AI-assisted monitoring system and method based on vector tile caching technology

Abstract

The application discloses an AI-assisted monitoring system and method based on a vector tile caching technology, and the AI-assisted monitoring system comprises a GIS vector tile generation unit, an AI edge operator, a time sequence caching scheduling unit, a business research and judgment unit, a multi-terminal intelligent display unit, a medical privacy desensitization unit and the like. The AI-assisted monitoring method comprises the steps of early modeling, video acquisition and edge analysis, time sequence caching scheduling and coordinate conversion, business linkage research and judgment, privacy protection and data security control, multi-terminal output and the like. The system and method of the application extract human feature attributes through linkage of AI edge operator visual analysis, and fuse GIS vector tile space-time data sets with time sequence passenger flow weights, generate a double business analysis engine, can complete face pre-booking comparison to prevent false detection, department queuing passenger flow statistics, congestion triage scheduling, dangerous situation spatial positioning alarm, medical track backtracking and the like, and the research and judgment results can be used for terminal user query, retrieval and display.

Description

Technical Field

[0001] This invention relates to AI-assisted monitoring models, and in particular to an AI-assisted monitoring system and method based on vector tile caching technology. It is especially suitable for applying vector tile caching algorithms to health check-up scenarios, and achieves lightweight, fast loading and high concurrency of health check-up data through logical slicing and multi-level caching. Background Technology

[0002] With the widespread adoption of smart healthcare and health checkups, the information technology for human health data monitoring is rapidly developing. Health checkup data exhibits complex characteristics such as massive volume, multi-dimensionality, high concurrency, and multi-terminal access. Existing health checkup solutions and intelligent monitoring technologies in hospital checkup centers generally suffer from defects such as poor business adaptability of caching algorithms, high computing costs, slow loading, offline unavailability, non-compliance with medical privacy, delayed emergency location, and visualization lag.

[0003] Vector tiles are a standardized, segmented, multi-scale pyramid-style GIS technology for transmitting vector geographic data, developed in contrast to raster tiles. Both raster and vector tiles essentially involve cropping and segmenting data before sending it to the front end, where it is then stitched together for display. While raster tiles render images as individual pixels, vector tiles contain geometric primitive information. They don't transmit images directly, only the geometric primitive information and attribute vector data, which are then rendered in real-time by the front end. Vector tile caching technology pre-computes or generates vector tiles in real-time and stores them across multiple storage media to address issues such as rendering lag and high database load on massive amounts of geographic data. It is commonly used for efficient distribution and rendering of geospatial data, effectively improving report and chart loading speeds, enabling offline data viewing, and supporting high-concurrency access and large-screen visualization. Essentially, vector tile caching trades space for time, balancing data real-time performance and access performance through a multi-level caching architecture and intelligent invalidation mechanisms, thereby improving the performance and efficiency of data display. Vector tile caching technology can break through the boundaries of traditional geographic information applications. In the field of smart health checkups, it can significantly improve data loading efficiency, system stability, and user experience, demonstrating strong practical value. This invention employs vector tile caching technology, eliminating the need for geographic latitude and longitude coordinates. Instead, it logically slices health checkup data into grids, with each tile representing lightweight structured data, thereby assisting in smart health checkups and regional health monitoring. Summary of the Invention

[0004] Given the shortcomings of existing AI-assisted monitoring solutions for health check-up centers, such as fragmented visual space, poor business adaptability of caching algorithms, high computing power costs, and delayed emergency storage and positioning, this invention provides an AI-assisted monitoring system based on vector tile data caching. It deeply integrates visual recognition with vector tile caching technology and adopts a hotspot preloading and time-series weighted LRU caching strategy to construct a lightweight data caching solution suitable for health check-up centers. This effectively achieves lightweight, fast loading, high concurrency support, and offline availability of check-up data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The AI-assisted monitoring system based on vector tile caching technology includes: a GIS vector tile generation unit, an AI edge operator, a time-series caching and scheduling unit, a HEIS health check database, a business analysis unit, and a multi-terminal intelligent display unit.

[0006] The GIS vector tile generation unit first constructs an indoor GIS vector dataset by surveying the floor layout of the health checkup center, completing the preliminary modeling. Then, it uses a quadtree algorithm to perform multi-level scaling and tile cutting on the GIS vector dataset, generating hierarchical Mapbox standard vector tiles, and dividing the tiles into three-level cache partitions (hot, warm, and cold) to generate a vector tile spatial heat dataset. This GIS tile generation unit is connected to the HEIS health checkup database and has pre-entered the time-series passenger flow weight parameters for health checkups in the relevant departments.

[0007] The AI ​​edge operator has a built-in edge vision perception unit, which is used to capture, collect and transmit indoor video streams of the physical examination center, perform edge computing and video analysis, extract human feature data, and output and provide feedback on the results.

[0008] The temporal cache scheduling unit is connected to the GIS vector tile generation unit and the AI ​​edge computing power box. It has a built-in LRU cache algorithm that integrates temporal passenger flow weights to couple the temporal passenger flow weights with the spatial heat data in the standard GIS vector tile dataset, complete the data cache scheduling, and generate a GIS vector tile spatiotemporal dataset that integrates temporal passenger flow weights. Then, it performs a two-way mapping linkage between human pixel coordinates and vector tile geographic coordinates to complete the adaptive scheduling of the GIS vector tile data exclusive to the hospital's physical examination center, generate a two-factor vector tile spatiotemporal dataset that is bound to the human characteristic attributes of the examinee, and store the generated two-factor vector tile spatiotemporal dataset in the HEIS database of the physical examination center.

[0009] The business analysis and early warning unit performs business data analysis by retrieving the three-factor GIS vector tile dataset in the HEIS database, and completes functions such as face appointment comparison to prevent proxy inspection, statistics of queuing patients in departments, congestion triage scheduling, spatial positioning alarm for dangerous situations, and backtracking analysis of physical examination trajectories.

[0010] The multi-terminal display application layer is used to query and / or display health heatmaps, anomaly alarms, trajectory tracking, resource scheduling, and trend prediction.

[0011] Preferably, the floor layout of the physical examination center surveyed in the early modeling stage of the GIS vector tile generation unit includes floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageways, resulting in a GIS vector data set of floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageway layouts. The quadtree algorithm is used to perform multi-level scaling tile cutting on the vector data set to generate hierarchical Mapbox standard vector tiles, resulting in a GIS vector tile standard dataset, which is then divided into hot, warm, and cold three-level tile partition caches.

[0012] Specifically, the GIS vector data set adopts a quadtree hierarchical cutting method, with scaling levels divided into 1-18 levels. Levels 1-6 are global outline tiles for floors, levels 7-12 are department zoning tiles, and levels 13-18 are workstation fine-grained tiles. Only the sub-tiles corresponding to the addition or removal of departments and the modification of workstations are modified, so as to achieve incremental tile updates of the entire map without reloading.

[0013] The AI ​​edge vision perception unit further includes: multiple high-definition cameras and edge computing gateways deployed at key passage points in the physical examination center, which perform real-time near-field human body detection and face capture of examinees, as well as identification of workstation occupancy in examination rooms and identification of fall and gathering intrusion into restricted areas.

[0014] Preferably, the edge computing gateway has a built-in lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model, which are deployed at key passage points in the physical examination center to perform real-time analysis and processing of video streams transmitted from multiple cameras, extract and output the examinee's pixel coordinates, ReID, posture, behavior and physiological characteristics, as well as human characteristic attributes of the examination room in idle state.

[0015] Preferably, the lightweight YOLO visual inference model uses the YOLO-Nano architecture to complete INT8 quantization compression, the edge gateway single-channel video inference latency is ≤50ms, and a medical-specific dark light occlusion dataset is added for iterative training in the dark light and personnel occlusion scenarios in the physical examination room, with a human body recognition accuracy of ≥98.5% and a fall abnormal behavior recognition accuracy of ≥97%.

[0016] Optionally, the improved LRU caching algorithm based on time-series passenger flow weighting includes: collecting passenger flow data from the hospital's historical physical examination periods over the past 7-30 days, dividing it into three time-series periods: morning peak (8:00-11:00), afternoon off-peak (11:00-14:00), and evening off-peak (14:00-17:30), coupling the dual factors of departmental spatial heat, assigning a weight coefficient of 1.8-2.0 to vector tiles of popular departments such as blood collection, color ultrasound, and CT, and assigning a weight coefficient of 0.4-0.8 to tiles of general internal medicine logistics channels, dynamically extending the cache residence time of high-weight tiles during peak periods, prioritizing the elimination of low-weight idle tiles, and adaptively balancing server memory usage and tile cache hit rate.

[0017] The mapping of human pixel coordinates to tile geographic coordinates includes: calibrating the internal and external parameters of the camera, establishing a conversion matrix between the image pixel plane and the indoor GIS vector plane, identifying and locking the feature ID of a single examinee using ReID, synchronizing the coordinate points in real time, overlaying them onto the cached vector tiles to generate a continuous and visualized physical examination route, and supplementing the trajectory breakpoints based on the path prediction of the passage tile when the person temporarily obscures the view.

[0018] The business analysis and early warning unit retrieves the three-factor vector tile spatiotemporal dataset from the HEIS database and performs business data analysis in conjunction with related information. The business data analysis links the human feature attributes extracted by AI edge operator video analysis with the GIS vector tile spatiotemporal dataset that incorporates time-series passenger flow weights to generate two major analysis engines: a physical examination feature health index analysis engine and a GIS vector spatial analysis engine, for use by the multi-terminal users for querying, searching, or display.

[0019] An AI-assisted monitoring method based on vector tile caching technology is used to implement the above system, and it includes the following steps: Step 1: Complete the preliminary modeling: Survey the floor layout of the physical examination center, construct an indoor GIS vector map, and then use the quadtree algorithm to cut and generate 1-18 level hierarchical vector tiles to divide the area into three levels of cold, warm, and hot buffer zones; enter the departmental time-series passenger flow weight parameters and connect to the hospital's HEIS database; Step 2: Employ an AI edge operator with a built-in edge vision perception unit to capture, collect, and transmit indoor video streams from the health checkup center, perform edge computing and video analysis, extract human feature attribute data, and output and feedback results. The edge vision perception unit includes multiple high-definition cameras and an edge computing gateway. The edge computing gateway has a built-in lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model. It is deployed at key passage points in the health checkup center to perform real-time analysis and processing of video streams transmitted from multiple high-definition cameras, extract and output the examinee's pixel coordinates, ReID, posture, behavior, and physiological characteristics, as well as human feature attributes in the idle state of the examination room. Step 3: By integrating an improved LRU caching algorithm with time-series passenger flow weights, the time-series passenger flow weights and vector tile spatial heat data are dually fused. A two-way mapping linkage between human pixel coordinates and vector tile geographic coordinates is performed, converting human pixel coordinates into vector tile geographic coordinates. The vector tile spatial heat data with time-series passenger flow weights is bound together to achieve adaptive scheduling of the hospital health check-up center's exclusive GIS vector tile data set, generating a three-factor GIS vector tile dataset, and recording it into the HEIS database. Step 4: Retrieve the three-factor GIS vector tile dataset from the HEIS database, perform business data analysis, complete facial recognition appointment comparison to prevent proxy examinations, statistics on patient flow in clinic queues, and analysis of physical examination trajectory backtracking; periodically collect and push available workstations to complete intelligent triage scheduling; and trigger pop-up alerts by identifying falls, gatherings, and intrusions into restricted areas, linked to GIS vector tile hotspot locations. Step 5: Output to multiple terminals, using large display screens, nurse tablets, and mobile terminals to query and display real-time heat maps of physical examination patients, workstation status, abnormal alarms, and trajectory tracking. The backend can trace the entire physical examination trajectory of the examinees as needed.

[0020] The beneficial effects of this invention: Vector tile caching technology was originally used for efficient rendering and distribution of geospatial data. This invention applies vector tile hierarchical caching technology to health checkup scenarios in hospital physical examination centers. Through logical slicing and multi-level caching, it achieves lightweight, fast loading, high concurrency support, and offline availability of physical examination data. Specifically, this invention creatively links human feature attributes extracted by AI edge operator visual analysis with a GIS vector tile spatiotemporal dataset that incorporates time-series passenger flow weights to generate two major analysis engines: a physical examination feature health index analysis engine and a GIS vector spatial analysis engine. These engines can perform business analysis such as facial recognition appointment comparison to prevent proxy checkups, departmental queue passenger flow statistics, congestion triage scheduling, spatial location alarm for emergencies, and physical examination trajectory backtracking. The analysis results can be queried, retrieved, and displayed by end users. Other significant beneficial effects include: (1) By overlaying tile hierarchical partitioning cache with edge local visual reasoning, there is no need to transmit high-definition video and full GIS map, which greatly reduces the server computing load and network bandwidth pressure; (2) Avoids re-cutting and serializing vector data every time the geographical location and movement trajectory of the examinee are requested, effectively reducing the repetitive calculation of the route; (3) Integrating the improved LRU caching algorithm with time-sequential passenger flow weight, compared with the traditional LRU caching algorithm, the cached tiles can be returned in milliseconds, the vector map scaling and panning are smoother, and the map rendering is smooth without lag or delay. (4) Greatly improves the efficiency of business management and can realize business optimization functions such as automatic anti-proxy examination, intelligent triage, second-level location of danger, and one-click trajectory back. (5) The medical privacy desensitization unit of the present invention can hierarchically manage face desensitization, timed data destruction, and operation log recording, which is fully compatible with the medical portrait data confidentiality and hospital network security compliance requirements. (6) The present invention can also effectively protect the data privacy of examinees and effectively prevent unauthorized access and attacks through the triple security mechanism of security authentication, data encryption and access control adopted by the AI ​​edge operator. In addition, the monitoring system of the present invention supports offline viewing and high-concurrency access, reliably realizes the visualization of personal health check-up reports, and can be adapted to regional health monitoring, mobile health check-up terminals, and expert business systems of health check-up centers. Attached Figure Description

[0021] The following figures are included as part of this invention to understand and explain the principles of the invention.

[0022] Figure 1 A schematic diagram of the overall technical architecture of a specific embodiment of the AI-assisted monitoring system according to the present invention is shown.

[0023] Figure 2 A simplified diagram of the organizational structure of a GIS vector tile generation unit, as disclosed in a specific embodiment of the AI-assisted monitoring system according to the present invention, is shown.

[0024] Figure 3 A simplified diagram of the time-series cache scheduling unit organization structure disclosed in a specific embodiment of the AI-assisted monitoring system according to the present invention is shown.

[0025] Figure 4 A flowchart illustrating a specific embodiment of the AI-assisted monitoring method according to the present invention is shown. Detailed Implementation

[0026] In the accompanying drawings, the same or similar reference numerals are used to denote the same or similar elements or elements having the same or similar functions. In the specific embodiments described below, exemplary embodiments of the invention will be described and illustrated in more detail with reference to the accompanying drawings.

[0027] This invention discloses an AI-assisted monitoring system based on vector tile caching technology. Figure 1 A schematic diagram of the overall architecture of the AI-assisted monitoring system according to the present invention is shown. (Combined with...) Figure 1 The AI-assisted monitoring system of this invention includes: a GIS vector tile generation unit, an AI edge operator, a time-series cache scheduling unit, a business analysis unit, and a multi-terminal intelligent display unit. Preferably, the system disclosed in this invention may further include a medical privacy desensitization unit, etc.

[0028] Figure 2 A structural analytical diagram of the GIS vector tile generation unit according to the present invention is shown. (Combined with...) Figure 2 The GIS vector tile generation unit of this invention completes preliminary modeling by pre-constructing an indoor GIS vector dataset of the entire floor of a medical examination center, i.e., constructing an indoor GIS vector map of the medical examination center. In one specific embodiment, the vector tile generation unit surveys the floor layout of the medical examination center, specifically involving surveying vector data including floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageway layouts. It then performs data cleaning and coordinate calibration work, such as coordinate transformation from CAD architectural drawings of the medical examination center to an indoor GIS vector map, topology verification, and attribute binding, to construct a complete indoor GIS vector dataset, thus completing the preliminary modeling. Next, a quadtree algorithm is used to perform multi-level scaling and tile cutting on the vector data such as floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageways using a vector tile cutting engine. This generates hierarchical Mapbox standard vector tiles, resulting in a standard GIS vector tile dataset. The dataset is then divided into hot, warm, and cold three-level cache partitions, and managed and automatically upgraded / downgraded using a time-series scheduling engine. In a preferred embodiment of the GIS vector tile generation unit of the present invention, the GIS vector dataset adopts a quadtree hierarchical cutting method, with scaling levels divided into 1-18 levels. Levels 1-6 are global floor outline tiles, levels 7-12 are department zoning tiles, and levels 13-18 are workstation fine-grained tiles. Only the sub-tiles corresponding to the addition or removal of departments and changes in workstations are modified, achieving incremental tile updates of the entire map without reloading, thus generating a GIS vector tile dataset. Simultaneously, the time-series passenger flow weight parameters for departmental physical examinations are entered and connected to the hospital's Health Examination Information System (HEIS). In a preferred embodiment of the present invention, the time-series passenger flow weight parameters are set as follows: Blood collection / imaging tiles during the morning peak (8:00-11:00) have a weight of 2.0, off-peak weight 1.2, and low-peak weight 0.6; the weight for the logistics corridor during all times is 0.4; the maximum cache memory threshold is 8G, and tiles exceeding the threshold are prioritized for elimination of low-weight tiles.

[0029] Furthermore, in the early modeling stage, the GIS vector tile generation unit of this invention requires pre-designing tile dimensions, levels, and ID rules; and establishing a three-level caching service to realize tile generation, updating, and eviction logic. In a specific embodiment of this invention, the GIS vector tile caching scheme uses (Z, X, Y) tile coordinates as the key, employs the LRU (Least Recently Used) algorithm to manage the memory cache, and combines SQLite / Redis / S3 for persistence; when a tile is not found, it is generated from the HEIS database in real time and written back to the cache, supporting partial failure updates. Specifically, in the embodiments of this invention, GIS vector tile cutting can also logically slice physical examination data according to four dimensions: time, indicators, population, and institution, without relying on geographical latitude and longitude. Each tile is lightweight structured data, including: indicator value, anomaly marker, unit, interpretation, timestamp, etc.

[0030] The AI ​​edge operator disclosed in this invention is used to capture, collect and transmit indoor video streams in a health checkup center, perform edge computing and video analysis, extract human feature data, and output and provide feedback on the results.

[0031] The AI ​​edge operator has a built-in edge vision perception unit, including multiple high-definition cameras and an edge computing gateway deployed at key passage points in the physical examination center. The edge computing gateway is deployed at key passage points in the physical examination center and has a built-in lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model. It is used to perform real-time analysis and processing of video streams transmitted from multiple cameras and extract features such as ReID, posture, behavior, and physiological characteristics.

[0032] In one specific embodiment, the AI ​​edge operator integrates an edge vision perception unit and includes multiple high-definition cameras, several sensors, an embedded edge computing gateway, and other edge devices, which are deployed at key locations within the health checkup center. In a preferred embodiment, the multiple high-definition cameras and the edge computing gateway are deployed at key locations such as the waiting hall, clinic entrances and exits, radiology department, blood collection area, and restricted area passages within the health checkup center. The multiple high-definition cameras are used to collect and transmit video streams of the health checkup area in real time.

[0033] The edge computing gateway incorporates a lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human pose recognition model. Multiple high-definition cameras transmit the captured video streams to the edge computing gateway in real time. The gateway locally performs human feature detection, face capture, physical examination station occupancy recognition, and behavior recognition such as falls, gatherings, and intrusions into restricted areas. It outputs the examinee's human feature data, including pixel coordinates, ReID, body posture features, facial features, behavior, and examination room vacancy status. In a preferred embodiment, the built-in lightweight YOLO visual inference model uses the YOLO-Nano architecture for INT8 quantization compression, achieving a single-channel video inference latency of ≤50ms for the edge gateway. Iterative training is conducted using a medical-specific low-light occlusion dataset for scenarios with dim lighting and personnel obstruction in the examination room, resulting in a human recognition accuracy of ≥98.5% and a fall / abnormal behavior recognition accuracy of ≥97%.

[0034] In specific embodiments, the AI ​​edge operator of this invention employs a high-performance processor with powerful computing and storage capabilities, facilitating the handling of complex tasks including image recognition and data analysis. It can perform localized data processing, analysis, and signal feedback, reducing data transmission latency and thus helping the AI-assisted monitoring system of this invention to promptly complete assessments and make effective decisions. If necessary, complex data can also be transmitted to the cloud for cloud computing before being transmitted back to the local machine. The aforementioned edge computing and analysis results generate corresponding output data and feedback signals, which are used to control intelligent devices or provided to the user interface for querying and display.

[0035] Figure 3 A structural diagram of the timing cache scheduling unit according to the present invention is shown. Combined with... Figure 1 , Figure 2 and Figure 3 The temporal cache scheduling unit disclosed in this invention connects the GIS vector tile generation unit and the AI ​​edge operator, and incorporates an improved LRU cache scheduling algorithm that integrates temporal passenger flow weights. In a specific embodiment of the temporal cache scheduling unit of this invention, firstly, the temporal passenger flow weights and spatial heat data in the GIS vector tile standard dataset are deeply coupled to complete data cache scheduling, generating a GIS vector tile spatiotemporal dataset that integrates temporal passenger flow weights. Then, a bidirectional mapping linkage between human pixel coordinates and tile geographic coordinates is performed to complete the adaptive scheduling of the GIS vector tile data specific to the hospital's health checkup center, generating a two-factor vector tile spatiotemporal dataset bound to the human characteristic attributes of the examinees, and storing the generated two-factor vector tile spatiotemporal dataset in the HEIS database of the health checkup center.

[0036] In one specific implementation, the process of mapping human pixel coordinates to tile geographic coordinates is as follows: calibrate the internal and external parameters of the camera, establish a conversion matrix between the image pixel plane and the indoor GIS vector plane, identify and lock the human feature ID of a single examinee using ReID, synchronize the coordinate points in real time, overlay them onto the cached vector tiles to generate a continuous and visualized physical examination route, and fill in the trajectory breakpoints based on the path prediction of the passage tile when the person temporarily obscures the view.

[0037] This invention integrates an improved LRU cache scheduling algorithm with time-series passenger flow weighting, creating a coupled optimization algorithm specifically for medical checkups. It abandons the logic of traditional LRU, which relies solely on access time for cache eviction. In one specific implementation, the LRU cache scheduling algorithm includes: collecting passenger flow data from 7-30 days of historical checkup periods at the hospital, dividing it into three time slots: morning peak (8:00-11:00), afternoon off-peak (11:00-14:00), and evening off-peak (14:00-17:30). It couples these with the dual factors of departmental spatial popularity, assigning a weight coefficient of 1.8-2.0 to vector tiles in popular departments such as blood collection, color Doppler ultrasound, and CT scans, and a weight coefficient of 0.4-0.8 to tiles in general internal medicine support corridors. It dynamically extends the cache residence time of high-weight tiles during peak hours, prioritizes the eviction of low-weight idle tiles, and adaptively balances server memory usage with tile cache hit rate. This improved LRU cache algorithm with time-series passenger flow weighting is a coupled optimization algorithm specifically for medical checkups, abandoning the logic of traditional LRU, which relies solely on access time for cache eviction. In a further specific embodiment of the present invention, the aforementioned business analysis and early warning unit performs business data analysis by retrieving the three-factor vector tile spatiotemporal dataset and related information from the HEIS database. This business data analysis is linked to the human feature attributes extracted by AI edge operator visual analysis and the GIS vector tile spatiotemporal dataset with time-series passenger flow weights, generating two major analysis engines: a physical examination feature health index analysis engine and a GIS vector spatial analysis engine. Through the physical examination feature health index analysis engine and the GIS vector spatiotemporal analysis engine, business analysis can be completed, such as facial recognition appointment comparison to prevent proxy examinations, statistics of passenger flow in department queues, congestion triage scheduling, spatial location alarms for emergencies, and backtracking of physical examination trajectories. The analysis results can be queried, retrieved, and displayed by end users.

[0038] The AI-assisted monitoring system of this invention also discloses a multi-level terminal interaction unit, including but not limited to a central control screen, nurse station tablets, security terminals, and a management backend. This unit is used to retrieve cached vector tiles to achieve low-latency map visualization, query, display / push alarm information, workstation status, and passenger flow heatmap data. It features intelligent triage, anti-proxy verification, accurate alarm for potential hazards, full-domain trajectory backtracking, and multi-terminal low-latency visual control, addressing three core requirements: low-cost deployment, medical compliance, and intranet compatibility.

[0039] In another preferred embodiment of the present invention, a medical privacy desensitization unit is disclosed, which can hierarchically control facial visibility permissions, retrieve complete facial data only during verification periods, automatically perform pixel blurring and desensitization on video faces during full-domain monitoring periods, isolate and store facial data from ordinary monitoring videos, and adapt to the data compliance requirements of hospital security level 2.0. In one specific embodiment, the medical privacy desensitization unit of the present invention sets up a three-level data retention mechanism: facial verification data is retained for 7 days, desensitized monitoring videos are retained for 30 days, and trajectory spatial data is retained for 90 days. All data reading, writing, retrieval, and export are logged throughout, and one-click destruction of expired medical personnel data is supported in the background. In another preferred embodiment, the AI ​​edge operator of the present invention adopts a triple security mechanism of security authentication, data encryption, and access control, effectively protecting the data privacy of examinees and effectively preventing unauthorized access and attacks. In addition, the AI ​​edge computing box also has good flexibility, scalability, and ease of deployment. Users can choose different functional modules and interfaces according to their own business needs to achieve personalized customization development.

[0040] Figure 4 A simplified flowchart illustrating the workflow of a specific embodiment of the AI-assisted monitoring method according to the present invention is shown to help reveal the implementation steps of the aforementioned AI-assisted monitoring system. (In conjunction with...) Figure 4 The AI-assisted monitoring method based on vector tile caching technology disclosed in this invention includes the following steps: Step 1: Complete the preliminary modeling: Survey the floor layout of the physical examination center and construct an indoor GIS vector map; use the quadtree algorithm to cut and generate 1-18 level hierarchical vector tiles, and divide them into three levels of cold, warm and hot buffer zones; enter the departmental time-series passenger flow weight parameters and connect to the hospital's HEIS database; Step 2: Video Acquisition and Edge Analysis: An AI edge operator with a built-in edge vision perception unit is used to capture, acquire, and transmit indoor video streams of the medical examination center, and perform edge computing and video analysis to extract human feature attribute data and output and feedback results. The edge vision perception unit includes multiple high-definition cameras and an edge computing gateway. The edge computing gateway has a built-in lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model. It is deployed at key passage points in the medical examination center to perform real-time analysis and processing of video streams transmitted from multiple high-definition cameras, extract and output the examinee's pixel coordinates, ReID, posture, behavior and physiological characteristics, as well as human feature attributes in the idle state of the examination room. Step 3: Temporal Cache Scheduling and Coordinate Transformation: By integrating an improved LRU cache algorithm with temporal passenger flow weights, the temporal passenger flow weights and vector tile spatial heat data are dually fused. A two-way mapping linkage between human pixel coordinates and vector tile geographic coordinates is performed to convert human pixel coordinates into vector tile geographic coordinates. The vector tile spatial heat data with temporal passenger flow weights is bound to achieve adaptive scheduling of the hospital health check-up center's exclusive GIS vector tile data set, generating a three-factor GIS vector tile dataset, which is then entered into the HEIS database. Step 4: Business linkage analysis: Retrieve the three-factor GIS vector tile dataset from the HEIS database, perform business data analysis, complete facial recognition appointment comparison to prevent proxy examinations, statistics on patient flow in clinic queues, and analysis of physical examination trajectory backtracking; periodically collect and push available workstations to complete intelligent triage scheduling; and trigger pop-up alerts by identifying falls, gatherings, and intrusions into restricted areas, and linking them to GIS vector tile hotspot locations.

[0041] Step 5: Privacy Protection and Data Security Management: A three-tiered data retention mechanism is adopted, with hierarchical control over facial recognition permissions and automatic anonymization of faces across the entire monitoring domain. Preferably, facial verification data is retained for 7 days, anonymized monitoring videos for 30 days, and trajectory spatial data for 90 days. All data reading, writing, retrieval, and export are logged in their entirety, and one-click destruction of expired medical personnel data is supported in the background. Preferably, the AI ​​edge operator also employs a triple security mechanism including security authentication, data encryption, and access control.

[0042] Step 6: Output to multiple terminals, using large display screens, nurse tablets, and mobile terminals to query and display real-time heat maps of physical examination patients, workstation status, abnormal alarms, and trajectory tracking. The backend can trace the entire physical examination trajectory of the examinees as needed.

[0043] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-assisted monitoring system based on vector tile caching technology, characterized in that, include: The system includes a GIS vector tile generation unit, an AI edge operator, a temporal cache scheduling unit, a HEIS health check database, a business analysis unit, and a multi-terminal intelligent display unit; among these components... The GIS vector tile generation unit first constructs an indoor GIS vector dataset by surveying the floor layout of the physical examination center; then, it uses a quadtree algorithm to perform multi-level scaling and tile cutting on the GIS vector dataset to generate hierarchical Mapbox standard vector tiles, and divides the tiles into three-level cache partitions of hot, warm and cold to generate a vector tile spatial heat dataset, thus completing the preliminary modeling work. The GIS tile generation unit is connected to the HEIS physical examination database and pre-enters the time-series passenger flow weight parameters of the business departments' physical examinations; The AI ​​edge operator has a built-in edge vision perception unit, which is used to capture, collect and transmit indoor video streams of the physical examination center, perform edge computing and video analysis, extract human feature data, and output and provide feedback on the results. The temporal cache scheduling unit is connected to the GIS vector tile generation unit and the AI ​​edge computing power box. It has a built-in LRU cache algorithm that integrates temporal passenger flow weights to couple the temporal passenger flow weights with the spatial heat data in the standard GIS vector tile dataset, complete the data cache scheduling, and generate a GIS vector tile spatiotemporal dataset that integrates temporal passenger flow weights. Then, it performs a two-way mapping linkage between human pixel coordinates and vector tile geographic coordinates to complete the adaptive scheduling of the GIS vector tile data exclusive to the hospital's physical examination center, generate a two-factor vector tile spatiotemporal dataset that is bound to the human body feature attributes of the examinee, and store the generated two-factor vector tile spatiotemporal dataset in the HEIS database of the physical examination center. The business analysis and early warning unit performs business data analysis by retrieving the three-factor GIS vector tile dataset in the HEIS database, and completes face appointment comparison to prevent proxy inspection, department queueing passenger flow statistics, congestion triage scheduling, spatial positioning alarm for dangerous situations, and physical examination trajectory backtracking analysis. The multi-terminal display application layer is used to query and / or display health heatmaps, anomaly alarms, trajectory tracking, resource scheduling, and trend prediction.

2. The monitoring system according to claim 1, characterized in that, The GIS vector tile generation unit surveys the floor layout of the physical examination center during the early modeling stage, including floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageways. This survey yields a GIS vector data set of floor outlines, clinic boundaries, examination workstations, electronic restricted areas, and passageway layouts. The quadtree algorithm is used to perform multi-level scaling and tile cutting on the vector data set to generate hierarchical Mapbox standard vector tiles, resulting in a standard GIS vector tile dataset. The dataset is then divided into three levels of tile partitioning: hot, warm, and cold.

3. The monitoring system according to claim 1 or 2, characterized in that, The GIS vector data set adopts a quadtree hierarchical cutting method, with scaling levels divided into 1-18 levels. Levels 1-6 are global floor outline tiles, levels 7-12 are department zoning tiles, and levels 13-18 are workstation fine-grained tiles. Only the sub-tiles corresponding to the addition or removal of departments and the modification of workstations are modified, so that incremental tile updates of the entire map can be achieved without reloading.

4. The monitoring system according to claim 1, characterized in that, The AI ​​edge vision perception unit includes: multiple high-definition cameras and edge computing gateways deployed at key passage points in the physical examination center, which perform real-time near-field human detection and face capture of examinees, as well as identification of workstation occupancy in examination rooms and identification of fall and gathering intrusion into restricted areas.

5. The monitoring system according to claim 3, characterized in that, The edge computing gateway incorporates a lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model. It is deployed at key passage points in the physical examination center to perform real-time analysis and processing of video streams transmitted from multiple cameras, extracting and outputting the examinee's pixel coordinates, ReID, posture, behavior, and physiological characteristics, as well as human characteristic attributes of the examination room in its idle state.

6. The monitoring system according to claim 4, characterized in that, The lightweight YOLO visual inference model uses the YOLO-Nano architecture to complete INT8 quantization compression, with a single-channel video inference latency of ≤50ms at the edge gateway. Iterative training is conducted using a medical-specific low-light occlusion dataset for scenarios such as dim lighting and personnel occlusion in the examination room. The accuracy rate for human body recognition is ≥98.5%, and the accuracy rate for recognizing abnormal behavior such as falling to the ground is ≥97%.

7. The monitoring system according to claim 1, characterized in that, The improved LRU caching algorithm based on time-series passenger flow weighting includes: collecting passenger flow data from the hospital's historical physical examination periods over the past 7-30 days, dividing it into three time-series periods: morning peak (8:00-11:00), afternoon off-peak (11:00-14:00), and evening off-peak (14:00-17:30). It couples these data with the dual factors of departmental spatial heat, assigning a weight coefficient of 1.8-2.0 to vector tiles in popular departments such as blood collection, color Doppler ultrasound, and CT scans, and a weight coefficient of 0.4-0.8 to tiles in the general internal medicine logistics corridor. It dynamically extends the cache dwell time of high-weight tiles during peak periods, prioritizes the elimination of low-weight idle tiles, and adaptively balances server memory usage with tile cache hit rate.

8. The monitoring system according to claim 1, characterized in that, The mapping of human pixel coordinates to tile geographic coordinates includes: calibrating the internal and external parameters of the camera, establishing a conversion matrix between the image pixel plane and the indoor GIS vector plane, identifying and locking the feature ID of a single examinee using ReID, synchronizing the coordinate points in real time, overlaying them onto the cached vector tiles to generate a continuous and visualized physical examination route, and supplementing the trajectory breakpoints based on the path prediction of the passage tile when the person temporarily obscures the view.

9. The monitoring system according to claim 1, characterized in that, The business analysis and early warning unit retrieves the three-factor vector tile spatiotemporal dataset from the HEIS database and performs business data analysis in conjunction with related information. The business data analysis links the human feature attributes extracted by AI edge operator video analysis with the GIS vector tile spatiotemporal dataset that incorporates time-series passenger flow weights to generate two major analysis engines: a physical examination feature health index analysis engine and a GIS vector spatial analysis engine, for use by the multi-terminal users for querying, searching, or display.

10. An AI-assisted monitoring method based on vector tile caching technology, used to implement the system described in claims 1-9, characterized in that, Includes the following steps: Step 1: Complete the preliminary modeling: Survey the floor layout of the physical examination center, construct an indoor GIS vector map, and then use the quadtree algorithm to cut and generate 1-18 level hierarchical vector tiles to divide the area into three levels of cold, warm, and hot buffer zones; enter the departmental time-series passenger flow weight parameters and connect to the hospital's HEIS database; Step 2: Employ an AI edge operator with a built-in edge vision perception unit to capture, collect, and transmit indoor video streams from the health checkup center, perform edge computing and video analysis, extract human feature attribute data, and output and feedback results. The edge vision perception unit includes multiple high-definition cameras and an edge computing gateway. The edge computing gateway has a built-in lightweight YOLO visual inference model, a pedestrian ReID cross-camera re-identification model, and a human posture recognition model. It is deployed at key passage points in the health checkup center to perform real-time analysis and processing of video streams transmitted from multiple high-definition cameras, extract and output the examinee's pixel coordinates, ReID, posture, behavior, and physiological characteristics, as well as human feature attributes in the idle state of the examination room. Step 3: By integrating an improved LRU caching algorithm with time-series passenger flow weights, the time-series passenger flow weights and vector tile spatial heat data are dually fused. A two-way mapping linkage between human pixel coordinates and vector tile geographic coordinates is performed, converting human pixel coordinates into vector tile geographic coordinates. The vector tile spatial heat data with time-series passenger flow weights is bound together to achieve adaptive scheduling of the hospital health check-up center's exclusive GIS vector tile data set, generating a three-factor GIS vector tile dataset, and recording it into the HEIS database. Step 4: Retrieve the three-factor GIS vector tile dataset from the HEIS database, perform business data analysis, complete facial recognition appointment comparison to prevent proxy examinations, statistics on patient flow in clinic queues, and analysis of physical examination trajectory backtracking; periodically collect and push available workstations to complete intelligent triage scheduling; and trigger pop-up alerts by identifying falls, gatherings, and intrusions into restricted areas, linked to GIS vector tile hotspot locations. Step 5: Output to multiple terminals, using large display screens, nurse tablets, and mobile terminals to query and display real-time heat maps of physical examination patients, workstation status, abnormal alarms, and trajectory tracking. The backend can trace the entire physical examination trajectory of the examinees as needed.

Images