ICU ward intelligent sensing control system based on vision

Abstract

The invention relates to the technical field of ICU ward sensing control, and discloses an ICU ward intelligent sensing control system based on vision. The system comprises an acquisition module, a dynamic segmentation module, a space-time correlation module, a collaborative correction module and a decision engine module. The acquisition module captures body surface temperature distribution of a patient, limb movement tracks of medical staff and surface contact frequency data of medical equipment in real time through a multispectral camera. The dynamic segmentation module divides a patient body surface infection risk area according to a temperature gradient threshold value, recognizes medical care violation operation in combination with the movement track acceleration change rate, and generates an equipment pollution level label. And the time-space association module establishes a data time-space mapping relation and outputs an infection risk topological graph with a three-dimensional position mark. And the collaborative correction module dynamically optimizes the evaluation standard according to the result, and the decision engine module generates a real-time sterilization instruction, a violation acousto-optic warning and an equipment isolation scheduling scheme. The system realizes multi-dimensional data fusion and dynamic intervention, and enhances the accuracy and timeliness of ICU infection prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ICU ward infection control, in particular to an ICU ward intelligent infection control system based on vision. BACKGROUND

[0002] As the core place for the treatment of critically ill patients in hospitals, ICU wards mainly treat high-risk groups with low immunity and multiple organ dysfunction. The environmental microbial load and the standardization of medical operations are directly related to the prognosis of patients. Infection prevention and control has always been a key issue in this field. Current ICU ward infection control mainly relies on traditional methods such as manual inspection, fixed-point monitoring and post-tracing. These methods have many limitations in practical application. Manual inspection relies on the experience of medical staff, which is highly subjective and has fixed inspection intervals. It is difficult to achieve real-time coverage of patient status, medical operation and equipment pollution, especially at night or during busy work periods of medical staff, which is prone to risk omissions. Fixed-point monitoring often uses single-function sensors, such as monitoring only the environmental temperature or the surface state of a single device, which cannot integrate multi-dimensional data to form a comprehensive infection risk awareness, resulting in one-sided risk identification.

[0003] In terms of patient infection risk assessment, the existing technology often uses average body temperature as the basis for judgment, ignoring the correlation between local temperature abnormalities and infection occurrence, and making it difficult to identify potential infection foci such as skin damage and pressure ulcers in advance. The standardization of medical staff's operation management also has shortcomings. Traditional video monitoring can only achieve post-playback and cannot provide real-time warning when irregular operations occur. The contamination level of medical equipment is often assessed by manual wiping and detection at fixed intervals. This static management mode is disconnected from the actual usage frequency of the equipment, and some high-frequency contact equipment becomes an important transmission medium for cross-infection due to not being handled in time.

[0004] The data processing of existing infection control systems is mostly isolated analysis, lacking the mining of the spatio-temporal correlation between patient risk areas, medical operation behavior and equipment contamination status, and unable to form a dynamic risk topology awareness. When an infection case occurs, tracing the source and transmission path of the infection requires a lot of time and labor costs, and it is difficult to quickly cut off the transmission chain. With the development of medical technology, the diagnosis and treatment activities in ICU wards are becoming increasingly complex, and traditional infection control methods cannot meet the needs of fine and real-time prevention and control. An intelligent infection control technology that can integrate multi-dimensional data, achieve dynamic assessment and precise intervention is needed to improve the initiative and effectiveness of infection prevention and control. SUMMARY

[0005] The purpose of the present application is to provide an ICU ward intelligent infection control system based on vision to solve the problems raised in the background.

[0006] In order to achieve the above object, the application provides a visual-based intelligent infection control system for ICU wards, which comprises: A collection module for acquiring patient body surface temperature distribution data, medical staff limb motion trajectory data and medical equipment surface contact frequency data in the ICU ward in real time through a multispectral camera; A dynamic segmentation module for dividing the patient body surface into high-risk infection areas and low-risk infection areas according to a temperature gradient threshold, identifying medical staff's illegal operation behaviors based on the motion trajectory acceleration change rate, and generating equipment pollution level labels in combination with the contact frequency data; A space-time correlation module for establishing a timestamp mapping relationship between the high-risk infection areas and the illegal operation behaviors, binding the pollution level labels with the spatial coordinates of the corresponding medical equipment, and outputting an infection risk topology map with three-dimensional position markers; A collaborative correction module for dynamically adjusting the temperature gradient threshold and the acceleration change rate determination standard according to the infection risk topology map, and synchronously updating the generation logic of the pollution level labels; A decision engine module for receiving the output data of the collaborative correction module, generating real-time sterilization instructions for the high-risk infection areas, audible and visual warning signals for the illegal operation behaviors, and isolation scheduling schemes for the contaminated equipment.

[0007] Preferably, the operation performed by the dynamic segmentation module includes: Adaptive window width algorithm is used to analyze the local extreme points in the temperature distribution data, the temperature decay rate is detected outwardly from the center of the extreme points, and the high-risk infection area boundary is determined when the decay rate exceeds the preset threshold; The joint angle change sequence in the medical staff's motion trajectory is extracted through the skeleton key point tracking technology, the number of angle mutations and the amplitude between consecutive frames are calculated, and if the mutation threshold and the amplitude threshold are met at the same time, it is determined as illegal operation; The number of independent IDs of different personnel contacting the surface of the medical equipment within a unit time is counted, and the numerical intensity of the pollution level label is calculated by weighted calculation combined with the contact duration.

[0008] Preferably, the operation performed by the space-time correlation module includes: The occurrence time of the high-risk infection area and the occurrence time of the illegal operation behavior are aligned at the millisecond level, and if the time difference between the two is less than the synchronization tolerance window, a strong correlation marker is established; The two-dimensional image coordinates of the medical equipment are converted into the three-dimensional space coordinates of the ward through the point cloud registration technology, so that the pollution level label and the physical position of the equipment form a rigid binding; The infection risk topology map is constructed by fusing the strong correlation marker and the rigid binding data, which contains risk areas coded with different colors, behavior events with timestamps, and device pollution labels displayed in suspension.

[0009] Preferably, the operations performed by the cooperative revision module include: Scaling the reference value of the temperature gradient threshold according to the distribution density of the high-risk infection area in the infection risk topology map; Based on the type distribution of historical violation operation behaviors, dynamically adjusting the combined conditions of the joint angle mutation frequency threshold and the amplitude threshold; According to the aggregation degree of the pollution level label in the three-dimensional space, the weight distribution proportion of the contact ID number and duration is recalculated.

[0010] Preferably, the operations performed by the decision engine module include: According to the real-time sterilization instruction, control the ultraviolet lamp array installed on the top of the sickbed to perform fan-shaped scanning irradiation on the high-risk infection area; Trigger the LED warning ring and directional loudspeaker located at the position where the violation operation occurs through the audible and visual warning signal, and the warning signal strength is proportional to the severity of the violation behavior; According to the isolation scheduling scheme, drive the automatic transport robot to move the contaminated equipment to the designated disinfection cabin, and update the available equipment distribution map in the ward.

[0011] Preferably, the acquisition module further includes: Depth sensors deployed at the four corners of the ward are used to supplement the contact frequency data on the back of the medical equipment, and the sampling frequency is an integer multiple of the frame rate of the multispectral camera; The embedded pre-processing unit is used to correct the non-uniformity of the original temperature distribution data and eliminate the measurement deviation caused by the camera's own thermal noise.

[0012] Preferably, the dynamic segmentation module further includes: The abnormal trajectory filtering unit is used to exclude false motion trajectories caused by camera obstruction or light reflection, and uses a confidence evaluation mechanism based on motion energy accumulation; The temperature compensation unit is used to correct the temperature measurement value according to the patient's body surface humidity data, which is provided by the ward environment sensor.

[0013] Preferably, the spatio-temporal correlation module further includes: The topology optimization unit is used to periodically compress the redundant nodes in the infection risk topology map, and merge adjacent high-risk infection areas with a spatial distance less than the merging radius; The behavior backtracking unit is used to store the complete motion trajectory video clip 30 seconds before the violation operation behavior occurs, and establish an index library in chronological order.

[0014] Preferably, the cooperative revision module further includes: An incremental learning unit is configured to record the change trend of the false positive rate and the false negative rate of the system after each threshold adjustment, and generate a parameter optimization direction suggestion. A conflict arbitration unit is configured to, when the temperature gradient threshold and the pollution level label update logic conflict, preferentially execute an adjustment scheme that has less impact on the identification accuracy of the high-risk infection area.

[0015] Preferably, the decision engine module further comprises: A sterilization effect verification unit is configured to re-scan the high-risk infection area after the ultraviolet lamp irradiation is completed, and compare whether the temperature distribution data change conforms to the expected decay curve. An emergency release switch is configured to, when the system misjudges and causes the equipment to be mis-isolated, allow medical staff to manually abort the transport robot task and reset the equipment state marker.

[0016] Compared with the prior art, the present application has the following beneficial effects: The acquisition module uses a multi-spectral camera as a core perception device, breaking through the functional limitations of a single sensor, and can simultaneously acquire three types of key data, namely, patient body surface temperature distribution, medical staff limb movement trajectory, and medical device surface contact frequency, thereby achieving comprehensive coverage of the core elements of ICU ward infection control and avoiding the risk blind spots caused by data fragmentation in traditional monitoring. This multi-dimensional data acquisition mode provides rich and comprehensive basic information for subsequent risk assessment, so that infection risk identification is no longer dependent on a single indicator.

[0017] The dynamic segmentation module establishes differentiated analysis logic for different monitoring objects, changing the traditional "one-size-fits-all" evaluation method in infection control. By dividing the patient body surface into risk areas using a temperature gradient threshold, the potential infection site of local temperature abnormalities can be accurately locked, which is more consistent with the local characteristics of infection occurrence than the method of only focusing on the average temperature. Based on the acceleration change rate of the movement trajectory, the system can identify irregular operations and achieve quantitative analysis of medical staff behavior, thereby eliminating the subjectivity of manual judgment and making operation standard management more objective and scientific. In combination with the contact frequency, the system generates a device pollution level label, which closely links device management with actual use, thereby avoiding the waste of resources and risk lag caused by fixed-period detection.

[0018] The spatio-temporal correlation module builds a correlation bridge for multi-dimensional risk data, correlates the time stamp of the high-risk infection area with the time stamp of the irregular operation, and binds the pollution level label with the device spatial coordinates. The final output is a three-dimensional position marker infection risk topology map, which converts abstract risk data into intuitive visual information. This presentation method allows medical management personnel to quickly locate the spatio-temporal distribution characteristics of risk points and clearly understand the transmission path and correlation of infection risks, thereby solving the problem of scattered risk information and difficulty in comprehensive judgment in traditional infection control.

[0019] The cooperative correction module enables the system to have self-optimization capability, and can dynamically adjust the temperature gradient threshold, the acceleration change rate determination standard and the pollution level label generation logic according to the real-time generated infection risk topology graph. This dynamic adjustment mechanism enables the system to adapt to dynamic scenes such as changes in patient conditions in the ICU ward, optimization of medical operation processes, changes in equipment usage frequency, etc., avoiding evaluation deviations of fixed parameters in complex environments, and always maintaining the matching degree of evaluation standards and actual needs.

[0020] The decision engine module converts the analysis results of the system into specific intervention measures. The real-time sterilization instruction can quickly respond to high-risk infection areas, reducing the probability of infection spread. The audible and visual warning signals immediately remind when irregular operations occur, helping medical staff to correct non-standard behavior in time, forming immediate operation constraints. The isolation scheduling scheme provides clear guidance for the treatment of contaminated equipment, avoiding cross-infection. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The working principle diagram of the visual-based ICU ward intelligent infection control system described in the application; Figure 2 The flowchart executed by the cooperative correction module; Figure 3 The flowchart executed by the decision engine module; Figure 4 The Gantt chart of the multi-robot device isolation task scheduling; Figure 5 The violin plot of the quadratic polynomial non-uniformity correction coefficient distribution. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0023] Please refer to Figure 1This invention provides a vision-based intelligent infection control system for ICU wards. The system includes: a multispectral camera that acquires real-time data on patient surface temperature distribution, limb movement trajectories of medical staff, and surface contact frequency of medical equipment within the ICU ward; a dynamic segmentation module that divides the patient's surface into high-risk and low-risk infection areas based on temperature gradient thresholds, and identifies violations by medical staff based on the rate of change of acceleration of movement trajectories; and generates equipment contamination level labels by combining contact frequency data. A spatiotemporal correlation module establishes a timestamp mapping relationship between high-risk infection areas and violations, and binds contamination level labels to the spatial coordinates of corresponding medical equipment, outputting an infection risk topology map with three-dimensional location markers. A collaborative correction module dynamically adjusts the temperature gradient threshold and acceleration change rate judgment criteria based on the infection risk topology map, synchronously updating the generation logic of the contamination level labels. A decision engine module receives the output data from the collaborative correction module and generates real-time sterilization instructions for high-risk infection areas, audible and visual warning signals for violations, and isolation and scheduling schemes for contaminated equipment.

[0024] Example 1: In specific implementation, the dynamic segmentation module uses an adaptive window width algorithm to analyze the real-time patient surface temperature distribution data acquired by the multispectral camera. The adaptive window width algorithm identifies local extreme points by scanning the temperature distribution data matrix. Local extreme points are defined as pixel positions where the temperature value is higher or lower than all neighboring points within a specified neighborhood. An initial analysis window is set with each local extreme point as the center. The window size is dynamically adjusted according to the variance of the temperature values ​​around the extreme point. When the variance is large, a smaller window width is used to capture detailed features, and when the variance is small, a larger window width is used to smooth noise. The temperature decay rate is calculated from the central extreme point outward in the radiation direction. The decay rate is obtained by calculating the temperature change per unit radial distance. When the decay rate exceeds a preset critical value, the current pixel position is marked as a boundary point of a high-risk infection area. All connected boundary points jointly delineate the outline of the high-risk infection area. In practical implementation, the preset critical value is set at a decrease of 0.5 degrees Celsius per second based on clinical infection risk standards. The window size adjustment of the adaptive window width algorithm is optimized using the gradient descent method to ensure that the window size converges to a minimum of 3x3 pixels in high-density extreme point regions and expands to a maximum of 15x15 pixels in low-density regions. The temperature decay rate is calculated using the Sobel operator for gradient estimation to enhance the robustness of boundary detection. In some embodiments, the adaptive window width algorithm introduces a multi-scale analysis strategy. First, it quickly locates potential high-risk areas on the low-resolution temperature map, and then refines the boundaries on the original high-resolution data to improve computational efficiency. It can be understood that the local extreme point detection of the adaptive window width algorithm relies on the smoothing preprocessing of the temperature distribution data, and Gaussian filtering is used to eliminate random noise interference.

[0025] The dynamic segmentation module extracts the joint angle change sequence from the limb movement trajectory data of medical staff using skeletal keypoint tracking technology. The skeletal keypoint tracking technology is based on a pre-trained deep neural network model. The model input is a sequence of video frames captured by a multispectral camera, and the output is the two-dimensional coordinates of 17 joint points of the medical staff's body in each frame. The joint points include the head, neck, shoulder, elbow, wrist, hip, knee, and ankle. The joint angle change sequence between consecutive frames is obtained by calculating the change in the direction angle of the line connecting adjacent joint points. For example, the elbow joint angle is represented by the vector angle formed by the shoulder, elbow, and wrist. The module calculates the number and magnitude of angle mutations between consecutive frames. The mutation count is the number of frames where the angle change exceeds a temporary threshold per unit time, and the mutation magnitude is the absolute value of the angle change in each mutation event. If both the mutation count threshold and the magnitude threshold are met, it is judged as a violation of the rules. In practical implementation, the mutation frequency threshold is set to 5 times per second, and the mutation amplitude threshold is set to 30 degrees. The neural network model for skeletal keypoint tracking technology adopts the HRNet architecture, which is fine-tuned on a dataset including ICU scenes to improve joint detection accuracy. Joint angle calculation uses the vector dot product formula to ensure that the angle value ranges from 0 to 180 degrees. Optionally, the skeletal keypoint tracking technology introduces temporal consistency verification, using optical flow information to compensate for short-term jitter in joint coordinates and reduce false detections. It can be understood that the extraction of joint angle change sequences depends on the stability of the video frame rate. The system ensures a constant frame rate of 30 frames per second output from the multispectral camera through a hardware synchronization mechanism.

[0026] The dynamic segmentation module counts the number of unique IDs (IDs) on the surface of medical equipment that are touched by different personnel within a unit of time. It then calculates the numerical intensity of the contamination level label by weighting the contact duration. The number of unique IDs distinguishes different medical personnel using facial recognition or RFID tag technology. The unit of time is set as one work shift (e.g., 8 hours). The contact duration is counted from the first contact between a person's hand and the equipment surface until the hand leaves the surface. The numerical intensity of the contamination level label is calculated using a weighted summation formula, with the weighting coefficient for the number of unique IDs being... The weighting factor for contact duration is Numerical intensity for: in For the number of unique IDs, For total contact duration, the weighting coefficient is... and The system is dynamically adjusted based on infection control strategies. In practice, the unit of time is divided into multiple sliding windows, each with a length of 15 minutes. The number of unique IDs is counted using a hash table to store identified IDs, avoiding duplicate counting. Contact duration is recorded using a high-precision timer with millisecond-level accuracy. The numerical intensity of the contamination level label is normalized to a range of 0-100 for easy comparison. In some embodiments, a time decay function is introduced into the weighted calculation, with recent contact events having a higher weight than historical events, reflecting the timeliness of contamination risk. Optionally, medical device surface contact detection uses a combination of background subtraction and morphological operations. First, the moving hand region is segmented from the video stream. Then, the overlap area between the hand and the device surface is calculated. When the overlap area exceeds a threshold, it is determined to be a contact event.

[0027] The dynamic segmentation module includes an abnormal trajectory filtering unit, which eliminates false motion trajectories caused by camera obstruction or light reflection. This unit employs a confidence assessment mechanism based on motion energy accumulation. Motion energy accumulation is achieved by integrating the displacement vector magnitude of specific joints in a continuous frame sequence. The displacement vector magnitude represents the Euclidean distance between adjacent frames, and the integration interval is set to the most recent 10 frames. The confidence score is defined as the ratio of the integrated value to a preset energy threshold. When the confidence score is below 0.7, the trajectory is considered a false signal and is discarded. In practice, the integral calculation of motion energy accumulation uses a trapezoidal rule approximation to improve calculation accuracy. The preset energy threshold is calibrated to 0.5 meters per second based on the normal movement speed of medical personnel. The abnormal trajectory filtering unit processes all relevant node trajectories in parallel, retaining only the trajectory chain with the highest confidence score for subsequent analysis. It is understood that the performance of the abnormal trajectory filtering unit depends on the accuracy of joint positioning. The system predicts joint motion trajectories using a Kalman filter to reduce random errors. Optionally, the abnormal trajectory filtering unit introduces multi-camera fusion verification. When the view of a single camera is obstructed, the authenticity of the trajectory is cross-verified using data from other cameras.

[0028] The dynamic segmentation module includes a temperature compensation unit. This unit corrects the temperature measurement based on the patient's surface humidity data. The humidity data is provided by a wireless humidity sensor network deployed around the bed, using the ZigBee protocol and uploading humidity readings once per second. The compensation algorithm built into the temperature compensation unit linearly corrects the original temperature measurement by looking up a pre-calibrated temperature-humidity correction coefficient table. This correction coefficient table is established through laboratory calibration experiments, measuring the surface temperature of a standard heat source at different humidity levels to establish a functional relationship between temperature error and humidity. In specific implementations, the temperature-humidity correction coefficient table is stored as a two-dimensional lookup table, with row indices representing humidity values ​​and column indices representing the original temperature values. The output is the corrected temperature value. The temperature compensation unit synchronously reads the current humidity sensor data after acquiring temperature distribution data and performs real-time table lookup correction. In some embodiments, the temperature compensation unit supports dynamically updating the correction coefficient table, maintaining measurement accuracy through periodic calibration. Optionally, the temperature compensation unit introduces a spatial interpolation algorithm to estimate the current area's humidity value using neighboring sensor readings when humidity sensor network data is missing.

[0029] Example 2: The spatiotemporal correlation module performs millisecond-level alignment between the occurrence time of high-risk infection areas and the occurrence time of violations. The occurrence time of high-risk infection areas is extracted from the timestamps output by the dynamic segmentation module, while the occurrence time of violations is obtained from the event logs detected by skeletal keypoint tracking technology. The alignment process employs a high-precision clock synchronization mechanism to ensure that the timestamp reference is unified to Coordinated Universal Time (UTC). If the time difference between the occurrence time of high-risk infection areas and the occurrence time of violations is less than the synchronization tolerance window, a strong correlation marker is established. The default value of the synchronization tolerance window is set to 500 milliseconds. The strong correlation marker includes the correlation type, time difference confidence score, and a unique identifier for the correlated event. In specific implementation, the millisecond-level alignment operation is implemented through a message queue middleware. High-risk infection area events and violation event events are injected into the queue as two independent message streams. The correlation engine matches them in chronological order using a sliding window scanning algorithm. The window size is the synchronization tolerance window. The confidence score of the strong correlation marker is calculated inversely proportional to the time difference; the smaller the time difference, the higher the score. In some embodiments, the size of the synchronization tolerance window can be dynamically adjusted according to the intensity of ward activity. The window is shortened during high-frequency operation periods to improve correlation accuracy, and expanded during low-frequency periods to increase correlation opportunities. It is understood that the accuracy of millisecond-level alignment depends on the time synchronization accuracy of all data acquisition devices, and the system uses a network time protocol to periodically calibrate the clocks of each device.

[0030] The spatiotemporal correlation module converts the two-dimensional image coordinates of medical equipment into three-dimensional spatial coordinates of the ward using point cloud registration technology. This point cloud registration technology employs an iterative nearest-neighbor algorithm. Input data includes the two-dimensional image coordinates of the medical equipment captured by a multispectral camera and the three-dimensional point cloud data of the ward collected by a depth sensor. The iterative nearest-neighbor algorithm solves for the optimal rigid body transformation matrix, including rotation and translation vectors, by minimizing the Euclidean distance error between feature points in the two-dimensional image and corresponding points in the three-dimensional point cloud. This transforms the two-dimensional image coordinates of the medical equipment into the global three-dimensional coordinate system of the ward, rigidly binding the contamination level label to the physical location of the equipment. In specific implementation, the iterative nearest-neighbor algorithm of the point cloud registration technology includes feature point extraction, corresponding point matching, transformation matrix calculation, and error convergence checking. Feature point extraction uses an accelerated robust feature algorithm to extract key points from the two-dimensional image. Corresponding point matching uses a k-dimensional tree data structure to accelerate nearest neighbor search. The transformation matrix calculation is achieved through singular value decomposition, and the error convergence threshold is set to 0.001 meters. Optionally, the point cloud registration technology introduces a multimodal fusion strategy, combining color image texture information to improve registration robustness. It is understandable that the performance of point cloud registration technology is affected by point cloud density and noise level. The system improves data quality by removing outliers through preprocessing filtering.

[0031] The spatiotemporal correlation module integrates strongly correlated markers and rigidly bound data to construct an infection risk topology map. This map uses a graph data structure, where nodes represent high-risk infection areas, unauthorized operational events, or medical devices, and edges represent spatiotemporal relationships. Node attributes include spatial coordinates, timestamps, and risk level labels, while edge attributes include correlation strength and time difference. The infection risk topology map is output in a 3D visualization format, including risk areas coded with different colors, timestamped behavioral events, and floating device contamination labels. In specific implementations, the construction process of the infection risk topology map includes graph initialization, node addition, edge connection, and layout optimization. Graph initialization creates an empty graph structure; node addition is dynamically inserted based on real-time data flow; edge connection is established based on strong correlation markers and spatial proximity; and the layout uses a force-directed algorithm to automatically adjust node positions to avoid overlap. In some embodiments, the infection risk topology map supports interactive queries, allowing users to view detailed information such as related event sequences by clicking on nodes. It is understood that real-time updates of the infection risk topology map require the support of an efficient graph database; the system uses an in-memory graph database to ensure low-latency response.

[0032] The spatiotemporal correlation module includes a topology optimization unit, which periodically compresses redundant nodes in the infection risk topology map. Redundant nodes are defined as high-risk infection area nodes that are spatially adjacent and have similar attributes. The topology optimization unit calculates the spatial distance and attribute similarity between nodes. If the spatial distance is less than the merging radius and the attribute similarity is higher than a threshold, a node merging operation is performed. The merging radius is dynamically adjusted according to the actual ward layout, using a smaller merging radius (e.g., 0.5 meters) in areas with dense medical equipment and a larger merging radius (e.g., 1.0 meter) in patient activity areas. In specific implementation, the compression operation cycle of the topology optimization unit is set to be once per minute. The node spatial distance is calculated based on three-dimensional Euclidean distance, and the attribute similarity is comprehensively evaluated by comparing the risk level label values ​​and timestamp differences. The attributes of the merged new node are taken as the maximum or average value of the original nodes. Optionally, the topology optimization unit supports manually setting the merging strategy, such as prioritizing the merging of nodes of the same type. It can be understood that the effectiveness of the topology optimization unit depends on the adaptive adjustment algorithm of the merging radius, which calculates the optimal radius in real time based on the node distribution density.

[0033] The spatiotemporal correlation module includes a behavior backtracking unit, which stores complete motion trajectory video clips from 30 seconds prior to the occurrence of the violation. These clips are extracted from the original video stream of the multispectral camera, with the extraction time range tracing back 30 seconds from the moment the violation occurred. The video clips are encoded in H.264 format to reduce storage space. The behavior backtracking unit builds an index database in chronological order of occurrence, using a B+ tree data structure for efficient retrieval. The key is the timestamp of the violation, and the value is the storage path of the video clip. In specific implementations, the behavior backtracking unit's storage management employs a circular buffer mechanism, with the buffer size set to store data from the most recent 24 hours. When storage space is insufficient, the oldest video clip is automatically overwritten. The B+ tree structure of the index database supports range queries and exact searches, with query latency controlled within milliseconds. In some embodiments, the behavior backtracking unit integrates video analysis functions, automatically extracting keyframes from the video as preview thumbnails. It is understood that the data integrity of the behavior backtracking unit depends on the continuous capture of the video stream, and the system ensures that video clips are not lost through redundant storage.

[0034] In practical implementation, the iterative nearest-point algorithm for point cloud registration requires initial transformation estimation during execution. This initial transformation estimation is achieved through coarse registration of the medical device's preset 3D model and the point cloud data. Coarse registration uses feature descriptor matching, such as the fast point feature histogram descriptor, to calculate initial rotation and translation parameters. The iterative nearest-point algorithm then runs in the refined registration stage until the error converges. Optionally, point cloud registration technology can incorporate semantic segmentation information to improve efficiency by registering only point clouds of medical devices. It is understood that the accuracy of point cloud registration directly affects the spatial positioning accuracy of pollution level labels; the system periodically verifies the registration error using ground control points.

[0035] In its implementation, the color coding scheme of the infection risk topology map follows international standards: high-risk infection areas are represented by red, medium-risk areas by yellow, and low-risk areas by green. Violation event icons use flashing warning signs to distinguish severity levels. The floating display of equipment contamination labels uses a gradient transparency effect, with lower transparency for higher contamination levels. Optionally, the infection risk topology map supports multi-view viewing, such as top view, front view, and 3D perspective view. Understandably, the visualization effect of the infection risk topology map requires accelerated rendering by the graphics processing unit to ensure a smooth interactive experience.

[0036] In practical implementation, the node merging operation of the topology optimization unit involves attribute fusion logic. For merged high-risk infected area nodes, the risk level of the new node is taken as the highest value of the original node's level, the timestamp is taken as the earliest occurrence time, and the spatial coordinates are taken as the weighted center of the original node's coordinates. Optionally, the topology optimization unit can be configured to retain merging history and allow the reversal of erroneous merging operations. It is understandable that the periodic execution frequency of the topology optimization unit needs to balance computational overhead and topology graph simplicity; excessively high frequency may lead to over-merging.

[0037] In its implementation, the behavior retrospective unit uses a distributed file system for video clip storage to improve data reliability and access speed. The B+ tree structure of the index is maintained in memory to accelerate queries and periodically persisted to disk to prevent data loss. Optionally, the behavior retrospective unit supports video clip export for later analysis and training. It is understood that the storage capacity planning for the behavior retrospective unit needs to consider video duration and resolution, and the system supports compressed archiving of historical data.

[0038] After the strong correlation markers in the spatiotemporal correlation module are established, a topology map update event is triggered. This update event notifies the collaborative correction module and the decision engine module via a publish-subscribe pattern, enabling system linkage. In implementation, the propagation of strong correlation markers uses an asynchronous message mechanism to avoid blocking the real-time data processing pipeline. The throughput performance of the spatiotemporal correlation module depends on the message processing architecture; the system uses multi-threading to process multiple correlation tasks in parallel. The 3D spatial coordinate binding process of the spatiotemporal correlation module includes a coordinate system normalization step, transforming local coordinates collected by different sensors to the global ward coordinate system. The transformation parameters are pre-determined through calibration experiments. In implementation, coordinate system normalization uses homogeneous coordinate transformation to ensure mathematical consistency of rotation and translation operations. Optionally, coordinate system normalization supports dynamic correction to handle sensor displacement or vibration. The accuracy of coordinate system normalization directly affects the accuracy of spatial correlation; the system automatically calibrates periodically using reflection markers. The infection risk topology map output interface of the spatiotemporal correlation module provides standardized data formats, such as JavaScript object notation or protocol buffers, facilitating integration with third-party systems. In practice, the output data includes metadata describing the topology map version and coordinate system definition. Optionally, the infection risk topology map can be exported to a common 3D model format such as point cloud data format or stereolithography format.

[0039] Example 3: See Figure 2 The collaborative correction module performs operations including scaling the baseline value of the temperature gradient threshold proportionally based on the distribution density of high-risk infection areas in the infection risk topology map. The distribution density is obtained by calculating the number of high-risk infection area nodes per unit volume of ward space, defined as cubic meters. The scaling operation uses a linear scaling function. The new baseline value of the temperature gradient threshold is equal to the original baseline value multiplied by the ratio of the distribution density to a preset reference density. The preset reference density is derived from historical infection data and represents the average risk level of the ICU ward. In specific implementations, the distribution density is calculated using a spatial grid division method, dividing the three-dimensional space of the ward into cubic grids with sides of 0.5 meters. The number of high-risk infection area nodes within each grid is counted, and the average value of all grids is taken as the overall distribution density. The scaling range of the baseline value of the temperature gradient threshold is limited to 0.5 to 2.0 times the original value to prevent over-adjustment. In some embodiments, a time decay factor is introduced into the calculation of the distribution density, giving higher weight to recently appearing high-risk infection area nodes to reflect the timeliness of the risk distribution. It is understandable that the dynamic adjustment of the temperature gradient threshold baseline value can enable the system to adapt to the monitoring needs of different infection risk periods, improve monitoring sensitivity during high-risk periods, and reduce false alarm rate during low-risk periods.

[0040] The collaborative correction module dynamically adjusts the combination of thresholds for the number and amplitude of joint angle mutations based on the distribution of historical violation types. This distribution is statistically derived from event records stored in the behavior backtracking unit, and includes types such as excessively fast operation speed, excessively large movement amplitude, and non-standard posture. The statistical period is set to the most recent 24 hours. The combination of thresholds for the number and amplitude of joint angle mutations is represented using a decision tree model. The node splitting conditions of the decision tree model are dynamically adjusted according to the frequency of the type distribution, with stricter threshold conditions corresponding to high-frequency violation types. In practice, the type distribution statistics use a sliding time window, updated every 5 minutes. The decision tree model is generated using the C4.5 algorithm, with each leaf node corresponding to a set of threshold combination conditions. The model update frequency is synchronized with the type distribution statistics. Optionally, a risk weight coefficient is introduced into the adjustment of the threshold combination conditions, assigning higher weights to violation types that may lead to serious infections. It can be understood that the dynamic optimization of the joint angle mutation threshold combination conditions makes the identification of violations more consistent with actual clinical operation patterns, reducing misjudgments of reasonable operations.

[0041] The collaborative correction module recalculates the weighting ratio of contact ID quantity to duration based on the clustering degree of pollution level labels in three-dimensional space. The clustering degree is evaluated using a spatial clustering algorithm. A density-based noise-applied spatial clustering algorithm is used to identify dense areas of pollution level labels. The new weighting ratio is calculated based on the clustering results, increasing the weight of contact ID quantity in highly clustered areas and increasing the weight of contact duration in dispersed areas. In specific implementation, the parameter settings for the density-based noise-applied spatial clustering algorithm are set as follows: neighborhood radius is 1 meter, minimum number of points is 3, and the weighting ratio is calculated using the following formula: in: and It's a new weight. and It is the basic weight. It's the learning rate. It is the number of labels within the cluster region. This is the total number of tags. and This is the maximum adjustment amount. In some embodiments, the weight adjustment process introduces a momentum term to avoid drastic fluctuations in the proportion value. It can be understood that recalculating the weight allocation ratio allows the pollution level assessment to more accurately reflect the infection risk characteristics under different scenarios.

[0042] The collaborative correction module includes an incremental learning unit. This unit records the trends in the false positive and false negative rates after each threshold adjustment, generating parameter optimization direction suggestions. The false positive rate is defined as the proportion of normal events incorrectly identified as high-risk infection areas or violations, while the false negative rate is defined as the proportion of real infection risk events not identified. The trends are derived from performance data after the most recent 20 adjustments via linear regression analysis. The parameter optimization direction suggestions are represented as vectors, pointing in the direction that simultaneously reduces both the false positive and false negative rates. In specific implementations, the performance data records of the incremental learning unit are stored in a time-series database. Each data point includes fields such as adjustment time, parameter value, false positive rate, and false negative rate. The linear regression analysis uses the least squares method to fit the trend line. The generation of parameter optimization direction suggestions is based on the gradient descent principle, calculating the negative gradient direction of the performance surface. In some embodiments, the incremental learning unit supports multi-objective optimization, balancing the trade-off between the false positive and false negative rates. It can be understood that the continuous learning capability of the incremental learning unit enables the system to gradually adapt to changes in the ICU environment and continuously improve monitoring accuracy.

[0043] The collaborative correction module includes a conflict arbitration unit. This unit prioritizes adjustments that have less impact on the accuracy of identifying high-risk infection areas when a conflict arises between the temperature gradient threshold and the pollution level label update logic. A conflict occurs when different parameter adjustment schemes have opposite effects on the overall system performance. The conflict arbitration unit makes decisions by evaluating the sensitivity index of each parameter to the accuracy of identifying high-risk infection areas. The sensitivity index is defined as the rate of change in identification accuracy caused by a unit change in the parameter. In implementation, the decision-making process of the conflict arbitration unit includes conflict detection, sensitivity calculation, scheme ranking, and execution selection. Conflict detection is achieved by comparing the predicted performance change directions of each scheme. Sensitivity calculation uses perturbation analysis, with the parameter unit change set at 10% of the standard deviation. Schemes are ranked in ascending order of sensitivity, prioritizing schemes with lower sensitivity. Optionally, the conflict arbitration unit introduces a voting mechanism, referencing historical decision effects when multiple schemes have similar sensitivities. It can be understood that the intervention of the conflict arbitration unit ensures the stability of system parameter adjustments and avoids performance degradation due to parameter oscillations.

[0044] In practical implementation, the calculation of distribution density needs to exclude temporary interference factors, such as false high-risk areas caused by temporary displacement of medical equipment. The system identifies stable distribution patterns through duration filtering, and only high-risk infection area nodes that have existed for more than 5 minutes participate in the density calculation. It is understandable that the accuracy of distribution density calculation directly affects the effectiveness of temperature gradient threshold adjustment; excessive filtering may lead to response delays, while insufficient filtering introduces noise interference. The dynamic adjustment of joint angle abrupt change threshold combinations needs to consider the specificities of different operational scenarios. For example, rapid operations in emergency situations may be misjudged as violations. The system performs contextual correction of thresholds by integrating ward alarm status information. It is understandable that contextual correction of threshold combinations enables the system to distinguish between emergency medical operations and genuine violations, improving identification specificity.

[0045] In practical implementation, the recalculation process of the pollution level label weight allocation ratio includes normalization processing to ensure that the sum of the new weights is 1, avoiding systematic bias in the evaluation results. This normalization ensures the continuity and comparability of pollution level values, facilitating the setting of a unified intervention threshold. The parameter optimization direction of the incremental learning unit needs to be converted into specific parameter adjustment amounts. The conversion process uses a constrained optimization algorithm, limiting the adjustment amount to ±10% of the current parameter value. This constraint prevents excessively large single adjustments from causing system instability, achieving a smooth transition. The sensitivity calculation of the conflict arbitration unit requires substantial historical data support. Initially, a preset default sensitivity value is used, gradually replaced by the actual calculated value over time. This initialization strategy of the conflict arbitration unit ensures reliable system operation in the absence of historical data, gradually optimizing decision quality during the learning process.

[0046] The components of the collaborative correction module exchange data via a message bus. Adjustment decisions are arbitrated by a conflict arbitration unit, generating a unified parameter update command which is then broadcast to all relevant modules in the system. In implementation, the message bus uses a publish-subscribe model, supporting many-to-many asynchronous communication to ensure real-time system response. This communication mechanism ensures data consistency between the collaborative correction module and other parts of the system, preventing parameter state confusion due to timing issues. The collaborative correction module maintains a parameter adjustment log, recording the decision basis, execution results, and performance changes for each adjustment, used for auditing and problem tracking. In implementation, the parameter adjustment log uses structured storage, including fields such as timestamp, operator, parameter name, old value, new value, and decision module. The complete record of the parameter adjustment log provides data support for system optimization, facilitating the analysis of the effectiveness of adjustment strategies.

[0047] Example 4: See Figure 3The decision engine module performs operations including controlling an ultraviolet (UV) lamp array mounted on the top of the bed to perform fan-shaped scanning irradiation on high-risk infection areas according to real-time sterilization commands. The real-time sterilization commands are received from the collaborative correction module and include the two-dimensional contour coordinates and risk level information of the target area. The UV lamp array consists of multiple UV light-emitting diode (LED) units arranged in a grid, each of which can be independently controlled to switch on / off and adjust its brightness. The fan-shaped scanning irradiation mode is achieved by sequentially activating UV LED unit groups in different directions. The irradiation angle, duration, and intensity are dynamically set according to the risk level of the high-risk infection area. In specific implementation, the UV lamp array's driving circuit uses pulse width modulation (PWM) technology to precisely control the brightness level of the UV LED units. The PWM frequency is set to 1000 Hz to avoid visible flicker. The fan-shaped scanning path planning algorithm calculates the optimal coverage path based on the contour coordinates of the high-risk infection area, ensuring that the UV spot completely covers the target area with an overlap rate of less than 10%. The correspondence between irradiation parameters and risk levels is pre-stored in a lookup table; the higher the risk level, the higher the brightness of the corresponding UV LED unit and the longer the irradiation duration. It is understandable that the fan-shaped scanning irradiation of the ultraviolet lamp array enables targeted disinfection of high-risk infection areas on the patient's body surface, minimizing ultraviolet exposure to non-target areas.

[0048] The decision engine module triggers an LED warning ring and a directional speaker located at the location of the violation via audible and visual warning signals. The generation of these signals is based on violation event data output by the spatiotemporal correlation module. This event data includes the violation type, location coordinates, and severity score. The LED warning ring, installed at a specific location on the ward ceiling, consists of red, green, and blue LED strips. The directional speaker uses an array of ultrasonic transducers to generate an audible sound beam. In practice, the LED warning ring's color display is determined by the violation severity score: a slow-flashing yellow warning appears when the score is below a threshold, while a fast-flashing red warning appears when the score is above the threshold. The directional sound wave emission of the directional speaker is achieved by adjusting the phase difference between the units in the ultrasonic transducer array. The sound beam width is set to ±15 degrees to ensure the sound is concentrated near the location of the violation. The intensity of the audible and visual warning signal is directly proportional to the severity of the violation; for each increase in severity score, the sound pressure level increases by 3 decibels, and the LED brightness increases by 20%. Optionally, the audible and visual warning signal supports multilingual voice prompts, with pre-set warning statements based on the violation type. It is understandable that the precise spatial positioning of audible and visual warning signals can effectively remind medical staff to correct improper operations without disturbing other people in the ward.

[0049] The decision engine module drives an automated transport robot to move contaminated equipment to a designated disinfection chamber based on an isolation scheduling plan. This plan, generated by the decision engine module based on the contamination level labels of medical equipment and real-time ward status information, includes the contaminated equipment number, the location of the target disinfection chamber, and the transport priority. The automated transport robot uses a differential drive chassis equipped with a robotic arm and navigation sensors. The designated disinfection chamber is a closed ultraviolet disinfection device located in a corner of the ward. In implementation, the algorithm for generating the isolation scheduling plan comprehensively considers multiple factors such as contamination level, equipment urgency, and robot availability, using the Hungarian algorithm to solve for optimal task allocation. The automated transport robot's navigation system integrates LiDAR simultaneous localization and mapping (SLAM) and visual odometry technology to build a real-time environmental map and plan a collision-free path. The robotic arm's end effector adaptively adjusts its grasping posture according to the equipment's shape. Optionally, the isolation scheduling plan supports a manual review mode, allowing the head nurse to confirm or modify the transport task before dispatch. In essence, the intervention of the automated transport robot enables rapid isolation of contaminated equipment, cutting off potential transmission routes for cross-infection.

[0050] The decision engine module includes a sterilization effectiveness verification unit. This unit rescans high-risk infection areas after UV lamp irradiation to compare temperature distribution data changes with the expected decay curve. The rescanning operation is performed by a multispectral camera five seconds after the UV lamp array is turned off to avoid interference. The expected decay curve is predicted and generated using a physical model based on UV lamp irradiation parameters and initial temperature data. In specific implementation, the sterilization effectiveness verification unit's execution flow includes three steps: data acquisition, curve fitting, and difference analysis. The data acquisition stage obtains the temperature distribution matrix of the high-risk infection area after irradiation. The curve fitting stage uses the least squares method to fit the actual temperature data into an exponential decay curve. The difference analysis stage calculates the coefficient of determination between the fitted curve and the expected decay curve. When the coefficient of determination is lower than a threshold, the sterilization effect is deemed unsatisfactory, triggering a re-irradiation process. In some embodiments, the sterilization effectiveness verification unit introduces a machine learning model to predict more complex temperature change patterns, improving verification accuracy. It can be understood that the sterilization effectiveness verification unit forms a closed-loop control of the disinfection process, ensuring the effectiveness of infection risk prevention and control measures.

[0051] The decision engine module includes an emergency release switch. This switch allows medical staff to manually halt the transport robot's mission and reset the device's status flag when a system misjudgment leads to unintended isolation. The emergency release switch employs a dual verification mechanism: a hardware button and software confirmation. The hardware button is installed at the nurse station control panel and ward entrance, while the software confirmation interface, displayed on a touchscreen, requires the user's ID and password. In implementation, the emergency release switch's trigger logic has two levels: pressing the hardware button pauses the automated transport robot's movement and issues an audible alert; after the medical staff selects the reason for termination on the software confirmation interface and submits, the transport mission is officially canceled, and the device's status flag is reset from isolated to available and updated to the central database. Optionally, the emergency release switch's operation log generates a safety event report for archiving and future reference. In essence, the emergency release switch provides a necessary channel for manual intervention, balancing system automation with clinical operational flexibility.

[0052] In practical implementation, the safety management of the ultraviolet lamp array includes multiple protection measures. Infrared human body sensors are installed to monitor the irradiation area; when personnel are detected approaching, the power supply to the ultraviolet LED unit is immediately cut off. The system log records the parameters and duration of each irradiation operation. It can be understood that the safety design of the ultraviolet lamp array complies with medical device electromagnetic radiation safety standards, ensuring that patients and medical staff are not accidentally exposed.

[0053] In practical implementation, the linkage control of the LED warning ring and directional speakers adopts an event-driven architecture. When multiple violations are detected simultaneously, warning signals are triggered sequentially according to a priority queue, avoiding overlapping and interference between sound and light signals. This priority management of the sound and light warning system ensures immediate response to high-risk violations. The task management of the automated transport robot supports parallel processing mode. Multiple robots work collaboratively through a central scheduling system, and their task status is displayed in real time on the ward monitoring screen. This cluster scheduling of automated transport robots improves the efficiency of contaminated equipment isolation and is suitable for large ICU wards. The threshold setting of the sterilization effect verification unit has adaptive capabilities, dynamically adjusting the determination coefficient threshold based on historical sterilization data from high-risk infection areas, improving the rationality of the verification standard. This adaptive threshold adjustment of the sterilization effect verification unit avoids misjudgments caused by changes in environmental factors. The access control of the emergency release switch adopts a role-based access control model. Medical staff of different ranks have different operating permissions, with higher-level permissions covering lower-level operations. This access control of the emergency release switch ensures the security of system operations and traceability of responsibility.

[0054] The decision engine module integrates with the hospital information system, synchronizing equipment status and patient information in real time. Sterilization instructions and isolation protocols are generated with reference to patient infection indicators in the electronic medical records. In implementation, the system interface uses a standard health information exchange protocol to achieve data interoperability, ensuring standardized and reliable information flow. Table 1 illustrates the correspondence between UV lamp irradiation parameters and the risk level of high-risk infection areas.

[0055] Table 1: Correspondence between UV lamp irradiation parameters and risk levels See Figure 4 This diagram illustrates the time and resource allocation of the decision engine module driving automated transport robots to perform contaminated equipment isolation scheduling tasks within a vision-based intelligent infection control system for ICU wards. The vertical axis (R1, R2, R3) represents three automated transport robots (the main task executors), while the horizontal time axis quantifies task execution time in minutes. Different colors correspond to task priorities: red for high priority, yellow for medium priority, and blue for low priority. Each task bar is labeled with a medical device (e.g., ventilator A, monitor B, etc.). For robot R1: the high-priority ventilator A isolation task is executed first, starting at 0 minutes and lasting 15 minutes; followed by the medium-priority infusion pump C isolation task, starting at 15 minutes and ending at 30 minutes. For robot R2: the high-priority monitor B isolation task starts at 0 minutes and lasts 20 minutes; followed by the medium-priority suction device E isolation task, starting at 20 minutes and ending at 35 minutes. For robot R3: the high-priority defibrillator D isolation task starts at 5 minutes and lasts for 20 minutes (ending at 25 minutes); then the low-priority monitor F isolation task is executed, starting at 25 minutes and ending at 40 minutes. Through a clear display of time and resource dimensions, the scheduling logic of multiple robots in equipment isolation tasks is intuitively presented. Combining factors such as equipment contamination level (reflected in task priority) and task duration, optimization techniques such as the Hungarian algorithm are used to achieve efficient task allocation among multiple robots, ensuring that contaminated equipment is quickly transferred to the disinfection chamber, cutting off the transmission route of cross-infection in the ICU ward. This demonstrates the system's precision and coordination in resource scheduling during infection risk prevention and control.

[0056] Example 5: The acquisition module includes depth sensors deployed at the four corners of the ward. These depth sensors supplement the contact frequency data on the back of the medical equipment. The depth sensors measure distance using the time-of-flight principle, which calculates the distance value by converting the time difference between the emission and reception of a light pulse. The sampling frequency of the depth sensors is an integer multiple of the frame rate of the multispectral camera. The multispectral camera frame rate is set to 30 frames per second, and the depth sensor sampling frequency is set to 60 frames per second, a multiple of 2. In practice, the installation positions of the depth sensors at the four corners of the ward are precisely calibrated. The four depth sensors are located at the four apex corners where the ceiling and walls meet, at a height of 2.5 meters. The pitch angle is adjusted to a downward tilt of 15 degrees to ensure coverage of the area behind the medical equipment. The frame rate synchronization between the depth sensors and the multispectral camera uses a master-slave clock synchronization scheme. The multispectral camera, as the master device, generates 30 synchronization pulses per second, and the depth sensors, as slave devices, acquire one frame of data on the rising and falling edges of each synchronization pulse, achieving synchronous sampling at twice the frame rate.

[0057] The depth sensor data from the acquisition module is used to generate 3D point cloud data for the back of the medical device. This 3D point cloud data includes spatial coordinate information and reflection intensity information. The spatial coordinate information is represented by a local coordinate system with the depth sensor as the origin, and the reflection intensity information reflects the light reflection characteristics of the object's surface. In specific implementation, the raw distance data acquired by the depth sensor is converted into a 3D point cloud through coordinate transformation. Each point cloud data point contains X, Y, and Z coordinate values ​​and a reflection intensity value. The point cloud density is set to 10,000 points per square meter. The point cloud data for the back of the medical device is segmented from the scene point cloud using a background subtraction algorithm. This algorithm detects foreground objects by comparing the current frame with the background model. Optionally, color information can be added to the depth sensor point cloud data through registration and fusion with color images from a multispectral camera. It can be understood that the 3D point cloud data provided by the depth sensor effectively compensates for the visual blind spots of the multispectral camera on the back of the medical device, forming complete surface contact monitoring coverage.

[0058] The acquisition module includes an embedded preprocessing unit. This unit performs non-uniformity correction on the raw temperature distribution data acquired by the multispectral camera. Non-uniformity correction aims to eliminate measurement deviations caused by the camera's own thermal noise. The embedded preprocessing unit employs a field-programmable gate array (FPGA) architecture to achieve parallel computing capabilities. The non-uniformity correction algorithm performs pixel-level compensation by reading the temperature sensor readings inside the multispectral camera and applying a polynomial fitting model. In specific implementation, the non-uniformity correction process of the embedded preprocessing unit is divided into two stages: reference frame acquisition and real-time correction. The reference frame acquisition stage is executed when the multispectral camera starts up, acquiring background noise distribution maps at different operating temperatures by obscuring the lens. The real-time correction stage selects the corresponding noise distribution map based on the current operating temperature of the multispectral camera and performs polynomial fitting compensation on the raw output value of each pixel. The polynomial coefficients are determined through laboratory calibration experiments. In some embodiments, the non-uniformity correction algorithm employs adaptive filtering technology, dynamically adjusting the filtering parameters according to the scene content. It can be understood that the non-uniformity correction of the embedded preprocessing unit significantly improves the accuracy and reliability of temperature measurement data, providing a high-quality data foundation for subsequent infection risk analysis.

[0059] The polynomial fitting model of the embedded preprocessing unit adopts a quadratic polynomial form, which includes a constant term, a first-order term, and quadratic coefficients. Each pixel has an independent set of polynomial coefficients, which are stored in the flash memory of the embedded preprocessing unit. In specific implementation, the coefficient calibration process of the polynomial fitting model is carried out in a constant-temperature laboratory. The multispectral camera is pointed at a blackbody radiation source, and the output value of each pixel is recorded at different temperatures of the blackbody radiation source. The optimal polynomial coefficients for each pixel are fitted using the least squares method. The update cycle of the polynomial fitting model is set to once every three months to ensure that the correction accuracy remains stable over time. Optionally, the polynomial fitting model supports online updates, automatically triggering a recalibration process when multispectral camera performance drift is detected. It can be understood that the refinement of the polynomial fitting model provides personalized correction parameters for each pixel, effectively overcoming the inherent non-uniformity defects of the sensor.

[0060] The data fusion of the depth sensor and multispectral camera in the acquisition module employs a timestamp alignment algorithm. This algorithm unifies the data streams acquired by different sensors onto the same time base, which is the GPS clock signal. In practice, the timestamp alignment algorithm is implemented in an embedded preprocessing unit. The algorithm first timestamps each data packet with a precision down to the microsecond level. Then, based on the principle of minimum time difference, it pairs the data packets from different sensors. The timestamp difference between the depth sensor data and the multispectral camera data must be less than a synchronization tolerance threshold, set to 1 millisecond, to ensure the time consistency of the data fusion. Optionally, the timestamp alignment algorithm supports dynamic adjustment of the synchronization tolerance threshold, automatically optimizing the alignment accuracy based on the data stream load. In essence, the high-precision implementation of the timestamp alignment algorithm provides a temporal guarantee for the effective fusion of multimodal sensor data.

[0061] The embedded preprocessing unit performs background subtraction and noise filtering on the depth sensor data. The background subtraction algorithm uses a Gaussian mixture model to build a scene background model, and the noise filtering uses a statistical outlier removal algorithm to eliminate outliers. In specific implementation, the Gaussian mixture model for background subtraction uses three Gaussian distributions to model the background probability of each pixel, with a model update rate of 5 times per second. The statistical outlier removal algorithm analyzes the distance distribution between each point and its neighbors, removing outliers whose mean distance exceeds three times the standard deviation. The processed point cloud data is used for subsequent contact frequency analysis. Optionally, the background subtraction algorithm supports manual background model reset, allowing for re-initialization of the background model when there are significant changes in the ward layout. It can be understood that the preprocessing workflow for depth sensor data ensures the quality of the point cloud data, laying the foundation for accurate contact detection.

[0062] The depth sensor point cloud data processing in the acquisition module includes a point cloud segmentation step. Point cloud segmentation separates the medical device point cloud from the scene point cloud. The point cloud segmentation algorithm is based on the principle of region growing. In specific implementation, the point cloud segmentation algorithm first removes the ground and wall point clouds through planar detection, then selects seed points from the remaining point cloud, and performs region growing based on the point cloud normal vector and curvature similarity. Each grown region corresponds to an independent medical device point cloud cluster. The point cloud segmentation result is used for subsequent contact frequency statistics. In some embodiments, the point cloud segmentation algorithm combines color information to improve segmentation accuracy. After the depth sensor point cloud is registered with the multispectral camera color image, color consistency is used as an additional constraint condition for region growing. It can be understood that the accuracy of point cloud segmentation directly affects the reliability of contact frequency statistics on the back of the medical device. A refined segmentation algorithm ensures that each device is independently identified and tracked.

[0063] The temperature distribution data non-uniformity correction in the embedded preprocessing unit includes a bad pixel repair function. Bad pixel repair performs interpolation compensation for pixels with abnormal responses in the multispectral camera. In practice, the bad pixel repair algorithm is executed after non-uniformity correction. The algorithm first identifies bad pixels whose response values ​​deviate from the average value of surrounding pixels by more than a threshold. Then, it uses bilinear interpolation to calculate replacement values ​​from neighboring normal pixels. The threshold for bad pixel repair is set to three standard deviations to ensure that only obviously abnormal pixels are repaired. Optionally, the bad pixel repair algorithm supports dynamic updates of the bad pixel map, periodically detecting and recording the locations of newly appearing bad pixels. It can be understood that the bad pixel repair function further improves the integrity and usability of temperature distribution data, preventing abnormal pixels from interfering with infection risk analysis.

[0064] The depth sensor data and multispectral camera data from the acquisition module are fused to generate complete surface contact frequency data for medical devices. Data fusion employs a feature-level fusion strategy in the embedded preprocessing unit. In practice, feature-level fusion extracts contact event features from both the depth sensor point cloud and the multispectral camera images. These features include contact location, contact area, and contact duration. Then, based on feature similarity, correlation fusion is performed to generate a unified contact frequency data record. This data includes fields such as timestamp, device number, contact person ID, and contact area coordinates. Optionally, feature-level fusion supports confidence weighting, assigning appropriate weights to features provided by different sensors. It can be understood that multi-sensor data fusion fully leverages the complementary advantages of different modalities, improving the comprehensiveness and accuracy of contact frequency monitoring.

[0065] The output data of the embedded preprocessing unit is encoded in a standardized format, which includes a data header, a payload, and a checksum. In practice, the data header contains metadata such as sensor number, timestamp, and data length; the payload stores the preprocessed sensor data; and the checksum uses a cyclic redundancy check algorithm to ensure data transmission integrity. The standardized format data is sent to the dynamic segmentation module for further processing via a gigabit Ethernet interface. Optionally, the standardized format supports data compression, automatically activating compressed transmission mode when network bandwidth is limited.

[0066] See Figure 5This figure illustrates the distribution characteristics of the coefficients of the polynomial fitting model in the embedded preprocessing unit. The polynomial fitting model, in quadratic polynomial form, includes constant, linear, and quadratic coefficients, used to correct the non-uniformity of the raw temperature distribution data acquired by the multispectral camera. The figure presents the numerical distribution of the three types of coefficients (constant, linear, and quadratic) in the form of a violin plot. The shape of the violin plot reflects the probability density distribution of the coefficient values, while the box plot in the middle shows the median (white dots), quartile range (black boxes), and extreme values ​​(black whiskers). The red dashed line represents the baseline where the coefficient value is 0. As can be seen from the figure, the constant term has the widest distribution range, followed by the linear term, while the quadratic term has the most concentrated distribution. This distribution characteristic reflects the differences in non-uniformity correction requirements among pixels of the multispectral camera. The large fluctuations in the constant term reflect individual differences in the basic pixel response, while the concentrated distribution of the quadratic term indicates that the non-linear changes in the pixel response are relatively consistent. By analyzing the distribution of these coefficients, the calibration effect of the polynomial fitting model can be evaluated, providing data support for subsequent model optimization (such as coefficient update cycle, calibration experimental design, etc.).

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A vision-based intelligent sensing and control system for ICU wards, characterized in that, The system includes: The data acquisition module is used to acquire real-time data on the distribution of patients' body surface temperature, the movement trajectory of medical staff's limbs, and the contact frequency of medical equipment surfaces in the ICU ward through a multispectral camera. The dynamic segmentation module is used to divide the patient's body surface into high-risk infection areas and low-risk infection areas based on temperature gradient thresholds, identify medical staff's violations based on the rate of change of acceleration of motion trajectory, and generate equipment contamination level labels by combining contact frequency data. The spatiotemporal correlation module is used to establish a timestamp mapping relationship between high-risk infection areas and violations of regulations, bind contamination level labels with the spatial coordinates of corresponding medical equipment, and output an infection risk topology map with three-dimensional location markers. The collaborative correction module is used to dynamically adjust the temperature gradient threshold and acceleration change rate judgment criteria based on the infection risk topology map, and synchronously update the generation logic of the pollution level label. The decision engine module receives output data from the collaborative correction module and generates real-time sterilization instructions for high-risk infection areas, audible and visual warning signals for violations, and isolation and scheduling schemes for contaminated equipment.

2. The vision-based intelligent sensing and control system for ICU wards according to claim 1, characterized in that, The operations performed by the dynamic segmentation module include: An adaptive window width algorithm is used to analyze local extreme points in temperature distribution data. The temperature decay rate is detected by radiating outward from the extreme point. When the decay rate exceeds a preset critical value, the boundary of the high-risk infection area is delineated. The joint angle change sequence in the movement trajectory of medical staff is extracted by skeletal key point tracking technology. The number and magnitude of angle mutations between consecutive frames are calculated. If the mutation number threshold and the magnitude threshold are met at the same time, it is judged as a violation of the operation. The number of unique IDs that are touched by different people on the surface of medical equipment per unit time is counted, and the numerical intensity of the contamination level label is calculated by weighting the contact duration.

3. The vision-based intelligent sensor control system for ICU wards according to claim 2, characterized in that, The operations performed by the spatiotemporal correlation module include: Align the emergence time of high-risk infection areas with the occurrence time of violations at the millisecond level. If the time difference between the two is less than the synchronization tolerance window, a strong correlation marker is established. Point cloud registration technology is used to convert the two-dimensional image coordinates of medical equipment into three-dimensional spatial coordinates of the ward, so that the contamination level label is rigidly bound to the physical location of the equipment. An infection risk topology map is constructed by integrating strongly correlated tags and rigidly bound data. The map includes risk areas coded with different colors, time-stamped behavioral events, and floating device contamination labels.

4. The vision-based intelligent sensor control system for ICU wards according to claim 3, characterized in that, The operations performed by the collaborative correction module include: Based on the distribution density of high-risk infection areas in the infection risk topology map, the baseline value of the temperature gradient threshold is scaled proportionally. Based on the distribution of historical violation types, dynamically adjust the combination of the threshold for the number of joint angle mutations and the amplitude threshold; Based on the degree of clustering of pollution level labels in three-dimensional space, the weighting ratio of the number of contact IDs and the duration is recalculated.

5. The vision-based intelligent sensing and control system for ICU wards according to claim 4, characterized in that, The operations performed by the decision engine module include: The ultraviolet lamp array installed on the top of the bed is controlled to perform a fan-shaped scan irradiation of high-risk infection areas according to the real-time sterilization command; The LED warning ring and directional speaker located at the location of the violation are triggered by audible and visual warning signals, and the intensity of the warning signal is proportional to the severity of the violation. Based on the isolation and scheduling plan, the automated transport robot is driven to move contaminated equipment to the designated disinfection chamber, while the distribution map of available equipment in the ward is updated.

6. The vision-based intelligent sensing and control system for ICU wards according to claim 5, characterized in that, The acquisition module also includes: Depth sensors deployed at the four corners of the ward are used to supplement the contact frequency data on the back of medical equipment, and their sampling frequency is kept as an integer multiple of the frame rate of the multispectral camera. An embedded preprocessing unit is used to correct the non-uniformity of the raw temperature distribution data and eliminate measurement deviations caused by the thermal noise of the camera itself.

7. The vision-based intelligent sensing and control system for ICU wards according to claim 6, characterized in that, The dynamic segmentation module also includes: An abnormal trajectory filtering unit is used to eliminate false motion trajectories caused by camera obstruction or light reflection. It adopts a confidence assessment mechanism based on the accumulation of motion energy. The temperature compensation unit is used to correct the temperature measurement value based on the patient's body surface humidity data, which is provided by the ward environment sensor.

8. The vision-based intelligent sensing and control system for ICU wards according to claim 7, characterized in that, The spatiotemporal correlation module also includes: Topology optimization unit is used to periodically compress redundant nodes in the infection risk topology graph and merge adjacent high-risk infection areas whose spatial distance is smaller than the merging radius. The behavior tracing unit is used to store the complete motion trajectory video clips 30 seconds before the violation occurred, and to build an index library in chronological order of occurrence.

9. The vision-based intelligent sensing and control system for ICU wards according to claim 8, characterized in that, The collaborative correction module also includes: The incremental learning unit is used to record the changing trends of the system's false alarm rate and false negative rate after each threshold adjustment, and to generate suggestions for parameter optimization. The conflict arbitration unit is used to prioritize the adjustment scheme that has less impact on the identification accuracy of high-risk infection areas when there is a conflict between the temperature gradient threshold and the pollution level label update logic.

10. The vision-based intelligent sensing and control system for ICU wards according to claim 9, characterized in that, The decision engine module also includes: The sterilization effect verification unit is used to rescan high-risk infection areas after UV lamp irradiation and compare whether the temperature distribution data changes conform to the expected decay curve. The emergency release switch is used to allow medical staff to manually stop the transport robot's mission and reset the equipment status flags when the system misjudges and causes the equipment to be mistakenly isolated.

Images