Motion function data acquisition method and system based on human-computer interaction

Abstract

The invention relates to a motion function data acquisition method and system based on human-computer interaction. The motion function analysis method is applied to a motion function analysis device and comprises the steps that in response to an initial position instruction of an abnormal object, initial motion analysis is conducted based on historical treatment data of the abnormal object, and initial motion data is presented; when the corresponding initial body image is detected, pose analysis is executed, and a body condition labeling box is displayed to a labeling person. After annotation confirmation is received, carrying out movement analysis in combination with annotation data and historical treatment data, and outputting action change data; and if the current body image corresponding to the current body image is detected, taking the current body image as a new initial image, and repeating the pose analysis and labeling process. And until an annotation stop instruction is received, the system summarizes each annotation and historical treatment data to carry out object function analysis, and motion function data of the abnormal object is generated. By adopting the method, the accuracy of understanding the current abnormal information of the abnormal personnel can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent interaction, in particular to a motion function data acquisition method and system based on human-computer interaction. BACKGROUND

[0002] In the traditional technology, the motion function data of patients with motor dysfunction before and after rehabilitation intervention is obtained by personnel observation and direct input of corresponding grade data on the computer, which leads to time-consuming and laborious scoring of abnormal personnel and inaccuracy, and in the case of medical resource shortage, the accuracy of understanding the current abnormal information of abnormal personnel is insufficient. SUMMARY

[0003] Therefore, it is necessary to provide a motion function data acquisition method and system based on human-computer interaction, which can effectively improve the accuracy of understanding the current abnormal information of abnormal personnel.

[0004] In a first aspect, the present application provides a motion function data acquisition method based on human-computer interaction, applied to a motion function analysis device, comprising: In response to an initial position determination instruction of an abnormal object, performing initial action analysis according to historical treatment data of the abnormal object, and presenting initial motion data to the abnormal object; When an initial body image corresponding to the initial motion data is detected, performing pose analysis on the initial body image, and presenting a body condition labeling box to a labeling personnel; In response to a labeling confirmation instruction of the body condition labeling box, performing movement analysis according to body condition labeling data of the body condition labeling box and the historical treatment data, and presenting action change data to the abnormal object; When a current body image corresponding to the action change data is detected, taking the current body image as the initial body image, and returning to perform the step of performing pose analysis on the initial body image and presenting a body condition labeling box to a labeling personnel; Until in response to a labeling stop instruction of the body condition labeling box, performing object function analysis according to each body condition labeling data and the historical treatment data, and obtaining motion function data of the abnormal object.

[0005] In a second aspect, the present application further provides a motion function data acquisition system based on human-computer interaction, comprising a motion function analysis device, a first interaction module and a second interaction module. The functional analysis device responds to the initial position determination instruction of the abnormal object acquired in the first interaction module, performs initial action analysis according to the historical treatment data of the abnormal object, and presents the initial motion data to the abnormal object through the first interaction module; In the case where the initial body image corresponding to the initial motion data is detected, the pose of the initial body image is analyzed, and a body condition labeling box is presented to the labeling personnel through the second interaction module; In response to the labeling confirmation instruction of the body condition labeling box acquired in the second interaction module, the movement analysis is performed according to the body condition labeling data of the body condition labeling box and the historical treatment data, and the action change data is presented to the abnormal object through the first interaction module; In the case where the current body image corresponding to the action change data is detected, the current body image is taken as the initial body image, and the step of performing the pose analysis according to the initial body image and presenting the body condition labeling box to the labeling personnel through the second interaction module is returned; Until in response to the labeling stop instruction of the body condition labeling box, the object function analysis is performed according to each body condition labeling data and the historical treatment data, and the motion function data of the abnormal object is obtained.

[0006] The motion function data acquisition method and system based on human-computer interaction can ensure initial pose recognition of the motion function analysis device and accurate matching of individual positions by triggering an initial position instruction in response to an abnormal object and combining the initial position instruction with historical treatment data to complete start action analysis and visual presentation. Then, when an initial body image corresponding to the start motion is detected, a body condition marking box is given to the marker to introduce human checking and key position positioning, so that the marker can accurately judge the abnormal condition of the abnormal object from a professional perspective. After receiving the marking confirmation, the marking data and the historical treatment data are fused to carry out movement analysis, and indexes such as joint angle, trajectory and stability are quantified, action change data is generated and fed back to the abnormal object in real time, so that the direction and distance of the next examination required by the abnormal object can be accurately calculated. When a new current body image is detected, it is set as a new initial image, and the cycle of "pose analysis-marking-movement analysis-feedback" is repeated, so that the abnormal object can continuously move in the direction of the determined motion function. After the marking stops, the object function comprehensive analysis is carried out on the marking data and the historical data of each round, and the traceable and quantifiable motion function data are output. Not only can the accuracy of understanding the current abnormal information of the abnormal person be effectively improved, but also the objectivity and stability of the motion evaluation result can be improved, the scientificity and efficiency of the rehabilitation scheme can be significantly improved, the burden of medical marking and review can be reduced, the patient participation and compliance can be enhanced due to the visual and understandable instant feedback, and the overall motion function of the abnormal object can be accurately quantified and individualized optimized. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0008] Figure 1 An application environment diagram of the motion function data acquisition system based on human-computer interaction in an embodiment; Figure 2 A flowchart of the motion function data acquisition method based on human-computer interaction in an embodiment; Figure 3 An internal structure diagram of the motion function analysis device in an embodiment. DETAILED DESCRIPTION

[0009] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0010] The method for acquiring motion function data based on human-computer interaction provided by the embodiments of the present application can be applied to a system environment as shown in Figure 1 . The first interaction module 102 and the second interaction module 103 communicate with the motion function analysis device 104 through a network. The data storage system can store data required to be processed by the motion function analysis device 104. The data storage system can be integrated on the motion function analysis device 104, or can be placed on the cloud or other network motion function analysis device. The motion function analysis device 104 can be realized by an independent motion function analysis device or a motion function analysis device cluster composed of multiple motion function analysis devices.

[0011] In an exemplary embodiment, as shown in Figure 2 , a method for acquiring motion function data based on human-computer interaction is provided. Taking the motion function analysis device in Figure 1 as an example, the method includes the following steps 202 to 210. Wherein:

[0012] Step 202, in response to the initial position determination instruction of the abnormal object, starting action analysis is performed according to the historical treatment data of the abnormal object of the motion function analysis device, and starting motion data is presented to the abnormal object of the motion function analysis device.

[0013] Step 204, in the case that the initial body image corresponding to the starting motion data of the motion function analysis device is detected, pose analysis is performed on the initial body image of the motion function analysis device, and a body condition labeling box is presented to the labeling personnel.

[0014] Step 206, in response to the labeling confirmation instruction of the body condition labeling box of the motion function analysis device, movement analysis is performed according to the body condition labeling data of the body condition labeling box of the motion function analysis device and the historical treatment data of the motion function analysis device, and action change data is presented to the abnormal object of the motion function analysis device.

[0015] Step 208, in the case that the current body image corresponding to the action change data of the motion function analysis device is detected, the current body image of the motion function analysis device is taken as the initial body image of the motion function analysis device, and the step of performing pose analysis on the initial body image of the motion function analysis device to present a body condition labeling box to the labeling personnel is returned.

[0016] Step 210, until the response to the motion function analysis device body condition annotation frame annotation stop instruction, according to each motion function analysis device body condition annotation data and motion function analysis device historical treatment data, the object function analysis is carried out, and the motion function data of the abnormal object of the motion function analysis device is obtained.

[0017] Wherein, the abnormal object is the target individual (patient or subject) receiving examination and evaluation.

[0018] Wherein, the initial position determination instruction is a control instruction for the abnormal object to start the current examination process.

[0019] Wherein, the historical treatment data is the past treatment, evaluation and follow-up record related to the abnormal object, including scale score, image / motion data, prescription and contraindication information.

[0020] Wherein, the starting action analysis is a strategy calculation process based on historical treatment data and the first evaluation action of the current environment constraint selection safety and maximum information.

[0021] Wherein, the starting motion data is a set of starting action instructions presented to the abnormal object, including target parts (covering key functions of core dimensions of motor function assessment, including: Reflex related parts: biceps tendon, triceps tendon, Achilles tendon, knee tendon; Upper limb joints and accessory parts: shoulder joint (including upper lifting, retraction, abduction, external rotation functional points), elbow joint (flexion functional points), forearm (pronation / supination functional points); Wrist core part: wrist joint (including dorsiflexion, flexion functional points); Hand function part: fingers (including metacarpophalangeal joint, proximal and distal interphalangeal joint), thumb (adduction functional point); Lower limb joints and accessory parts: hip joint (including flexion, extension, adduction functional points), knee joint (including flexion, extension functional points), ankle joint (including dorsiflexion, plantar flexion functional points), direction / amplitude / rhythm / holding time and safety boundary parameters.

[0022] The above target parts are strictly matched with the core observation dimensions of motor function assessment, which ensures that the starting action can accurately cover the key motor functions such as reflex activity, synergistic movement, separation movement and stability, and the specified objective evaluation index, including target parts, direction / amplitude / rhythm / holding time and safety boundary parameters, lays a foundation consistent with the evaluation system for subsequent pose analysis and function data acquisition.

[0023] Wherein, the initial body image is an image or short sequence collected in the time window corresponding to the starting motion data for the first pose analysis.

[0024] wherein, pose analysis is the process of keypoint detection, skeleton reconstruction and joint parameter estimation on initial body image with uncertainty output.

[0025] wherein, annotator is the professional (e.g. doctor / therapist) who confirms, corrects or supplements the output of the motion function analysis device through the annotation second interaction module.

[0026] wherein, body condition annotation box is a parameterized annotation box or deformable envelope superimposed on the image to define the target area and present the body condition according to the pose analysis.

[0027] wherein, annotation confirmation instruction is the interactive instruction of the annotator to confirm / submit the body condition annotation box and its content through the annotation second interaction module.

[0028] wherein, body condition annotation data is structured data composed of body condition annotation box and its associated labels, including the position / shape of the box, timestamp, confidence and explanatory annotation.

[0029] wherein, movement analysis is the analysis process of motion estimation on continuous images within the annotation box defined area combined with individualized priori evaluation of deviation and trend.

[0030] wherein, action change data is the next step action adjustment suggestion for abnormal objects based on movement analysis, including direction, amplitude, speed, rhythm and alternative action, etc.

[0031] wherein, current body image is the image or short sequence collected within the corresponding time window after the abnormal object performs according to the action change data, which is used for the next round of analysis.

[0032] wherein, annotation stop instruction is the interactive instruction to terminate the current cycle and enter the overall function analysis stage.

[0033] wherein, object function analysis is the analysis process of generating function representation after time sequence alignment, deformation invariant calculation and index normalization on multiple rounds of annotation and motion data.

[0034] wherein, motion function data is the final output of object function analysis, including structured results of each function index value, confidence interval and standardized score.

[0035] Specifically, the motion function analysis device responds to the initial position determination instruction obtained by the abnormal subject operating the first interaction module, first reads its historical treatment data (including past scores, joint range of motion ROM, forbidden actions, dominant side, fatigue threshold and recent symptoms), generates individualized starting action priori. Under this priori constraint, combined with the self-check of the environment at that time (device field of view / lighting / station stability / occlusion), a candidate starting action set is generated, and the safety risk score (falling / pain / forbidden conflict) and information benefit (uncertainty reduction of target function dimension) are calculated for each candidate. Optimal action is selected by risk-constrained information maximization criterion. Then the selected action is organized as "starting motion data" and delivered, including: target joints and motion direction / amplitude / rhythm / holding time, visual demonstration (skeleton / trajectory / safety zone), allowable error threshold and safety boundary, expected duration and early stop condition. The terminal presents in the first interaction module synchronously with visual superposition + voice password + (optional) tactile prompt, and generates the session identifier and time window of this cycle.

[0036] The motion function analysis device detects the initial body image within the session identifier and time window set by the starting motion data after identifying it. First, quality and synchronization verification (frame consistency / occlusion / blur / lighting / timestamp consistency) is performed, and if it is qualified, the target part and safety zone in the starting motion data are used to automatically crop the interest region of the trunk / related joints and complete the camera internal and external parameter correction. In the interest region, the skeleton key point and confidence are calculated by the key point heat map network, the hierarchical consistency check and joint angle, barycenter position, etc. Basic index calculation are performed combining monocular depth / geometric priori, and the uncertainty estimate of each key point is output. On this basis, the minimum risk external envelope (or topologically constrained deformable envelope) of the key point set of the trunk and related joints is generated according to the risk-constrained principle, forming the body condition analysis data with timestamp and confidence label, and the corresponding body condition annotation box is generated using the body condition analysis data and presented in the second interaction module (including automatic prompt and adjustable control points). If the quality verification fails, the repositioning / light supplement / repositioning prompt is given and the annotation box is not generated, and in the privacy mode only the skeleton frame and mask contour are presented without displaying the original pixels.

[0037] The motion function analysis device first locks the target region image sequence with the annotation box and its timestamp and completes quality verification (out-of-frame / occlusion / blur / synchronization consistency) in response to the annotation confirmation instruction from the second interaction module. When qualified, it performs dense optical flow and / or key point tracking on consecutive frames within the annotation box or the entire region according to the bone topology, obtains the local motion field, and extracts the observed joint trajectory (displacement, velocity, angular velocity). Subsequently, the motion function analysis device calls the individualized reference motion band formed by the historical treatment data (phase / amplitude aligned with this session), compares the observed joint trajectory with the reference motion band, calculates the out-of-bound deviation and direction (insufficient / excessive / time lag), and generates action change data within the safety threshold to guide the abnormal object to move. It includes the direction / amplitude / speed / rhythm / holding time that needs to be adjusted and (if necessary) alternative actions and pause prompts, accompanied by synchronized presentation of visual arrows and voice commands; if the quality verification fails or the risk assessment exceeds the threshold, output a degraded instruction (reduce amplitude, slow down, change to a sitting position, etc.) and request re-sampling. The motion function analysis device simultaneously writes the motion field summary, deviation index, and action change data of this round into the session buffer for strategy decision and audit trace. Among them, the action change data is the same as the motion start data, which is a set of subsequent action instructions presented to the abnormal object, including target parts (covering key functions of core dimensions of motor function assessment, including: reflex-related parts: biceps tendon, triceps tendon, Achilles tendon, knee tendon; upper limb joints and accessory parts: shoulder joint (including upper lift, posterior shrink, abduction, external rotation functional points), elbow joint (flexion functional point), forearm (pronation / supination functional point); wrist core part: wrist joint (including dorsiflexion, flexion functional point); hand function part: fingers (including metacarpophalangeal joint, proximal and distal interphalangeal joint), thumb (adduction functional point); lower limb joints and accessory parts: hip joint (including flexion, extension, adduction functional points), knee joint (including flexion, extension functional points), ankle joint (including dorsiflexion, plantar flexion functional points), direction / amplitude / rhythm / holding time and safety boundary, etc. However, the action change data can be determined from the motion start data by fixed movement guidance for the abnormal object, or determined by analyzing the motion start data for the abnormal object. Similarly, the subsequent action change data can be determined from the previous action change data by fixed movement guidance for the abnormal object, or determined by analyzing the previous action change data for the abnormal object.

[0038] The motion function analysis device detects the execution time window of the action change data to capture the corresponding current body image (frame / short sequence), first with session ID and timestamp to complete matching verification, and then quality and synchronization check (out of frame / occlusion / fog / exposure abnormality / frame rate drift). When the verification is qualified, the fast pose coarse solution (lightweight key point reasoning or incremental optical flow) is performed under the consistent ROI and calibration parameters of the previous round, and the execution degree and risk indicators (completion degree, remaining deviation, fall / pain risk) are calculated. If the completion degree is greater than or equal to the threshold and the risk is controllable, the current body image is registered as a new initial body image, and the loop state is updated synchronously, including the target ROI, camera calibration cache, individualized priori, and strategy memory (information gain accumulation, fatigue counter, early stop criterion). Return to step two to continue pose analysis, if any verification is not qualified or the risk exceeds the threshold, enter the error correction path (prompt repositioning / lighting / re-sampling, or automatically downgrade to low amplitude, low speed, alternative posture), and record the failure reason and sensor state for audit. If consecutive failures are greater than or equal to N times or the early stop / fatigue threshold is triggered, the current action is suspended and the label termination or low-intensity evaluation is requested. In privacy mode, only desensitized skeletons / features and timestamps are cached to ensure loop traceability.

[0039] The motion function analysis device receives a label stop instruction from the labeler or meets the early stop condition (information gain / uncertainty / fatigue threshold) in the second interaction module, and summarizes the body condition label data of each round in the current session and the historical treatment data, and performs de-duplication and quality screening according to the session ID and timestamp. Based on DTW / phase anchor, the key posture / joint angle-time sequence is phase-aligned and time length-normalized, and the individual body proportion, ROM upper limit and other priori are used for amplitude standardization to form a function trajectory set. Then the joint angle range (ROM), left-right symmetry, coordination / phase consistency, stability / center of gravity fluctuation, rhythm and tremor spectrum are calculated in the relative configuration, and the point estimate+confidence interval is obtained by uncertainty propagation. The above indexes are mapped to the standardized score (according to the age / gender / disease reference interval and hospital scale calibration) to generate the motion function data and conclusion label (normal / mildly restricted / severely restricted, etc.), and a structured report (including key frames, index curves and criterion explanation) is output simultaneously and written into the audit log; In privacy mode, only desensitized skeletons and derived features are saved, and original pixels are not exported.

[0040] In the motion function data acquisition method based on human-computer interaction, the initial position trigger instruction of the abnormal object is responded, the initial action analysis and visual presentation are completed combined with the historical treatment data of the abnormal object, the initial pose recognition of the motion function analysis device is ensured, and the individual position is accurately matched. Then, when the initial body image corresponding to the initial motion is detected, the body condition labeling box is given to the labeling personnel to introduce the human checking and key part positioning, the abnormal condition of the abnormal object is accurately judged from the professional angle of the labeling personnel. After receiving the labeling confirmation, the moving analysis is carried out by fusing the labeling data and the historical treatment data, the indexes such as joint angle, trajectory and stability are quantified, the action change data is generated and fed back to the abnormal object in real time, so that the direction and distance of the next examination required by the abnormal object are accurately calculated. When a new current body image is detected, the new initial image is set, and the cycle of “pose analysis-labeling-moving analysis-feedback” is repeated, so that the abnormal object can continuously move in the direction of the required motion function. When the labeling stops, the object function comprehensive analysis is carried out on the labeling data and the historical data of each round, and the traceable and quantifiable motion function data are output. Not only the accuracy of understanding the current abnormal information of the abnormal personnel can be effectively improved, but also the objectivity and stability of the motion evaluation result can be improved, the scientificity and efficiency of the rehabilitation scheme making and adjustment can be significantly improved, the burden of medical care labeling and review can be reduced, the participation and compliance of patients can be enhanced due to the visual and understandable instant feedback, and the overall motion function of the abnormal object is accurately quantified and individualized optimized.

[0041] In an exemplary embodiment, the motion function analysis device performs pose analysis on the initial body image of the motion function analysis device, presents a body condition labeling box to the labeling personnel, including steps 302 to 306. Among them:

[0042] Step 302, the initial body image of the motion function analysis device is analyzed, the body segmentation data of the motion function analysis device is presented to the abnormal object of the motion function analysis device.

[0043] Step 304, in response to the optimization operation instruction of the abnormal object of the motion function analysis device to the body segmentation data of the motion function analysis device, the image acquisition optimization of the motion function analysis device is controlled, the segmentation data annotation box and the optimized segmentation data of the motion function analysis device are presented to the abnormal object of the motion function analysis device.

[0044] Step 306, in response to the segmentation data auxiliary instruction of the segmentation data annotation box of the motion function analysis device, the pose analysis is performed on the optimized segmentation data of the motion function analysis device, and the body condition labeling box of the motion function analysis device is presented to the labeling personnel of the motion function analysis device.

[0045] Wherein, the body segmentation analysis is a process of performing human semantic segmentation and part detection on the initial body image to generate the mask, boundary and confidence of each anatomical segment.

[0046] Wherein, the body segmentation data is the structured data obtained from the segmentation analysis, including segment ID, mask / boundary coordinates, confidence, timestamp and session identification, etc.

[0047] Wherein, the optimization operation instruction is an interactive instruction issued by the abnormal object based on the presented segmentation result to improve the collection conditions such as perspective, lighting and station position, etc.

[0048] Wherein, the image collection optimization is a process of automatically adjusting the camera perspective, exposure / white balance, light compensation and posture guidance, etc. to obtain higher quality images according to the optimization operation instruction.

[0049] Wherein, the segmentation data annotation box is a labeling box / spline envelope that is automatically generated around the segmentation boundary, can be dragged and edited, and is used for fine-tuning and confirming the segmentation data.

[0050] Wherein, the optimized segmentation data is the segmentation structured data obtained by reanalyzing after the image collection optimization and annotation, reflecting the latest boundary and confidence.

[0051] Wherein, the segmentation data auxiliary instruction is auxiliary data or instruction for adjusting the existing problems of the optimized segmentation data in the segmentation data annotation box to optimize the pose analysis.

[0052] Specifically, the motion function analysis device performs synchronization and quality check (out of frame, occlusion, blur, exposure / white balance abnormality) on the initial body image. When the check is qualified, the initial body image is pixel-level divided and boundary refined based on the human semantic segmentation and part detection model in the whole frame range, generating segmentation mask, boundary polyline and confidence of head and neck, trunk, upper limb, lower limb and key joint neighborhood, and summarizing to form body segmentation data (including segment ID, coverage ratio, boundary coordinates, confidence and "need to optimize" prompt mark) with session identification and timestamp. Then the body segmentation data is presented in the first interaction module with the abnormal object in a semi-transparent superposition and brief text prompt manner, as the input basis for subsequent optimization operation instruction and pose analysis; if the quality check fails, the patient is prompted to reposition, compensate light or re-collect, and the segmentation data is not generated.

[0053] The motion function analysis device detects abnormal objects and issues optimization operation instructions (touch / voice / gesture, such as "too dark", "a little to the left", "closer") for the presented body segmentation data, and generates a collection optimization scheme according to the instruction type and the current image quality indicators (field coverage, occlusion rate, clarity, exposure / white balance), and executes in the order of "priority camera, then posture, and then environment", wherein the camera reframes (pan / zoom / tilt), adaptive exposure and white balance, local fill light / noise reduction / HDR synthesis; if the threshold is still not reached, give posture guidance (screen arrows / voice prompts to fine-tune the position and orientation). After completing the optimized collection, the motion function analysis device re-executes the body segmentation analysis based on the new frame and generates optimized segmentation data (updated segmentation mask, boundary and confidence, timestamp), and simultaneously combines the specific form of the optimized segmentation data to present the segmentation data annotation box in the first interaction module. If any key indicator of the optimized segmentation data still does not meet the standard, feedback the specific correction prompt (such as "half step to the right" "raise the camera by 5°") and allow the collection optimization to be triggered again.

[0054] The motion function analysis device receives segmentation data annotation box issued segmentation data auxiliary instruction, with session ID and timestamp to verify the data and code of the instruction for the optimized segmentation data, and if qualified, it is used as a constraint to combine the optimized segmentation data in the corresponding segmentation region of interest to obtain the skeleton key point and confidence using the key point heat map reasoning, and to perform consistency check combined with the segmentation boundary and anatomical level (trunk-thorax-pelvis), and to calculate the joint angle, barycenter position and its uncertainty. The key point set and confidence interval obtained from the above data generate body condition analysis data, and the body condition analysis data generate body condition annotation box (minimum risk bounding box or topologically constrained deformable bounding box), with timestamp and confidence label, and the annotation box is pushed to the annotation personnel terminal together with the key frame and key point description for confirmation / tuning. If the quality indicators of the pose analysis (out of frame, occlusion, confidence threshold) do not meet the standard, give specific re-collection prompt and return to step two to trigger collection optimization; in privacy mode, only the desensitized skeleton and annotation box outline are presented to the annotation end, and the original pixels are not displayed.

[0055] In this embodiment, by presenting the body segmentation data to the abnormal object, the subject can immediately correct the position and shielding, etc., significantly reducing the invalid frames caused by framing, lighting, background interference, incorrect action, etc.; then based on the optimization operation instruction, the image acquisition optimization such as framing / exposure / lighting is automatically completed by the motion function analysis device, and the segmentation data annotation box and optimized segmentation data are generated, realizing the double-channel quality control of "patient self-correction + device self-adjustment", improving the signal-to-noise ratio and segmentation boundary accuracy from the source, reducing the cost of re-sampling and artificial communication; finally, under the constraints of segmentation prior and annotation, the pose analysis is carried out and only the body condition annotation box is presented to the labeling personnel, which greatly reduces the search space of the solution, improves the key point reliability and interpretability, and forms an auditable timestamp and version link, which can improve data consistency, annotation efficiency and clinical usability.

[0056] In one exemplary embodiment, the motion function analysis device performs pose analysis on the motion function analysis device optimized segmentation data in response to the segmentation data assisted instruction of the motion function analysis device segmentation data annotation box, and presents the motion function analysis device body condition annotation box to the motion function analysis device labeling personnel, including steps 402 to 408. Wherein:

[0057] Step 402, according to the motion function analysis device segmentation data assisted instruction, monocular depth reconstruction is performed on the motion function analysis device optimized segmentation data to obtain body surface point cloud data.

[0058] Step 404, the local curvature consistency data in the motion function analysis device body surface point cloud data is subjected to quasi-conformal normalization mapping to obtain body normalized graph data.

[0059] Step 406, the motion function analysis device body normalized graph data is subjected to key point heat map reasoning to obtain body key point coordinate data.

[0060] Step 408, the relative configuration data in the motion function analysis device body key point coordinate data is subjected to morphing invariant function analysis to obtain motion function analysis device body condition analysis data.

[0061] Wherein, monocular depth reconstruction is a process of estimating pixel depth based on a single / monocular image combined with camera internal participation scale prior to project dense three-dimensional surface.

[0062] Wherein, the body surface point cloud data is a three-dimensional point set formed by depth and pixel coordinates back projection to the camera coordinate system, containing point coordinates, normal and confidence, etc.

[0063] Wherein, the local curvature consistency data is a consistency measure of the principal curvature direction and amplitude in the point cloud neighborhood, which is used as a weight to suppress noise and folding.

[0064] Wherein, the quasi-conformal normalized mapping is a mapping process that parameterizes a three-dimensional body surface to a two-dimensional domain while keeping the angles as much as possible, allowing limited controllable stretching.

[0065] Wherein, the body normalized atlas data is a two-dimensional parameter domain image obtained by quasi-conformal mapping, and its pixel / three-dimensional bidirectional correspondence, confidence mask, and distortion index.

[0066] Wherein, the key point heat map inference is a process of using a deep network to output a probability heat map of each anatomical key point on an image / atlas and obtain key point coordinates after decoding.

[0067] Wherein, the body key point coordinate data is a structured result containing key points in parameter domain / pixel / three-dimensional coordinates, covariance (uncertainty), and visibility markers.

[0068] Wherein, the relative configuration data is an angle, distance ratio, and topological relationship that is independent of rigid transformation after removing translation, rotation, and scale from the key point set.

[0069] Wherein, the deformation invariant functional analysis is an analysis of calculating function indicators (such as ROM, symmetry, and phase consistency) that are robust to slight deformation on the relative configuration and its time series, and propagating uncertainty.

[0070] Wherein, the body condition analysis data is a set of interpretable function indicators and their confidence intervals output by the deformation invariant functional analysis, which is used to generate body condition bounding boxes.

[0071] Specifically, the motion function analysis device receives the segmented data auxiliary instruction, verifies and locks the corresponding optimized segmented data with session ID and timestamp, calls the camera internal participation distortion parameter to complete geometric correction within the given segmented boundary and region of interest, and performs scale disambiguation according to historical treatment data and this time segment proportion / ground height priori. After the above operations, the segmented mask guided edge preserving deep network is used to perform monocular depth inference on the initial body image, and temporal consistency and confidence fusion are performed on adjacent frames to suppress noise and occlusion artifacts, and finally the calibrated depth and pixel coordinates are back projected to the camera coordinate system, combined with normal estimation and outlier rejection, output the body surface point cloud data with three-dimensional coordinates, normal and confidence.

[0072] Denoise and normal consistency processing are performed on the point cloud data of the body surface, and the local curvature consistency weight (the curvature direction and amplitude are more consistent in the neighborhood, and the weight is higher) is calculated according to the neighborhood principal curvature estimation. According to this, stable anatomical anchor points (such as the sternal angle, left and right anterior superior iliac spines) and low curvature lines are selected as boundary / suture constraints. Then, the point cloud is aligned in posture (determine the front / up direction) with the pelvis center→sternum vector, and the surface is quasi-conformally flattened (LSCM / ARAP, etc.) under the guidance of the weight field. The high-consistency area is preferentially maintained in angle, the small-consistency area is allowed to be stretched to a limited extent, and smoothing regularization is applied to the occluded / sparse area to avoid folding. After mapping, the body standardized atlas data is output, which includes the parametric domain coordinates (u, v) of each point, the bidirectional mapping (u, v ↔ 3D) to three dimensions, the confidence mask, and the local distortion indicator (Jacobian / conformal distortion boundary), and the session ID and timestamp are recorded for subsequent key point reasoning and traceable audit. If the distortion or folding is above the threshold, automatically fallback to the local multi-atlas scheme (shoulder / trunk / pelvis block mapping) and merge its confidence domain.

[0073] The motion function analysis device uses the parametric domain coordinates (u, v) provided by the body standardized atlas data, the confidence mask, and the bidirectional mapping (u, v ↔ pixel / three-dimensional) to constrain the parametric domain. First, a multi-scale high-resolution network is used to output a probability heat map of each anatomical key point in the mask. Then, the heat map is weighted corrected by the distortion indicator of the atlas. Then, soft decoding (soft-argmax / NMS) is performed on the key points of each weighted corrected probability heat map to obtain the parametric domain coordinates and the second moment covariance (as uncertainty). The results are back-projected to pixel coordinates and camera three-dimensional coordinates through the inverse mapping of the atlas, and the uncertainty is first propagated to the 2D / 3D space using the Jacobian. Anatomical / motion consistency verification (bone segment length ratio, joint angle boundary, and adjacent frame short window smoothing) is performed on the back-projected results. Low-confidence or constraint-violating points are given visibility markers or MAP corrections. The body key point coordinate data with parametric / pixel / three-dimensional coordinates, covariance, confidence, and visibility markers are output, along with session ID and timestamp.

[0074] According to the visibility and uncertainty threshold, the valid points of the relative configuration data (relative configuration) of the body key point coordinate data (including 2D / 3D coordinates, covariance, visibility) are screened, and Procrustes alignment is performed to eliminate the effects of translation / rotation / scale. Then, the device / size-independent invariant features (joint angle and angular velocity, bone segment length ratio, left-right symmetry, principal axis orientation, phase / correlation indicators, etc.) are calculated on the "relative configuration", and the key point covariance is propagated through the Jacobian to obtain the point estimates and confidence intervals of each indicator, and the structured body condition analysis data is summarized. On this basis, further according to the mean-variance distribution of the part key points and the anatomical topological constraints, the minimum risk envelope / variable envelope with a coverage probability not lower than a set threshold is solved, the body condition analysis data is obtained, and the body condition bounding box (including timestamp and confidence label) for presentation and subsequent interaction is generated, and the analysis data and the basis for generating the box are recorded into the session audit.

[0075] In this embodiment, the body surface point cloud is obtained by monocular depth reconstruction, and the surface is stably flattened under the premise of maintaining the angle structure by local curvature consistency noise suppression and constraint unfolding. Subsequently, key point heat map reasoning and soft decoding are performed in a standard domain with controlled distortion and more easily isolated occlusions, significantly reducing false positives and false negatives and improving key point confidence. After obtaining the key points, relative configuration is carried out for functional analysis, and the uncertainty is propagated from the first order of the key points to the functional indicators, obtaining body condition analysis data that is independent of the device / size and comparable across time and devices. The body condition bounding box is generated in the last step, avoiding search bias and error amplification caused by pre-bounding. It can improve the robustness, interpretability and consistency of posture / function evaluation under the premise of only requiring monocular hardware, reduce resampling and manual intervention, and optimize clinical labeling and decision-making efficiency.

[0076] In one exemplary embodiment, the motion function analysis device performs movement analysis on the motion function analysis device abnormal object according to the body condition annotation data of the motion function analysis device body condition bounding box and the motion function analysis device historical treatment data, and presents action change data to the motion function analysis device abnormal object, including steps 502 to 506. Among them:

[0077] Step 502, according to the motion function analysis device body condition analysis data and the motion function analysis device historical treatment data, individual prior modeling is performed on the action template space of the motion function analysis device abnormal object, and action intention prior data is obtained.

[0078] Step 504, dense optical flow calculation is performed on the target region data in the motion function analysis device body condition annotation data, and local motion field data is obtained.

[0079] Step 506, according to the motion function analysis device action intention prior data and the motion function analysis device local motion field data, the action difference domain of the motion function analysis device abnormal object is analyzed by opportunity constraint fusion, and the action change data is obtained.

[0080] Among them, the action template space is the action distribution and parameter set that can be acted on the abnormal object obtained by aligning the historical treatment data and the population template based on the body condition analysis this time.

[0081] Among them, individualized prior modeling is the process of aligning the historical trajectory of the subject and the population template according to the phase and amplitude, and estimating the mean / covariance and quantile band to generate individual prior.

[0082] Among them, the action intention prior data is a set of prior parameters for comparative evaluation, such as allowed interval, mean / covariance, cooperative constraint and safety threshold, given by joint-time index.

[0083] Among them, the target region data is an image sequence or ROI information defined by the body condition bounding box and time stamped as the input range of subsequent motion estimation.

[0084] Among them, the dense optical flow calculation is a process of estimating the displacement vector field of adjacent frames pixel by pixel on the target region data.

[0085] Among them, the local motion field data is a set of time series quantities such as joint displacement and velocity and their confidence information obtained by projecting / integrating the dense optical flow in the joint neighborhood.

[0086] Among them, the action difference domain is a set of deviations that exceed the allowed band with respect to the observed joint motion and are connected in time-joint.

[0087] Among them, the opportunity constraint fusion analysis is a process of aggregating the conditional value at risk (CVaR) criteria of residual cost under a given confidence / risk threshold to robustly determine the difference and generate the basis for action adjustment.

[0088] Specifically, joint-time trajectories corresponding to the action to be evaluated are filtered from historical treatment data, combined with the body condition analysis data of the current session for data preparation (outlier rejection, noise smoothing) and phase alignment / duration normalization (DTW or rhythm anchor points), and then amplitude normalized according to individual ROM and limb segment proportions. On this standardized basis, an aligned action template space is constructed, and the mean-covariance and quantile bands [L(t), U(t)] are calculated for each joint at each time index, and the cross-joint coordination constraint matrix (such as hip-knee-ankle phase coupling), velocity / acceleration boundary and safety limit (taboo angle, pain threshold, fatigue threshold) are estimated, and if necessary, the population template is introduced for Bayesian update to improve robustness. Finally, the action intention prior data is calculated, which includes the allowed domain and confidence of each joint over time (mean μ, covariance Σ, quantile band), coordination and safety constraints, risk weight, and early stop conditions.

[0089] According to the target region and time window provided by the body condition annotation data, sequence quality inspection is performed (out-of-frame / occlusion / foggy / exposure and timestamp consistency), and after passing the quality inspection, a multi-scale pyramid is constructed on the target region data corresponding to the target region, and dense optical flow estimation is performed according to the energy model of "brightness consistency + smoothing regularization" or learning optical flow network, and consistency verification is performed on the forward / backward optical flow to generate confidence masks and occlusion masks. Then, combined with the segmentation boundary and skeletal topology, boundary conditions are set in the joint neighborhood (link direction priority, local rotation allowed in joint neighborhood, cross-segment normal flux constraint), the initial optical flow is optimized with constraints and robustly smoothed in a short time window, and finally the pixel-level flow field is integrated / projected according to the joint neighborhood to obtain observed joint displacement, velocity and angular velocity and their uncertainties, which are fused with the optical flow tensor, confidence / occlusion mask and timestamp to form local motion field data.

[0090] The observed joint trajectories in the local motion field data are aligned to the action intention prior data (quantile band [L(t), U(t)] / mean μ-covariance Σ / coordination constraint) according to the session ID and timestamp, and the out-of-bound deviation is calculated at each time step within the confidence and occlusion mask to generate a residual cost map (zero within the band, penalty according to hinge or Mahalanobis distance outside the band, and weighted by coordination constraints). Then, the cost is aggregated conditionally in danger (CVaR) on "joint x time" to extract the action difference domain of abnormal objects (spatiotemporal connected components and intensity levels of continuous super-risk). On this basis, action change data for the subject is automatically synthesized according to the direction and amplitude of the difference domain (amplitude Δθ, velocity Δω, rhythm / phase Δτ to be increased / decreased, and holding time), and alternative actions / degradation suggestions and early stop prompts are given when safety boundaries are triggered, and the difference domain and instruction parameters are written into the session state to drive the next cycle and audit traces.

[0091] In this embodiment, by individualizing prior modeling of the motion template space of the abnormal object based on body condition analysis data and historical treatment data, forming motion intention prior data containing quantile bands and collaborative constraints, the subsequent judgment is based on the feasible motion distribution of the patient himself, avoiding misjudgment caused by general threshold; then performing dense optical flow calculation in the target region defined by the body condition annotation data and projecting it as joint time series, obtaining high-resolution, position-accurate and more robust local motion field data regardless of occlusion, reducing background interference and invalid frames; finally, the risk perception fusion analysis of prior and observation residuals is performed by opportunity constraints (such as CVaR), and the motion difference domain is robustly extracted and quantitative, executable motion change data (direction / amplitude / speed / rhythm / alternative suggestions) are generated under the premise of ensuring given confidence / false alarm rate, so as to realize real-time adaptive correction of abnormal objects, improve sensitivity and specificity, reduce repeated collection and manual parameter adjustment costs, and enhance process explainability and auditability.

[0092] In an exemplary embodiment, the motion function analysis device performs dense optical flow calculation on the target region data in the motion function analysis device body condition annotation data to obtain local motion field data, including steps 602 to 610. Among them:

[0093] Step 602, boundary flow initialization processing is performed on the key joint neighborhood data in the target region data of the motion function analysis device to obtain joint boundary condition field data.

[0094] Step 604, Laplace multi-scale pyramid calculation is performed on the target region data of the motion function analysis device to obtain multi-scale pyramid image data.

[0095] Step 606, according to the joint boundary condition field data of the motion function analysis device, dense optical flow calculation is performed on the multi-scale pyramid image data of the motion function analysis device to obtain initial optical flow field data.

[0096] Step 608, the divergence component data and the vorticity component data in the initial optical flow field data of the motion function analysis device are projected by the skeletal topology Hodge to obtain topologically consistent optical flow field data.

[0097] Step 610, forward and backward consistency verification is performed on the topologically consistent optical flow field data of the motion function analysis device to obtain local motion field data of the motion function analysis device.

[0098] Among them, the key joint neighborhood data is a local pixel / voxel set and its mask, scale and main direction information determined adaptively around the key joint center according to the bone segment length and uncertainty.

[0099] The boundary flow initialization processing is an initialization step of setting a tangential initial flow, a normal flux weight, and a finite rotation permissibility at a joint based on a joint neighborhood and a segmented boundary.

[0100] The joint boundary condition field data is field data formed by rasterizing the tangential reference vector, the normal flux weight, and the rotation permissibility mask and attaching a time stamp and a confidence level.

[0101] The Laplacian multi-scale pyramid calculation is a processing of constructing a Gaussian pyramid and then performing inter-layer difference to obtain a Laplacian representation of each scale on the target region sequence.

[0102] The multi-scale pyramid image data is a pyramid data set composed of images, gradients, masks, and aligned boundary condition fields of each layer from coarse to fine.

[0103] The initial optical flow field data is a pixel-wise displacement vector field estimated under multi-scale constraints, as well as a residual map, a confidence / occlusion mask, and a time identifier.

[0104] The flow divergence component data is scalar component data calculated from the optical flow field and representing local convergence or divergence intensity.

[0105] The rotation component data is scalar component data calculated from the optical flow field and representing local rotation intensity and direction.

[0106] The skeletal topology Hodge projection is a processing of orthogonally decomposing the optical flow into irrotational, non-divergent, and necessary harmonic components under the constraint of skeletal link / joint / segmented boundary and recombining them.

[0107] The topologically consistent optical flow field data is optical flow vector field data that satisfies anatomical flux and rotation constraints after being corrected by the skeletal topology Hodge projection and updates the confidence level.

[0108] The forward and backward consistency check is a bidirectional consistency check of pixel reliability and occlusion based on the reprojection error and photometric residual of the forward and reverse optical flow.

[0109] Specifically, after frame sequence operation and timestamp verification on target region data, according to the key joint center, segmented boundary and bone segment connection relationship given by the body condition annotation data, an adaptive neighborhood is constructed for each joint (the radius is scaled with the bone segment length and uncertainty, and an anisotropic elliptical kernel along the bone segment axis is used). The bone segment tangent vector and cross-segment normal direction are calculated in any neighborhood, and the initial flow amplitude of the tangent is set based on the joint speed of the previous frame (zero or motion prior prediction if there is no history), while the cross-segment normal is set to zero / weak flux weight, and a limited vorticity permission mask is set for the joint minimum radius range. The above tangent reference vector, normal flux weight, vorticity permission mask and confidence are rasterized to a field with the same size as the image, and are output as joint boundary condition field data together with session ID and timestamp.

[0110] After photometric normalization and denoising (histogram matching / time domain equalization, light bilateral filtering) on the target region data, a Gaussian pyramid is constructed layer by layer with a scale factor of 1 / 2 (each layer Gaussian low-pass→downsampling), and a corresponding Laplacian pyramid is formed by the difference between adjacent layers (current layer original image - upsampled next layer Gaussian image), and the joint boundary condition field data and the confidence / occlusion mask generated by quality inspection are downsampled and interpolated at the same proportion to realize layer-by-layer alignment. At each scale layer, the image, gradient (∂x / ∂y), boundary condition (tangent reference, normal flux weight, vorticity permission mask) and mask weight are recorded, and the session ID and timestamp are attached, and finally encapsulated as multi-scale pyramid image data containing multi-layer data blocks from coarse to fine.

[0111] According to the joint boundary condition field data, the multi-scale pyramid image data is iterated from the coarsest layer of the self-pyramid to the finest layer, that is, at each layer of the multi-scale pyramid image data, the flow or zero field of the previous layer is first initialized, an energy function containing data items with consistent brightness / gradient and anisotropic total variation smoothing items is constructed, and the boundary conditions are explicitly incorporated (add prior items along the tangent of the bone segment, apply flux suppression across the segmented normal, and add vorticity relaxation penalty in the joint circle domain). Then perform resampling (warping) - linearization - solving cycle (such as TV-L1 / Gauss-Newton), after each round of update, do bilateral edge-preserving regularization and sub-pixel refinement at the current layer, get the hierarchical flow field and upsample it to a finer layer. In the finest layer, the forward and reverse directions are solved respectively and the joint residual error is evaluated, and the confidence / occlusion mask and residual error map are generated, and finally the pixel-by-pixel displacement vector, confidence / occlusion mask, residual error map are output as initial optical flow field data together with session ID and timestamp.

[0112] After computing the flow change of each pixel in the confidence and occlusion mask of the initial optical flow field, two types of description are made: one represents the degree of "convergence or divergence" (divergence), and the other represents the degree of "rotation" (rotation); and the skeletal topology is established. According to the established skeletal topology of the abnormal object, the image is divided into three types of regions: link regions, joint circle domains, and cross-segment boundaries, and anatomically consistent boundary conditions are set, wherein the link region is preferentially non-divergent, the joint circle domain allows limited rotation, and the cross-segment boundary is subjected to zero normal flux constraint. Under the above conditions, two scalar potential fields are solved to obtain the corresponding gradient flow and rotational flow (i.e. solving two types of Poisson equations to obtain potential functions (gradient potential and flow function)), the original optical flow is orthogonally decomposed into non-rotational and non-divergent parts, and a small amount of harmonic component is added when necessary to meet the global flux or consistent around. Finally, adaptive weighted reorganization is performed according to the region to suppress the cross-segment normal leakage and abnormal rotation outside the joint, forming a topologically consistent optical flow field data that meets the skeletal topology and boundary flux constraint, while outputting the residual divergence / rotation index and the updated confidence.

[0113] The forward and reverse flows of the adjacent frames of the topologically consistent optical flow are calculated, and the pixels of the previous frame are projected to the next frame according to the forward flow and then back projected to the previous frame according to the reverse flow. By comparing the deviation of returning to the original position and the corresponding brightness residual, a consistency confidence mask that only retains the reliable pixels and an occlusion mask that marks the occluded regions are generated. For the regions detected to be inconsistent or occluded, directional repair is performed according to the skeletal topology, i.e. tangential interpolation or extrapolation is used along the bone segment direction to continue the reasonable joint motion, the normal leakage is suppressed and the boundary is clear across the bone segment direction, and a small range of rotational component consistent with the physiological rotation is allowed in the joint circle domain. Subsequently, the obtained optical flow is robustly time-smoothed in a short time window, making it continuous in the time axis and noise-resistant, and further integrating and projecting the flow lines of the optical flow in the neighborhood of each joint, the trajectories of the displacement, velocity and angular velocity of the joints over time are extracted, and the confidence of each trajectory is given according to the forward and backward consistency and the residual history. If the overall consistency is lower than the threshold for a long time or the occlusion ratio is too high, the back sampling prompt or the degradation strategy will be triggered to ensure safety and reliability. Finally, the local motion field data is obtained, which includes the corrected pixel-level optical flow, the confidence and occlusion mask, the time sequence trajectory of each joint and its confidence information, and is accompanied by session identification and time stamp.

[0114] In this embodiment, the tangential / normal boundary flow initialization of the joint neighborhood provides a directional prior for large displacement and rotation at the joint, significantly reducing cross-segment normal leakage and initial uncertainty; the multi-scale pyramid ensures convergence and stability under weak texture, lighting changes, and fast motion conditions; the dense optical flow calculation under boundary conditions naturally fits the initial flow field to the axial and joint rotation mode of the bone segment; the subsequent skeletal topology Hoche projection orthogically corrects the flow field according to the physiological constraints of the connecting rod, joint, and cross-segment boundary, suppressing local components that do not conform to anatomical accessibility; finally, the output local motion field data is spatially and temporally continuous, occlusion-robust, and confidence-quantifiable, not only reducing false positives and false negatives and reducing resampling and manual intervention, but also providing high signal-to-noise ratio, interpretable, and auditable basic data.

[0115] In one exemplary embodiment, the motion function analysis device performs opportunity constraint fusion analysis on the motion difference domain of the abnormal object of the motion function analysis device according to the motion function analysis device action intention prior data and the motion function analysis device local motion field data, and obtains motion change data, including steps 702 to 706. Among them:

[0116] Step 702, according to the motion function analysis device action intention prior data, the joint-time prior trajectory set of the motion function analysis device historical treatment data is modeled by quantile pipeline, and the opportunity constraint prior domain data is obtained.

[0117] Step 704, the joint neighborhood displacement data in the motion function analysis device local motion field data is integrated by flow line, and the observed joint trajectory data is obtained.

[0118] Step 706, the residual cost calculation is performed on the out-of-bound deviation data in the motion function analysis device observed joint trajectory data, and the residual cost data is obtained.

[0119] Step 708, according to the residual cost data of the motion function analysis device, the condition in danger is aggregated for each joint of the abnormal object of the motion function analysis device, and the motion difference domain data of the motion function analysis device is obtained; Step 710, the connected component time series data in the motion difference domain data of the motion function analysis device is statistically summarized, and the motion change data of the motion function analysis device is obtained.

[0120] Among them, the joint-time prior trajectory set is the reference trajectory set of the joint changing with time extracted from the historical treatment data and (optional) population template and aligned by phase and amplitude.

[0121] Among them, the quantile pipeline modeling is a modeling process of estimating upper and lower quantile bands, mean and covariance of the prior trajectory according to "joint x time", and forming a confidence interval evolving with time.

[0122] wherein, the opportunity constraint prior domain data are the allowable value intervals and their risk weights and safety thresholds of each joint at each time instance under a given confidence level.

[0123] wherein, the joint neighborhood displacement data are the pixel-level displacement vectors and their confidence information projected / sampled from the dense optical flow at each joint neighborhood.

[0124] wherein, the streamline integral is the integral process of accumulating displacement along the optical flow vector field over time to generate the joint centroid position, velocity and angular velocity trajectories.

[0125] wherein, the observed joint trajectory data are the sequences of joint position / velocity / angular velocity over time and their confidence labels obtained from the streamline integral and the mask constraint.

[0126] wherein, the out-of-bound deviation data are the deviation samples of the observed joint trajectory at a time instance beyond the allowable band, containing the information of deviation magnitude and direction.

[0127] wherein, the residual cost calculation is the process of converting the out-of-bound deviation into a uniform scale of cost value (with possible collaborative and safety penalties) according to its magnitude, direction and confidence.

[0128] wherein, the residual cost data are the cost value matrices (with direction labels and confidence intervals) organized by “joint x time”, used for subsequent risk aggregation.

[0129] wherein, the conditional value-at-risk aggregation is the process of aggregating only the tail worst deviations with conditional value-at-risk as the criterion, and risk-aware summarizing the residual cost in joint and time dimensions.

[0130] wherein, the action difference domain data are the continuous out-of-bound spatio-temporal segments and their involved joints, severities and confidences obtained from the risk aggregation and connectivity determination.

[0131] wherein, the connected component time series data are the continuous segment sequences and their start and end times and statistical features obtained from the decomposition of the action difference domain along the time axis.

[0132] Specifically, with the action intent prior data of the current session as a constraint, the joint-time prior trajectory set matching the to-be-evaluated action is screened out from the historical treatment data, and data preparation (missing data completion, denoising, and outlier rejection), phase alignment / time length normalization (DTW or rhythm anchor point), and amplitude normalization (scaling according to individual ROM and limb segment proportion) are sequentially completed, and weighted fusion is performed between individual and population sources (individual priority, and introduction of population template robustness when there are few samples). Subsequently, quantile pipeline modeling is performed according to the "joint x time" index, and the lower / middle / upper quantile bands and covariance at each time are output, and the upper limit of velocity / acceleration and the inter-joint coordination threshold are generated, and safety boundaries such as taboo angles, pain / fatigue thresholds are superimposed, and finally an opportunity constraint prior domain data (including allowable interval, risk weight, and early stop condition) is formed with a set confidence, accompanied by version number and time stamp.

[0133] Frame sequence alignment is performed on the local motion field data according to the session ID and timestamp, and the seed points and neighborhood sampling points of each joint are selected under the constraints of joint neighborhood mask and confidence / occlusion mask. Streamline integration and re-projection are performed along the pixel-level optical flow on the time axis, and the inter-frame displacement is accumulated into the joint centroid trajectory, and the velocity, angular velocity and their expressions in the camera / world coordinates are calculated synchronously. Kalman / spline is used for robust interpolation and short window smoothing for occluded or inconsistent sections, and adaptive restart is performed as necessary according to the last valid state. After completion, confidence and missing markers (derived from forward and backward consistency and residual history) are generated for each trajectory, and structured observed joint trajectory data (joint ID, position / velocity / angular velocity sequence over time, mask and confidence interval, version number and timestamp) are output for subsequent opportunity constraint comparison.

[0134] The opportunity constraint prior domain data and observed joint trajectory data are aligned by session ID and timestamp for each joint and each time; samples within the prior allowable band (position / velocity / acceleration and coordination threshold) are recorded as zero; out-of-band deviations are generated according to the deviation amplitude, deviation direction and observation confidence, and the residual cost is calculated after scaling the dimensions to a uniform scale (position and velocity components are standardized respectively). For time points with coordination violations (such as hip-knee phase coupling mismatch) and safety boundary triggers (taboo angles, fatigue thresholds), additional penalty weights are added, and short-time isolated out-of-boundaries are robustly filtered out through time neighborhood to suppress burrs. Finally, the residual cost data organized according to "joint x time" (including cost value, direction label, confidence interval and mask, version number and timestamp) is output.

[0135] According to the residual cost data, the risk confidence level and the minimum duration are set, and the cost of the "worst tail" is preferentially aggregated on the time axis of each joint (only samples higher than the quantile threshold are counted), to obtain a conditional at-risk risk score reflecting the severity. Connectivity and persistence are determined in the joint-time plane, that is, continuous over-risk time periods are merged, sporadic burrs are suppressed, and mutually coupled joint over-risk segments within the same period are merged into a single abnormal segment according to the cross-joint coordination relationship. The start and end time, involved joint set, main deviation direction (amplitude / speed / phase), risk level and confidence of each segment are calculated, and if the safety boundary (taboo angle / fatigue threshold) is triggered, the level is raised and the emergency attribute is marked. Finally, structured motion difference domain data (including segment metadata, risk score and timestamp / version number) is obtained.

[0136] The motion difference domain data is first de-merged and sequentially arranged according to the time axis, the main deviation direction (amplitude / speed / phase), peak and median deviation, duration and involved joint set of each component are calculated, and the processing priority is determined in combination with the risk level and safety boundary marker. The deviation elements are mapped to executable adjustment parameters for the subject (such as amplitude increment / decrement Δθ, speed slow down / increase Δω, advance / lag rhythm Δτ, adjustment holding time, or alternative action / pause suggestion when the taboo / fatigue threshold is triggered), and numerical targets, allowable errors and application time windows are added to each instruction according to the confidence and historical response effect. Structured motion change data (adjustment list, priority and safety prompt for each joint / period) is generated, supporting the synchronous presentation of screen arrows / words and voice commands in two channels, and the version number and timestamp are recorded for the next cycle and audit traceability.

[0137] In this embodiment, the joint-time prior trajectory set is first compressed into opportunity constraint prior domain data by quantile pipeline modeling, so that the judgment reference fits the patient's own ability boundary, avoiding misjudgment caused by general threshold. Then the observed joint trajectory data is obtained from the local motion field by streamline integration, which can still obtain continuous and physically consistent joint timing under pixel layer noise and occlusion conditions. Subsequently, the out-of-bound samples are converted into residual cost data, and the conditional at-risk aggregation only focuses on the worst tail deviation, robustly suppressing short burrs and incidental noise, forming motion difference domain data with severity and confidence markers. Finally, the connected components are time-series statistically summarized to output motion change data (executable adjustments of direction, amplitude, speed, rhythm and duration), so as to realize real-time adaptive correction under the premise of ensuring safety boundary and false alarm rate, significantly improve sensitivity and specificity, reduce re-sampling and manual parameter adjustment costs, and provide full-link traceable audit and clinical interpretability.

[0138] In an exemplary embodiment, the motion function analysis device obtains motion function data of an abnormal subject of the motion function analysis device according to the body condition annotation data of the motion function analysis device and the historical treatment data of the motion function analysis device, including steps 802 to 806. Wherein:

[0139] Step 802, according to the historical treatment data of the motion function analysis device, the key posture sequence data in the body condition annotation data of the motion function analysis device is phase aligned to obtain function trajectory data.

[0140] Step 804, the function trajectory data is analyzed by persistent homology topology to obtain deformation invariant function topology summary data.

[0141] Step 806, the index vector data in the deformation invariant function topology summary data of the motion function analysis device is scored by scale normalization to obtain motion function data of the motion function analysis device.

[0142] Wherein, the key posture sequence data is the continuous value sequence of the key joint / posture on the time axis extracted from the body condition annotation data, containing time stamp and confidence.

[0143] Wherein, the phase alignment is to align multiple action sequences to the same motion phase and unify the length by rhythm anchor point or dynamic time warping.

[0144] Wherein, the function trajectory data is a standardized joint position / angle / speed sequence set with time after phase alignment and amplitude (body proportion / ROM) normalization.

[0145] Wherein, the persistent homology topology analysis is to embed the function trajectory into phase space, construct filter and calculate persistent barcode / persistent graph to extract stable topological features.

[0146] Wherein, the deformation invariant function topology summary data is a feature summary that is robust to translation / rotation / scale and time reparameterization, which is obtained by vectorizing / image coding the persistent homology results and metric alignment.

[0147] Wherein, the index vector data is a fixed-length numerical feature vector (such as persistence distribution, landscape energy, topological stability, etc.) selected and summarized from the topology summary for scoring and comparison.

[0148] Wherein, the scale normalization scoring is a process of standardizing the index vector according to hierarchical reference (age / sex / height / weight / disease, etc.) and mapping to clinical scale or percentile score.

[0149] Specifically, the reference rhythm of the current action is determined with historical treatment data and the template, and the key posture / joint angle-time sequence in each round of body condition labeling data is taken as the alignment object. After quality preparation (timestamp verification, denoising, missing data completion, and outlier removal), the rhythm anchor point is selected according to the action type or dynamic time warping (DTW) is used to complete phase alignment and time length unification. The amplitude is normalized according to the individual body proportion and the upper limit of the current ROM, and the left and right sides are mirrored / symmetrized. The aligned index time sequence is output as functional trajectory data in the form of "index ID-time-value (including confidence)".

[0150] The functional trajectory data is subjected to data normalization and time synchronization verification, and then a phase space point cloud is constructed in two modes of "single index time delay / multiple index joint". According to the local phase consistency data and noise level, the metric and weight are adaptively set, and then a weighted Vietoris-Rips filter is constructed on the point cloud to calculate the persistent barcode / persistent diagram of each dimension to extract stable ring and cavity structures. In order to facilitate downstream learning and comparison, the persistent result obtained by the above calculation is vectorized (persistent landscape or persistent image), and the optimal transport alignment is used to eliminate the influence of scale change and time reparameterization, while the stability evaluation of noise and sampling disturbance is output, and finally the deformation-invariant functional topology summary data (including main topology features, persistence metric and confidence information) with session identification and timestamp is generated.

[0151] According to the age / gender / height / weight / disease and side in the hospital or population reference library, the corresponding stratified reference is selected, and the scale normalization (unit unification, Z-score / percentile conversion, extreme value truncation and missing data interpolation) is performed on each topology index (such as the persistence of ring, landscape energy, and cooperative stability) in the deformation-invariant functional topology summary data, and the normalized results are mapped according to the calibrated clinical scale and aggregated into sub-dimensions (ROM, symmetry, rhythm / collaboration, stability / tremor, etc.) and total score according to the weight scheme. At the same time, based on the previous uncertainty, the confidence interval of each index is propagated and consistency checked, and when the safety threshold / warning rule is triggered, the risk label and explanation label are added, and finally the structured motor function data (total score and sub-item score, confidence interval, stratified reference interval / percentile, risk prompt and suggestion) is output.

[0152] In this embodiment, the key posture sequence is phase-aligned and time-unified by historical treatment data as anchor points, significantly weakening the interference of individual rhythm differences and sampling jitter; then the persistent homology topology analysis is performed on the functional trajectory, and the topology invariants robust to translation, rotation, scale and time reparameterization are extracted, avoiding misjudgment caused by threshold or single-point features; finally, the topology summary vector is scaled and scored according to the age / sex / height / weight / disease stratification criteria, obtaining motion function data (including confidence and risk prompt) that can be directly compared across devices and time periods, thereby improving sensitivity and specificity, reducing artificial subjectivity and repeated collection, and facilitating long-term follow-up and clinical decision support.

[0153] In an exemplary embodiment, the motion function analysis device performs persistent homology topology analysis on the functional trajectory data to obtain deformation-invariant functional topology summary data, including steps 902 to 910. Among them:

[0154] Step 902, time delay embedding processing is performed on the functional trajectory data of the motion function analysis device to obtain phase space point cloud data.

[0155] Step 904, the local phase consistency data in the phase space point cloud data of the motion function analysis device is adaptively transformed to obtain the coherent stable weight data.

[0156] Step 906, the phase space point cloud data of the motion function analysis device and the coherent stable weight data of the motion function analysis device are analyzed by weighted simplicial complex filtering to obtain persistent barcode data.

[0157] Step 908, the persistent barcode data of the motion function analysis device is encoded by persistent landscape to obtain topology descriptor data; Step 910, the scale and reparameterization sensitive components in the topology descriptor data of the motion function analysis device are optimally transmitted and aligned to obtain the deformation-invariant functional topology summary data of the motion function analysis device.

[0158] Wherein, the time delay embedding processing is the process of reconstructing the single / multi-index time series into high-dimensional state vector sequence in the sliding window according to the selected delay and embedding dimension.

[0159] Wherein, the phase space point cloud data is a high-dimensional state vector set and its time index and meta information formed after time delay embedding.

[0160] Wherein, the local phase consistency data is a set of indicators that measure whether the instantaneous phase of each trajectory is coherent (coherent) in a short neighborhood.

[0161] Wherein, the adaptive metric transformation is a process of dynamically rescaling the point pair distance and neighborhood scale according to the phase consistency, noise and density.

[0162] wherein the coherent stability weight data is a set of parameters for improving the stability of persistent homology results, including the weight given to each point (or neighbor), the neighborhood radius, and the metric correction factor.

[0163] wherein the weighted simplicial complex filtering analysis is a process of constructing weighted Vietoris-Rips simplicial complex family and calculating persistent homology under the increasing of weighted distance and threshold.

[0164] wherein the persistent barcode data is a barcode-like result recording the "birth-death" interval of topological features at filtering scales and its stability.

[0165] wherein the persistent landscape encoding is a mapping of persistent barcode to a function / channel representation on a uniform grid for comparison and vectorization process for downstream learning.

[0166] wherein the topological descriptor data is a fixed-length numerical / rasterized feature derived from persistent landscape (or persistent image) to represent the topological structure of trajectories.

[0167] wherein the scale and reparameterization sensitive component is the part of the topological descriptor that still changes significantly under scale variation or time axis resampling / velocity deformation.

[0168] wherein the optimal transport alignment is an alignment method that minimizes the Wasserstein cost as a criterion to register the descriptor, thus eliminating the scale and time reparameterization differences.

[0169] Specifically, after data preparation (denoising, standardization, timestamp verification, and missing value completion) is performed on the functional trajectory data as input, the delay time is selected by autocorrelation / mutual information, the embedding dimension is determined by the false nearest neighbor criterion, and the phase space vector is reconstructed in a sliding window according to the strategy of single-index delay embedding or multi-index joint embedding. The edge samples and abnormal segments at the beginning and end of the window are removed or robustly interpolated, the mapping relationship between each reconstructed vector and the original time index, sequence source, and confidence label is preserved, and finally the phase space point cloud data with session ID and timestamp is output, which contains point coordinates, corresponding trajectory identification, and local density, etc.

[0170] For each trajectory, the phase consistency and local noise / sampling density indicators (e.g. Hilbert phase, short window coherence, and k-neighbor density) are estimated within a short window of phase space point cloud data, based on which the point-to-point distance and neighborhood scale are adaptively re-scaled. In phase-consistent and densely sampled regions, the distance is compressed and the neighborhood radius is relaxed to preserve details, while in phase-mismatched or noisy regions, the distance is enlarged and the neighborhood is tightened to suppress false structures. Meanwhile, the neighborhood relations across trajectories and phases are penalized to avoid improper edge connections, and if necessary, the abnormal isolated points are down-weighted. Finally, for each point, the homology-stable weight data containing the weight, neighborhood radius, metric correction factor, and confidence label are generated.

[0171] The coordinates and metric correction factors of the phase space point cloud data and the homology-stable weight data are uniformly normalized, and optionally landmark-witness point sparsification is performed according to the weight and local density to control complexity. Based on the re-scaled weighted distance, the neighborhood relations are established and the scale parameter sequence is determined, and the weighted simplicial complex filter is constructed from "coarse to fine" in stages, i.e. at each scale, the edges are connected according to the threshold value and the high-dimensional simplex is generated, while carrying the stability label from the weight field. During the filtering process, the persistent homology is calculated in real time, the birth-death events of each dimension (connected components, loops, cavities) are tracked, and the representative generators and their mapping back to the original trajectory / time index are recorded. Finally, short-lived noise removal and segment merging are performed according to the adaptive threshold of the weight, and the persistent barcode data containing multi-dimensional barcode, stability score, representative simplex, and metadata (session ID, timestamp) are output.

[0172] The persistent barcode data is processed according to the homology dimension respectively, and the specific calculation process is as follows: first, the truncation and denoising threshold is adaptively set according to the lifetime and stability score of the barcode, and the short-lived noise segment is removed and the representative generator is retained; then, in the unified scale interval and fixed sampling grid, each barcode is converted into a persistent landscape channel (each dimension corresponds to a set of landscape layers), and the landscape amplitude is weighted according to the stability and lifetime; the obtained multi-channel landscape is subjected to light smoothing and amplitude normalization to ensure that the outputs of different sequences are fixed-length and comparable numerical representations; finally, each channel is concatenated / grided in order of dimension and level, together with the encoding metadata (scale interval, grid resolution, dimension label, session ID, and timestamp) to output as topological descriptor data.

[0173] The data of topological descriptors is selected to align with a reference (in-hospital benchmark or the best period of the object history), and pre-normalization is completed (uniform scale interval and sampling grid, time axis re-parameterization candidate set). Then, the cost metric between the descriptors is constructed (amplitude and position difference between landscape grid / channels, diagonal / virtual point cost for unmatched components), and the optimal transport alignment is solved under the constraints of entropy regularization and mass conservation to obtain the transport mapping from the target to the reference. According to the mapping, the multi-channel landscape is resampled / reweighted for registration, and a consistent representation robust to scale variation and time re-parameterization is output, with alignment residuals and stability scores given. Finally, a deformation-invariant functional topological summary data (including feature vectors after alignment, residual indicators and alignment metadata) with session ID and timestamp is formed.

[0174] In this embodiment, the action dynamics is reconstructed in phase space by delay embedding to avoid looking at only local fluctuations of the original time series; adaptive metrics based on local phase consistency "calibrate" the adjacency relationship before filtering to suppress false structures and improve the stability of persistent homology; weighted simplicial complex filtering preserves important global structures such as loops / cavities and weakens short-lived noise; persistent landscape encoding normalizes the barcode into a fixed-length vector for easy comparison and downstream learning; optimal transport alignment further eliminates scale and time axis differences caused by different devices / rhythm / speed, making cross-period and cross-scene results directly comparable and interpretable, reducing false positives and false negatives and subjective threshold dependence, and improving the consistency, sensitivity and clinical usability of functional assessment.

[0175] In an exemplary embodiment, a motion function analysis device is provided, the internal structure diagram of which can be as shown in Figure 3 The motion function analysis device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Those skilled in the art can understand that the structure shown in Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the motion function analysis device to which the scheme of the present application is applied. The specific motion function analysis device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0176] In one embodiment, a motion function data acquisition system based on human-computer interaction is also provided, which includes a motion function analysis device, a first interaction module and a second interaction module. The motion function analysis device performs initial motion analysis according to the historical treatment data of the abnormal object in response to the initial position determination instruction of the abnormal object acquired in the first interaction module, and presents the initial motion data to the abnormal object through the first interaction module; In a case where an initial body image corresponding to the initial motion data is detected, pose analysis is performed on the initial body image, and a body condition labeling frame is presented to a labeler through the second interaction module; In response to a labeling confirmation instruction of the body condition labeling frame acquired in the second interaction module, movement analysis is performed according to the body condition labeling data of the body condition labeling frame and the historical treatment data, and motion change data is presented to the abnormal object through the first interaction module; In a case where a current body image corresponding to the motion change data is detected, the current body image is taken as the initial body image, and the step of performing pose analysis according to the initial body image and presenting a body condition labeling frame to a labeler through the second interaction module is returned to perform; Until in response to a labeling stop instruction of the body condition labeling frame, object function analysis is performed according to each body condition labeling data and the historical treatment data, and motion function data of the abnormal object is obtained.

[0177] In one embodiment, the motion function analysis device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0178] In one embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0179] In one embodiment, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. The processor of the motion function analysis device reads the computer instructions from the computer readable storage medium. The processor executes the computer instructions, so that the motion function analysis device performs the steps in the above method embodiments.

[0180] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0181] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware, and the computer program can be stored in a nonvolatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above-mentioned embodiment methods.

[0182] Any combination of the technical features in the above embodiments can be made, and in order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the description.

[0183] The above embodiments only express several implementation manners of the present application, the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for acquiring motion function data based on human-computer interaction, characterized in that, The method, applied to a motion function analysis device, includes: In response to the command to determine the initial position of the abnormal object, the starting motion analysis is performed based on the historical treatment data of the abnormal object, and the starting motion data is presented to the abnormal object. If the initial body image corresponding to the initial motion data is detected, the pose analysis of the initial body image is performed, and a body condition annotation box is presented to the annotator. In response to the annotation confirmation command of the body condition annotation box, a movement analysis is performed based on the body condition annotation data of the body condition annotation box and the historical treatment data, and the movement change data is presented to the abnormal object; If the current body image corresponding to the motion change data is detected, the current body image is used as the initial body image, and the process returns to the step of performing pose analysis based on the initial body image and presenting a body condition annotation box to the annotator. Until the annotation stop command of the body condition annotation box is received, the object function analysis is performed based on the body condition annotation data and the historical treatment data to obtain the motion function data of the abnormal object.

2. The method according to claim 1, characterized in that, The step of performing pose analysis on the initial body image and presenting body condition annotation boxes to the annotators includes: The initial body image is segmented and parsed to present the segmented body data to the abnormal object; In response to the abnormal object’s optimization operation command for the body segment data, the motion function analysis device is controlled to perform image acquisition optimization, and segment data annotation boxes and optimized segment data are presented to the abnormal object. In response to the segmented data auxiliary instruction of the segmented data annotation box, pose analysis is performed on the optimized segmented data, and the body condition annotation box is presented to the annotator.

3. The method according to claim 2, characterized in that, The step of responding to the segmented data auxiliary instruction in response to the segmented data annotation box, performing pose analysis on the optimized segmented data, and presenting the body condition annotation box to the annotator includes: According to the segmented data auxiliary instructions, monocular depth reconstruction is performed on the optimized segmented data to obtain body surface point cloud data; Quasi-conformal normalization mapping is performed on the local curvature consistency data in the body surface point cloud data to obtain normalized body atlas data; The body's standardized atlas data is subjected to key point heatmap reasoning to obtain the coordinate data of key body points. Deformation-invariant functional analysis is performed on the relative configuration data in the body key point coordinate data to obtain body condition analysis data; the body condition analysis data is used to generate the body condition annotation box.

4. The method according to claim 1, characterized in that, The step of performing motion analysis based on the body condition annotation data in the body condition annotation box and the historical treatment data, and presenting motion change data to the abnormal object, includes: Based on the physical condition analysis data and the historical treatment data, an individualized prior model is performed on the action template space of the abnormal object to obtain action intention prior data. Dense optical flow calculations are performed on the target region data in the body condition annotation data to obtain local motion field data; Based on the prior data of the action intent and the local motion field data, a chance-constrained fusion analysis is performed on the action difference domain of the abnormal object to obtain the action change data.

5. The method according to claim 4, characterized in that, The process of performing dense optical flow calculations on the target region data in the body condition annotation data to obtain local motion field data includes: Boundary flow initialization processing is performed on the key joint neighborhood data in the target region data to obtain joint boundary condition field data; Laplacian multi-scale pyramid calculation is performed on the target region data to obtain multi-scale pyramid image data; Based on the joint boundary condition field data, dense optical flow calculation is performed on the multi-scale pyramid image data to obtain initial optical flow field data; Skeletal topological Hodge projection is performed on the divergence component data and curl component data in the initial optical flow field data to obtain topologically consistent optical flow field data. The local motion field data is obtained by performing forward and backward consistency checks on the topologically consistent optical flow field data.

6. The method according to claim 4, characterized in that, The step involves performing chance-constrained fusion analysis on the action difference domain of the abnormal object based on the prior data of the action intent and the local motion field data to obtain action change data, including: Based on the prior data of the action intention, quantile pipeline modeling is performed on the joint-time prior trajectory set of the historical treatment data to obtain chance-constrained prior domain data. Streamline integration is performed on the joint neighborhood displacement data in the local motion field data to obtain the observed joint trajectory data; The residual cost is calculated by performing residual cost calculation on the out-of-bounds deviation data in the observed joint trajectory data; Based on the residual cost data, conditional risk aggregation is performed on each joint of the abnormal object to obtain action difference domain data; The action change data is obtained by statistically summarizing the time-series data of the connected components in the action difference domain data.

7. The method according to claim 1, characterized in that, The step of performing object function analysis based on the labeled physical condition data and the historical treatment data to obtain the motor function data of the abnormal object includes: Based on the historical treatment data, the key posture sequence data in the body condition annotation data are phase-aligned to obtain functional trajectory data; Persistent homology topology analysis is performed on the functional trajectory data to obtain deformation-invariant functional topology summary data; The motion function data is obtained by scaling and scoring the index vector data in the deformation-invariant functional topology summary data.

8. The method according to claim 7, characterized in that, The persistent cohomological topology analysis of the functional trajectory data yields deformation-invariant functional topology summary data, including: The functional trajectory data is processed using time delay embedding to obtain phase space point cloud data; An adaptive metric transformation is performed on the local phase consistency data in the phase space point cloud data to obtain homogeneous stable weight data. Weighted simple complex filtering analysis is performed on the phase space point cloud data and the homology stable weight data to obtain persistent barcode data; Persistent landscape encoding is performed on the persistent barcode data to obtain topological descriptor data; The scale and reparameterization sensitive components in the topology descriptor data are optimally transport-aligned to obtain the deformation-invariant functional topology summary data.

9. A motion function data acquisition system based on human-computer interaction, characterized in that, The system includes a motion function analysis device, a first interaction module, and a second interaction module; The motion function analysis device responds to the initial position determination instruction of the abnormal object obtained in the first interaction module, performs initial motion analysis based on the historical treatment data of the abnormal object, and presents the initial motion data to the abnormal object through the first interaction module; If the initial body image corresponding to the initial motion data is detected, the pose analysis of the initial body image is performed, and the body condition annotation box is presented to the annotator through the second interaction module; In response to the annotation confirmation instruction of the body condition annotation box obtained in the second interaction module, a movement analysis is performed based on the body condition annotation data of the body condition annotation box and the historical treatment data, and the movement change data is presented to the abnormal object through the first interaction module; If the current body image corresponding to the motion change data is detected, the current body image is used as the initial body image, and the process of performing pose analysis based on the initial body image and presenting the body condition annotation box to the annotator through the second interaction module is returned. Until the annotation stop command of the body condition annotation box is received, the object function analysis is performed based on the body condition annotation data and the historical treatment data to obtain the motion function data of the abnormal object.

10. The system according to claim 9, wherein the motion function analysis device comprises a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images