behavior recognition technology, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire visible light images, infrared thermal imaging data, and environmental parameter data of the target. A disease feature knowledge base is used to store the mapping relationship between specific clinical manifestations and behavioral patterns of various highly infectious diseases. The specific clinical manifestations include joint pain features of Chikungunya fever, retinitis-related behavioral features of Rift Valley fever, and facial flushing features of Crimean-Congo hemorrhagic fever. The AI ​​visual analysis module preprocesses the visible light image, uses an improved U-Net network to perform pixel-level segmentation of the eye and rash areas, and combines Wood's lamp auxiliary light source to enhance the morphological and color features of the rash. A 3D pose estimation algorithm was used to extract key points of the whole skeleton and to quantitatively analyze joint protective posture and limp characteristics. The behavior recognition module is used to analyze the temporal characteristics of the target's actions based on continuous frame images, and to identify behaviors related to systemic poisoning symptoms, gastrointestinal symptoms, and abnormal nervous system behaviors. The risk assessment and early warning module is used to input the visual features extracted by the AI ​​visual analysis module and the behavioral features identified by the behavior recognition module into the disease feature knowledge base for matching, calculate the risk level and generate a structured screening report.

2. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The AI ​​visual analysis module includes a facial thermal distribution analysis unit. The facial thermal distribution analysis unit constructs a facial thermal distribution map using infrared thermal imaging data, calculates the temperature gradient of different areas of the face, identifies a fever pattern of high fever in the trunk accompanied by cold extremities, and a specific facial flushing pattern. The specific facial flushing pattern includes the flushing features of a drunken appearance corresponding to Crimean-Congo hemorrhagic fever.

3. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The 3D pose estimation algorithm uses the OpenPose algorithm to extract 33 key points of the whole body skeleton. The quantitative analysis of joint protective posture includes: calculating the joint angle change rate, limb movement trajectory curvature and gait symmetry parameters. When a specific joint range of motion is detected to be lower than a preset threshold and there is compensatory movement, it is determined to be a joint protective posture.

4. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The behavior recognition module includes a respiratory symptom recognition unit. The respiratory symptom recognition unit uses a dual-stream convolutional neural network to extract spatial and temporal features, and fuses and analyzes the frequency of chest undulation, the timing of mouth opening actions, and neck vibration features in continuous frames to distinguish between sudden cough, persistent cough, and violent sneezing behaviors, and to identify occlusion compensation features in the case of wearing a mask.

5. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The system also includes a multimodal data fusion module, which is used to align the timestamps and spatial coordinates of the visible light image and infrared thermal imaging data, introduce environmental parameter data as a weighting factor into the feature fusion layer, and construct a two-dimensional feature vector of symptoms and behaviors. The environmental parameter data includes light intensity, population density, and temperature and humidity data.

6. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The risk assessment and early warning module refers to a risk assessment model built based on the analytic hierarchy process (AHP), which quantifies symptom indicators, behavioral indicators, and environmental indicators into risk scores, sets graded early warning thresholds, and automatically generates a risk tracing report containing suspected disease types, abnormal signs, and movement trajectories when the risk score reaches the high-risk threshold.

7. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The system also includes a privacy protection module, which performs blurring and desensitization processing on human faces in visible light images during the data acquisition stage; adopts SSL / TLS encryption protocol during the data transmission stage; and adopts AES-256 encryption algorithm during the data storage stage, and sets a data storage period, automatically desensitizing and deleting non-critical data that exceeds the period.

8. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The improved U-Net network introduces an attention mechanism module at the encoder end to enhance the perception of areas with blurred rash boundaries and subtle features of conjunctival congestion. The Wood's lamp auxiliary light source excites the fluorescence response of the rash area with ultraviolet light with wavelengths from 320 nm to 400 nm.

9. The highly infectious disease screening system based on AI visual analysis and behavior recognition technology according to claim 1, characterized in that, The disease feature knowledge base supports online iterative updates of the model, receives clinical feature data of emerging infectious diseases, adjusts the model parameters of the AI ​​visual analysis module through transfer learning, and updates the mapping relationship in the disease feature knowledge base to adapt to the screening needs of emerging infectious diseases.

Images