Orthopedic patient movement monitoring method

Abstract

The invention relates to the field of medical big data analysis and orthopedics department data monitoring, and discloses an orthopedics department patient movement monitoring method, which comprises the following steps: S1, carrying out interframe image processing on the collected data, and extracting the gait characteristics of a patient; s2, processing the collected adjacent frame images to obtain an original grayscale image, a histogram equalization image and a de-noising result image; s3, performing feature fusion on the original grayscale image, the histogram equalization image and the de-noising result image to obtain a pre-processed image; s4, performing registration and difference calculation processing on the preprocessed image; s5, training and marking the data by adopting an inter-frame difference method; s6, constructing an orthopedic patient movement monitoring model; and S7, obtaining a monitoring result based on the mobile monitoring model. The method has the advantages that through multi-dimensional dynamic analysis of motion details of the patient, whole-process and refined tracking of motion behaviors of the patient is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of medical big data analysis and orthopedic data monitoring, and in particular to a method for monitoring and early warning of the movement of orthopedic patients. Background Art

[0002] During the recovery process after orthopedic fracture fixation surgery, especially for special groups such as osteoporosis patients and some elderly people with fragile bones, measures such as restricting the movement of these fragile special groups or forcing them to stay in bed are generally adopted to reduce the risk of re-injury.

[0003] Traditional patient activity assessment mainly relies on manual observation, regular inspections or fixed-point measurements based on wearable sensors. These methods have problems such as strong subjectivity, poor continuity, and low patient compliance, making it difficult to achieve full-process and detailed tracking of patients' movement behaviors. Summary of the Invention

[0004] In order to address the deficiencies in the prior art, the present invention provides a method for monitoring the movement of orthopedic patients, which realizes full-process and refined tracking of the patient's movement behavior through multi-dimensional dynamic analysis of the patient's movement details.

[0005] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0006] A method for monitoring the movement of orthopedic patients, comprising the following steps:

[0007] S1. Collect patient movement data, perform inter-frame image processing on the collected data, and extract the patient's gait characteristics;

[0008] S2. Processing the collected adjacent frame images to obtain an original grayscale image, a histogram equalization image, and a denoising result image;

[0009] S3, performing feature fusion on the original grayscale image, the histogram equalization image, and the denoising result image to obtain a preprocessed image;

[0010] S4, performing registration and difference calculation on the pre-processed images;

[0011] S5. Use the frame difference method to train and label the data;

[0012] S6. Construct a mobile monitoring model for orthopedic patients;

[0013] S7. Obtain monitoring results based on the mobile monitoring model.

[0014] Furthermore, in S2, the grayscale value is calculated using the weighted average method to obtain the original grayscale image;

[0015] The cumulative distribution function is used to map the original grayscale value to the new distribution to obtain a histogram equalization image;

[0016] The non-local mean denoising method is used to perform weighted averaging denoising using the redundant information of similar structures in the image, retaining edge details and matching the range of similar blocks to obtain the denoised image.

[0017] Furthermore, the feature fusion method described in S3 is obtained by adding the pixel values of the corresponding pixel positions and then taking the average.

[0018] Furthermore, S4 specifically includes:

[0019] Align the pre-processed images corresponding to two consecutive frames to the same spatial coordinate system to eliminate the geometric offset caused by differences in shooting angle, position or time;

[0020] By traversing all feature descriptors of the two images, calculating the Hamming distance, and retaining the top 50% matching pairs with the smallest distance, the interference of false matches is reduced;

[0021] Quantify the difference between two consecutive registered frames and identify obvious displacement features.

[0022] The data of two consecutive frames of original images are processed using the inter-frame difference method to obtain the movement change data of the first orthopedic patient;

[0023] The original grayscale image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the second orthopedic patient;

[0024] The histogram equalization image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the third orthopedic patient;

[0025] The denoised image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fourth orthopedic patient;

[0026] The pre-processed image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fifth orthopedic patient;

[0027] The inter-frame difference method is used to process the Canny edge detection image data corresponding to two consecutive frames of images to obtain the movement change data of the sixth orthopedic patient;

[0028] The ORB feature point image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the seventh orthopedic patient.

[0029] Furthermore, the first orthopedic patient movement change data, the second orthopedic patient movement change data, the third orthopedic patient movement change data, the fourth orthopedic patient movement change data, the fifth orthopedic patient movement change data, the sixth orthopedic patient movement change data, the seventh orthopedic patient movement change data and the pixel-level difference heat map are feature fused to obtain orthopedic patient movement monitoring training sample data.

[0030] Furthermore, S6 builds an orthopedic patient mobility monitoring model including:

[0031] Conduct topological manifold kernel density estimation on orthopedic patient mobility monitoring data;

[0032] Adaptively select the bandwidth parameter by minimizing the error function;

[0033] Perform forward propagation processing of data for convolutional neural network training;

[0034] Update convolution kernel weights based on topological manifold kernel density estimation;

[0035] Calculate the loss function in electromagnetic force-guided optimization;

[0036] Adaptive adjustment of learning rate based on changes in the loss function of the convolutional neural network;

[0037] The iterative stopping condition of convolutional neural network training is judged. When the relative change rate of the validation set loss function is less than the threshold in three consecutive iterations, the model is judged to have converged, and the iteration is stopped, indicating that the convolutional neural network model training is completed.

[0038] Furthermore, the movement data is derived from video recordings of the patient's gait and movement patterns, and during the data collection process, the patient is asked to follow instructions and walk in a standardized gait analysis area.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. The present invention performs inter-frame difference on multi-source image data such as original images, grayscale images, histogram equalization images, denoised images, edge detection images, and ORB feature maps, and fuses them into a unified training sample to achieve multi-dimensional dynamic analysis of patient movement details.

[0041] 2. This paper uses kernel density estimation to analyze the manifold structure of data in high-dimensional feature space, thereby dynamically focusing on learning details in dense areas. Compared with traditional convolutional neural network methods, this method can avoid training bias under high-dimensional sparse data and significantly improve the model's sensitivity to subtle gait changes.

[0042] 3. The present invention optimizes the weight update path by simulating the multi-force controlled behavior of particles in the electromagnetic field, prevents falling into the local optimum, guides the network to perform more effective global search in high-dimensional space, and improves training robustness.

[0043] 4. The present invention automatically adjusts the kernel density estimation bandwidth based on minimizing the error function, taking into account both local details and global smoothness. Compared with the fixed bandwidth method, it is more adaptable to changes in data-dense / sparse areas and improves the model generalization ability.

[0044] 5. The present invention dynamically adjusts the learning rate according to the rate of change of the loss function, effectively suppressing oscillation and overfitting in the late stage of training, and compensating for the problem that the conventional fixed attenuation strategy is insufficiently responsive to complex data scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Attachment Figure 1 It is a flowchart of the present invention;

[0046] Attachment Figure 2 This is an example of two consecutive frames of original image data, showing the movement process of an orthopedic patient;

[0047] Attachment Figure 3 The original grayscale image, histogram equalization image, and denoising result image corresponding to the first frame of the original image;

[0048] Attachment Figure 4 The original grayscale image, histogram equalization image, and denoising result image corresponding to the second frame original image;

[0049] Attachment Figure 5 It is the preprocessed image, Canny edge detection image, and ORB feature point (green) image corresponding to the first frame of the original image;

[0050] Attachment Figure 6 It is the preprocessed image, Canny edge detection image, and ORB feature point (green) image corresponding to the second frame original image;

[0051] Attachment Figure 7 It is a pixel-level difference heat map obtained from the data of two consecutive frames of images;

[0052] Attachment Figure 8 It is a three-dimensional heat map;

[0053] Attachment Figure 9 It is a contour map

[0054] Attachment Figure 10 This is a comparison chart of the relationship between dynamic adjustment of learning rate and training loss. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0056] To facilitate understanding of this embodiment, a method for monitoring the movement of orthopedic patients disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 A flow chart of a method for monitoring movement of orthopedic patients disclosed in an embodiment of the present invention is shown. Figure 1 As shown, the orthopedic patient movement monitoring method includes the following steps:

[0057] S1. Collect patient movement data, perform inter-frame image processing on the collected data, and extract the patient's gait characteristics;

[0058] In the orthopedic patient mobility monitoring method proposed in the present invention, the data acquisition method relies on video recording of the patient's gait and movement pattern, using a fixed camera or handheld camera equipment to shoot in a standardized laboratory environment or rehabilitation center.

[0059] Video capture uses an RGB three-channel camera, and the frame rate is generally set at 60fps to ensure the continuity of gait movement and the clarity of details;

[0060] During the data collection process, patients will be asked to follow instructions and walk in a standardized gait analysis area. The camera can be set at different angles such as the front, side, and back to capture complete gait data. The laboratory environment is equipped with fixed markers or reference objects. Figure 2 The data examples of two consecutive frames of original images are shown, which show the movement process of an orthopedic patient.

[0061] After acquisition is completed, the video data will undergo inter-frame image processing, that is, continuous frames will be analyzed to extract the patient's gait characteristics. Specifically, the key frame extraction method is adopted, that is, key frames are selected at specific moments in the gait cycle (such as the sole of the foot touching the ground, leaving the ground, swinging, etc.) to reduce redundant information.

[0062] To eliminate imaging interference, unify image representation, and enhance the recognizability of key structures (skeleton / marker points), the data is preprocessed as follows:

[0063] S2. Processing the collected adjacent frame images to obtain an original grayscale image, a histogram equalization image, and a denoising result image;

[0064] Specifically, the collected adjacent frame images are first grayscaled to convert the RGB three-channel image into a single-channel grayscale image, removing color redundant information, simplifying subsequent calculation complexity, and reducing memory usage;

[0065] The present invention uses the weighted average method to calculate the grayscale value to obtain the original grayscale image, which is expressed as:

[0066] I gray =0.299R img +0.587G img +0.114B img

[0067] Where, I gray is the grayscale pixel value; R img is the pixel value of the first channel of the image; G img is the pixel value of the second channel of the image; B img The pixel value of the third channel of the image

[0068] Furthermore, histogram equalization is performed to expand the dynamic range and enhance the contrast by redistributing the pixel grayscale values;

[0069] The present invention uses a cumulative distribution function to map the original grayscale value to a new distribution to obtain a histogram equalization image.

[0070] Furthermore, the present invention adopts a non-local mean denoising method, which uses the redundant information of similar structures in the image to perform weighted average denoising and retain edge details. The search window of the non-local mean denoising method is 21×21 pixels, matching the similar block range to obtain the denoised result image, such as Figure 3 and Figure 4 shown.

[0071] S3, performing feature fusion on the original grayscale image, the histogram equalization image, and the denoising result image to obtain a preprocessed image;

[0072] The feature fusion method is obtained by weighted feature processing, specifically by adding pixel values of corresponding pixel positions and then taking the average.

[0073] Furthermore, in order to locate key feature points, the Canny edge detection method is used to extract continuous edges from the preprocessed image to obtain a Canny edge detection image.

[0074] Furthermore, the ORB feature point detection (Oriented FAST and Rotated BRIEF) method is used to process the preprocessed image to quickly locate corner points (such as bone corners and marker centers), and binary encoding is performed to describe the texture feature extraction around the feature points.

[0075] Examples of preprocessed images, Canny edge detection images, and ORB feature point (green) images are shown below: Figure 5 and Figure 6 shown.

[0076] S4, performing registration and difference calculation on the pre-processed images;

[0077] Align the pre-processed images corresponding to two consecutive frames to the same spatial coordinate system to eliminate the geometric offset caused by differences in shooting angle, position or time;

[0078] By traversing all feature descriptors of the two images, calculating the Hamming distance (ORB is a binary descriptor), retaining the top 50% matching pairs with the smallest distance, and reducing false matching interference.

[0079] Furthermore, the difference between the two consecutive frames of images after registration is quantified to identify obvious displacement features. The present invention uses a pixel-level difference processing method to process the two consecutive frames of images after registration to obtain a pixel-level difference heat map, such as Figure 7 As shown in the figure, the difference values are specifically coded using jet colors, with warm colors (red / yellow) indicating high-difference areas. The distribution statistics of the difference values are obtained by statistical histograms to determine whether there are concentrated abnormal areas (such as tailing at the right tail of the histogram).

[0080] S5. Use the frame difference method to train and label the data;

[0081] The data of two consecutive frames of original images are processed using the inter-frame difference method to obtain the movement change data of the first orthopedic patient;

[0082] The original grayscale image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the second orthopedic patient;

[0083] The histogram equalization image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the third orthopedic patient;

[0084] The denoised image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fourth orthopedic patient;

[0085] The pre-processed image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fifth orthopedic patient;

[0086] The inter-frame difference method is used to process the Canny edge detection image data corresponding to two consecutive frames of images to obtain the movement change data of the sixth orthopedic patient;

[0087] The ORB feature point (green) image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the seventh orthopedic patient;

[0088] The first orthopedic patient movement change data, the second orthopedic patient movement change data, the third orthopedic patient movement change data, the fourth orthopedic patient movement change data, the fifth orthopedic patient movement change data, the sixth orthopedic patient movement change data, the seventh orthopedic patient movement change data and the pixel-level difference heat map are feature fused to obtain orthopedic patient movement monitoring training sample data.

[0089] The feature fusion method is obtained by weighted feature processing, specifically by adding pixel values of corresponding pixel positions and then taking the average.

[0090] The inter-frame difference method is a method for obtaining the contour of a moving target by performing a differential operation on two adjacent frames or images separated by several frames in a video image sequence. The pixel values of two adjacent frames or images separated by several frames in a video stream are subtracted, and the subtracted image is thresholded to extract the moving area.

[0091] Furthermore, the orthopedic patient movement monitoring training sample data was annotated based on two consecutive frames of original images through manual labeling. The labeling categories included: small amplitude (safety), medium amplitude (warning), and large amplitude (high risk), a total of 3 categories.

[0092] S6. Orthopedic patient movement monitoring model training

[0093] The present invention uses a convolutional neural network to monitor the movement of orthopedic patients. The training process of the convolutional neural network is as follows:

[0094] S601. Perform topological manifold kernel density estimation on orthopedic patient movement monitoring data.

[0095] The training dataset composed of orthopedic patient mobility monitoring data is vectorized to form a high-dimensional feature space. These data often have the characteristics of uneven distribution and local high-density aggregation in orthopedic patient mobility monitoring.

[0096] Conventional technologies usually directly use overall statistical distribution or simple distance measurement methods to process data, but it is difficult to accurately grasp the key areas when faced with high-dimensional sparsity and local aggregation, thereby ignoring the subtle distribution differences in patient movement characteristics.

[0097] This paper uses topological manifold kernel density estimation to characterize the local density of each data point in high-dimensional space, thereby identifying the potential manifold structure. By performing kernel density estimation on data points, subsequent feature extraction and classification are more targeted. The calculation method is expressed as:

[0098]

[0099] Where, is the kernel density estimate of the data point; x is the input of the function, representing the data point input to the convolutional neural network; n is the total number of training data; x i is the eigenvector of the i-th data point; i is a positive integer; K(·) is the kernel function (Gaussian kernel is often used); h is the bandwidth parameter.

[0100] The topological manifold kernel density estimation refers to the use of kernel density estimation methods to characterize the local density of each data point in the high-dimensional data space, targeting the complex and changeable local characteristics of the mobile monitoring data of orthopedic patients, and revealing the intrinsic topological structure of the data by identifying areas where data points are densely clustered.

[0101] The manifold structure refers to the essential geometric shape or layout of the movement monitoring data of orthopedic patients in high-dimensional space. These structures are manifested through the local similarity and continuity of the data.

[0102] The data point refers to a point representing a single observation instance in a high-dimensional feature space, specifically a piece of orthopedic patient movement monitoring data.

[0103] s602. Optimization selection of bandwidth parameters.

[0104] In the high-dimensional scenario of orthopedic patient mobility monitoring, the bandwidth parameter directly determines the degree of local and global balance of the convolutional neural network kernel function. Different bandwidths will affect the smoothness and detail retention of density estimation.

[0105] Conventional techniques usually fix an empirical bandwidth or select it through a simple cross-validation method, which can easily lead to inaccurate estimation in sparse areas and over-smoothing in dense areas, and cannot take into account both local refinement and global overview.

[0106] The present invention adaptively selects bandwidth parameters by minimizing the error function, ensuring that important features are preserved while avoiding oversmoothing in the high-dimensional distribution of orthopedic patient movement data. The calculation method is expressed as:

[0107]

[0108] Where, is the optimized bandwidth parameter; x j is the eigenvector of the jth data point; j is a positive integer; is the true or prior density estimate for the i-th data point.

[0109] In one embodiment, Figure 8 and Figure 9As shown in the figure, through the composite visualization of three-dimensional heat map and contour lines, the kernel density estimation error distribution of the traditional fixed bandwidth selection method and the adaptive bandwidth optimization of the present invention in different data density areas is compared. The experiment focuses on analyzing the error control ability of orthopedic patient mobile monitoring data in high-density clustered areas and sparse areas. Conventional methods are prone to excessive smoothing errors in data-dense areas and estimation distortion in sparse areas. The adaptive mechanism of the present invention establishes a dynamic association between bandwidth parameters and local density, which effectively suppresses noise interference in sparse areas while maintaining the detail accuracy of high-density areas. The distribution of advantageous areas in the three-dimensional heat map intuitively demonstrates the method's adaptive matching ability for complex data structures, and the position of the optimal working point in the contour map verifies the rationality of the parameter optimization strategy.

[0110] Furthermore, the bandwidth search objective is quantified by the bandwidth optimization loss function, which is calculated as follows:

[0111]

[0112] Where, L h The loss function is optimized for bandwidth. By minimizing this value, the bandwidth that best suits the distribution of orthopedic patients' movement characteristics can be obtained. is the kernel density estimate of the ith data point.

[0113] S603: Perform forward propagation processing on the data for convolutional neural network training.

[0114] The present invention uses a convolutional neural network to achieve deep feature learning of inter-frame differential feature maps. It relies on the combination of convolutional layers and pooling layers to capture key areas and perform dimensionality reduction at multiple levels. The activation function of the fully connected layer uses ReLU activation to retain motion details while reducing irrelevant information. The implementation of convolution operation, pooling operation and ReLU activation is expressed as follows:

[0115]

[0116] p(x)=max(P ch *x)

[0117] h(x)=max(0,x)

[0118] Where f(x) is the output of the convolutional layer; x is the input of the function, representing the data point input to the convolutional neural network; W k is the weight of the kth convolution kernel; K is the number of convolution kernels; k is a positive integer; * represents the convolution operation; b is the bias term; p(x) is the output of the pooling layer; P ch is the pooling window size; h(x) is the result after being processed by the ReLU activation function.

[0119] s604. Update the convolution kernel weights based on topological manifold kernel density estimation.

[0120] When training convolutional neural networks, the differences in the distribution density of orthopedic patient movement data in different areas will cause traditional gradient updates to ignore local details, because ordinary gradient descent is often guided by the overall error and cannot take into account the information value brought by different local densities.

[0121] Conventional methods rely purely on the gradient of backpropagation. Although they can gradually learn key information, they often fail to adjust in a timely manner when the local distribution is particularly dense or sparse, and may result in insufficient feature learning of relatively important areas.

[0122] This paper incorporates the gradient information obtained from the topological manifold kernel density estimation into the convolution kernel update, allowing the network to adaptively focus on the key moving areas with high density in the manifold. The convolution kernel weight update is achieved by superimposing the manifold gradient after the standard gradient descent term, which is expressed as:

[0123]

[0124] Where, is the convolution kernel weight of the kth convolution kernel at the t+1th iteration; is the convolution kernel weight of the kth convolution kernel at the tth iteration; η t is the learning rate of the t-th iteration of the convolutional neural network; is the gradient of the loss function of the convolutional neural network with respect to the weight of the kth convolution kernel; is the gradient term based on the topological manifold kernel density estimation; L is the loss function of the convolutional neural network; λ is the manifold density influence factor. Preferably, λ is set to 0.1 and η is set to 0.01.

[0125] The gradient term based on topological manifold kernel density estimation refers to an additional gradient added to the weight update of the convolutional neural network. The additional gradient is based on the local density estimation of each data point in the manifold. This gradient term allows the network to pay more attention to key areas with high density values in the distribution of orthopedic patient mobility monitoring data during training, thereby improving the recognition and learning capabilities of these areas.

[0126] s605. Calculate the loss function of electromagnetic field force guided optimization.

[0127] The high-dimensional parameter space in orthopedic patient mobility monitoring is extremely complex. Relying on conventional gradient descent methods may fall into local optimality or converge to certain weight configurations too early during convergence, making it impossible to fully explore the global solution space.

[0128] Although conventional methods can increase the search range to a certain extent through means such as momentum terms, they may still be unable to escape the constraints of local minima in complex high-dimensional distributions or strong noise environments.

[0129] The present invention simulates the dynamic behavior of particles during training by guiding optimization through electromagnetic field forces, allowing weights to obtain more flexible search power in high-dimensional space. This allows the convolutional neural network to take into account both global and local information during training, thereby better avoiding falling into local optimality, which can be expressed as:

[0130]

[0131] Where F is the comprehensive force on the particle, which represents the multi-directional guiding force on the weight parameters of the convolutional neural network during the high-dimensional space optimization process, including electric field force (driving the convolution kernel weight parameters of the convolutional neural network to move in the direction of loss reduction), magnetic field force (constraining the divergence of the update direction of the convolution kernel weight parameters of the convolutional neural network), manifold density gradient force (adjusting the search path according to the density of orthopedic patient movement monitoring data) and weight inertia force (retaining the historical update trend of the convolution kernel weight of the convolutional neural network); q is the charge of the particle, which represents the weight parameter in the current optimization stage. The importance weight is dynamically calculated based on its feature contribution in the convolutional layer; E is the electric field strength, which represents the pulling strength of the gradient direction of the convolutional neural network's loss function on the update of the convolution kernel weight parameters; v is the particle velocity, which represents the update momentum of the convolutional neural network's convolution kernel weight parameters in historical iterations; B is the magnetic field strength, which represents the orthogonal constraint component of the convolutional neural network's convolution kernel weight parameter update direction, used to avoid falling into a local optimal trajectory; γ is the manifold density gradient influence coefficient; α is the convolution kernel weight adjustment parameter; ΔW is the change in the convolution kernel weight between this iteration and the previous iteration. Preferably, γ is set to 0.3 and α is set to 0.2.

[0132] The electromagnetic field force-guided optimization refers to the process of training convolutional neural networks using orthopedic patient movement monitoring data. By simulating the motion laws of charged particles in the electromagnetic field, the optimization process of the convolution kernel weights is transformed into a dynamic system driven by multiple physical forces, thereby achieving a more robust global search in the high-dimensional parameter space.

[0133] The dynamic behavior of the particles refers to the fact that under the complex and changeable distribution of motion patterns in orthopedic data, the convolution kernel weight parameters of the convolutional neural network adjust their update direction and step size according to the real-time calculated comprehensive forces (electric field, magnetic field, manifold force) during training, retaining historical momentum to maintain stability.

[0134] The electric field strength characterizes the traction strength of the gradient direction of the loss function of the convolutional neural network on the update of the convolution kernel weight parameters. It simultaneously considers the impact of global error reduction and local data density on weight update, ensuring that the traction effect of the gradient is appropriately amplified in areas with high local density of orthopedic patient movement monitoring data, thereby making the weight update more sensitive to key areas. The calculation method is expressed as:

[0135]

[0136] Where, is the gradient of the loss function of the convolutional neural network with respect to the convolution kernel weight; is the kernel density estimate of the data point; δ ce is the density weighting coefficient. Preferably, δ ce Set to 0.2.

[0137] The particle velocity represents the update momentum of the convolution kernel weight parameters of the convolutional neural network in the historical iteration. In order to smooth the update process and retain historical information, an exponentially weighted moving average method can be used for updating. The update method is expressed as:

[0138] v t =μ ce ·v t-1 +(1-μ)·ΔW

[0139] Where, v t-1 is the particle velocity at the t-1th iteration; v t is the particle velocity at the tth iteration; μ ce is the momentum retention coefficient; ΔW is the change in the convolution kernel weight between this iteration and the previous iteration. ce Set to 0.9.

[0140] The magnetic field strength represents the orthogonal constraint component of the convolution kernel weight parameter update direction of the convolutional neural network, which is used to avoid falling into the local optimal trajectory. It is achieved by extracting the orthogonal component of the loss gradient in the current velocity direction, stripping the component along the existing momentum direction in the gradient, and retaining only the part orthogonal to the velocity direction, thereby providing a direction correction for the weight during the update and avoiding excessive dependence on a single gradient direction. The calculation method is expressed as:

[0141]

[0142] Where <·,·> represents the vector inner product operation.

[0143] The charge of the particle reflects the contribution of the current convolution kernel in feature extraction. Its calculation takes into account the gradient response strength of the convolution kernel and the local density information of the corresponding data point. It is specifically implemented through a normalized calculation method based on the Sigmoid function, which is expressed as:

[0144]

[0145] Where Sig(·) is the Sigmoid activation function; Represents the gradient of the convolution kernel with respect to the weight at the current input, reflecting the strength of feature contribution; a va is the first adjustment coefficient for balancing gradient response and data density; a vb is a second adjustment coefficient for balancing gradient response and data density. va Set to 0.2, a vb Set to 0.3.

[0146] Furthermore, the convolutional neural network loss function is calculated by combining the comprehensive force on the particle and the bandwidth optimization loss function. By combining the manifold density weighted term and the electromagnetic field force regularization term, the joint optimization of the high-dimensional parameter space is achieved. The calculation method is expressed as:

[0147]

[0148] Where, L h is the bandwidth optimization loss function; L is the loss function of the convolutional neural network; β is the electromagnetic field force regularization coefficient, which controls the contribution of the comprehensive force to the total loss; F k is the comprehensive electromagnetic force corresponding to the kth convolution kernel; ||·|| is the L2 norm; ∈ is the classification loss weight; y is the true label of the orthopedic patient movement monitoring data; Predicting labels for orthopedic patient mobility monitoring data using convolutional neural networks; is the cross entropy loss function. Preferably, β is set to 0.2 and ∈ is set to 0.3.

[0149] s606. Adaptive adjustment of learning rate based on changes in the loss function of the convolutional neural network.

[0150] When using a training dataset consisting of orthopedic patient movement monitoring data for model training, noise interference or abnormal data distribution may occur in the later stage of model training (when the model is close to fitting), which may lead to training oscillation or overfitting if a large learning rate is maintained.

[0151] Conventional methods usually set a fixed decay strategy or artificially reduce the learning rate at certain predetermined stages of training. They lack a fine-grained perception of real-time loss changes and cannot cope well with scenarios with large fluctuations.

[0152] The present invention adaptively adjusts the learning rate based on the change of the loss function of the convolutional neural network, so that the network can gradually reduce the step size and perform more detailed parameter search in the later stage. Specifically, the learning rate of the next iteration is adjusted by the relative change rate of the loss function and the gradient information, so as to achieve flexible response to unstable situations, which is expressed as:

[0153]

[0154] Where η t is the learning rate of the tth iteration of the convolutional neural network; η t+1 is the learning rate of the t+1th iteration of the convolutional neural network; λ is the first learning rate adjustment factor; μ is the second learning rate adjustment factor; ΔL t is the change in the loss function at the tth iteration; L t is the loss function of the t-th iteration; is the gradient of the loss function at iteration t. Preferably, λ is set to 0.3 and μ is set to 0.5.

[0155] In one embodiment, Figure 10 As shown in the figure, the dynamic coupling relationship between the learning rate adjustment strategy and the loss reduction is analyzed through the dual-axis curve. The conventional method uses mechanical attenuation to maintain a high learning rate in the later stage of training (40-60 rounds), resulting in loss oscillation. The present invention dynamically adjusts the loss change rate through real-time perception, automatically reduces the step size when the model is close to convergence, and retains the ability to correct the gradient direction. It effectively suppresses overshoot in the key stage (marked area), making the loss curve present a more stable exponential decay form.

[0156] S607: Determine the iterative stopping condition of the convolutional neural network training.

[0157] When the relative change rate of the validation set loss function is less than a threshold (such as 0.1%) in three consecutive iterations, the model is judged to have converged, and the iteration is stopped, indicating that the convolutional neural network model training is completed.

[0158] S7. Obtain monitoring results based on the mobile monitoring model

[0159] The trained mobile monitoring model, namely the convolutional neural network model, is used to classify the mobile monitoring of orthopedic patients, and the classification results correspond to three categories: small amplitude (safe), medium amplitude (warning), and large amplitude (high risk).

[0160] In one embodiment, the structure of the convolutional neural network is:

[0161] Input layer: The input layer is the orthopedic patient movement monitoring data, each piece of data represents a specific movement behavior.

[0162] Convolutional layer and pooling layer: The convolutional neural network contains multiple (such as 3) convolutional layers, and each layer uses multiple (such as 4) convolution kernels to extract features from the input data;

[0163] Each convolutional layer is followed by a pooling layer to reduce the size of the feature map while retaining key features.

[0164] Fully connected layer: The high-order features output by the pooling layer are flattened into a one-dimensional vector and fed into one or two layers of fully connected networks to further integrate the features.

[0165] Output layer: The output layer contains three neurons, corresponding to the three categories of small amplitude (safe), medium amplitude (warning), and large amplitude (high risk);

[0166] Through normalization processing (such as Softmax function), the output of the network is converted into three probability values, and the category corresponding to the neuron with the largest probability value is taken as the final output category.

[0167] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for monitoring the movement of orthopedic patients, characterized in that: The following steps are involved: S1. Collect patient movement data, perform inter-frame image processing on the collected data, and extract the patient's gait characteristics; S2. Processing the collected adjacent frame images to obtain an original grayscale image, a histogram equalization image, and a denoising result image; S3, performing feature fusion on the original grayscale image, the histogram equalization image, and the denoising result image to obtain a preprocessed image; S4, performing registration and difference calculation on the pre-processed images; S5. Use the frame difference method to train and label the data; S6. Construct a mobile monitoring model for orthopedic patients; S7. Obtain monitoring results based on the mobile monitoring model.

2. The method for monitoring the movement of orthopedic patients according to claim 1, characterized in that: In S2, the weighted average method is used to calculate the grayscale value to obtain the original grayscale image; The cumulative distribution function is used to map the original grayscale value to the new distribution to obtain a histogram equalization image; The non-local mean denoising method is used to perform weighted average denoising using the redundant information of similar structures in the image, retaining edge details and matching the range of similar blocks to obtain the denoised result image.

3. The method for monitoring the movement of orthopedic patients according to claim 1, characterized in that: The feature fusion method described in S3 is obtained by adding the pixel values of corresponding pixel positions and then taking the average.

4. The method for monitoring the movement of orthopedic patients according to claim 1, characterized in that: S4 specifically includes: Align the pre-processed images corresponding to two consecutive frames to the same spatial coordinate system to eliminate the geometric offset caused by differences in shooting angle, position or time; By traversing all feature descriptors of the two images, calculating the Hamming distance, and retaining the top 50% matching pairs with the smallest distance, the interference of false matches is reduced; Quantify the difference between two consecutive registered frames and identify obvious displacement features.

5. The method for monitoring the movement of orthopedic patients according to claim 1, characterized in that: S5 specifically includes: The data of two consecutive frames of original images are processed using the inter-frame difference method to obtain the movement change data of the first orthopedic patient; The original grayscale image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the second orthopedic patient; The histogram equalization image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the third orthopedic patient; The denoised image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fourth orthopedic patient; The pre-processed image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the fifth orthopedic patient; The inter-frame difference method is used to process the Canny edge detection image data corresponding to two consecutive frames of images to obtain the movement change data of the sixth orthopedic patient; The ORB feature point image data corresponding to two consecutive frames of images are processed using the inter-frame difference method to obtain the movement change data of the seventh orthopedic patient.

6. The method for monitoring movement of orthopedic patients according to claim 5, characterized in that: The first orthopedic patient movement change data, the second orthopedic patient movement change data, the third orthopedic patient movement change data, the fourth orthopedic patient movement change data, the fifth orthopedic patient movement change data, the sixth orthopedic patient movement change data, the seventh orthopedic patient movement change data and the pixel-level difference heat map are feature fused to obtain orthopedic patient movement monitoring training sample data.

7. The method for monitoring movement of orthopedic patients according to claim 1, characterized in that: S6 builds an orthopedic patient mobility monitoring model including: Conduct topological manifold kernel density estimation on orthopedic patient mobility monitoring data; Adaptively select the bandwidth parameter by minimizing the error function; Perform forward propagation processing of data for convolutional neural network training; Update convolution kernel weights based on topological manifold kernel density estimation; Calculate the loss function in electromagnetic force-guided optimization; Adaptive adjustment of learning rate based on changes in the loss function of the convolutional neural network; The iterative stopping condition of convolutional neural network training is judged. When the relative change rate of the validation set loss function is less than the threshold in three consecutive iterations, the model is judged to have converged, and the iteration is stopped, indicating that the convolutional neural network model training is completed.

8. The method for monitoring movement of orthopedic patients according to claim 1, characterized in that: The movement data is derived from video recordings of the patient's gait and movement patterns, and during the data collection process, the patient is asked to follow instructions and walk in a standardized gait analysis area.

Images