Intensive care unit infusion monitoring system and method based on machine vision

Abstract

The application discloses an ICU infusion monitoring system and method based on machine vision, and aims at the problems that the existing contact type infusion monitoring is easy to interfere with the infusion pipeline, does not meet the requirements of ICU sterile diagnosis and treatment, and the monitoring precision of the general visual monitoring scheme is insufficient and it is difficult to adapt to low-speed micro-precision infusion monitoring, adopts a non-contact visual collection and edge local reasoning architecture, the system comprises a perception collection module, a network transmission module, an edge reasoning module and a business application module which are sequentially connected in communication, the method establishes a pixel and physical mapping relationship of an infusion monitoring region of interest (ROI) through initialization calibration, sequentially completes target detection, sub-region clipping, semantic segmentation, infusion drop speed and residual amount calculation and multi-feature fusion anomaly judgment on the edge side, and outputs graded alarm and trace data, and the application can realize non-contact and high-precision infusion monitoring in the ICU scene, adapt to the requirements of sterile diagnosis and treatment, and improve the monitoring precision and abnormal response efficiency.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare and intensive care technology, specifically to an ICU infusion monitoring system and method based on machine vision. Background Technology

[0002] Intravenous infusion is one of the core routes of clinical drug administration for critically ill patients in the ICU. The rate control, residual volume monitoring, and abnormal early warning during the infusion process are directly related to patient medication safety. Currently, commonly used infusion monitoring technologies can be mainly divided into the following two categories:

[0003] One type is contact sensor infusion monitoring technology. This solution uses contact sensors such as infrared beams and gravity weighing to collect signals. The sensors are clipped to the infusion tubing or suspended below the infusion container to realize infusion drip rate statistics and residual volume calculation. The monitoring data is uploaded to the nurse station terminal and triggers abnormal alarms. This type of solution has a simple structure and low deployment cost, and is widely used in routine infusion scenarios in general wards. However, it requires physical contact with the infusion tubing and drip chamber, which can easily cause displacement of precision tubing and increase the risk of contamination. It is difficult to adapt to the special requirements of aseptic diagnosis and treatment in ICU.

[0004] Another type is AI visual infusion monitoring technology for general wards. This solution uses civilian cameras with general visual algorithms to collect ward video streams in a large-scale centralized monitoring manner. It achieves infusion drip rate and remaining volume identification through target detection and frame difference counting. It has the characteristics of being contactless and requiring no tubing modification. However, this type of solution is mainly suitable for conventional speed infusion scenarios, with limited monitoring accuracy. It is difficult to stably identify low-speed micro-droplets and small changes in liquid level, and cannot meet the monitoring needs of precise infusion for critically ill ICU patients.

[0005] Therefore, it is essential to design an ICU infusion monitoring system and method based on machine vision. Summary of the Invention

[0006] The purpose of this invention is to provide an ICU infusion monitoring system and method based on machine vision, so as to solve the problems of contact monitoring schemes that are prone to interfering with infusion lines and do not meet the requirements of aseptic diagnosis and treatment, as well as the problems of ordinary visual monitoring schemes having insufficient monitoring accuracy and being unable to adapt to low-speed micro-volume precision infusion.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] Firstly, a machine vision-based ICU infusion monitoring system is provided, comprising a perception and acquisition module, a network transmission module, an edge inference module, and a business application module:

[0009] The sensing and acquisition module is installed above the ICU ward and is used to collect real-time video streams of the infusion monitoring area;

[0010] The network transmission module adopts an intra-hospital local area network architecture to transmit the real-time video stream output by the sensing and acquisition module to the edge inference module, and transmit the monitoring data and alarm data output by the edge inference module to the business application module.

[0011] The edge inference module is used to pre-store calibration data. After receiving the real-time video stream, it calls the calibration data to perform target detection within the Region of Interest (ROI) for infusion monitoring to identify infusion monitoring targets. Based on the spatial position of the infusion monitoring targets in the video frame, it crops the infusion monitoring sub-regions. Pixel-level semantic segmentation is performed on the infusion monitoring sub-regions to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber outline, a multi-target tracking algorithm is used to obtain the continuous trajectory tracking results of the droplets. Based on the continuous trajectory tracking results, the number of droplets per unit time is calculated to obtain the infusion drip rate. Based on the liquid surface boundary position and the pixel-to-physical mapping relationship, the remaining amount of medication and the remaining infusion time are calculated. The infusion drip rate, remaining amount of medication, infusion tubing morphology features, infusion pump operating status, and occlusion degree of the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level. The calibration data includes the Region of Interest (ROI) for infusion monitoring, pixel-to-physical mapping relationship, and infusion baseline state parameters. The infusion monitoring targets include the drip chamber, infusion tubing, infusion container, and infusion pump.

[0012] The business application module is used to receive monitoring data and abnormal results output by the edge inference module, generate corresponding alarm information, record alarm handling information, generate infusion process traceability data, and perform storage.

[0013] As a further technical solution of the present invention, the sensing and acquisition module is a camera device that conforms to the electromagnetic compatibility standards of medical electrical equipment. It is embedded in the ceiling of the ICU ward and has a disinfection-proof shell and wide dynamic range imaging capability. The edge inference module is an embedded gateway with local AI computing power, physically isolated from the hospital's internal and external networks, and equipped with local storage media. The network transmission module adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at a fixed period. When a heartbeat is lost for a preset number of consecutive times, it automatically switches to the backup link. After the main link recovers stability and completes status verification, it automatically switches back. In the case of network outage, the edge inference module can cache monitoring data through local storage media, and the data is stored using a cyclic overwrite strategy.

[0014] As a further technical solution of the present invention, the edge reasoning module has a built-in library of commonly used ICU infusion container specifications, and pre-stores the inherent size parameters and volume conversion parameters of various containers; when performing infusion drip rate calculation, a unique tracking identifier is assigned to each droplet, and droplets with the same tracking identifier are counted only once, and static filtering rules are set to remove static background interference; when performing drug remaining volume calculation, the corresponding calculation model is matched according to the container type, and a hue and saturation joint judgment method is used to identify the liquid surface boundary for dark-colored drugs and light-proof drugs; a total of 8 types of infusion abnormalities can be judged, including infusion drip rate overspeed abnormality, infusion drip rate too slow or drip stop abnormality, drug remaining volume warning, infusion completion abnormality, tubing bending or pressure abnormality, tubing displacement or detachment abnormality, infusion monitoring sub-area obstruction abnormality, and pump working status abnormality.

[0015] The alarm information of the business application module is divided into three levels according to risk level. Low-risk anomalies are indicated by large screen annotation and log recording. Medium-risk anomalies are indicated by large screen pop-ups, mobile push notifications and local sound alerts. High-risk anomalies are indicated by large screen pop-ups, mobile sound and light alarms and ward linkage emergency alarms. Data transmission and local storage are encrypted. Sensitive field keys are managed by the hardware security module. Preset abnormal event data is marked as read-only and cannot be overwritten or tampered with.

[0016] Secondly, a machine vision-based method for monitoring ICU infusions is provided, comprising the following steps:

[0017] S1 System Deployment: An image acquisition unit is deployed above the ICU ward. The image acquisition unit is used to acquire real-time video streams of the infusion monitoring area. An edge computing unit is configured for local AI algorithm inference. The image acquisition unit transmits the acquired real-time video streams to the edge computing unit through the hospital's local area network.

[0018] S2 Initialization Calibration: Before the system runs, the initial monitoring screen is acquired through the image acquisition unit. The region of interest (ROI) for infusion monitoring is defined in the initial monitoring screen. The pixel-to-physical mapping relationship between the image pixels and physical dimensions within the ROI is established. The reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions are extracted as infusion reference state parameters. The three calibration data, namely the ROI, the pixel-to-physical mapping relationship, and the infusion reference state parameters, are stored in the edge computing unit.

[0019] S3 Local Inference Processing: The edge computing unit receives the real-time video stream and calls the calibration data. Within the ROI, it performs target detection to identify infusion monitoring targets, including drip chambers, infusion tubing, infusion containers, and infusion pumps. Within the ROI, the video frame is cropped based on the spatial position of the infusion monitoring target within the video frame to obtain an infusion monitoring sub-region. Pixel-level semantic segmentation is performed on the infusion monitoring sub-region to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber outline, a multi-target tracking algorithm is used to obtain the droplet continuous trajectory tracking result. Based on the droplet continuous trajectory tracking result, the number of droplets per unit time is calculated to obtain the infusion drip rate. The remaining drug volume and remaining infusion time are calculated based on the liquid surface boundary position and the pixel-to-physical mapping relationship. The infusion drip rate, remaining drug volume, infusion tubing morphological features extracted based on the infusion tubing outline, infusion pump operating status obtained based on target detection, and occlusion degree of the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level.

[0020] S4 Alarm and Data Management: The edge computing unit generates alarm information based on the risk level, records alarm handling information, generates infusion process traceability data, and performs local storage.

[0021] As a further technical solution of the present invention, S1 specifically includes:

[0022] S1.1 An image acquisition unit is embedded in the ceiling of the ICU ward. The image acquisition unit is a camera device that meets the electromagnetic compatibility standards of medical electrical equipment and has a disinfection-proof shell and wide dynamic range imaging capability.

[0023] S1.2 Configure an embedded gateway with local AI computing power as an edge computing unit. The edge computing unit is physically isolated from the internal and external networks of the hospital and is equipped with local storage media.

[0024] S1.3 The image acquisition unit and the edge computing unit are connected to the hospital's local area network. The hospital's local area network adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at a fixed period. When the heartbeat is lost for a preset number of consecutive times, it automatically switches to the backup link. After the main link recovers stability and completes the status verification, it automatically switches back. In the case of network disconnection, the edge computing unit can cache monitoring data, and the data is stored using a cyclic overwrite strategy.

[0025] As a further technical solution of the present invention, S2 specifically includes:

[0026] S2.1 Acquire the initial monitoring screen, delineate the region of interest (ROI) for infusion monitoring in the screen, and determine the baseline range for algorithm operation;

[0027] S2.2 By calculating the pixel-to-physical conversion coefficient by the ratio of the actual effective height of the container to the corresponding image pixel height, the pixel-to-physical mapping relationship is established.

[0028] S2.3 Extract the reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions, as well as the reference parameters of ambient light, and use them together as the reference state parameters of infusion for subsequent anomaly comparison and light deviation correction.

[0029] S2.4 stores the Region of Interest (ROI), pixel-to-physical mapping relationship, and infusion baseline state parameters of the infusion monitoring as calibration data in the edge computing unit; the system has a built-in library of commonly used ICU infusion container specifications, and pre-stores the inherent size parameters and volume conversion parameters of various containers; the calibration results are completed once and remain effective for a long time.

[0030] As a further technical solution of the present invention, S3 specifically includes:

[0031] S3.1 Target Detection and Region Cropping: Locate the infusion monitoring target within the Region of Interest (ROI) for infusion monitoring and filter background interference. Cropping is performed based on the coordinates of the infusion monitoring target to obtain the infusion monitoring sub-region. S3.2 Pixel-Level Semantic Segmentation: Perform semantic segmentation on the infusion monitoring sub-region to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline.

[0032] S3.3 Calculation of infusion drip rate and remaining drug volume: The continuous trajectory of the droplet is obtained by a multi-target tracking algorithm and the infusion drip rate is calculated. The remaining drug volume and remaining infusion time are calculated by the liquid surface boundary position and the relationship between pixels and physical mapping.

[0033] S3.4 Multidimensional Anomaly Judgment: Integrates multiple monitoring features for comprehensive processing, and outputs the infusion anomaly type and corresponding risk level.

[0034] As a further technical solution of the present invention, S3 also satisfies:

[0035] S3.3.1 When calculating the infusion drip rate, a unique tracking identifier is assigned to each droplet, and droplets with the same tracking identifier are counted only once; a static filtering rule is set to determine the area without morphological changes for a consecutive preset number of frames as static background and remove it; when the infusion drip rate is detected to be lower than the preset infusion drip rate threshold, the pixel weight of the drip chamber area is increased to improve the recognition sensitivity of tiny droplets.

[0036] S3.3.2 When calculating the remaining volume of the liquid, the corresponding calculation model is matched according to the container type. For regular cylindrical containers, the remaining volume is calculated by the liquid level and the pre-stored cross-sectional area. For micro-pump syringes, the remaining volume is calculated by converting the piston position into the propulsion distance. For irregular soft bags, the pre-stored multi-segment fitted volume comparison table is called to correct the deformation error. For dark-colored liquids and light-proof liquids, the liquid surface boundary is identified by a combination of hue and saturation.

[0037] S3.4.1 When determining abnormalities, there are 8 categories of infusion abnormalities that can be determined: Infusion drip rate exceeding the upper limit set by the doctor's order and lasting for more than the preset duration; Infusion drip rate too slow or stopping abnormality: Infusion drip rate less than the lower limit set by the doctor's order, or no droplets falling for a continuous preset duration; Insufficient remaining medication warning: The remaining medication volume is less than the preset grading threshold; Infusion completion abnormality: The liquid level is lower than the minimum effective liquid level of the container and there is no medication in the drip chamber; Tubing bending or pressure abnormality: The radius of curvature of the infusion tubing profile is less than the preset threshold; Tubing displacement or detachment abnormality: The coordinates of the infusion tubing deviate from the reference position by more than the preset distance; Infusion monitoring sub-area obstruction abnormality: The obstructed area of ​​the infusion monitoring sub-area is greater than the preset proportion and lasts for more than the preset duration; Pump operating status abnormality: The operating status characteristics of the infusion pump disappear and a simultaneous stop in drip is detected.

[0038] As a further technical solution of the present invention, in S4, the alarm information is divided into three levels according to the risk level. Low-risk anomalies are indicated by large screen annotation and log recording. Medium-risk anomalies are indicated by large screen pop-ups, mobile terminal push notifications and local sound alerts. High-risk anomalies are indicated by large screen pop-ups, mobile terminal sound and light alarms and ward linkage emergency alarms. Data transmission and local storage are both encrypted. Sensitive field keys are managed by the hardware security module. Preset abnormal event data is marked as read-only and cannot be overwritten or tampered with.

[0039] As a further technical solution of the present invention, it also includes:

[0040] S5.1 Periodic Calibration: Automatically performs global calibration during low-load periods each day, updating pixel-to-physical mapping relationships, resetting infusion baseline status parameters, and correcting light deviations; triggers manual on-site calibration to recalibrate parameters when hardware is adjusted, containers are replaced, or lenses are shifted; all calibration operations generate log records and are stored.

[0041] S5.2 Privacy Desensitization: In step S3, privacy desensitization processing is also performed on the video frames to identify the patient's face and body privacy areas in real time, blur the privacy areas, and retain only the effective area for infusion monitoring.

[0042] Compared with existing technologies, the beneficial effects of this machine vision-based ICU infusion monitoring system and method are:

[0043] Adapted to the sterile treatment environment of the ICU, reducing the risk of infusion interference: Through the non-contact acquisition method of the image acquisition unit deployed above, the entire monitoring can be completed without physical contact with the infusion tubing, drip chamber and infusion container. This avoids displacement, bending and contamination of the precision infusion tubing caused by contact hardware, does not damage the sterile treatment environment of the ward, and meets the sterile operation requirements of ICU intensive care.

[0044] Improve the accuracy of infusion monitoring and adapt to low-speed precision infusion scenarios: The calculation range is narrowed by two-level region processing of the region of interest (ROI) and the sub-region of infusion monitoring. Pixel-level semantic segmentation is used to accurately extract the outline of the drip chamber, the liquid surface boundary, and the infusion tubing. The infusion drip rate is calculated based on continuous droplet trajectory tracking. The remaining amount of medication and the remaining infusion time are calculated by combining the pixel-physical mapping relationship and the multi-type container adaptation model. This can stably identify slow droplets and minute liquid surface changes, adapting to the precision monitoring needs of low-speed micro-infusion in the ICU.

[0045] Enhanced environmental adaptability and guaranteed system stability: The system adopts acquisition equipment with electromagnetic interference resistance and a dual-link transmission architecture, which can adapt to the complex electromagnetic environment of the ICU, variable lighting conditions and personnel obstruction scenarios; combined with static filtering rules to remove background interference, and with a periodic calibration mechanism to correct long-term operating deviations, the system reduces identification errors caused by environmental factors and ensures the long-term stability and reliability of monitoring results.

[0046] Achieving low-latency anomaly response and improving monitoring and treatment efficiency: Relying on edge computing units to complete all algorithm inference and anomaly judgment, without relying on cloud computing, shortening anomaly response time; By integrating multi-dimensional features such as infusion drip rate, remaining drug volume, infusion tubing morphology, infusion pump operating status and degree of obstruction, the type of infusion anomaly and corresponding risk level are determined, and combined with a graded alarm mechanism to achieve differentiated push for different risk anomalies, matching the treatment needs of different urgency levels in the ICU, and improving the response and treatment efficiency of infusion anomalies;

[0047] To ensure data security and compliance and meet the requirements for diagnosis and treatment traceability: It has the ability to process video frames for privacy and can blur the patient's privacy area to protect the patient's privacy; it adopts a full-link encryption mechanism for transmission and storage, retains alarm and handling information throughout the process and generates traceability data for the infusion process, forming a complete diagnosis and treatment data chain to meet the relevant requirements for medical data security management and diagnosis and treatment quality control traceability;

[0048] Simplify deployment and maintenance processes and reduce monitoring and management costs: There is no need to install a large number of contact-type sensor hardware on a single bed, deployment is convenient and there are few vulnerable parts, resulting in low daily maintenance workload; centralized monitoring and closed-loop alarm handling mode can reduce the burden of repetitive bedside inspections for nursing staff and improve the overall management efficiency of ICU infusion monitoring. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0050] Figure 2 This is an example diagram of the ICU ward location coverage plan of the present invention;

[0051] Figure 3 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figures 1-2 The present invention provides an embodiment 1: an ICU infusion monitoring system based on machine vision, comprising a sensing and acquisition module, a network transmission module, an edge inference module, and a business application module:

[0054] The sensing and acquisition module is installed above the ICU ward to collect real-time video streams from the infusion monitoring area;

[0055] The network transmission module adopts an intra-hospital local area network architecture to transmit the real-time video stream output by the sensing and acquisition module to the edge inference module, and transmit the monitoring data and alarm data output by the edge inference module to the business application module.

[0056] The edge inference module is used to pre-store calibration data. After receiving the real-time video stream, it calls the calibration data to perform target detection within the Region of Interest (ROI) for infusion monitoring to identify infusion monitoring targets. Based on the spatial position of the infusion monitoring targets in the video frame, it crops the infusion monitoring sub-regions. Pixel-level semantic segmentation is performed on the infusion monitoring sub-regions to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber outline, a multi-target tracking algorithm is used to obtain the continuous trajectory tracking results of the droplets. Based on the continuous trajectory tracking results, the number of droplets per unit time is calculated to obtain the infusion drip rate. Based on the liquid surface boundary position and the pixel-to-physical mapping relationship, the remaining amount of medication and the remaining infusion time are calculated. The infusion drip rate, remaining medication, infusion tubing morphology features, infusion pump operating status, and the degree of occlusion in the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level. The calibration data includes the ROI for infusion monitoring, pixel-to-physical mapping relationship, and infusion baseline state parameters. The infusion monitoring targets include the drip chamber, infusion tubing, infusion container, and infusion pump.

[0057] The business application module is used to receive monitoring data and abnormal results output by the edge inference module, generate corresponding alarm information, record alarm handling information, generate infusion process traceability data and perform storage.

[0058] Furthermore, the sensing and acquisition module is a camera device that complies with the electromagnetic compatibility standards of medical electrical equipment. It is embedded in the ceiling of the ICU ward and has a disinfection-proof shell and wide dynamic range imaging capabilities. The edge inference module is an embedded gateway with local AI computing power, physically isolated from the hospital's internal and external networks, and equipped with local storage media. The network transmission module adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at fixed intervals. If a heartbeat is lost for a preset number of consecutive times, it automatically switches to the backup link. After the main link recovers stability and completes status verification, it automatically switches back. In the event of a network outage, the edge inference module can cache monitoring data through local storage media, and the data is stored using a cyclic overwrite strategy.

[0059] Furthermore, the edge reasoning module has a built-in library of commonly used ICU infusion container specifications, pre-stores the inherent size parameters and volume conversion parameters of various containers; when performing infusion drip rate calculation, each droplet is assigned a unique tracking identifier, and droplets with the same tracking identifier are counted only once, and static filtering rules are set to remove static background interference; when performing drug remaining volume calculation, the corresponding calculation model is matched according to the container type, and a hue and saturation joint judgment method is used to identify the liquid surface boundary for dark-colored drugs and light-proof drugs; a total of 8 types of infusion abnormalities can be judged, including infusion drip rate overspeed abnormality, infusion drip rate too slow or drip stop abnormality, drug remaining volume warning, infusion completion abnormality, tubing bending or pressure abnormality, tubing displacement or detachment abnormality, infusion monitoring sub-area obstruction abnormality, and pump working status abnormality;

[0060] Alarm information for business application modules is divided into three levels according to risk level. Low-risk anomalies are indicated by large screen annotations and log recordings. Medium-risk anomalies are indicated by large screen pop-ups, mobile push notifications, and local sound alerts. High-risk anomalies are indicated by large screen pop-ups, mobile sound and light alarms, and ward-linked emergency alarms. Data transmission and local storage are encrypted. Sensitive field keys are managed by the hardware security module. Pre-set abnormal event data is marked as read-only and cannot be overwritten or tampered with.

[0061] Please see Figure 3 The present invention provides an embodiment 2: an ICU infusion monitoring method based on machine vision, comprising the following steps:

[0062] S1 System Deployment: An image acquisition unit is deployed above the ICU ward to acquire real-time video streams of the infusion monitoring area; an edge computing unit is configured for local AI algorithm inference, and the image acquisition unit transmits the acquired real-time video streams to the edge computing unit through the hospital's local area network;

[0063] Furthermore, S1 specifically includes:

[0064] S1.1 An image acquisition unit is embedded in the ceiling of the ICU ward. The image acquisition unit is a camera device that complies with the electromagnetic compatibility standards of medical electrical equipment and has a disinfection-proof shell and wide dynamic range imaging capability.

[0065] S1.2 is configured with an embedded gateway with local AI computing power as an edge computing unit. The edge computing unit is physically isolated from the internal and external networks of the hospital and is equipped with local storage media.

[0066] S1.3 connects the image acquisition unit and the edge computing unit to the hospital's local area network. The hospital's local area network adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at a fixed period. If the heartbeat is lost for a preset number of consecutive times, it will automatically switch to the backup link. After the main link recovers stability and completes the status verification, it will automatically switch back. In the case of network outage, the edge computing unit can cache monitoring data, and the data is stored using a cyclic overwrite strategy.

[0067] S2 Initialization Calibration: Before the system runs, the initial monitoring screen is acquired through the image acquisition unit. The region of interest (ROI) for infusion monitoring is defined in the initial monitoring screen. The pixel-to-physical mapping relationship between the image pixels and physical dimensions within the ROI is established. The reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions are extracted as infusion reference state parameters. The three calibration data, namely ROI, pixel-to-physical mapping relationship, and infusion reference state parameters, are stored in the edge computing unit.

[0068] Furthermore, S2 specifically includes:

[0069] S2.1 Acquire the initial monitoring screen, delineate the region of interest (ROI) for infusion monitoring in the screen, and determine the baseline range for algorithm operation;

[0070] S2.2 By calculating the pixel-to-physical conversion coefficient by the ratio of the actual effective height of the container to the corresponding image pixel height, the pixel-to-physical mapping relationship is established.

[0071] S2.3 Extract the reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions, as well as the reference parameters of ambient light, and use them together as the reference state parameters of infusion for subsequent anomaly comparison and light deviation correction.

[0072] S2.4 stores the Region of Interest (ROI), pixel-to-physical mapping relationship, and infusion baseline state parameters of the infusion monitoring as calibration data in the edge computing unit; the system has a built-in library of commonly used ICU infusion container specifications, and pre-stores the inherent size parameters and volume conversion parameters of various containers; the calibration results are completed once and remain effective for a long time;

[0073] S3 Local Inference Processing: The edge computing unit receives real-time video streams and calls calibration data. Within the ROI, it performs target detection to identify infusion monitoring targets, including drip chambers, infusion tubing, infusion containers, and infusion pumps. Within the ROI, the video frames are cropped based on the spatial position of the infusion monitoring targets in the video frames to obtain infusion monitoring sub-regions. Pixel-level semantic segmentation is performed on the infusion monitoring sub-regions to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber outline, a multi-target tracking algorithm is used to obtain the continuous trajectory tracking results of the droplets. Based on the continuous trajectory tracking results, the number of droplets per unit time is calculated to obtain the infusion drip rate. The remaining amount of medication and the remaining infusion time are calculated based on the liquid surface boundary position and the pixel-to-physical mapping relationship. The infusion drip rate, remaining medication, infusion tubing morphological features extracted based on the infusion tubing outline, infusion pump operating status obtained based on target detection, and occlusion degree of the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level.

[0074] Furthermore, S3 specifically includes:

[0075] S3.1 Target Detection and Region Cropping: Locate the infusion monitoring target within the Region of Interest (ROI) for infusion monitoring and filter background interference. Cropping is performed based on the coordinates of the infusion monitoring target to obtain the infusion monitoring sub-region. S3.2 Pixel-Level Semantic Segmentation: Perform semantic segmentation on the infusion monitoring sub-region to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline.

[0076] S3.3 Calculation of infusion drip rate and remaining drug volume: The continuous trajectory of the droplet is obtained by a multi-target tracking algorithm and the infusion drip rate is calculated. The remaining drug volume and remaining infusion time are calculated by the liquid surface boundary position and the relationship between pixels and physical mapping.

[0077] S3.4 Multi-dimensional anomaly determination: Integrates multiple monitoring features for comprehensive processing, and outputs the infusion anomaly type and corresponding risk level;

[0078] Furthermore, S3 also satisfies:

[0079] S3.3.1 When calculating the infusion drip rate, a unique tracking identifier is assigned to each droplet, and droplets with the same tracking identifier are counted only once; a static filtering rule is set to determine the area without morphological changes for a consecutive preset number of frames as static background and remove it; when the infusion drip rate is detected to be lower than the preset infusion drip rate threshold, the pixel weight of the drip chamber area is increased to improve the recognition sensitivity of tiny droplets.

[0080] S3.3.2 When calculating the remaining volume of the liquid, the corresponding calculation model is matched according to the container type. For regular cylindrical containers, the remaining volume is calculated by the liquid level and the pre-stored cross-sectional area. For micro-pump syringes, the remaining volume is calculated by converting the piston position into the propulsion distance. For irregular soft bags, the pre-stored multi-segment fitted volume comparison table is called to correct the deformation error. For dark-colored liquids and light-proof liquids, the liquid surface boundary is identified by a combination of hue and saturation.

[0081] S3.4.1 When determining abnormalities, there are 8 categories of infusion abnormalities that can be determined: Infusion drip rate exceeding the upper limit set by the doctor's order and the duration exceeding the preset time; Infusion drip rate too slow or stopping abnormality: Infusion drip rate less than the lower limit set by the doctor's order, or no dripping for a continuous preset time; Insufficient remaining medication warning: The remaining medication volume is less than the preset grading threshold; Infusion completion abnormality: The liquid level is lower than the minimum effective liquid level of the container and there is no medication in the drip chamber; Tubing bending or pressure abnormality: The radius of curvature of the infusion tubing profile is less than the preset threshold; Tubing displacement or detachment abnormality: The coordinates of the infusion tubing deviate from the reference position by a preset distance; Infusion monitoring sub-area obstruction abnormality: The obstructed area of ​​the infusion monitoring sub-area is greater than the preset proportion and the duration exceeds the preset time; Pump operating status abnormality: The operating status characteristics of the infusion pump disappear and a simultaneous stop-drip is detected.

[0082] S4 Alarm and Data Management: The edge computing unit generates alarm information based on risk level, records alarm handling information, generates infusion process traceability data and performs local storage;

[0083] Furthermore, in S4, alarm information is divided into three levels according to risk level. Low-risk anomalies are indicated by large screen annotations and log recordings. Medium-risk anomalies are indicated by large screen pop-ups, mobile push notifications, and local sound alerts. High-risk anomalies are indicated by large screen pop-ups, mobile sound and light alarms, and ward-linked emergency alarms. Data transmission and local storage are both encrypted. Sensitive field keys are managed by the hardware security module. Preset abnormal event data is marked as read-only and cannot be overwritten or tampered with.

[0084] Furthermore, it also includes:

[0085] S5.1 Periodic Calibration: Automatically performs global calibration during low-load periods each day, updating pixel-to-physical mapping relationships, resetting infusion baseline status parameters, and correcting light deviations; triggers manual on-site calibration to recalibrate parameters when hardware is adjusted, containers are replaced, or lenses are shifted; all calibration operations generate log records and are stored.

[0086] S5.2 Privacy Desensitization: In step S3, privacy desensitization processing is also performed on the video frames to identify the patient's face and body privacy areas in real time, blur the privacy areas, and retain only the effective area for infusion monitoring.

[0087] Example 3: An ICU infusion monitoring system based on machine vision:

[0088] This system comprises a sensing and acquisition module, a network transmission module, an edge inference module, and a business application module, which are connected in sequence via communication. The specific implementation methods of each module are as follows:

[0089] 1. The sensing and acquisition module is embedded in the ceiling directly above the IV stand in the ICU ward, with a fixed 30° downward angle and the lens facing the IV monitoring area. The module is a medical-grade 4K camera, and its electromagnetic compatibility meets the YY0505-2012 standard: electrostatic discharge immunity ±8kV for contact discharge and ±15kV for air discharge; radio frequency radiation immunity 10V / m (80MHz~2.7GHz); electrical fast transient / burst immunity ±2kV (power line) and ±1kV (signal line), resisting electromagnetic interference from ICU monitors, ventilators, high-frequency electrosurgical units, and other equipment. The device casing is made of medical-grade ABS+PC flame-retardant material. The device has an IP66 protection rating and can withstand disinfection by wiping with 75% medical alcohol, chlorine-containing disinfectant with an effective chlorine concentration of 500mg / L, and quaternary ammonium salt disinfectant at least four times a day. It can withstand continuous use for five years without shell aging, coating peeling, or lens fogging. Imaging parameters include: 3840×2160 resolution, fixed frame rate of 30fps, and 120dB ultra-wide dynamic range. It is equipped with a non-red-explosive infrared supplementary light unit, enabling normal imaging under 0.001 lux illumination. The horizontal field of view is 90°. It uses a national standard POE power supply module and anti-interference shielded cable. The sensing and acquisition module collects real-time video streams from the infusion monitoring area and transmits them to the edge inference module via a network transmission module.

[0090] 2. The network transmission module adopts an intra-hospital local area network architecture with physical isolation from the external network; video streams are transmitted using the RTSP protocol, while structured monitoring data and alarm data are transmitted using the MQTT protocol; this embodiment employs a dual transmission architecture consisting of a PoE wired main link and an intra-network WiFi 6 backup link: the main link heartbeat detection cycle is 500ms, and if three consecutive heartbeats are lost, it automatically and seamlessly switches to the backup link within 1.5s, with no packet loss during the switching process; after the main link recovers, it automatically switches back to the main link after confirming stable operation for 30s and that there is no untransmitted data on the backup link; data transmission uses TLS1.3+AES-256-GCM encryption mechanism;

[0091] 3. The edge inference module is an embedded industrial gateway with a built-in NPU computing power of no less than 16 TOPS, a built-in 128GSSD local storage medium, and runs an embedded Linux industrial system. It has no external network interface and is physically isolated from the external network. The edge inference module pre-stores calibration data, including the region of interest (ROI) for infusion monitoring, pixel-to-physical mapping relationship, and infusion baseline state parameters. After receiving the real-time video stream, it calls the calibration data, performs inference operations according to a fixed time sequence, and outputs monitoring data and anomaly judgment results. The module has a built-in library of commonly used ICU infusion container specifications, pre-stores the inherent size parameters and volume conversion parameters of hard glass bottles, standard soft bags, micro-infusion pump syringes, and irregular soft bags. Sensitive fields such as patient ID and medication information in local storage are encrypted using AES-256-CBC+HMAC-SHA256, and the key is managed by the device's security chip and is not written to disk in plaintext. In the offline state, a single channel can cache no less than 72 hours of monitoring data at full load, using a FIFO cyclic overwrite strategy. Pre-set abnormal event data is automatically marked as read-only and will not be overwritten.

[0092] 4. The business application module includes a centralized monitoring screen at the nurse station and a handheld PDA terminal for nursing staff. These are used to receive monitoring data and abnormal results output by the edge inference module, generate corresponding alarm information, record alarm handling information, generate infusion process traceability data, and execute and store it. Alarm information is divided into three levels according to risk level, corresponding to different push and prompt methods. All handling operations and monitoring data are fully traceable.

[0093] Example 4: A machine vision-based method for monitoring infusion in the ICU:

[0094] This method is based on the above system and follows the step logic of the method claims throughout. All calculation formulas strictly adopt the definitions and forms given in the raw materials.

[0095] S1 system deployment:

[0096] S1.1 An image acquisition unit is embedded in the ceiling of the ICU ward, with the lens tilted to 30° to fully cover the infusion monitoring area containing the IV stand, drip chamber, infusion container, and infusion pump. A PoE power supply module and shielded cable are installed, with the cable running inside the ceiling and no exposed wiring. S1.2 An edge computing unit is deployed in the ward's low-voltage room, connected to the hospital's local area network, and configured with an IP whitelist and port isolation. S1.3 A dual-transmission architecture is established, consisting of a PoE wired main link and an internal WiFi 6 backup link. The heartbeat detection cycle is configured to 500ms, automatically switching to the backup link after three consecutive heartbeat losses, and automatically switching back to the main link 30 seconds after it stabilizes. The real-time video stream acquired by the image acquisition unit is transmitted to the edge computing unit via the hospital's local area network.

[0097] S2 initialization calibration:

[0098] A one-time calibration is performed before the system goes live, and the calibration results are valid indefinitely. The specific process is as follows:

[0099] S2.1 The control image acquisition unit acquires the initial monitoring image of the current infusion monitoring area, delineates the region of interest (ROI) for infusion monitoring in the initial monitoring image, and selects the range to cover all infusion monitoring targets, including the drip chamber, infusion tubing, infusion container, and infusion pump body. Irrelevant background areas are removed, and the baseline range for algorithm operation is determined.

[0100] S2.2 Establishing the pixel-to-physical mapping relationship: Select the currently used standard infusion container, and set the effective total height of the container as... (Unit: mm), corresponding to the image pixel height is (Unit: pixel) Calculate the pixel to physical conversion factor using the following formula. : In this embodiment, a standard 500ml soft infusion bag is used. It is 200mm. For 1000 pixels, the calculation is as follows This establishes a pixel-to-physical mapping relationship;

[0101] S2.3 Under normal and stable infusion conditions, extract the current infusion tubing reference morphological parameters, drip chamber reference state parameters, and ambient light reference parameters, and use them together as infusion reference state parameters for subsequent anomaly comparison and light deviation correction;

[0102] S2.4 stores the region of interest (ROI) for infusion monitoring, the pixel-to-physical mapping relationship, and the infusion baseline state parameters as calibration data in the edge computing unit;

[0103] S3 local inference processing:

[0104] The edge computing unit receives real-time video streams and calls calibration data, performing inference according to a fixed timing sequence:

[0105] S3.1 Target Detection and Region Cropping: Target detection is performed within the Region of Interest (ROI) for infusion monitoring to identify and locate infusion monitoring targets, including drip chambers, infusion tubing, infusion containers, and infusion pumps, while filtering out background interference targets such as medical personnel and equipment supports; based on the spatial position of the infusion monitoring targets in the video frame, the video frame is cropped within the ROI to obtain the infusion monitoring sub-region;

[0106] S3.2 Pixel-level semantic segmentation: Pixel-level semantic segmentation is performed on the infusion monitoring sub-region to distinguish the boundaries of liquid, tubing and air pixel by pixel, and to extract the outline of the drip chamber, the boundary of the liquid surface and the outline of the infusion tubing.

[0107] S3.3 Calculation of infusion drip rate and remaining drug volume: Two calculations are executed in parallel, and the results are output synchronously.

[0108] (1) Infusion drip rate calculation: Based on the droplet targets detected within the drip chamber contour, a multi-target tracking algorithm is used to assign a unique tracking identifier to each droplet and continuously track the droplet's falling trajectory; a static filtering rule is set to determine and remove areas with no morphological changes for 20 consecutive frames as static background, filtering out reflections and bubble interference; the statistical duration is set to be... (Unit: minutes) The total number of effective droplets within the time period is Calculate the infusion drip rate using the following formula. (Unit: drops / minute): Droplets with the same tracking identifier are counted only once to avoid duplicate counting and missed counting; when the infusion drip rate is detected to be less than 20 drops / min, the algorithm automatically amplifies the pixel weight of the drip chamber area to improve the recognition sensitivity of tiny droplets;

[0109] (2) Calculation of remaining drug volume and remaining infusion time: Assume the current liquid level pixel height is Based on the pixel-to-physical mapping relationship, the actual physical height of the current liquid surface is calculated using the following formula. : Match the corresponding computation model based on the container type:

[0110] Regular cylindrical containers (hard glass bottles, standard soft bags): Let the cross-sectional area of ​​the container be... (Pre-stored in the container specification library), calculate the remaining liquid volume using the following formula. :

[0111] Micro-infusion pump syringe: Identify piston pixel position, calculate piston advance distance, and determine remaining volume;

[0112] Irregular soft bags: The pre-stored multi-segment fitted volume comparison table is called to match the volume corresponding to the current liquid level and correct the deformation error; for transparent liquids, the brightness difference is used to determine the liquid surface boundary; for dark liquids and light-proof liquids, hue + saturation are used to determine the liquid surface boundary; finally, the remaining liquid volume is calculated and the remaining infusion time is dynamically estimated in combination with the real-time infusion drip rate.

[0113] S3.4 Multi-dimensional Anomaly Detection: This function extracts five types of features: infusion drip rate, remaining medication volume, infusion tubing morphology features extracted from the tubing outline, infusion pump operating status identified from target detection results, and the degree of occlusion in the infusion monitoring sub-region. These multi-dimensional features are fused and matched against preset judgment rules to output the infusion anomaly type and corresponding risk level. Eight types of infusion anomalies can be identified, with the following judgment rules: ① Infusion drip rate exceeding the prescribed limit: The measured infusion drip rate exceeds the upper limit set by the doctor, and the duration exceeds 3 seconds; ② Infusion drip rate too slow or stopped: The measured infusion drip rate is less than the upper limit set by the doctor. ① Limit, or no droplets falling for 10 consecutive seconds; ② Insufficient medication remaining warning: The remaining medication volume is less than the preset classification threshold; ③ Infusion completion abnormality: The liquid level is lower than the lowest effective liquid level of the container, and there is no medication in the drip chamber; ④ Abnormality of tubing bending or pressure: The radius of curvature of the infusion tubing profile is less than the preset threshold; ⑤ Abnormality of tubing displacement or detachment: The coordinates of the infusion tubing deviate from the reference position by more than the preset pixel distance; ⑥ Abnormality of infusion monitoring sub-area occlusion: The occluded area of ​​the infusion monitoring sub-area is greater than 60%, and the duration exceeds 5 seconds; ⑦ Abnormality of pump working status: The imaging features of the infusion pump operation indicator area disappear, and dripping stop is detected simultaneously;

[0114] As a preferred implementation, privacy desensitization processing is performed on video frames simultaneously during the inference process: real-time identification of the patient's facial and body privacy areas, blurring of the privacy areas, and retention of only the effective area for infusion monitoring;

[0115] S4 Alarm and Data Management:

[0116] The edge computing unit generates corresponding alarm information based on the determined risk level and pushes it to the business application module:

[0117] Level 1 Warning (Low Risk): Corresponds to 20% of the remaining medication, slight deviation of the infusion rate from the doctor's order (deviation < ±2 drops / min), and slight local obstruction. This is indicated by light-colored markings on the nurse station screen and only recorded in the log.

[0118] Level 2 Warning (Medium Risk): Corresponds to 10% of the remaining medication, a significant deviation of the infusion drip rate from the doctor's order (deviation ≥ ±2 drops / min), and a slight bend in the tubing. The warning will be issued via a high-brightness pop-up window on a large screen, text push notifications on the nursing PDA, and short beep alerts on local devices.

[0119] Level 3 Emergency (High Risk): Corresponds to infusion completion, tubing detachment / severe bending, complete cessation of dripping, pump shutdown, and large-area obstruction. This is achieved through a large-screen pop-up window, nursing PDA audible and visual alarms, and ward-side audible and visual linkage. After nursing staff complete the procedure, they enter the results, and the system automatically links the alarm event with the procedure information to form a complete alarm and procedure record. After each infusion, complete infusion process traceability data is generated, encrypted, and stored locally on the edge computing unit.

[0120] Preferred supplementary steps: Regular calibration:

[0121] Automatic scheduled verification: Every day at 02:00 AM during the low-load period of the ICU, the system automatically performs a global verification, recalibrates the pixel-physical mapping relationship, resets the infusion baseline status parameters, and corrects the recognition deviation caused by changes in day and night light. After the verification is completed, a log is automatically generated, and an operation and maintenance reminder is automatically pushed when the deviation exceeds the preset threshold.

[0122] Manual calibration: When scenarios such as replacing the image acquisition unit, adjusting the installation angle, using a new infusion container, or lens shifting after disinfection occur, manual on-site calibration is triggered to update the calibration data; all calibration operations are fully recorded.

[0123] Example 5, Technical Effect Verification:

[0124] The testing environment, testing methods, and judgment criteria in this verification embodiment are all derived from the raw materials, and are reproducible and verifiable. General testing conditions: simulating a standard single-patient ICU ward in a tertiary hospital, with electromagnetic interference from monitors and ventilators, covering both standard indoor lighting and nighttime infrared modes; testing equipment includes this monitoring system, a medical stopwatch, a high-precision electronic scale, standard measuring cups, and a medical micro-infusion pump; test samples cover standard medications, dark-colored chemotherapy solutions, light-protected medications, as well as 500ml / 250ml / 100ml infusion bags / bottles and 50ml micro-infusion pump syringes; manual comparison and statistical analysis were performed by two certified ICU nurses.

[0125] 1. Infusion drip rate accuracy verification test conditions: The micro-infusion pump was set with infusion drip rates of 5 drops / min, 10 drops / min, and 18 drops / min, covering the mainstream low-speed range in the ICU; Test procedure: After the system ran stably for 5 minutes, the timing started, and the total number of drops in 10 minutes was counted manually with a stopwatch, and the infusion drip rate data output by the system was read simultaneously; Each test condition was repeated 5 times, and the average value was taken; Test results: The maximum difference between the manually measured average value and the system output average value was 0.4 drops / min, which meets the accuracy requirement of error ≤ ±0.5 drops / min;

[0126] 2. Verification Test Conditions for Remaining Drug Volume Measurement Accuracy: A 500ml soft bag and a 50ml micro-syringe were selected, and tests were conducted at remaining liquid levels of 80%, 50%, and 20%, respectively. Test Procedure: The drug was emptied or added to the designated level, and the weight was converted to a standard volume using an electronic scale as the true value. The remaining drug volume calculated by the system's visual measurement was read. The measurement accuracy was calculated using the following formula: Test results: The calculation accuracy under different containers and different liquid levels is ≥98%;

[0127] 3. Target Recognition Accuracy Verification Test Sample: 10,000 frames of video were continuously collected from the ICU over 24 hours, including scenes with personnel movement, equipment obstruction, and strong / weak light; Statistical objects: three types of infusion monitoring targets: drip chambers, infusion lines, and infusion containers; Test results: Target recognition accuracy ≥ 99.5%;

[0128] 4. System response delay verification test objects: three types of anomalies: pipe detachment, dripping stoppage, and insufficient reserve; test method: use a high-speed timer to record the time difference between the occurrence of the anomaly and the system output alarm; test results: the maximum response delay for all three types of anomalies is ≤200ms;

[0129] False alarm / missed alarm rate verification test conditions: Under normal ICU treatment environment, slight obstruction, light changes and electromagnetic interference were artificially simulated and the system was run continuously for 72 hours; Statistical results: All real abnormalities were identified within 72 hours, and the missed alarm rate was 0; The number of invalid alarms was extremely small, and the false alarm rate was far lower than that of traditional contact monitoring equipment.

[0130] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An ICU infusion monitoring system based on machine vision, characterized in that, It includes a perception and acquisition module, a network transmission module, an edge inference module, and a business application module: The sensing and acquisition module is installed above the ICU ward and is used to collect real-time video streams of the infusion monitoring area; The network transmission module adopts an intra-hospital local area network architecture to transmit the real-time video stream output by the sensing and acquisition module to the edge inference module, and transmit the monitoring data and alarm data output by the edge inference module to the business application module. The edge inference module is used to pre-store calibration data. After receiving the real-time video stream, it calls the calibration data to perform target detection within the Region of Interest (ROI) for infusion monitoring to identify infusion monitoring targets. Based on the spatial position of the infusion monitoring targets in the video frame, it crops the infusion monitoring sub-regions. Pixel-level semantic segmentation is performed on the infusion monitoring sub-regions to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber outline, a multi-target tracking algorithm is used to obtain the continuous trajectory tracking results of the droplets. Based on the continuous trajectory tracking results, the number of droplets per unit time is calculated to obtain the infusion drip rate. Based on the liquid surface boundary position and the pixel-to-physical mapping relationship, the remaining amount of medication and the remaining infusion time are calculated. The infusion drip rate, remaining amount of medication, infusion tubing morphology features, infusion pump operating status, and occlusion degree of the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level. The calibration data includes the Region of Interest (ROI) for infusion monitoring, pixel-to-physical mapping relationship, and infusion baseline state parameters. The infusion monitoring targets include the drip chamber, infusion tubing, infusion container, and infusion pump. The business application module is used to receive monitoring data and abnormal results output by the edge inference module, generate corresponding alarm information, record alarm handling information, generate infusion process traceability data, and perform storage.

2. The ICU infusion monitoring system based on machine vision according to claim 1, characterized in that, The sensing and acquisition module is a camera device that complies with the electromagnetic compatibility standards of medical electrical equipment. It is embedded in the ceiling of the ICU ward and has a disinfection-proof shell and wide dynamic range imaging capabilities. The edge inference module is an embedded gateway with local AI computing power, physically isolated from the hospital's internal and external networks, and equipped with local storage media. The network transmission module adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at a fixed period. If a heartbeat is lost for a preset number of consecutive times, it automatically switches to the backup link. After the main link recovers stability and completes status verification, it automatically switches back. In the offline state, the edge inference module can cache monitoring data through local storage media, and the data is stored using a cyclic overwrite strategy.

3. The ICU infusion monitoring system based on machine vision according to claim 1, characterized in that, The edge reasoning module has a built-in library of commonly used ICU infusion container specifications, and pre-stores the inherent size parameters and volume conversion parameters of various containers. When performing infusion drip rate calculation, a unique tracking identifier is assigned to each droplet, and droplets with the same tracking identifier are counted only once. Static filtering rules are set to remove static background interference. When calculating the remaining amount of medication, the corresponding calculation model is matched according to the container type. For dark-colored and light-proof medications, a combined hue and saturation judgment method is used to identify the liquid surface boundary. There are a total of 8 types of infusion abnormalities that can be judged, including infusion drip rate exceeding the normal speed, infusion drip rate being too slow or stopping, warning of insufficient remaining medication, infusion completion abnormality, tubing bending or pressure abnormality, tubing displacement or detachment abnormality, obstruction of infusion monitoring sub-area abnormality, and pump working status abnormality. The alarm information of the business application module is divided into three levels according to the risk level. Low-risk anomalies are indicated by large screen annotation and log recording. Medium-risk anomalies are indicated by large screen pop-up, mobile terminal push and local prompt sound. High-risk anomalies are indicated by large screen top pop-up, mobile terminal sound and light alarm and ward linkage emergency alarm. Both data transmission and local storage employ encryption mechanisms. Sensitive field keys are managed by a hardware security module, and abnormal event data is pre-marked as read-only and cannot be repeatedly overwritten or tampered with.

4. A machine vision-based method for monitoring infusion in the ICU, characterized in that, Includes the following steps: S1 System Deployment: An image acquisition unit is deployed above the ICU ward. The image acquisition unit is used to acquire real-time video streams of the infusion monitoring area. An edge computing unit is configured for local AI algorithm inference. The image acquisition unit transmits the acquired real-time video streams to the edge computing unit through the hospital's local area network. S2 Initialization Calibration: Before the system runs, the initial monitoring screen is acquired through the image acquisition unit. The region of interest (ROI) for infusion monitoring is defined in the initial monitoring screen. The pixel-to-physical mapping relationship between the image pixels and physical dimensions within the ROI is established. The reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions are extracted as infusion reference state parameters. The three calibration data, namely the ROI, the pixel-to-physical mapping relationship, and the infusion reference state parameters, are stored in the edge computing unit. S3 Local Inference Processing: The edge computing unit receives the real-time video stream and calls the calibration data. Within the ROI, it performs target detection to identify infusion monitoring targets, including drip chambers, infusion tubing, infusion containers, and infusion pumps. Within the ROI, the video frame is cropped based on the spatial position of the infusion monitoring target in the video frame to obtain an infusion monitoring sub-region. Pixel-level semantic segmentation is performed on the infusion monitoring sub-region to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. Based on the droplet targets detected within the drip chamber contour, a multi-target tracking algorithm is used to obtain the continuous trajectory tracking results of the droplets. The number of droplets per unit time is calculated based on the continuous trajectory tracking results to obtain the infusion drip rate. The remaining amount of medicine and the remaining infusion time are calculated based on the liquid surface boundary position and the pixel-physical mapping relationship. The infusion drip rate, the remaining amount of medicine, the morphological features of the infusion tubing extracted based on the infusion tubing contour, the operating status of the infusion pump obtained based on target detection, and the degree of occlusion of the infusion monitoring sub-region are fused to output the infusion anomaly type and corresponding risk level. S4 Alarm and Data Management: The edge computing unit generates alarm information based on the risk level, records alarm handling information, generates infusion process traceability data, and performs local storage.

5. The ICU infusion monitoring method based on machine vision according to claim 4, characterized in that, S1 specifically includes: S1.1 An image acquisition unit is embedded in the ceiling of the ICU ward. The image acquisition unit is a camera device that meets the electromagnetic compatibility standards of medical electrical equipment and has a disinfection-proof shell and wide dynamic range imaging capability. S1.2 Configure an embedded gateway with local AI computing power as an edge computing unit. The edge computing unit is physically isolated from the internal and external networks of the hospital and is equipped with local storage media. S1.3 The image acquisition unit and the edge computing unit are connected to the hospital's local area network. The hospital's local area network adopts a dual transmission architecture consisting of a main link and a backup link. The main link performs heartbeat detection at a fixed period. When the heartbeat is lost for a preset number of consecutive times, it automatically switches to the backup link. After the main link recovers stability and completes the status verification, it automatically switches back. In the case of network disconnection, the edge computing unit can cache monitoring data, and the data is stored using a cyclic overwrite strategy.

6. The ICU infusion monitoring method based on machine vision according to claim 4, characterized in that, S2 specifically includes: S2.1 Acquire the initial monitoring screen, delineate the region of interest (ROI) for infusion monitoring in the screen, and determine the baseline range for algorithm operation; S2.2 By calculating the pixel-to-physical conversion coefficient by the ratio of the actual effective height of the container to the corresponding image pixel height, the pixel-to-physical mapping relationship is established. S2.3 Extract the reference morphological parameters of the infusion tubing and the reference state parameters of the drip chamber under normal infusion conditions, as well as the reference parameters of ambient light, and use them together as the reference state parameters of infusion for subsequent anomaly comparison and light deviation correction. S2.4 stores the Region of Interest (ROI), pixel-to-physical mapping relationship, and infusion baseline state parameters of the infusion monitoring as calibration data in the edge computing unit; the system has a built-in library of commonly used ICU infusion container specifications, and pre-stores the inherent size parameters and volume conversion parameters of various containers; the calibration results are completed once and remain effective for a long time.

7. The ICU infusion monitoring method based on machine vision according to claim 4, characterized in that, S3 specifically includes: S3.1 Target Detection and Region Cropping: Locate the infusion monitoring target within the Region of Interest (ROI) for infusion monitoring and filter background interference. Cropping is performed based on the coordinates of the infusion monitoring target to obtain the infusion monitoring sub-region. S3.2 Pixel-Level Semantic Segmentation: Perform semantic segmentation on the infusion monitoring sub-region to extract the drip chamber outline, liquid surface boundary, and infusion tubing outline. S3.3 Calculation of infusion drip rate and remaining drug volume: The continuous trajectory of the droplet is obtained by a multi-target tracking algorithm and the infusion drip rate is calculated. The remaining drug volume and remaining infusion time are calculated by the liquid surface boundary position and the relationship between pixels and physical mapping. S3.4 Multidimensional Anomaly Judgment: Integrates multiple monitoring features for comprehensive processing, and outputs the infusion anomaly type and corresponding risk level.

8. The ICU infusion monitoring method based on machine vision according to claim 7, characterized in that, S3 also satisfies: S3.3.1 When calculating the infusion drip rate, a unique tracking identifier is assigned to each droplet, and droplets with the same tracking identifier are counted only once; a static filtering rule is set to determine the area without morphological changes for a consecutive preset number of frames as static background and remove it; when the infusion drip rate is detected to be lower than the preset infusion drip rate threshold, the pixel weight of the drip chamber area is increased to improve the recognition sensitivity of tiny droplets. S3.3.2 When calculating the remaining volume of the liquid, the corresponding calculation model is matched according to the container type. For regular cylindrical containers, the remaining volume is calculated by the liquid level and the pre-stored cross-sectional area. For micro-pump syringes, the remaining volume is calculated by converting the piston position into the propulsion distance. For irregular soft bags, the pre-stored multi-segment fitted volume comparison table is called to correct the deformation error. For dark-colored liquids and light-proof liquids, the liquid surface boundary is identified by a combination of hue and saturation. S3.4.1 When determining abnormalities, there are 8 categories of infusion abnormalities that can be determined: Infusion drip rate exceeding the upper limit set by the doctor's order and lasting for more than the preset duration; Infusion drip rate too slow or stopping abnormality: Infusion drip rate less than the lower limit set by the doctor's order, or no droplets falling for a continuous preset duration; Insufficient remaining medication warning: The remaining medication volume is less than the preset grading threshold; Infusion completion abnormality: The liquid level is lower than the minimum effective liquid level of the container and there is no medication in the drip chamber; Tubing bending or pressure abnormality: The radius of curvature of the infusion tubing profile is less than the preset threshold; Tubing displacement or detachment abnormality: The coordinates of the infusion tubing deviate from the reference position by more than the preset distance; Infusion monitoring sub-area obstruction abnormality: The obstructed area of ​​the infusion monitoring sub-area is greater than the preset proportion and lasts for more than the preset duration; Pump operating status abnormality: The operating status characteristics of the infusion pump disappear and a simultaneous stop in drip is detected.

9. The ICU infusion monitoring method based on machine vision according to claim 4, characterized in that, In S4, alarm information is divided into three levels according to risk level. Low-risk anomalies are indicated by large screen labeling and log recording. Medium-risk anomalies are indicated by large screen pop-up, mobile push and local sound alerts. High-risk anomalies are indicated by large screen top pop-up, mobile sound and light alarm and ward linkage emergency alarm. Both data transmission and local storage employ encryption mechanisms. Sensitive field keys are managed by a hardware security module, and abnormal event data is pre-marked as read-only and cannot be repeatedly overwritten or tampered with.

10. The ICU infusion monitoring method based on machine vision according to claim 4, characterized in that, Also includes: S5.1 Periodic Calibration: Automatically performs global calibration during low-load periods each day, updating pixel-to-physical mapping relationships, resetting infusion baseline status parameters, and correcting light deviations; triggers manual on-site calibration to recalibrate parameters when hardware is adjusted, containers are replaced, or lenses are shifted; all calibration operations generate log records and are stored. S5.2 Privacy Desensitization: In step S3, privacy desensitization processing is also performed on the video frames to identify the patient's face and body privacy areas in real time, blur the privacy areas, and retain only the effective area for infusion monitoring.

Images