Non-contact monitoring method and system for rehabilitation activity of neurosurgical patient

Abstract

The invention provides a neurosurgery patient rehabilitation activity non-contact monitoring method and system, and the method comprises the steps: collecting patient surface data, and inputting the patient surface data into a skeleton point estimation model to obtain a skeleton sequence tensor; identifying an action intention expressed by the skeleton sequence tensor, and marking a time range of the action intention on a time dimension; the action intention is obtained through recognition of a double-branch model; the double-branch model comprises an action intention recognition channel and an intention time positioning channel; evaluating the action intention and the time range thereof to obtain a consistency score; and evaluating a compliance score of the action intention on rehabilitation training, wherein the compliance score comprises a physical amplitude achievement degree, a time matching score and a consistency score.

Description

Technical Field

[0001] This invention belongs to the field of medical monitoring, and in particular relates to a non-contact monitoring method and system for the rehabilitation activities of neurosurgical patients. Background Technology

[0002] The rehabilitation phase after neurosurgery directly determines the final quality of a patient's recovery in motor, cognitive, and daily living abilities. In the early postoperative period, patients often experience confusion, muscle weakness, and medication effects, resulting in limited, irregular, or even intermittent rehabilitation movements. These movements are often accompanied by epileptic-like tics or involuntary tremors, making it difficult for healthcare professionals to accurately assess the patient's behavioral intentions and training effectiveness visually. Traditional contact-based monitoring (ECG patches, pressure pads, etc.) generally suffers from poor compliance due to the need for long-term application and cannot analyze specific movement patterns. Existing non-contact visual solutions rely heavily on large-amplitude postural changes or simple motion thresholds, lacking the ability to perceive low-amplitude, short-term skeletal deformities and failing to distinguish between "whether the patient wants to actively train" and "whether the movement itself meets the standards." Furthermore, existing systems often lack a unified assessment logic to close the loop of "movement detected → motivation understood → compliance assessed," leading to frequent false alarms and missed alarms, thus limiting their value for early rehabilitation planning and risk warning. Summary of the Invention

[0003] The purpose of this invention is to design a non-contact monitoring method and system for the rehabilitation activities of neurosurgical patients. This method avoids the dependence on patient contact with traditional equipment and overcomes the requirements of visual solutions for large movements. It can provide real-time early warning and efficacy quantification support with high sensitivity and low misjudgment rate in the early rehabilitation stage, significantly improving the safety and effectiveness of individualized rehabilitation management for neurosurgical patients.

[0004] To achieve the above objectives, a first aspect of the present invention provides a non-contact monitoring method for rehabilitation activities of neurosurgical patients, the method comprising:

[0005] Patient surface data is collected and input into a skeleton point estimation model to obtain a skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder for the depth map and a heatmap generation and coordinate regression module.

[0006] The action intent expressed by the skeleton sequence tensor is identified, and its time range is marked in the time dimension; the action intent is identified through a two-branch model; the two-branch model includes an action intent recognition channel and an intent time localization channel.

[0007] A consistency score is obtained by evaluating the action intention and its time range. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions: time period action structure characteristics, rehabilitation phase plan rule constraints, and spatial environment semantic constraints.

[0008] The compliance score for the motor intention assessment in rehabilitation training includes physical amplitude achievement, time matching score, and consistency score.

[0009] Furthermore, the skeleton sequence tensor contains skeleton data for T frames, where T is a preset number of frames; the skeleton data includes a set of skeleton keypoint coordinates after smoothing each frame.

[0010] Furthermore, the working steps of the skeleton point estimation model specifically include: inputting the patient surface data into the local feature encoder to extract the local gradient changes of the human body contour, and sending the extracted local gradient changes of the human body contour into the spatial attention module to generate a two-dimensional heat map of the skeleton points. Using the position of the maximum response point as the initial coordinates, and combining it with the distance data of the corresponding points in the depth map, the three-dimensional coordinates of each skeleton point are obtained by back projection and combined into a single frame skeleton structure.

[0011] Furthermore, the skeleton point estimation model also introduces a confidence filtering and point completion mechanism. This mechanism retains only points with a confidence level exceeding a preset threshold as valid outputs in the heatmap response of each frame, while low-confidence points are filled by linear interpolation between the previous and next frames to construct a complete skeleton frame.

[0012] Furthermore, a learnable temporal attention mechanism is introduced into the dual-branch model. This learnable temporal attention mechanism semantically evaluates the degree of matching between a frame and the overall intent and physically evaluates whether there are motion fluctuations in a frame, thereby improving the model's ability to locate the true intent segment.

[0013] Furthermore, the action intent recognition channel obtains a global representation by applying temporal global average pooling to the hidden sequence, and then feeds it into a set of three fully connected layers to predict the rehabilitation action category. The hidden sequence is obtained by extracting the dynamic context features of each frame skeleton by a skeleton sequence tensor temporal encoder.

[0014] Furthermore, the time-segment action structure features are obtained by extracting the average joint velocity and trajectory concentration within the time segment, where the trajectory concentration represents the three-dimensional spatial variance of a point within the corresponding time segment.

[0015] Furthermore, a motion rhythm variance penalty term is introduced into the consistency score to suppress the misjudgment of high-frequency, non-rhythmic human errors as rehabilitation intentions.

[0016] Furthermore, the physical achievement of the movement indicates whether a certain skeletal sequence has achieved the key joint displacement requirements corresponding to the movement; the time matching score is used to determine whether the movement is performed within the time range allowed by the patient's rehabilitation plan.

[0017] In a second aspect of the invention, a non-contact monitoring system for the rehabilitation activities of neurosurgical patients is provided, the system comprising:

[0018] The data acquisition unit is used to acquire patient surface data and input the patient surface data into the skeleton point estimation model to obtain a skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder of the depth map and a heat map generation and coordinate regression module.

[0019] An intent recognition unit is used to recognize the action intent expressed by the skeleton sequence tensor and mark its time range in the time dimension; the action intent is recognized by a dual-branch model; the dual-branch model includes an action intent recognition channel and an intent time localization channel.

[0020] The standard judgment unit is used to evaluate the action intention and its time range to obtain a consistency score. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions, including: time period action structure characteristics, rehabilitation stage plan rule constraints, and spatial environment semantic constraints.

[0021] The behavioral assessment unit is used to assess the compliance score of the said movement intention in rehabilitation training. The compliance score includes physical amplitude achievement, time matching score, and consistency score.

[0022] The beneficial technical effects of the present invention are at least as follows:

[0023] To address the aforementioned issues, this invention provides a non-contact monitoring method and system for neurosurgical patients' rehabilitation activities. It acquires the patient's skeletal structure in real time using an overhead RGB-D or millimeter-wave sensing device and utilizes temporal smoothing and confidence completion techniques to ensure stable input during the micro-movement phase. Subsequently, a dual-branch temporal network distinguishes active rehabilitation intentions from background clutter within milliseconds, and accurate intention timing is achieved using entropy regularization and physical amplitude weighting. Based on this, the system combines rehabilitation phase plans, spatial posture rules, and rhythmic consistency penalties to output a cognitive consistency score and further calculates a comprehensive compliance score including movement amplitude, time window, and cognitive cooperation.

[0024] Through the above mechanism, the system avoids the dependence on patient contact with traditional equipment and overcomes the requirements of visual solutions for large movements. It can provide real-time early warning and efficacy quantification support with high sensitivity and low misjudgment rate in the early rehabilitation stage, significantly improving the safety and effectiveness of individualized rehabilitation management for neurosurgical patients. Attached Figure Description

[0025] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0026] Figure 1 This is a flowchart of a non-contact monitoring method for the rehabilitation activities of neurosurgical patients according to the present invention.

[0027] Figure 2 This is a framework diagram of a non-contact monitoring system for the rehabilitation activities of neurosurgical patients according to the present invention. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0029] In one or more embodiments, such as Figure 1 As shown, a non-contact monitoring method for the rehabilitation activities of neurosurgical patients is disclosed, the method comprising the following:

[0030] S1: Collect patient surface data, and input the patient surface data into the skeleton point estimation model to obtain the skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder of the depth map and a heat map generation and coordinate regression module.

[0031] Specifically, in this embodiment, this step aims to provide the basic data input for the entire rehabilitation monitoring system: a structured and temporally consistent sequence of human skeletal points over a continuous period of time. This skeletal data must be obtained entirely non-contactly and should possess high temporal accuracy, sensitivity to micro-movements, and robustness to occlusion and environmental factors. In post-neurosurgery scenarios, patients are typically supine or lateral, with limited range of motion, minimal posture changes, and some limbs exhibiting neurological sluggishness or even loss of control. Therefore, ordinary cameras or standard human posture estimation algorithms often suffer from incomplete key point recognition, skeletal frame skipping, and spatial drift, failing to support subsequent judgment of rehabilitation intentions. This step, through a specially constructed data acquisition structure, a skeletal estimation network specifically for neurological patients, and a skeletal point smoothing mechanism with confidence weighting, achieves a non-contact human structural information extraction scheme optimized for this type of scenario.

[0032] Specifically, the data acquisition module is fixedly installed above or diagonally opposite the hospital bed, and includes a set of RGB-D depth cameras (such as Azure Kinect DK) or a 77GHz millimeter-wave radar sensing array, acquiring 15 to 30 frames of depth image sequences per second. Each image frame is approximately 640×480 pixels in size, with a depth range of 0.5 meters to 3.5 meters. The depth images are calibrated by the system to ensure that the viewing angle is orthogonal to or above the bed, avoiding estimation distortion caused by posture tilt.

[0033] In this embodiment, Azure Kinect is used. Its D-camera can stably acquire patient surface depth data in dark environments and can penetrate thin blankets of a certain thickness to extract contour changes. During acquisition, the system feeds data in batches into the downstream skeleton estimation network in fixed-length time windows (e.g., 10 seconds). For the radar scheme, continuous spectrum signals are synthesized through delay to form depth contour frames, which can also be converted into approximate depth map processing logic. The two input structures are ultimately unified as follows: For network processing.

[0034] Furthermore, the skeleton point estimation adopts a two-stage structure: the first stage is a local feature encoder for the depth map, and the second stage is a heatmap generation and coordinate regression module. The encoder part uses a stacked three convolutional layers, with each layer having 64, 128, and 256 kernels, a kernel size of 3×3, and the activation function being ReLU. This structure can extract local gradient changes of the human contour while maintaining the model's lightweight nature, and is suitable for inputs with relatively blurred facial and limb contours in a supine position.

[0035] The convolutional output is fed into a spatial attention module to guide the model to enhance responses in key regions such as the torso and joints, thereby generating a 2D heatmap of 15 skeleton points (each channel corresponds to a joint). The system uses the location of the point with the maximum response as the initial coordinates. Then, by combining the distance data of corresponding points in the depth map, the three-dimensional coordinates of each skeleton point are obtained through back projection. Combined into a single-frame skeleton structure .

[0036] Because neurosurgical patients often experience involuntary tremors or even brief muscle contractions after surgery, the skeleton points output by the model frequently exhibit temporal fluctuations or drifts. Therefore, a temporal filtering mechanism is introduced here to perform short-term smoothing on the skeleton points of each frame.

[0037] ;

[0038] in, Indicates the first The frame is composed of skeleton point coordinates predicted by a pose estimation network, with dimensions of [dimension number missing]. ; For the first time window The weighting coefficients of the frames are generated and normalized by Gaussian templates; For the first The final smoothing result of frame skeleton points.

[0039] The system also addresses the issues of occlusion and low-confidence keypoints by implementing a confidence filtering and point completion mechanism. In the heatmap response of each frame, only points with a confidence level greater than 0.6 are retained as valid outputs, while low-confidence points are filled by linear interpolation between the preceding and following frames to construct a structurally complete skeleton frame.

[0040] The final constructed skeleton sequence tensor is shown below:

[0041] ;

[0042] in, Represents a complete skeleton sequence tensor, containing Frame skeleton data; This is the set of skeleton keypoint coordinates for each frame after smoothing. The total number of key points is set to 15 in this system.

[0043] S2: Identify the action intent expressed by the skeleton sequence tensor and mark its time range in the time dimension; the action intent is identified through a dual-branch model; the dual-branch model includes an action intent recognition channel and an intent time localization channel;

[0044] Specifically, in this embodiment, this step is used to identify the specific motor intentions expressed by neurosurgical patients during rehabilitation, and further accurately mark the interval in which the intention occurs in the time dimension. The monitoring scenario targeted by this system is highly specific: patients are often still or moving slowly in bed, and their expressed rehabilitation movements have characteristics such as "small amplitude, unstable rhythm, unclear start and end, and mixed non-intentional movements". Therefore, compared with traditional gesture recognition systems or ordinary video action recognition systems, a recognition mechanism with a structure more adapted to this type of low-intensity, continuous skeletal data needs to be introduced.

[0045] Specifically, this step revolves around the skeleton sequence tensor output in step 1. A dual-branch model structure based on "micro-motion perception" and "temporal interpretability" is constructed: on the one hand, semantic intent recognition of low-amplitude movements is achieved; on the other hand, intent localization is realized through an attention mechanism with a temporal regularization structure. The key points are: 1) introducing a local perturbation filtering mechanism based on the consistency of skeleton point motion; 2) using the minimization of the entropy of the temporal attention distribution as regularization, forcing the model to focus on recognizing real action segments in time sequence; 3) using the physical motion amplitude and network response intensity together as judgment factors to guide the model to mine potential intents from low-amplitude movements.

[0046] The input is the skeleton sequence tensor output in step 1. ,in The number of skeleton points. The sequence length is given. This data has undergone occlusion point completion and temporal filtering, possesses sufficient structural stability, and can be directly used as input features for the action recognition module.

[0047] Input tensor First, a temporal encoder is fed in to extract the dynamic context features of the skeleton for each frame. The encoder uses a two-layer bidirectional GRU structure, with each layer having a hidden dimension of 64. This structure outputs the hidden sequence. ,in Indicates the feature dimension.

[0048] Furthermore, in the dual-branch module:

[0049] The first branch is the action intent recognition channel. This involves the hidden sequence. Global representation is obtained by applying time-domain global average pooling. And feed in a set of three fully connected layers (64→32→ , (Predict rehabilitation movement categories based on the number of movement categories) .

[0050] The second branch is used for temporal localization of intent. A learnable temporal attention mechanism is introduced here, with its core output being a frame-level weight distribution. each Indicates the first The probability that a frame belongs to the current intended action. To enhance its interpretability and physical consistency, the following fusion scoring mechanism is designed here:

[0051] ;

[0052] in, For the first Frame coding features; The projection matrix; This is a global representation vector; For the first Frame number Coordinates of the skeleton points; For the Sigmoid function; The first term is the amplitude adjustment factor (between 0.3 and 0.6); the second term is the average skeleton point motion amplitude, which serves as a weighting factor for the intensity of micro-motions. This formula combines two parts of information: first, it semantically assesses the degree of matching between the frame and the overall intent; second, it physically evaluates whether there are motion fluctuations in the frame, thereby improving the model's ability to locate segments of true intent, and is especially suitable for "non-uniform rhythm" scenarios.

[0053] To further suppress the model's blurred response to background frames, an entropy regularization term is introduced into the temporal attention distribution to encourage the model to focus on short-term concentrated intent regions:

[0054] ;

[0055] in, This is a small constant (to prevent numerical instability). This term is added to the total loss function during training, prompting... The distribution tends to be sparse, thereby reducing interference from still frames.

[0056] In actual deployment, the system sets an attention score threshold. (e.g., 0.4), when in consecutive frames Then the paragraph is marked as the interval in which an intention to act occurred. Multiple intentions can be identified sequentially within a sequence.

[0057] S3: The consistency score is obtained by evaluating the action intention and its time range. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions, including: time period action structure characteristics, rehabilitation phase plan rule constraints, and spatial environment semantic constraints.

[0058] Specifically, in this embodiment, this step aims to perform a "cognitive consistency judgment" on the rehabilitation movement intentions and their time range identified in the previous stage. This involves assessing whether the movement aligns with the patient's current rehabilitation stage, physiological state, and environmental conditions, excluding signals that might mislead rehabilitation intervention, such as "unconscious movements," "misidentified movements," and "inappropriate movements." In the post-neurosurgical monitoring environment, patients often experience confusion, delayed cognitive command transmission, and motor control impairments. While some movements may appear as standard patterns at the skeletal level, they may not necessarily stem from the patient's true intentions. Therefore, skeletal trajectories alone are insufficient for comprehensive behavioral understanding; a cross-consistency judgment method combining temporal movement characteristics, rehabilitation stage plans, and spatial constraints must be introduced. This step is a crucial relay link in the entire system; its judgment results not only provide a basis for subsequent rehabilitation responses but also serve as a core guarantee for the reliability of the monitoring system.

[0059] The system assesses the consistency of motor intention from three dimensions: temporal motor structure characteristics, rehabilitation phase planning rules, and spatial environmental semantic constraints. These are ultimately integrated into a scoring function to generate a consistency score. The entire decision-making logic has a modular structure, making it easy to deploy to edge devices.

[0060] First, the average joint velocity within this time period is extracted here. and trajectory concentration This is used to reflect whether an action has a genuine intention to be performed. This reflects whether the patient has made sufficient proactive movements. Then determine whether the action has a consistent goal or direction.

[0061] ;

[0062] in, Indicates the first Frame number The three-dimensional coordinates of the skeleton points This refers to the number of skeletal points. This metric measures the average speed of skeletal movement of the patient within the intended segment.

[0063] Recalculate the trajectory concentration index:

[0064] ;

[0065] in It represents the three-dimensional spatial variance of a point in that time period, reflecting whether the action is concentrated in a certain area. For example, "clenching a fist" should have a hand aggregation process, rather than a highly random drifting action.

[0066] Next, a stage rationality matrix is ​​introduced here. Used to assess whether the current movement matches the rehabilitation phase. . The system assigns a number to the patient's current postoperative recovery stage, which is updated in real time by the electronic medical record system or recovery planning module (e.g., day 3 post-surgery). Matrix Manually set, for example It indicates that "knee-bending" movements are prohibited during the first stage after surgery.

[0067] In addition, this step incorporates a spatial semantic rationality function. Used to classify actions and the actual spatial distribution of the skeleton It determines whether an action occurs within a semantically appropriate region. For example, the action of "sitting up" should be characterized by an increase in the height difference between the head and pelvis, and an overall rise in the center of gravity; the action of "raising a leg" should be characterized by the key points of one lower limb lifting off the bed surface. This part is implemented using a rule-driven geometric analysis module and requires no training.

[0068] To improve the sensitivity and robustness of the judgment, an innovative regularization term is introduced into the final consistency scoring function: a motion rhythm variance penalty term. Its design aims to prevent high-frequency, non-rhythmic maneuvers (such as tremors and shaking) from being misinterpreted as rehabilitation intentions.

[0069] ;

[0070] in, This is a set of skeletal point indices most relevant to the action (e.g., "clenching a fist" corresponds to the fingertips and wrist key points). This item distinguishes between "twitches" and "active intentions" by controlling the rhythmic consistency of skeletal micro-movements.

[0071] The final consistency scoring model is:

[0072] ;

[0073] in, This is the normalized output of the Sigmoid function; This is a local minimum constant used to avoid division by zero; These are adjustable hyperparameters (e.g., 0.25, 0.2, 0.2, 0.15), which can be optimized according to the scenario. The formula structure clearly integrates three types of structural information: temporal features, semantic rules, and action rhythm.

[0074] The final consistency score is obtained. This indicates whether the current intentional action is cognitively reasonable. If If the value is 0.6, the intention is judged as cognitively consistent; otherwise, it is recorded as a deviated intention or misidentification and sent to the system's early warning and statistics module.

[0075] S4: Assess the compliance score of the stated movement intention in rehabilitation training, the compliance score including physical amplitude achievement, time matching score, and consistency score.

[0076] Specifically, in this embodiment, the goal of this step is to further evaluate the compliance of the action in rehabilitation training after identifying the patient's rehabilitation intention (step 2) and completing the cognitive consistency assessment (step 3). This involves determining whether the action has rehabilitation value, meets the basic physical requirements of the action goal, and occurs within a reasonable timeframe. This step directly addresses a common problem in post-neurosurgery monitoring: even if some patients perform actions with "intention" and "cognitive consistency," the range of motion, body posture, or timing may not conform to the basic training goals of the rehabilitation plan, thereby reducing rehabilitation efficiency or creating risks. This step uses simple, clear, and compliant quantitative standards to score the compliance of each action segment, serving as the core indicator for the final structured monitoring output.

[0077] To achieve compliance assessment, a score is constructed based on two core dimensions: the degree of achievement of the physical characteristics of rehabilitation movements and the matching of the training plan's time window, outputting a comprehensive compliance score. This serves as the core indicator for the final structured monitoring output. The entire method closely aligns with actual rehabilitation needs, without introducing any scenarios-irrelevant models or external information.

[0078] First, calculate the motion physics achievement. This indicates whether the skeleton sequence has implemented the action. The corresponding key joint displacement requirements. Specifically, for each type of rehabilitation movement... Predefine a set of key skeleton points and minimum displacement threshold Determine whether the maximum displacement of these points meets the requirements:

[0079] ;

[0080] in, Indicates in Time of the first The three-dimensional coordinates of the key points; For action The corresponding set of key points; Minimum amplitude requirement; To prevent division by zero of constants.

[0081] Furthermore, the system determines whether the action is performed within the timeframe allowed by the patient's rehabilitation plan. Each action category... Corresponding to a predefined planned execution time period The system extracts the center point of the action time. Then calculate the time matching degree. :

[0082] ;

[0083] in, Indicates the recommended time point for action execution. This refers to the width of the recommended time period.

[0084] Ultimately, the compliance score is calculated based on three dimensions:

[0085] ;

[0086] in, For the degree of physical amplitude of the action, For time matching score, The cognitive consistency score is calculated (output of step 3). For weighted parameters, satisfying It can be set according to the actual scenario, and the default value is recommended to be 0.4, 0.3, 0.3.

[0087] The output is a structured compliance score result, including: compliance score. Sub-rating (Physical amplitude achievement); Sub-rating (Time matching); Sub-scoring (Cognitive consistency score); corresponding action number and time period .

[0088] The results will be recorded in the monitoring system database as part of the behavior monitoring report, and used for purposes such as physician retrospective analysis, rehabilitation trend assessment, and abnormal early warning.

[0089] Specific examples are as follows:

[0090] Action Number: c Action Time: t_start ~ t_end Amplitude Score (a_c): 0.82 Time Matching Degree (m_c): 0.91 Cognitive Consistency Score (r): 0.87 Final Compliance Score (R): 0.86

[0091] In one or more embodiments, such as Figure 2 As shown, a non-contact monitoring system for the rehabilitation activities of neurosurgical patients is disclosed, the system comprising:

[0092] The data acquisition unit is used to acquire patient surface data and input the patient surface data into the skeleton point estimation model to obtain a skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder of the depth map and a heat map generation and coordinate regression module.

[0093] An intent recognition unit is used to recognize the action intent expressed by the skeleton sequence tensor and mark its time range in the time dimension; the action intent is recognized by a dual-branch model; the dual-branch model includes an action intent recognition channel and an intent time localization channel.

[0094] The standard judgment unit is used to evaluate the action intention and its time range to obtain a consistency score. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions, including: time period action structure characteristics, rehabilitation stage plan rule constraints, and spatial environment semantic constraints.

[0095] The behavioral assessment unit is used to assess the compliance score of the said movement intention in rehabilitation training. The compliance score includes physical amplitude achievement, time matching score, and consistency score.

[0096] It is worth noting that the specific workflow of the non-contact monitoring system for neurosurgical patient rehabilitation activities provided in this embodiment of the invention is the same as that of the non-contact monitoring method for neurosurgical patient rehabilitation activities described in the above embodiment, and will not be repeated here.

[0097] This invention also provides a non-contact monitoring device for neurosurgical patient rehabilitation activities, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of a non-contact monitoring method for neurosurgical patient rehabilitation activities, for example... Figure 1 The steps S1 to S4 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0098] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the non-contact monitoring device for neurosurgical patient rehabilitation activities.

[0099] The aforementioned non-contact monitoring device for neurosurgical patient rehabilitation activities can be a desktop computer, laptop, handheld computer, or cloud server, etc. This device may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the device may also include input / output devices, network access devices, buses, etc.

[0100] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the non-contact monitoring device for neurosurgical patient rehabilitation activities, connecting all parts of the device via various interfaces and lines.

[0101] The memory can be used to store the computer programs and / or modules. The processor, by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory, realizes various functions of the non-contact monitoring device for neurosurgical patient rehabilitation activities. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the operation of the air conditioning controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, SmartMemory Card (SMC), Secure Digital Card (SD), Flash Card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0102] The module integrated into the non-contact monitoring device for neurosurgical patient rehabilitation activities, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0103] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0104] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A non-contact monitoring method for the rehabilitation activities of neurosurgical patients, characterized in that, The method includes: Patient surface data is collected and input into a skeleton point estimation model to obtain a skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder for the depth map and a heatmap generation and coordinate regression module. The action intent expressed by the skeleton sequence tensor is identified, and its time range is marked in the time dimension; the action intent is identified through a two-branch model; the two-branch model includes an action intent recognition channel and an intent time localization channel. A consistency score is obtained by evaluating the action intention and its time range. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions: time period action structure characteristics, rehabilitation phase plan rule constraints, and spatial environment semantic constraints. The compliance score for the motor intention assessment in rehabilitation training includes physical amplitude achievement, time matching score, and consistency score.

2. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The skeleton sequence tensor contains skeleton data for T frames, where T is a preset number of frames; the skeleton data includes a set of skeleton keypoint coordinates after smoothing each frame.

3. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The working steps of the skeleton point estimation model specifically include: inputting the patient surface data into the local feature encoder to extract the local gradient changes of the human body contour, and sending the extracted local gradient changes of the human body contour into the spatial attention module to generate a two-dimensional heat map of the skeleton points. Using the position of the maximum response point as the initial coordinates, and combining it with the distance data of the corresponding points in the depth map, the three-dimensional coordinates of each skeleton point are obtained by back projection and combined into a single frame skeleton structure.

4. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 3, characterized in that, The skeleton point estimation model also introduces a confidence filtering and point completion mechanism. This mechanism retains only points with a confidence level exceeding a preset threshold as valid outputs in the heatmap response of each frame, while low-confidence points are filled by linear interpolation between the previous and next frames to construct a complete skeleton frame.

5. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The dual-branch model introduces a learnable temporal attention mechanism, which semantically evaluates the degree of matching between a frame and the overall intent and physically evaluates whether there are motion fluctuations in a frame, thereby improving the model's ability to locate the true intent segment.

6. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The action intent recognition channel obtains a global representation by applying temporal global average pooling to the hidden sequence, and then feeds it into a set of three fully connected layers to predict the rehabilitation action category. The hidden sequence is obtained by extracting the dynamic context features of each frame skeleton by a skeleton sequence tensor temporal encoder.

7. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The time-segment action structure features are obtained by extracting the average joint velocity and trajectory concentration within the time period, where the trajectory concentration represents the three-dimensional spatial variance of a point in the corresponding time period.

8. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The consistency score incorporates a motion rhythm variance penalty term to suppress the misjudgment of high-frequency, non-rhythmic human errors as rehabilitation intentions.

9. The non-contact monitoring method for neurosurgical patient rehabilitation activities according to claim 1, characterized in that, The physical achievement of the movement indicates whether a certain skeletal sequence has achieved the key joint displacement requirements corresponding to the movement; the time matching score is used to determine whether the movement is performed within the time range allowed by the patient's rehabilitation plan.

10. A non-contact monitoring system for the rehabilitation activities of neurosurgical patients, characterized in that, The system includes: The data acquisition unit is used to acquire patient surface data and input the patient surface data into the skeleton point estimation model to obtain a skeleton sequence tensor. The skeleton point estimation model includes a local feature encoder of the depth map and a heat map generation and coordinate regression module. An intent recognition unit is used to recognize the action intent expressed by the skeleton sequence tensor and mark its time range in the time dimension; the action intent is recognized by a dual-branch model; the dual-branch model includes an action intent recognition channel and an intent time localization channel. The standard judgment unit is used to evaluate the action intention and its time range to obtain a consistency score. The consistency score is obtained by evaluating the consistency of the action intention from three dimensions, including: time period action structure characteristics, rehabilitation stage plan rule constraints, and spatial environment semantic constraints. The behavioral assessment unit is used to assess the compliance score of the said movement intention in rehabilitation training. The compliance score includes physical amplitude achievement, time matching score, and consistency score.

Images