A video compression and transmission method and system for an intelligent monitoring system in a hospital ward

By acquiring the baseline coordinates and width of the human posture bounding box in the intelligent monitoring system of the ward, dynamically limiting the effective tracking area, constructing a diffusion transfer matrix and assigning differentiated quantization parameters, the problem of video compression and blurring caused by the lag in visual model updates is solved, and the accurate transmission of key actions and continuous monitoring of fall events are realized.

CN123123582APending Publication Date: 2026-11-17NORTHWEST WOMEN & CHILDREN HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611296405.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent monitoring systems for wards cannot guarantee the continuity of key actions during the lag in visual model updates, resulting in blurred and distorted edges of key actions during video compression, which affects the accuracy of fall monitoring.

Method used

By obtaining the reference coordinates and width of the human pose bounding box, the effective tracking area is dynamically defined. A diffusion transfer matrix is ​​constructed based on the underlying downward motion vector to generate a probability distribution vector. Differentiated quantization parameters are assigned for video compression to ensure accurate transmission of key actions.

Benefits of technology

It improves the efficiency and accuracy of motion feature extraction, ensures the continuous monitoring and alarm accuracy of the cloud server for fall events, and eliminates the key motion ambiguity caused by visual update lag.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123582A_ABST
    Figure CN123123582A_ABST
Patent Text Reader

Abstract

This invention relates to the field of video compression technology, specifically to a video compression and transmission method and system for an intelligent monitoring system in a hospital ward. The method obtains the reference coordinates and width of the human body's pose bounding box, defines the effective tracking area during the encoding period without visual detection, obtains the estimated downward motion value based on the underlying downward motion vector within the effective tracking area, and then accumulates and updates the downward offset value; constructs a macroblock mesh, configures upward blocking and downward dynamically weighted connection weights to generate a diffusion transfer matrix, uses the human body's center node as the source point, and performs cyclic deduction using the diffusion transfer matrix to generate a probability distribution vector; truncates and zeroes the probability distribution vector based on the downward offset value to obtain a retained weight matrix; and assigns quantization parameters to the macroblocks based on their positions and corresponding weights and performs video compression encapsulation. This invention effectively solves the key motion blur defect caused by visual update lag and achieves precise allocation of bandwidth resources.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of video compression technology, specifically to a video compression and transmission method and system for an intelligent monitoring system in a hospital ward. Background Technology

[0002] With the development of smart healthcare technology, edge intelligent monitoring systems are typically deployed in hospital wards. Limited by the bandwidth constraints of the hospital's local area network, these systems usually employ an "edge compression-cloud recognition" architecture. This means that edge cameras compress video streams before transmitting them to a cloud server, which then identifies abnormal behaviors such as falls. To ensure image quality in critical areas with limited bandwidth, existing technologies typically employ a region-of-interest (ROI) dynamic encoding strategy based on visual models. This involves using a visual model to detect human bounding boxes, applying low quantization parameters within the bounding boxes to ensure image clarity, and applying high quantization parameters to the static background area outside the bounding boxes to significantly discard data and conserve network bandwidth.

[0003] However, due to the limited computing power of edge monitoring devices, the detection update frequency of visual models is usually much lower than the high frame output frequency of hardware video encoders, resulting in a long visual update stagnation period between consecutive encoded frames. When a patient in the ward falls rapidly downward, their limbs and torso quickly leave the initial bounding box protection area defined by the previous visual cycle. Because the underlying video encoder lacks timely target position updates, it will incorrectly identify the falling torso and limbs that have left the original boundary area as static background, thereby performing a high-ratio blind compression on them, resulting in severe blurring and distortion of the edges of key actions.

[0004] To compensate for the aforementioned visual lag, existing technologies typically employ a defensive strategy of proportionally enlarging the initial human body bounding box in two dimensions. However, this strategy severely deviates from the actual physical form of a person falling, resulting in the limited bandwidth being equally allocated to a large number of invalid areas around the enlarged box. Under constant bit rate constraints, the available bandwidth for core moving limbs is severely diluted by a large amount of useless background, leading to a break in the continuity of skeleton extraction on the backend cloud server and missed detection of fall events, greatly reducing the reliability of the ward intelligent monitoring system. Summary of the Invention

[0005] To address the technical problems of existing intelligent ward monitoring videos being unable to guarantee the continuity of key actions and accurately compress video during the visual model update lag period, the present invention aims to provide a video compression and transmission method and system for an intelligent ward monitoring system. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides a video compression and transmission method for an intelligent monitoring system in a ward, the method comprising the following steps: Obtain the base coordinates and width of the human pose bounding box, and initialize the downward offset value; During the non-visual detection encoding period, the effective tracking area is dynamically defined based on the reference coordinates, width, and downward offset value. Based on the underlying downward motion vector within the effective tracking area, the downward motion estimate for each frame is obtained, and then the downward offset value corresponding to each frame is accumulated and updated. A macroblock mesh is constructed. Based on the positional relationship between mesh nodes and the estimated downward motion, connection weights that block upward movement and dynamically weight downward movement are configured to generate a diffusion transfer matrix for each frame. Using the human body center node as the source point, a cyclic deduction is performed using the diffusion transfer matrix to generate the probability distribution vector corresponding to each frame. Based on the baseline coordinates and downward offset values, the bottom probability values ​​of the leading extension in the probability distribution vector are truncated and cleared to zero to obtain the retained weight matrix. Based on the position of the macroblock within the human pose bounding box and the magnitude of the corresponding value in the weighting matrix, differentiated quantization parameters are assigned to the macroblock, and video compression and encapsulation are performed.

[0006] Furthermore, the method for obtaining the reference coordinates and width of the human pose bounding box and initializing the downward offset value is as follows: The center x-coordinate and the y-coordinate of the bottom right corner of the human pose bounding box are used as the reference coordinates. The width is the distance between the left and right boundaries of the human pose bounding box. The downward offset value used to record the total downward displacement of the human body is initialized to zero.

[0007] Furthermore, the method for obtaining the effective tracking area is as follows: For any video frame during the non-visual detection coding period, based on the central horizontal coordinate, extend half of the width to the left and right sides respectively to obtain the horizontal span of the effective tracking area in the frame. Using the vertical coordinate of the lower right corner vertex as a reference, extend the width upwards and extend the downward offset value corresponding to the previous adjacent frame downwards to obtain the vertical span of the effective tracking area in that frame.

[0008] Furthermore, the method for obtaining the downward motion estimate for each frame and then accumulating and updating the corresponding downward offset value for each frame is as follows: Arrange the magnitudes of the downward-directed bottom-level motion vectors within the effective tracking region of the current frame in descending order to obtain a magnitude sequence; When the number of elements in the modulus sequence is greater than or equal to the preset extraction number, the average of the first preset extraction number of modulus in the modulus sequence is used as the downward motion estimate of the current frame. When the number of elements in the modulus sequence is less than the preset extraction number, the average value of all modulus in the modulus sequence is used as the downward motion estimate of the current frame. When the number of elements in the modulus sequence is zero, the downward motion estimate of the current frame is set to zero; The downward motion estimate of the current frame is added to the downward offset value of the previous adjacent frame to update the downward offset value of the current frame.

[0009] Furthermore, the method for constructing the macroblock mesh is as follows: Divide the ordinate of the lower right corner vertex by the preset macroblock size and round down to obtain the upper boundary of the macroblock mesh. Divide the total height of the video image by the preset macroblock size and round down to obtain the lower boundary of the macroblock grid; The difference between the center x-coordinate and the width is compared with 0 to extract the maximum value. The extracted maximum value is divided by the preset macroblock size and rounded down to obtain the left boundary of the macroblock mesh. The sum of the center x-coordinate and the width is compared with the total width of the video image to extract the minimum value. The extracted minimum value is divided by the preset macroblock size and rounded down to obtain the right boundary of the macroblock grid. The macroblocks within the area enclosed by the upper boundary, lower boundary, left boundary, and right boundary are constructed into a macroblock mesh; where one macroblock within the macroblock mesh corresponds to one mesh node.

[0010] Furthermore, the method for obtaining the diffusion transfer matrix is ​​as follows: For any grid node within the macroblock mesh, establish directed connections to adjacent grid nodes in the four directions of up, down, left, and right; Configure the initial weight of directed connection edges pointing outside the macroblock mesh to zero; For directed connection edges that do not point outside the macroblock mesh: the initial weight of the adjacent mesh node pointing directly above is set to zero, the initial weight of the adjacent mesh nodes pointing directly to the left and right is set to a preset constant, and the initial weight of the adjacent mesh node pointing directly below is set to the sum of the downward motion estimate of the current frame and the preset constant. The ratio of the initial weight of any grid node in any direction to the sum of the initial weights of that grid node in all four directions is used as the transfer weight in that direction. The transfer weights between all grid nodes are mapped into a square matrix according to the node index, forming the diffusion transfer matrix of the current frame.

[0011] Furthermore, the method for obtaining the probability distribution vector is as follows: The physical position formed by combining the central horizontal coordinate and the vertical coordinate of the lower right corner vertex is used as the human body reference point. Calculate the distance between the center of each grid node in the macroblock mesh and the human body reference point, and take the grid node with the smallest distance as the human body center node; Construct an initial vector with the same number of elements as the total number of grid nodes, according to the order of the grid nodes in the macroblock grid; assign 1 to the element corresponding to the human body center node in the initial vector, and assign 0 to the rest of the elements; The adaptive step size is obtained by dividing the downward offset value of the current frame by the preset macroblock size and rounding it up, and then adding it to the preset basic inference constant. Starting from the initial vector, perform iterative deductions one by one: multiply the difference between 1 and the preset probability retention coefficient, the diffusion transition matrix of the current frame, and the vector output from the previous iteration to obtain the first vector; multiply the preset probability retention coefficient with the initial vector to obtain the second vector; add the first vector and the second vector to obtain the vector output from this iteration. The derivation stops when the number of iterations equals the adaptive step size, and the final output vector is used as the probability distribution vector corresponding to the current frame.

[0012] Furthermore, the method for obtaining the retained weight matrix is ​​as follows: The ordinate of the lower right corner vertex is added to the downward offset value corresponding to the current frame to obtain the dynamic tangent coordinates; Obtain the vertical pixel coordinates corresponding to the geometric center of each grid node in the probability distribution vector corresponding to the current frame. For grid nodes whose vertical pixel coordinates are greater than the dynamic tangent coordinates, truncate and clear the corresponding probability values ​​in the probability distribution vector corresponding to the current frame. For grid nodes whose vertical pixel coordinates are less than or equal to the dynamic tangent coordinates, the corresponding probability values ​​in the probability distribution vector of the current frame are retained. The values ​​in the probability distribution vector corresponding to the processed current frame are mapped to a two-dimensional matrix according to the position index of each grid node to obtain the retained weight matrix corresponding to the current frame.

[0013] Furthermore, the method of assigning differentiated quantization parameters to macroblocks based on their positions within the human pose bounding box and the magnitudes of their corresponding values ​​in the retention weight matrix, and then performing video compression and encapsulation, is as follows: For any macroblock in the current frame, when the macroblock is located inside the human pose bounding box, or when the macroblock belongs to the macroblock mesh and the corresponding weight value in the retained weight matrix is ​​greater than the preset denoising probability threshold, a second preset quantization parameter is assigned to the macroblock. When the macroblock is not located inside the human pose bounding box and does not belong to the macroblock mesh, or when its corresponding weight value in the preserved weight matrix is ​​less than or equal to the preset denoising probability threshold, the macroblock is assigned a first preset quantization parameter. The compression ratio represented by the first preset quantization parameter is greater than the compression ratio represented by the second preset quantization parameter. The video compression calculation is performed on the current frame based on the quantization parameters assigned to each macroblock, and the compressed video stream is then encapsulated and output.

[0014] Secondly, another embodiment of the present invention provides a video compression and transmission system for an intelligent monitoring system in a ward. The system includes: a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above methods.

[0015] The present invention has the following beneficial effects: This invention dynamically defines the effective tracking area based on reference coordinates, width, and downward offset during the encoding period without visual detection. This accurately determines the feature reading range closely related to the fallen entity, which helps to eliminate interference from moving objects in the distant and irrelevant background areas and improves the efficiency of motion feature extraction. To obtain a smooth and realistic overall descent trend, the invention obtains the downward motion estimate for each frame based on the underlying downward motion vector within the effective tracking area. This accurately reflects the overall degree of fall of the torso and limbs, and helps to transform the discrete motion data of the underlying hardware into a reliable continuous tracking basis. Considering that isolated single-frame features cannot characterize the overall fall state, the invention accumulates and updates the corresponding downward offset value for each frame based on the downward motion estimate, realizing the trend estimation of fall depth in the blind zone without detection using the underlying encoded data. To ensure that the physical scale of subsequent region weight calculation is consistent with the hardware processing standard of the underlying video encoder, a macroblock grid is constructed to complete the dimensionality reduction mapping from pixel space to two-dimensional macroblock index space. Based on the positional relationship between grid nodes and the downward motion estimate, upward blocking and downward dynamically weighted connection weights are configured to generate the diffusion transfer matrix for each frame, accurately reflecting the weight. The physical guidance trend of a patient's torso falling rapidly under the action of force and the arms naturally spreading outward; in order to generate a smooth high-probability region with extremely low computational cost, a diffusion transfer matrix is ​​used as the source point to perform cyclic deduction, generating a probability distribution vector corresponding to each frame, and quickly outlining the inverted trapezoidal theoretical spatial distribution map reflecting the expansion of fall characteristics within the macroblock grid; in order to prevent a large amount of bandwidth from being allocated to the unrelated background at the bottom of the screen in the early stage of a fall, the bottom probability values ​​that extend ahead of the frame are truncated to zero based on the reference coordinates and downward offset values, and a retained weight matrix is ​​obtained. This achieves a bottom-based, one-size-fits-all numerical defense that combines dynamic displacement trends. To accurately direct limited network bandwidth to high-frequency motion characteristic regions under the constraint of a constant bit rate in the local area network, differentiated quantization parameters are assigned to macroblocks based on their position within the human posture bounding box and the corresponding values ​​in the retention weight matrix. This establishes a cross-regional resource reorganization mechanism that sacrifices background for core fidelity, thereby accurately performing video compression and encapsulation. This effectively eliminates the key motion blur defects caused by visual update lag, ensuring the continuity and accuracy of fall monitoring data extracted by the cloud server. Attached Figure Description

[0016] Figure 1 A schematic flowchart illustrating a video compression and transmission method for an intelligent monitoring system in a ward, provided as an embodiment of the present invention; Figure 2 This is a structural diagram of a video compression and transmission system for an intelligent monitoring system in a ward, provided as an embodiment of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, details a video compression and transmission method and a specific system scheme for an intelligent monitoring system for hospital wards provided by the present invention. Example 1:

[0018] This invention proposes a video compression and transmission method for an intelligent monitoring system in a ward. Please refer to [link / reference]. Figure 1 The diagram illustrates a schematic flowchart of a video compression and transmission method for an intelligent monitoring system in a ward according to an embodiment of the present invention. The method includes the following steps: Step S1: Obtain the reference coordinates and width of the human pose bounding box, and initialize the downward offset value.

[0019] Specifically, considering that the visual model detection update frequency of edge monitoring devices is usually much lower than the high-frequency frame output frequency of video encoders, resulting in a long visual update stagnation period between consecutive encoded frames, it is easy to fail to capture sudden actions such as patients falling rapidly downwards in time; in order to eliminate the temporal misalignment between low-frequency visual detection and high-frequency video encoding, this embodiment establishes a master-slave dual-cycle temporal update mechanism, that is, triggering a baseline state reset each time a low-frequency visual detection result is received, ensuring that the video encoder has a reliable calculation starting point in the visual blind zone; specifically: first, by parsing the bounding box of the target detection result transmitted by the edge device, the human body pose bounding box output by the visual model of the edge device at the current detection cycle node is obtained; Secondly, in order to establish the absolute reference benchmark between the target tracking region and the macroblock mesh generation in subsequent high-frequency video frames, and thus determine the reference coordinates and width of the human pose bounding box, this embodiment uses the center x-coordinate and the y-coordinate of the lower right corner vertex of the human pose bounding box as the reference coordinates to determine the horizontal reference point and initial lower edge position of the human body in the image; the distance between the left and right boundaries of the human pose bounding box is used as the width to define the default physical range for subsequent lateral feature tracking; so that the displacement trend of high-frequency video frames during the non-visual detection coding period can be calculated and spatially mapped based on the above fixed static reference parameters. Finally, considering the inter-frame cumulative effect of displacement caused by the patient's fall, in order to accurately characterize the actual depth of the patient's torso as it gradually falls downwards and to block interference from historical dirty data from the previous visual detection cycle, the downward offset value of the current detection cycle is initialized. That is, the downward offset value used to record the total downward displacement of the human body is forcibly initialized to zero, ensuring that the latest acquired baseline data and the zeroed offset variable can serve as the independent starting point for subsequent frame-by-frame derivation.

[0020] Step S2: During the non-visual detection encoding period, the effective tracking area is dynamically defined based on the reference coordinates, width, and downward offset value; based on the underlying downward motion vector within the effective tracking area, the downward motion estimation value of each frame is obtained, and then the downward offset value corresponding to each frame is accumulated and updated.

[0021] Specifically, to fill the monitoring gap when the visual model does not output results, the underlying encoded data of high-frequency video frames is analyzed during the non-visual detection encoding period (i.e., the time period between two visual detection result updates) to ensure continuous capture of human motion trends even during visual tracking pauses. To eliminate interference from moving objects in distant and irrelevant background areas, the effective tracking area is dynamically defined based on the reference coordinates, width, and downward offset value. This accurately determines the feature reading range closely related to the fallen entity, which helps improve the signal-to-noise ratio of motion feature extraction. Considering that a single descent feature may have local extreme noise, to obtain a smooth and realistic overall descent trend, the downward motion estimate for each frame is obtained based on the underlying downward motion vector within the effective tracking area, accurately reflecting the overall degree of fall of the torso and limbs. Considering the high-frequency characteristics of video frames and the continuity of physical displacement, to establish a mathematical relationship between single-frame motion and macroscopic distance, the downward offset value corresponding to each frame is updated by accumulating the downward motion estimate, achieving continuous estimation of fall depth. It should be noted that this embodiment obtains the bottom-level downward motion vector within the effective tracking area by parsing the hardware video encoder, and forces the extracted motion vector to be rounded up to the whole pixel dimension, thus avoiding the dimension misalignment and depth estimation inflation caused by directly substituting the original sub-pixel or quarter-pixel precision of the hardware.

[0022] Preferably, in one feasible method of this embodiment, the method for obtaining the effective tracking area is as follows: Considering that when a human falls, there is a continuous falling action of arms outstretched and torso downward, in order to accurately encompass the trajectory of this action without introducing unnecessary areas, for any video frame within the visual detection encoding period, based on the central horizontal coordinate, half the width is extended to the left and right sides respectively to obtain the horizontal span of the effective tracking area in that frame, thus limiting the tracking range to the horizontal physical width of the human body; at the same time, based on the vertical coordinate of the lower right corner vertex, the width is extended upward and downward, and the downward offset value corresponding to the previous adjacent frame is extended downward to obtain the vertical span of the effective tracking area in that frame. By extending upward to encompass possible waving actions and extending downward to closely follow the dynamic depth of the torso's fall, motion noise from irrelevant interference objects deep in or above the image is completely isolated vertically.

[0023] Preferably, in one feasible method of this embodiment, the method for obtaining the estimated downward motion value of each frame and then accumulating and updating the corresponding downward offset value of each frame is as follows: Arrange the magnitudes of the downward-directed bottom-level motion vectors within the effective tracking area of ​​the current frame in descending order to obtain a magnitude sequence; remove purely horizontal or upward invalid motions to highlight the core downward displacement features; considering that local limbs (such as a rapid instantaneous arm swing) may produce extreme values, in order to neutralize the data inflation caused by extreme value noise, when the number of elements in the magnitude sequence is greater than or equal to the preset extraction number, it indicates that there are enough falling feature blocks within the effective tracking area. In order to filter out occasional extreme values ​​and extract representative concentrated displacement degrees, the first preset extraction number of magnitudes in the magnitude sequence are then... The average value of all moduli in the moduli sequence is used as the downward motion estimate for the current frame. When the number of elements in the moduli sequence is less than the preset extraction number, it indicates that the local falling action in the effective tracking area is small or there are few feature blocks. In order to prevent array out-of-bounds calculation crash caused by forcibly reading a fixed number, the average value of all moduli in the moduli sequence is used as the downward motion estimate for the current frame. In this embodiment, the preset extraction number is set to 5 to ensure that the concurrent computing power limit of edge devices and the noise reduction requirements are taken into account. Implementers can set the preset extraction number according to the screen resolution and device computing power, which is not limited here. In addition, when the number of elements in the moduli sequence is zero, it indicates that the human body in the screen is absolutely still or only moves horizontally in the frame, lacking falling features. In this case, the downward motion estimate for the current frame is assigned to zero. Considering that isolated single-frame features cannot characterize the overall fall state, in order to transform high-frequency single-frame displacement features into a coherent macroscopic displacement state, the downward motion estimate of the current frame is added to the downward offset value corresponding to the previous adjacent frame to update the downward offset value corresponding to the current frame. This realizes the substitution of trend estimation of fall depth in the absence of detection blind zone using the underlying encoded data, avoiding the measurement paradox that may be caused by directly introducing real physical distance coordinates.

[0024] Step S3: Construct a macroblock mesh. Based on the positional relationship between mesh nodes and the estimated downward motion, configure upward blocking and downward dynamically weighted connection weights to generate a diffusion transfer matrix for each frame. Using the human body center node as the source point, perform iterative deduction using the diffusion transfer matrix to generate the probability distribution vector corresponding to each frame.

[0025] Specifically, considering that both the downward motion estimate and the downward offset value belong to the fine-grained pixel dimension, in order to ensure that the physical scale of the subsequent region weight calculation is consistent with the macroblock processing standard of the underlying video encoder, this embodiment constructs a macroblock mesh through coordinate transformation and edge limiting processing. This completes the dimensionality reduction mapping from pixel space to independent two-dimensional macroblock index space, which helps to ensure that the subsequently generated spatial probability distribution data can be seamlessly integrated and directly drive the macroblock-level parameter allocation of the hardware encoder. In order to mathematically simulate the force law of the human body falling and constrain the direction of numerical flow and diffusion pattern, connection weights that block upwards and dynamically weight downwards are configured based on the positional relationship between mesh nodes and the downward motion estimate. This process generates a diffusion transfer matrix for each frame, accurately reflecting the dynamic guiding trend of the patient's torso falling rapidly under gravity and the arms naturally extending outward. This helps to transform the objective fall laws at the engineering level into a rigorous algorithmic foundation. Considering that the limbs always rely on the natural outward extension of the torso when the human body falls, in order to generate smooth, high-probability regions with extremely low computational overhead, the diffusion transfer matrix is ​​used as the source point to perform cyclic deduction, generating the probability distribution vector corresponding to each frame. This accurately reflects the occupancy probability of the core features of the fall action in the monitoring screen, which helps to delineate the inverted trapezoidal theoretical spatial distribution map reflecting the expansion of fall features within the macroblock grid, providing solid data support for subsequent accurate reallocation of video compression bandwidth.

[0026] Preferably, in one feasible embodiment of this method, the method for constructing the macroblock grid is as follows: In order to be compatible with the general encoder standard and balance the computing power requirements, this embodiment sets the preset macroblock size to 16 to ensure that the conversion unit of the coordinate system strictly conforms to the fixed macroblock side length pixel value in the encoding protocol. The implementer can set the preset macroblock size according to the specific hardware support format, which is not limited here; In order to determine the vertical and horizontal cross-boundary reference of the computation grid and ensure that the probability inference process is strictly limited to the effective and legal physical boundary of the image, the vertical coordinate of the lower right corner vertex is divided by the preset macroblock size and rounded down to obtain the upper boundary of the macroblock grid, accurately determining the vertical starting reference for anti-backflow and avoiding invalid spread of values ​​above the initial position of the human body; The total height of the video image is divided by the preset macroblock size and rounded down to obtain the lower boundary of the macroblock grid, accurately determining the extreme physical position of the grid extension and avoiding memory overflow of the lower boundary of the image in subsequent probability inference; The difference between the center horizontal coordinate and the width is compared with 0. The maximum value is extracted, and a left edge limiting operation is performed to ensure that the left extension coordinate does not exceed the physical origin of the image. Then, the extracted maximum value is divided by the preset macroblock size and rounded down to obtain the left boundary of the macroblock grid, accurately determining the left tracking limit to accommodate the outward movement of the left arm, and avoiding the removal of the edge features of the arm extending outward when falling due to the grid being too narrow. The sum of the center horizontal coordinate and the width is compared with the total width of the video image to extract the minimum value, and a right edge limiting operation is performed to ensure that the right extension coordinate does not exceed the maximum width of the video image. Then, the extracted minimum value is divided by the preset macroblock size and rounded down to obtain the right boundary of the macroblock grid, accurately determining the right tracking limit to accommodate the outward movement of the right arm, and avoiding the hard truncation of limb features when extending to the right. It should be noted that the total height and total width of the video image are global physical resolution parameters predetermined by the hardware inherent properties of the edge monitoring camera device or the underlying video encoding protocol. These parameters are fixed in the system memory during the system initialization phase and can be directly called. To establish a data structure platform for transmitting parameters, macroblocks within the region enclosed by the upper, lower, left, and right boundaries are constructed as a macroblock grid, i.e., an independent two-dimensional indexed rectangular space is created. Each macroblock within the macroblock grid corresponds to one grid node.

[0027] Preferably, in one feasible embodiment of this method, the diffusion transfer matrix is ​​obtained as follows: In order to construct the physical pipeline for data flow within the grid and accurately analyze the flow path of human fall characteristics between macroblocks, for any grid node within the macroblock grid, directed connection edges are established pointing to adjacent grid nodes in the four directions of up, down, left, and right. Considering that edge and corner nodes may lead to data overflow, in order to perform out-of-bounds blocking defense operations, the initial weight of the directed connection edges pointing to the outside of the macroblock grid is configured to zero, ensuring that the values ​​only flow within the defined effective grid domain. For directed connection edges not pointing to the outside of the macroblock grid: the initial weight pointing to the adjacent grid node directly above is configured to zero, forcibly blocking the values ​​from flowing to the outside of the macroblock grid. The upward movement aligns with the natural law that a person cannot defy gravity and rebounds when falling. The initial weights of adjacent grid nodes pointing to the left and right are configured as preset constants, providing a baseline coefficient for the even distribution of values ​​to both sides, representing the basic lateral diffusion support of the outstretched arms. The initial weights of adjacent grid nodes pointing directly downwards are configured as the sum of the estimated downward movement value of the current frame and the preset constant, establishing a positive feedback weighted adjustment based on dynamic displacement characteristics. The more intense the estimated displacement trend, the greater the downward weighted potential energy. In this embodiment, the preset constant is set to 1 to ensure that the basic connection channel remains unobstructed even when the current frame is completely still. Implementers can set the preset constant according to the strength of the diffusion intention; no limitation is imposed here. To unify the overall distribution ratio of outward radiated data from each grid node and maintain a constant outflow, the ratio of the initial weight of any grid node in any direction to the sum of its initial weights in all four directions is used as the transfer weight in that direction, thus limiting the sum of all outflow weights of that grid node to 1. It should be noted that, to avoid the situation where the initial weights of a grid node in all four directions are 0, resulting in a meaningless denominator, this embodiment uses the sum of the initial weights in the four directions plus a preset minimum positive number as the denominator. The preset minimum positive number is set to... To ensure smooth floating-point fault tolerance when dealing with zero-connection nodes such as corners, implementers can set a preset minimum positive number based on the processor's floating-point arithmetic precision, without any limitation here; To form a global rule graph that can directly participate in subsequent linear algebraic derivations, the transition weights between all grid nodes are mapped into a square matrix according to the node index. That is, a high-dimensional structure representing the global circulation probability matrix is ​​constructed based on the spatial topology of the original grid, forming the diffusion transition matrix of the current frame. It should be noted that the diffusion transition matrix is ​​an asymmetric square matrix with a dimension equal to the square of the total number of grid nodes. It fully encapsulates the directional blocking and dynamic weighting rules of all local directions within the macroblock grid, serving as the underlying computational basis for subsequent spatial probability iterative evolution.

[0028] Preferably, in one feasible method of this embodiment, the probability distribution vector is obtained as follows: the physical position formed by combining the central x-coordinate and the y-coordinate of the lower right corner vertex is used as the human body reference point to establish the absolute reference coordinates of the origin; in order to seamlessly introduce the coordinate origin into the macroblock computation space, the Euclidean distance between the center of each grid node in the macroblock grid and the human body reference point is calculated (a well-known concept, not elaborated here), and the grid node with the smallest distance is used as the human body center node to complete the position calibration of the starting source point; in order to provide a container that matches the dimension of the diffusion transfer matrix, an initial vector with the same number of elements as the total number of grid nodes is constructed according to the position arrangement order of the grid nodes in the macroblock grid, so as to facilitate matrix multiplication with the diffusion transfer matrix in the same dimension. In order to initiate the first injection of probability values ​​in the grid, the element corresponding to the human body center node in the initial vector is assigned a value of 1, and the remaining elements are assigned a value of 0, so that this single non-zero value becomes the only absolute value source for subsequent operations, ensuring that the probability diffusion path generated later strictly conforms to the single physical origin law when the patient falls; Considering that using a fixed number of shallow simulation steps would cause the patient's actual torso to fall outside the coverage limit of the probability grid during deep falls, in order to make the extreme value of the simulation depth adaptively sink in sync with the extension of the actual estimated displacement depth, the downward offset value of the current frame is divided by the preset macroblock size and rounded up, then added to the preset basic simulation constant to obtain the adaptive step size. This ensures that the number of simulation layers is highly matched with the physical sinking depth of the fallen entity, avoiding the core area of ​​the fall from being missed and incorrectly compressed due to insufficient iteration steps. In this embodiment, the preset basic simulation constant is set to 3 to ensure that the basic coverage thickness is guaranteed even without significant displacement. The implementer can set the preset basic simulation constant according to the delay tolerance requirements of the environment, which is not limited here. In order to simulate the grid diffusion behavior of dense center and sparse edge decay, iterative simulation is performed successively starting from the initial vector: the difference between 1 and the preset probability retention coefficient, the diffusion transfer matrix of the current frame, and the vector output from the previous iteration are multiplied together to obtain the first vector. The quantity accurately reflects the feedforward flow trend of the historically accumulated probability values ​​spreading to the peripheral adjacent macroblocks along the set direction (blocking upward, leaning downward, equally dividing left and right), realizing the spatial extension of the fall feature to the surrounding grid. In addition, in order to establish the source point energy anchoring mechanism, the preset probability retention coefficient is further multiplied with the initial vector to obtain the second vector, which accurately reflects the inherent dwell probability that continues to exist at the center position of the torso. In order to superimpose the dual effects of flow spread and center replenishment, the first vector and the second vector are added to obtain the vector output by this iteration, which accurately reflects the gradient distribution law of the fall action being dense in the center and sparse and decreasing in the periphery, ensuring that each generation of the cycle stably backtracks and injects the probability base while spreading outward. In this embodiment, the preset probability retention coefficient is set to 0.15 to ensure that the overall distribution does not collapse rapidly to the edge of the screen without restriction, maintaining the funnel-shaped distribution morphology. The implementer can set the preset probability retention coefficient according to the contrast of the screen features, which is not limited here. When the number of iterations equals the adaptive step size, the derivation stops, the loop-avoidance lock is broken, and the final output vector is used as the probability distribution vector corresponding to the current frame. This enables the rapid delineation of the theoretical spatial distribution map reflecting the expansion of fall features in the memory macroblock grid with extremely low computational overhead, providing an accurate data foundation for subsequent tangent truncation and resource reallocation.

[0029] Step S4: Based on the baseline coordinates and downward offset values, truncate and clear the bottom probability values ​​that extend ahead in the probability distribution vector to obtain the retained weight matrix.

[0030] Specifically, considering that the probability distribution vector extrapolation algorithm is executed independently on a static grid map that extends directly to the bottom of the screen, as the number of iterations increases, the resulting probability distribution will naturally extend downwards. In order to prevent a large amount of bandwidth from being allocated to irrelevant areas at the bottom of the screen that have not yet been reached in the early stages of the fall, the bottom probability values ​​that extend ahead in the probability distribution vector are truncated and cleared to zero based on the reference coordinates and downward offset values. This yields a retained weight matrix, which realizes a one-size-fits-all numerical defense at the bottom by combining the dynamic bottom feature displacement trend, thus avoiding the problem of the theoretical extrapolation being out of sync with the actual fall depth.

[0031] Preferably, in one feasible method of this embodiment, the method for obtaining the weight matrix is ​​as follows: Considering that a single high-frequency displacement estimation cannot be directly used for boundary delineation of the image space, in order to construct a dynamic warning line that gradually extends downward with the coding frame rate in the image space, the ordinate of the lower right corner vertex is added to the downward offset value corresponding to the current frame to obtain the dynamic tangent coordinates, accurately anchoring the downward limit position of the human body in the current frame; in order to remove the interference areas not involved, the vertical pixel coordinates corresponding to the geometric center of each grid node in the probability distribution vector corresponding to the current frame are obtained, for the vertical image... Grid nodes with pixel coordinates greater than dynamic tangent coordinates indicate that the corresponding macroblock is below the tangent warning depth, i.e., in a clean background medium before the patient's torso has landed. In this case, the probability value in the probability distribution vector of the current frame is truncated and cleared to zero, stripping away the logical basis for subsequently assigning image quality tilt protection. For grid nodes with vertical pixel coordinates less than or equal to dynamic tangent coordinates, it indicates that the corresponding macroblock is in a high-risk area where the fall displacement trajectory has occurred or is occurring. In this case, the probability value in the probability distribution vector of the current frame is retained to ensure that the key feature weights left by the torso movement are intact. In order to generate a data format that perfectly matches the macroblock arrangement of the underlying video encoder, the values ​​in the probability distribution vector corresponding to the current frame after processing are mapped to a two-dimensional matrix according to the position index of each grid node to obtain the retained weight matrix corresponding to the current frame, so as to be handed over to the data reconstruction and allocation pipeline for underlying parameter overwriting.

[0032] Step S5: Based on the position of the macroblock in the human pose bounding box and the magnitude of the corresponding value in the weight matrix, assign differentiated quantization parameters to the macroblock and perform video compression and encapsulation.

[0033] Specifically, given that local area network bandwidth is typically strictly limited, and standard video coding protocols natively provide macroblock-level quantization parameter adjustment interfaces for regions of interest, in order to accurately direct limited network transmission bandwidth to high-frequency regions of motion features under constant bit rate constraints, differentiated quantization parameters are assigned to macroblocks based on their position within the human posture bounding box and the corresponding values ​​in the weighting matrix. This enables fine-grained reconfiguration of the underlying encoded data capacity, avoids the problem of irrelevant backgrounds distributing core resources, and thus accurately performs video compression and encapsulation, outputting locally differentiated high-frequency compressed frames. This ensures the continuity of fall monitoring data extraction and the accuracy of alarms from the backend server.

[0034] In order to filter out extremely weak and invalid diffusion values ​​that spread outward during the probability distribution derivation process, and to eliminate implicit small floating-point errors generated at the hardware level, a clear image quality intervention boundary is established. Therefore, for any macroblock in the current frame, if the macroblock is located inside the human pose bounding box, or if the macroblock belongs to the macroblock mesh and the corresponding weight value in the weight retention matrix is ​​greater than the preset denoising probability threshold, it indicates that the macroblock is within the motion occupancy area of ​​the high probability distribution, covering the key feature trajectory of the patient's falling torso and the arms moving downward. In order to ensure that the edge texture and geometric features of the key actions are not destroyed, a second preset quantization parameter is assigned to the macroblock, that is, a small quantization step size representing low-ratio compression is injected, prompting the encoder to concentrate resources on the macroblock. In this embodiment, the preset denoising probability threshold is set to 0.01 to ensure effective stripping of weak and diffuse redundant data. The implementer can set the preset denoising probability threshold according to the noise situation of the image, which is not limited here. When a macroblock is not located inside the human pose bounding box and does not belong to the macroblock mesh, or its corresponding weight value in the weight preservation matrix is ​​less than or equal to the preset denoising probability threshold, it indicates that the pixel space where the macroblock is located belongs to a static background area or a clean area that has not been touched by the fall displacement trajectory. In order to discard the static background to free up idle data capacity, a first preset quantization parameter is assigned to the macroblock, that is, a maximum step value upper limit representing a high compression ratio is injected, which causes the underlying encoder to discard most of the high-frequency discrete cosine transform coefficients inside the macroblock, releasing a massive amount of idle bandwidth quota; where the compression ratio represented by the first preset quantization parameter is greater than the compression ratio represented by the second preset quantization parameter, so that a resource cross-region reorganization mechanism of "background sacrifice for core fidelity" is established at the bottom layer; The system completes the overwriting of differentiated parameters for all macroblocks globally, enabling precise locking and guidance of all available and core bit quotas within the current video frame. To generate video images with a significantly non-linear resource distribution pattern, the system performs video compression calculations on the current frame based on the quantization parameters allocated to each macroblock. This significantly blurs and smooths the irrelevant background around and below the image, while clearly preserving the action edges within the funnel-shaped area along the estimated fall depth. The system then calls a standard streaming media encapsulation protocol to encapsulate and output the compressed video stream, allowing it to be transmitted to the cloud via a limited network pipeline. This enables the cloud-based motion capture model to stably extract the complete patient's fall skeleton and successfully trigger an alarm.

[0035] In summary, this embodiment obtains the reference coordinates and width of the human pose bounding box, defines the effective tracking area during the non-visual detection encoding period, obtains the estimated downward motion value based on the bottom-level downward motion vector within the effective tracking area, and then accumulates and updates the downward offset value; constructs a macroblock mesh, configures upward blocking and downward dynamically weighted connection weights to generate a diffusion transfer matrix, uses the human center node as the source point, and performs cyclic deduction using the diffusion transfer matrix to generate a probability distribution vector; truncates and zeroes the probability distribution vector based on the downward offset value to obtain the retained weight matrix; and assigns quantization parameters to the macroblock based on the macroblock position and corresponding weights and performs video compression and encapsulation. This invention effectively solves the critical motion blur defect caused by visual update lag and achieves precise bandwidth resource allocation. Example 2:

[0036] This invention also proposes a video compression and transmission system for an intelligent monitoring system in a ward. Please refer to [link / reference]. Figure 2 The diagram illustrates a video compression and transmission system structure of an intelligent monitoring system for a ward provided by an embodiment of the present invention. The system includes: a data acquisition module 10, a downward offset value acquisition module 20, a probability distribution vector acquisition module 30, a weight retention matrix acquisition module 40, and a video compression module 50.

[0037] The data acquisition module 10 is used to acquire the reference coordinates and width of the human pose bounding box and initialize the downward offset value.

[0038] The downward offset value acquisition module 20 is used to dynamically limit the effective tracking area based on the reference coordinates, width and downward offset value during the encoding period without visual detection; based on the underlying downward motion vector in the effective tracking area, the downward motion estimation value of each frame is obtained, and then the downward offset value corresponding to each frame is accumulated and updated.

[0039] The probability distribution vector acquisition module 30 is used to construct a macroblock mesh. Based on the positional relationship between mesh nodes and the estimated downward motion, it configures the connection weights that block upwards and dynamically weight downwards to generate the diffusion transfer matrix for each frame. Taking the human body center node as the source point, it uses the diffusion transfer matrix to perform cyclic deduction to generate the probability distribution vector corresponding to each frame.

[0040] The retained weight matrix acquisition module 40 is used to truncate and clear the bottom probability values ​​of the leading extension in the probability distribution vector to zero based on the reference coordinates and downward offset values, so as to obtain the retained weight matrix.

[0041] The video compression module 50 is used to assign differentiated quantization parameters to macroblocks based on their positions in the human pose bounding box and the magnitude of their corresponding values ​​in the retention weight matrix, and to perform video compression encapsulation.

[0042] It should be noted that the video compression and transmission system of the intelligent ward monitoring system and the video compression and transmission method of the intelligent ward monitoring system provided in the above embodiments belong to the same concept. The specific implementation process can be found in the method embodiments, and will not be repeated here.

Claims

1. A video compression and transmission method for an intelligent monitoring system in a ward, characterized in that, The method includes the following steps: Obtain the base coordinates and width of the human pose bounding box, and initialize the downward offset value; During the non-visual detection encoding period, the effective tracking area is dynamically defined based on the reference coordinates, width, and downward offset value. Based on the underlying downward motion vector within the effective tracking area, the downward motion estimate for each frame is obtained, and then the downward offset value corresponding to each frame is accumulated and updated. A macroblock mesh is constructed. Based on the positional relationship between mesh nodes and the estimated downward motion, connection weights that block upward movement and dynamically weight downward movement are configured to generate a diffusion transfer matrix for each frame. Using the human body center node as the source point, a cyclic deduction is performed using the diffusion transfer matrix to generate the probability distribution vector corresponding to each frame. Based on the baseline coordinates and downward offset values, the bottom probability values ​​of the leading extension in the probability distribution vector are truncated and cleared to zero to obtain the retained weight matrix. Based on the position of the macroblock within the human pose bounding box and the magnitude of the corresponding value in the weighting matrix, differentiated quantization parameters are assigned to the macroblock, and video compression and encapsulation are performed.

2. The video compression and transmission method of a ward intelligent monitoring system as described in claim 1, characterized in that, The method for obtaining the reference coordinates and width of the human pose bounding box and initializing the downward offset value is as follows: The center x-coordinate and the y-coordinate of the bottom right corner of the human pose bounding box are used as the reference coordinates. The width is the distance between the left and right boundaries of the human pose bounding box. The downward offset value used to record the total downward displacement of the human body is initialized to zero.

3. The video compression and transmission method of a ward intelligent monitoring system as described in claim 2, characterized in that, The method for obtaining the effective tracking area is as follows: For any video frame during the non-visual detection coding period, based on the central horizontal coordinate, extend half of the width to the left and right sides respectively to obtain the horizontal span of the effective tracking area in the frame. Using the vertical coordinate of the lower right corner vertex as a reference, extend the width upwards and extend the downward offset value corresponding to the previous adjacent frame downwards to obtain the vertical span of the effective tracking area in that frame.

4. The video compression and transmission method of a ward intelligent monitoring system as described in claim 1, characterized in that, The method for obtaining the estimated downward motion value for each frame and then accumulating and updating the corresponding downward offset value for each frame is as follows: Arrange the magnitudes of the downward-directed bottom-level motion vectors within the effective tracking region of the current frame in descending order to obtain a magnitude sequence; When the number of elements in the modulus sequence is greater than or equal to the preset extraction number, the average of the first preset extraction number of modulus in the modulus sequence is used as the downward motion estimate of the current frame. When the number of elements in the modulus sequence is less than the preset extraction number, the average value of all modulus in the modulus sequence is used as the downward motion estimate of the current frame. When the number of elements in the modulus sequence is zero, the downward motion estimate of the current frame is set to zero; The downward motion estimate of the current frame is added to the downward offset value of the previous adjacent frame to update the downward offset value of the current frame.

5. The video compression and transmission method of a ward intelligent monitoring system as described in claim 2, characterized in that, The method for constructing macroblock meshes is as follows: Divide the ordinate of the lower right corner vertex by the preset macroblock size and round down to obtain the upper boundary of the macroblock mesh. Divide the total height of the video image by the preset macroblock size and round down to obtain the lower boundary of the macroblock grid; The difference between the center x-coordinate and the width is compared with 0 to extract the maximum value. The extracted maximum value is divided by the preset macroblock size and rounded down to obtain the left boundary of the macroblock mesh. The sum of the center x-coordinate and the width is compared with the total width of the video image to extract the minimum value. The extracted minimum value is divided by the preset macroblock size and rounded down to obtain the right boundary of the macroblock grid. The macroblocks within the area enclosed by the upper boundary, lower boundary, left boundary, and right boundary are constructed into a macroblock mesh; where one macroblock within the macroblock mesh corresponds to one mesh node.

6. The video compression and transmission method of a ward intelligent monitoring system as described in claim 5, characterized in that, The method for obtaining the diffusion transfer matrix is ​​as follows: For any grid node within the macroblock mesh, establish directed connections to adjacent grid nodes in the four directions of up, down, left, and right; Configure the initial weight of directed connection edges pointing outside the macroblock mesh to zero; For directed connection edges that do not point outside the macroblock mesh: the initial weight of the adjacent mesh node pointing directly above is set to zero, the initial weight of the adjacent mesh nodes pointing directly to the left and right is set to a preset constant, and the initial weight of the adjacent mesh node pointing directly below is set to the sum of the downward motion estimate of the current frame and the preset constant. The ratio of the initial weight of any grid node in any direction to the sum of the initial weights of that grid node in all four directions is used as the transfer weight in that direction. The transfer weights between all grid nodes are mapped into a square matrix according to the node index, forming the diffusion transfer matrix of the current frame.

7. The video compression and transmission method of a ward intelligent monitoring system as described in claim 5, characterized in that, The method for obtaining the probability distribution vector is as follows: The physical position formed by combining the central horizontal coordinate and the vertical coordinate of the lower right corner vertex is used as the human body reference point. Calculate the distance between the center of each grid node in the macroblock mesh and the human body reference point, and take the grid node with the smallest distance as the human body center node; Construct an initial vector with the same number of elements as the total number of grid nodes, according to the order of the grid nodes in the macroblock grid; assign 1 to the element corresponding to the human body center node in the initial vector, and assign 0 to the rest of the elements; The adaptive step size is obtained by dividing the downward offset value of the current frame by the preset macroblock size and rounding it up, and then adding it to the preset basic inference constant. Starting from the initial vector, perform iterative deductions one by one: multiply the difference between 1 and the preset probability retention coefficient, the diffusion transition matrix of the current frame, and the vector output from the previous iteration to obtain the first vector; multiply the preset probability retention coefficient with the initial vector to obtain the second vector; add the first vector and the second vector to obtain the vector output from this iteration. The derivation stops when the number of iterations equals the adaptive step size, and the final output vector is used as the probability distribution vector corresponding to the current frame.

8. The video compression and transmission method of a ward intelligent monitoring system as described in claim 2, characterized in that, The method for obtaining the retained weight matrix is ​​as follows: The ordinate of the lower right corner vertex is added to the downward offset value corresponding to the current frame to obtain the dynamic tangent coordinates; Obtain the vertical pixel coordinates corresponding to the geometric center of each grid node in the probability distribution vector corresponding to the current frame. For grid nodes whose vertical pixel coordinates are greater than the dynamic tangent coordinates, truncate and clear the corresponding probability values ​​in the probability distribution vector corresponding to the current frame. For grid nodes whose vertical pixel coordinates are less than or equal to the dynamic tangent coordinates, the corresponding probability values ​​in the probability distribution vector of the current frame are retained. The values ​​in the probability distribution vector corresponding to the processed current frame are mapped to a two-dimensional matrix according to the position index of each grid node to obtain the retained weight matrix corresponding to the current frame.

9. The video compression and transmission method of a ward intelligent monitoring system as described in claim 5, characterized in that, The method for assigning differentiated quantization parameters to macroblocks based on their positions within the human pose bounding box and the magnitudes of their corresponding values ​​in the retention weight matrix, and then performing video compression and encapsulation, is as follows: For any macroblock in the current frame, when the macroblock is located inside the human pose bounding box, or when the macroblock belongs to the macroblock mesh and the corresponding weight value in the retained weight matrix is ​​greater than the preset denoising probability threshold, a second preset quantization parameter is assigned to the macroblock. When the macroblock is not located inside the human pose bounding box and does not belong to the macroblock mesh, or when its corresponding weight value in the preserved weight matrix is ​​less than or equal to the preset denoising probability threshold, the macroblock is assigned a first preset quantization parameter. The compression ratio represented by the first preset quantization parameter is greater than the compression ratio represented by the second preset quantization parameter. The video compression calculation is performed on the current frame based on the quantization parameters assigned to each macroblock, and the compressed video stream is then encapsulated and output.

10. A video compression and transmission system for an intelligent monitoring system in a ward, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the video compression and transmission method of the intelligent monitoring system for wards as described in any one of claims 1-9.