Motion capture method and system based on physical training

Abstract

The invention relates to the field of motion capture of physical training, in particular to a motion capture method and system based on physical training. The method comprises the following steps: based on a physical training scene, obtaining a training item, a computing power index and a camera parameter, and generating a session configuration packet containing a resource budget parameter group and an action stage template skeleton; laying guidance, self-calibration and reliability modeling are completed under the constraint of a preview video stream, and a probe registration packet is formed; performing accurate measurement triggering and sampling arrangement, constructing a key point virtual probe atlas and generating a key point atlas packet; a kinematics event sequence is extracted, stage slicing is completed, and a stage slicing package is formed; performing weighted alignment on the stage features to obtain an alignment residual packet; and finally, deviation deconstruction is completed based on the alignment residual packet, and an error correction instruction packet is generated. According to the method, consistent organization and stable judgment in the motion capture process are realized through a through link of session-level configuration and reliability modeling, and the traceability of motion analysis and error correction output in physical training is improved.

Description

Technical Field

[0001] This invention relates to the field of motion capture in sports training, and more particularly to motion capture methods and systems based on sports training. Background Technology

[0002] In the field of motion capture for sports training, existing solutions for motion capture methods and systems based on sports training typically involve keypoint precision measurement and stage segmentation processing around the preview video stream. Alignment and judgment are then completed under the constraints of the motion stage template skeleton associated with the training project. However, these solutions suffer from limitations such as inconsistencies in the temporal accuracy between the session configuration package and the preview video stream, a lack of a unified registration carrier for deployment guidance and self-calibration, and a lack of structured expression in reliability modeling. Existing methods often rely on fixed acquisition and processing links driven by camera parameters, neglecting the resource budget parameter constraints corresponding to computing power indicators. In scenarios where training projects change and the quality of the preview video stream fluctuates, it is easy to encounter problems such as difficulty in stably arranging precision measurement trigger sequences and sampling plans, lack of traceable input basis for the consistency verification and repair marking of keypoint virtual probe maps, and difficulty in stably generating alignment residual tensors and error correction instruction packages. For the joint processing of session configuration packages, probe registration packages, precision measurement and orchestration packages, key point atlas packages, stage slice packages, and alignment residual packages, existing technologies generally lack a cross-module index consistency mechanism constrained by probe reliability heatmaps and joint visibility state machines. This makes it difficult to form a consistent process in sports training, including generating session configuration packages, probe registration packages, precision measurement and orchestration packages, key point atlas packages, stage slice packages, and alignment residual packages; performing deviation deconstruction and generating anomaly judgment records and evidence indexes according to the rule engine; generating error correction instruction packages; and updating individualized rule parameter groups. This results in unstable correlation between anomaly judgment records and evidence indexes and error correction instruction packages. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a motion capture method based on sports training, comprising:

[0004] Acquire training projects, computing power metrics, and camera parameters to generate a session configuration package containing resource budget parameter sets and action phase template skeletons;

[0005] Based on the session configuration package and preview video stream, deployment guidance, self-calibration, and reliability modeling are performed to generate a probe registration package containing a probe geometry relationship registration table, a probe reliability heatmap, and a joint visibility state machine.

[0006] Based on the probe registration package, perform coarse screening judgment, fine measurement triggering, and key stage sampling arrangement to generate a fine measurement arrangement package containing fine measurement triggering sequence and sampling plan;

[0007] Based on the precision measurement and arrangement package, perform precision measurement of key points, construct a virtual probe map of key points, and perform consistency verification and repair marking to generate a key point map package;

[0008] Based on the keypoint map package, kinematic event sequences are extracted and matched with stage boundary rules to generate a stage slice package containing a stage slice set and stage credibility.

[0009] Based on the stage slice package, stage feature tokens are generated and aligned according to the probe reliability heatmap to generate an alignment residual package containing the alignment residual tensor and the missing test label set;

[0010] Based on the aligned residual package, deviation deconstruction is performed and anomaly judgment records and evidence indexes are generated according to the rule engine. Error correction instruction packages containing priority, timing, channel, and verification window fields are generated and individualized rule parameter groups are updated.

[0011] Furthermore, the resource budget parameter set includes a single-frame inference latency threshold, a precision frame quota, an output refresh cycle threshold, a stage sampling rate table, and a cache occupancy threshold; the action stage template skeleton includes a stage number table, a stage boundary event rule table, and a stage feature token dictionary.

[0012] Furthermore, the deployment guidance includes: extracting the character proportion score, screen shake score, backlight probability score, and occlusion ratio score from the preview video stream, and generating a deployment prompt record;

[0013] The self-calibration includes: estimating the field of view coverage boundary, human scale consistency parameters, and key joint visibility distribution within the guided action segment, and updating the probe geometric relationship registration table;

[0014] The joint visibility state machine includes a visible state, a suspected occlusion state, a missing state, and a jump state.

[0015] Furthermore, the coarse screening determination includes:

[0016] The real-time video stream is downsampled and low-resolution human body region localization is performed, motion energy sequences are calculated, and effective action marker sequences are generated.

[0017] The precision measurement trigger is based on the effective action marker sequence and the joint visibility state machine to generate a gated marker sequence, and the precision measurement trigger sequence is generated according to the gated marker sequence;

[0018] The key-stage sampling orchestration is based on the stage sampling multiplier table to generate a key-stage sampling plan.

[0019] Furthermore, the key point virtual probe map includes a set of joint nodes, a set of bone segments and edges, a node confidence field, and an edge constraint field;

[0020] The consistency check includes: calculating the consistency residual based on the length interval constraint, angle interval constraint and node velocity smoothing constraint of the bone segment edge set;

[0021] The repair markers include: generating a repair mask matrix for nodes whose consistency residuals exceed a threshold and writing it into the key point map package.

[0022] Furthermore, the kinematic event sequence consists of joint angle extreme value events, angular velocity zero-crossing events, support switching events, and trunk stability events;

[0023] The stage boundary rules consist of the event code sequence pattern, the duration frame threshold, and the event integrity threshold.

[0024] The stage credibility is synthesized by the mean weight of the probe reliability heatmap in the stage slice set and the event completeness score.

[0025] Furthermore, the stage feature token is composed of joint angle vector, angular velocity vector, trajectory curvature vector, left-right symmetry coefficient, and rhythm count;

[0026] The weighted alignment includes: generating an alignment index table based on stage boundary rules, generating a weight vector based on probe reliability heatmap, and calculating the alignment residual tensor according to the alignment index table and the weight vector;

[0027] The missing label set is generated by the state transition encoding of the joint visibility state machine.

[0028] Furthermore, the deviation deconstruction includes: generating a temporal deviation vector based on the alignment index table, generating an amplitude deviation vector based on the amplitude statistics of the alignment residual tensor, generating a collaborative deviation vector based on the multi-joint related residuals, and generating a stability deviation vector based on the high-frequency jitter residuals; the anomaly determination record includes an anomaly type code, severity code, stage number, and evidence index set.

[0029] Furthermore, the error correction instruction package consists of a set of instruction fragments, each of which includes a target stage number, a target deviation type code, a priority field, an output timing field, an output channel field, and a verification window field.

[0030] The individualized rule parameter group includes a threshold table, a priority mapping table, and a channel mapping table. The update recalibrates the threshold table and the priority mapping table by statistically analyzing the residual fall-off magnitude within the verification window.

[0031] Furthermore, a motion capture system based on sports training, applied to the method described in any one of claims 1-9, includes:

[0032] The session configuration module generates a session configuration package and sends it to the probe registration module;

[0033] The probe registration module generates a probe registration package based on the session configuration package and transmits it to the precision measurement orchestration module.

[0034] The precision measurement and arrangement module generates a precision measurement and arrangement package based on the probe registration package and transmits it to the key point map module;

[0035] The key point mapping module generates a key point mapping package based on the precision measurement and arrangement package and transmits it to the stage slicing module;

[0036] The stage slicing module generates a stage slicing package based on the key point map package and sends it to the alignment residual module;

[0037] The alignment residual module generates an alignment residual package based on the stage slice package and sends it to the decision strategy module;

[0038] The decision strategy module generates an error correction instruction package based on the alignment residual package and sends it to the instruction output module, while simultaneously outputting an individualized rule parameter group to the parameter update module.

[0039] The parameter update module updates the individualized rule parameter group and writes it back to the session configuration module.

[0040] The key innovations of this invention include:

[0041] (1) Based on the session configuration package and the preview video stream, perform deployment guidance, self-calibration, and reliability modeling to generate a probe registration package containing a probe geometric relationship registration table, a probe reliability heat map, and a joint visibility state machine. The probe reliability heat map and the joint visibility state machine are then used in the link constraints of the subsequent precision measurement arrangement package, key point map package, stage slice package, and alignment residual package.

[0042] (2) Based on the stage slice package, generate stage feature tokens and align them according to the probe reliability heatmap to generate an alignment residual package containing an alignment residual tensor and a set of missing test labels, wherein the set of missing test labels is organized by the state transition record of the joint visibility state machine and is indexed and associated with the alignment residual tensor.

[0043] (3) Based on the alignment residual package, perform deviation deconstruction and generate anomaly judgment records and evidence index according to the rule engine, generate error correction instruction package containing priority, timing, channel and verification window fields and update individualized rule parameter group, wherein the evidence index establishes a correspondence with the verification window field of the error correction instruction package and writes back to the generation link of the session configuration package.

[0044] The following are its main beneficial effects:

[0045] (1) In response to the problems of inconsistent timing between the session configuration package and the preview video stream, lack of unified registration carrier for deployment guidance and self-calibration, and lack of structured expression in reliability modeling in the existing scheme, the present invention achieves the following: by uniformly registering and transferring the probe geometric relationship registration table, probe reliability heat map, and joint visibility state machine in the probe registration package, the precision measurement trigger sequence and sampling plan form a consistent input caliber under the constraints of the same session configuration package, and the consistency verification and repair mark of the key point virtual probe map have the same source constraint information and traceability basis, so that the serial link from the key point map package to the stage slice package remains consistent.

[0046] (2) In view of the existing solutions, the lack of computable constraints on the quality fluctuation of the preview video stream and the lack of stable intermediate expression between the alignment residual tensor and the error correction instruction package in the alignment and judgment process in sports training scenarios caused by the drift of the judgment caliber, this invention incorporates the stage boundary rules and stage credibility of the stage slice set into the alignment input by organizing the stage feature tokens, and injects the reliability modeling results into the alignment residual tensor by weighted alignment using the probe reliability heatmap. At the same time, the state transition of the joint visibility state machine is structurally marked by the missing test label set and associated with the alignment residual tensor index, so that the alignment residual package forms a verifiable and consistent difference expression within the alignment window.

[0047] (3) In response to the problem of unstable correlation between anomaly determination records and evidence indexes and error correction instruction packages in existing solutions, which makes it difficult to form a consistent process from generating session configuration packages to updating individualized rule parameter groups, resulting in the breakage of the closed-loop control link, this invention generates anomaly determination records and evidence indexes through deviation deconstruction and rule engine, and establishes a correspondence between the priority, timing, channel, and verification window fields of the error correction instruction package and the evidence index. Then, the updated individualized rule parameter group is written back to the generation link of the session configuration package, so that the subsequent session configuration package and probe registration package are called under the same parameter caliber, thereby making the anomaly determination records and error correction instruction packages maintain a traceable and consistent correlation within the session-level link. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a motion capture method based on sports training provided in an embodiment of this application;

[0049] Figure 2 This is a structural block diagram of a motion capture system based on sports training, provided as an embodiment of this application. Detailed Implementation

[0050] Example 1: Refer to Figure 1This is a flowchart illustrating a motion capture method based on sports training provided in an embodiment of the present invention. The flowchart may include at least steps S100-S700:

[0051] S100: Obtain training items, computing power indicators, and camera parameters, and generate a session configuration package containing resource budget parameter groups and action phase template skeletons;

[0052] S200, based on session configuration package and preview video stream, performs deployment guidance, self-calibration, and reliability modeling, and generates a probe registration package containing probe geometry relationship registration table, probe reliability heat map, and joint visibility state machine;

[0053] S300, based on the probe registration package, performs coarse screening judgment, fine measurement triggering, and key stage sampling arrangement, and generates a fine measurement arrangement package containing fine measurement triggering sequence and sampling plan;

[0054] S400: Based on the precision measurement and arrangement package, perform precision measurement of key points, construct virtual probe maps of key points, perform consistency verification and repair marking, and generate key point map packages;

[0055] S500: Based on the key point map package, extract kinematic event sequences and match stage boundary rules to generate a stage slice package containing a stage slice set and stage credibility.

[0056] S600: Based on the stage slice package, generate stage feature tokens and align them according to the probe reliability heatmap to generate an alignment residual package containing the alignment residual tensor and the missing test label set;

[0057] S700, based on the aligned residual package, performs deviation deconstruction and generates anomaly judgment records and evidence indexes according to the rule engine, generates error correction instruction packages containing priority, timing, channel, and verification window fields, and updates individualized rule parameter groups.

[0058] S100: Obtain training items, computing power indicators, and camera parameters, and generate a session configuration package containing resource budget parameter groups and action phase template skeletons;

[0059] Specifically, in this step, when the training session is initiated or the training item is switched, the session configuration module receives external input and completes its structured solidification. The external input includes training items, computing power indicators, and camera parameters. The training items are a combination of the current sports training item identifier and action type identifier, and are associated with training scene constraint information. The training scene constraint information includes a description of the training space boundary, a lighting condition marker, and an output channel declaration. The computing power indicators are a portrait of the computing resources of the mobile terminal or edge computing device at the time of the session, including at least the available load level of the Central Processing Unit (CPU), the available queue level of the Graphics Processing Unit (GPU), the available operator set marker of the Neural Processing Unit (NPU), and the power status level. The camera parameters are a portrait of the imaging and acquisition capabilities of the camera at the time of the session, including at least the resolution level, frame rate level, shutter type marker, equivalent focal length marker, distortion model marker, and timestamp source marker. This step performs field normalization, caliber solidification, and consistency verification on the training project, the computing power index, and the camera parameters. Field normalization includes converting the frame rate unit, resolution description, and timestamp precision description returned by different device interfaces into a unified caliber record. Calibration solidification includes writing a field version number and field source mark for each type of field. Consistency verification includes performing bandwidth usage estimation on the combination of resolution level and frame rate level and checking it in relation to the cache usage threshold, performing monotonic sampling check on the timestamp source mark, and writing a time anchor trusted mark. When the consistency verification triggers a missing or conflicting condition, an abnormal configuration record is generated and written to the change record item of the session configuration module. At the same time, a preset downgrade level is selected and written to the level field of the camera parameters. The downgrade level is jointly determined by the available queue level and the power status level of the computing power index.

[0060] Further, after the field normalization and consistency verification are completed, this step generates a resource budget parameter set and solidifies it as the constraint input for subsequent steps. The resource budget parameter set is a parameter set that combines real-time constraints, precision measurement resource constraints, and cache resource constraints. Its minimum set consists of a single-frame inference latency threshold, a precision measurement frame quota, an output refresh cycle threshold, a stage sampling multiplier table, and a cache occupancy threshold. The single-frame inference latency threshold is obtained by mapping the computing power index and is verified in association with the frame rate level of the camera parameters. The precision measurement frame quota is used to limit the upper limit of the frame count that triggers precision measurement within the same output refresh cycle and is associated with the available operator set of the computing power index. The output refresh cycle threshold is used to limit the output beat of the error correction instruction packet and is associated with the output channel declaration. The stage sampling multiplier table is used to describe the sampling density multiplier of different action stages and is bound to the training project. The cache occupancy threshold is used to limit the ring cache capacity of the key point map packet and the stage slice packet and is associated with the resolution level of the camera parameters. To ensure auditability, the resource budget parameter group is written with a version number, generation timestamp, and parameter digest check code when it is generated. The parameter digest check code is calculated by a secure hash algorithm on the field sequence of the resource budget parameter group and written together with the session number into the header field of the session configuration package.

[0061] Furthermore, after the resource budget parameter group is solidified, this step generates an action stage template skeleton, which serves as the rule input for subsequent stage slices. The action stage template skeleton is a combination of a stage number table, a stage boundary event rule table, and a stage feature token dictionary. The stage number table records the stage set corresponding to the training project and its stage sequence index. The stage boundary event rule table records the boundary event code sequence pattern, duration frame threshold, and event integrity threshold for each stage. The stage feature token dictionary records the set of stage feature token fields associated with each stage. The sources of the action stage template skeleton include a local preset template library and historical snapshots of individualized rule parameter groups. This step performs version matching on the training project identifier when reading the template library and performs consistency verification on the parameter digest checksum when reading historical snapshots, and writes the matching result into the template source mark field. When the template library is missing a corresponding training project identifier, this step writes a template missing mark and generates a default stage number table. The default stage number table is mapped from the action type identifier of the training project, and the default mapping rule number is written into the change record item, so that subsequent steps have traceable source information when reading the action stage template skeleton. Understandably, the action phase template skeleton does not contain specific athlete sample trajectory parameters. Instead, it solidifies the phase boundary constraints and fields of interest within the phase into a structured skeleton for rule matching and invocation when extracting kinematic event sequences from the keypoint atlas package.

[0062] After generating the resource budget parameter group and the action phase template skeleton, this step encapsulates them together with the training project, the computing power index, and the camera parameters to form a session configuration package. The session configuration package contains a session number, configuration version number, device identifier summary, and output channel declaration. The session number is generated by the session configuration module when the training session is triggered and is associated with the historical session link. The configuration version number is incremented with each change record entry. The device identifier summary is generated by normalizing the hardware capability fields returned by the device interface. The output channel declaration records the enabled status flags and priority mapping fields for the display, voice, vibration, and headphone channels. After generation, the session configuration package is written to local persistent storage and simultaneously written to the circular cache index. Subsequently, the session configuration package is used as input to S200 and read by the probe registration module, enabling S200 to perform deployment guidance, self-calibration, and reliability modeling based on the session configuration package and the preview video stream.

[0063] In summary, the technical effects of this step are as follows: Through the structured encapsulation of the session configuration package, training projects, computing power indicators, and camera parameters are solidified under the same caliber and have version traceability attributes. Resource budget parameters and action stage template skeletons form the pre-constraint inputs for subsequent steps, enabling cross-step processing links to operate seamlessly within the same budget and stage rule framework.

[0064] S200, based on session configuration package and preview video stream, performs deployment guidance, self-calibration, and reliability modeling, and generates a probe registration package containing probe geometry relationship registration table, probe reliability heat map, and joint visibility state machine;

[0065] Specifically, in this step, the probe registration module reads the session configuration package output by S100 when the session is initiated and simultaneously connects to the preview video stream output by the camera to establish a preview processing pipeline. The session configuration package carries at least a resource budget parameter set and an action phase template skeleton. In this step, the resource budget parameter set is parsed into constraints for the preview processing frequency, buffer usage threshold, and self-calibration trigger beat. The action phase template skeleton is parsed into the set of joints of interest for subsequent joint visibility monitoring. Before entering the processing pipeline, the preview video stream undergoes timestamp source marker verification and frame sequence continuity verification. Timestamp source marker verification focuses on monotonically increasing detection and jitter interval detection, while frame sequence continuity verification focuses on dropped frame count and duplicate frame count. When an anomaly is detected during verification, a preview anomaly record is written, and the anomaly type code and the occurrence frame number are written to the session log area of ​​the probe registration module. The session log area is bound to the session number and archived incrementally with the configuration version number. Understandably, the "probe" referred to in this step refers to a video acquisition probe, which is a combination of video acquisition hardware such as mobile phones, tablets or cameras and their drivers and installation posture information. The probe registration module treats the probe as a computable object and maintains the probe status, scoring status and reconfiguration prompt status within the same session.

[0066] During the deployment guidance phase, the probe registration module extracts preview frames frame by frame based on the preview video stream or according to the sparse frame beats defined by the resource budget parameter group. For each preview frame, it performs coarse human body region localization and image quality measurement, and generates deployment prompt records accordingly. Among them, the coarse human body region localization uses a combination of foreground segmentation and human body bounding box regression at low resolution to obtain the human body proportion field. The human body proportion field represents the ratio of the human body bounding box area to the image area and is written into the human body proportion score. The image quality measurement includes at least the image jitter score, backlight probability score, and occlusion ratio score. The image jitter score is obtained by statistically analyzing the variance of the global motion vectors of adjacent preview frames. The backlight probability score is obtained by combining the brightness distribution difference between the human body region and the background region, the proportion of local highlight saturation segments, and the proportion of edge gradient decay segments. The occlusion ratio score is obtained by combining the texture loss rate of key areas within the human body bounding box and the incompleteness rate of the human body contour. The deployment prompt record is generated by writing the prompt frame number, prompt timestamp, and prompt type code. The prompt type code is obtained by matching the character proportion score, screen jitter score, backlight probability score, and occlusion ratio score with a threshold table. The threshold table is taken from the session configuration package or generated and written to the session log area by the probe registration module when the first session is triggered. In this step, the deployment prompt record is only archived as a component field of the probe registration package, and in the system embodiment, it is mapped to the distance adjustment mark, angle adjustment mark, height adjustment mark, and landscape / portrait mark on the terminal interface, thus forming a consistent session number link with the subsequent precision measurement and arrangement module.

[0067] During the self-calibration phase, the probe registration module extracts guided motion segments from the preview video stream and performs geometric parameter estimation and scale consistency estimation within the segments, thereby updating the probe geometric relationship registration table. The guided motion segments are triggered by the probe registration module according to the output channel declaration in the session configuration package and executed by the trainer. The start and end of the segments are jointly determined by motion energy mutation detection and human bounding box stable continuous frame count detection. Geometric parameter estimation includes at least field of view coverage boundary estimation and extrinsic parameter summary estimation. The field of view coverage boundary is calculated from the envelope trajectory of the four boundaries of the human bounding box within the guided motion segment and written into the field of field of view coverage boundary. The extrinsic parameter summary is obtained by fusing device attitude information and human principal axis direction estimation and written into the pitch angle, roll angle, and horizontal angle fields. Scale consistency estimation generates human scale consistency parameters, which are obtained by jointly calculating the steady-state interval of the human bounding box height sequence within the guided motion segment and the steady-state interval of the relative length ratio between joints and written into the human scale consistency parameter field. The minimum set of fields in the probe geometry relationship registration table includes probe number, resolution level, frame rate level, field of view coverage boundary, human scale consistency parameter, and timestamp source marker. These fields are generated jointly by the session configuration package and the guided action fragment. When device attitude information is missing, the pitch angle field, roll angle field, and horizontal angle field are marked as missing and a snapshot of the external parameter summary corresponding to the previous configuration version number is retained, so that the probe geometry relationship registration table has an auditable version chain.

[0068] During the reliability modeling phase, the probe registration module establishes a continuous update mechanism around the probe reliability heatmap and the joint visibility state machine. The probe reliability heatmap represents the availability weights of the preview video stream in terms of spatial location and joint category dimensions, while the joint visibility state machine represents the visibility state transition rules for each joint in the temporal dimension. The generation process of the probe reliability heatmap includes heat grid division, heat unit scoring, and time decay updates. Heat grid division divides the image into fixed-size grid units based on resolution levels and binds them to the human bounding box coordinate system. Heat unit scoring maps the human proportion score, image jitter score, backlight probability score, and occlusion ratio score into heat weights and writes them into the heat weight field. Time decay updates use a sliding window accumulation and exponential decay fusion to obtain stable heat weights, thereby reducing weight jitter caused by instantaneous flicker. The joint visibility state machine includes a visible state, a suspected occlusion state, a missing state, and a jump state. State transitions are jointly driven by the joint detection confidence sequence, the joint position change amplitude sequence, and the occlusion ratio score: when the joint detection confidence sequence is continuously below a threshold and the occlusion ratio score is above a threshold, the state transitions to the suspected occlusion state; when the joint detection confidence sequence is marked as missing and the number of consecutive frames exceeds a threshold, the state transitions to the missing state; and when the joint position change amplitude sequence exceeds a threshold and the image jitter score is in the low jitter range, the state transitions to the jump state. Each transition is written into a state transition record, which includes the joint number, source state, target state, trigger frame number, and trigger score summary. This record, along with the mean weight of the probe reliability heatmap within the same time window, is written into the probe registration package, thus providing gating input for the subsequent precision measurement and orchestration module.

[0069] In the engineering implementation, after the trainee starts a training session in the track scene, the probe registration module reads the session configuration package and accesses the preview video stream under the session start trigger, and completes the initialization of the preview anomaly record; when the character proportion score is in the low proportion range or the backlight probability score crosses the threshold, the deployment prompt record is written into the probe registration package and the corresponding adjustment mark is displayed on the display channel declared and indicated by the terminal output channel; then the trainee executes the guided action segment, and the probe registration module updates the field of view coverage boundary and human scale consistency parameter field of the probe geometric relationship registration table in the segment, and simultaneously refreshes the heat weight field of the probe reliability heatmap; when the trainee's occlusion ratio score increases during the starting action, causing the hip and knee joint states to migrate to the suspected occlusion state, the joint visibility state machine generates the corresponding state migration record and writes it into the probe registration package. Finally, the probe registration module encapsulates the probe geometric relationship registration table, probe reliability heat map, and joint visibility state machine into a probe registration package, and writes the session number, configuration version number, and generation timestamp fields into it. Then, the probe registration package is used as input to the S300 and read by the precision measurement orchestration module, so that the S300 performs coarse screening judgment, precision measurement triggering, and key stage sampling orchestration based on the probe registration package.

[0070] Summary of the technical effects of this step: This step incorporates the preview acquisition quality and geometric consistency into the main link, and forms a structured probe registration package through the probe geometric relationship registration table, probe reliability heat map and joint visibility state machine, so that the subsequent fine measurement arrangement and key point fine measurement have traceable gating input and version record.

[0071] S300, based on the probe registration package, performs coarse screening judgment, fine measurement triggering, and key stage sampling arrangement, and generates a fine measurement arrangement package containing fine measurement triggering sequence and sampling plan;

[0072] Specifically, after receiving the probe registration packet, the precision measurement orchestration module reads the probe geometric relationship registration table, the probe reliability heatmap, and the joint visibility state machine, and synchronously connects to the video acquisition channel with the probe number consistent with the probe geometric relationship registration table to form a real-time video stream processing link; the probe registration packet contains a session number and a configuration version number, which are used to associate and archive the coarse screening judgment log, the precision measurement trigger log, and the sampling plan version chain in this step.

[0073] The coarse screening judgment is a low-overhead discrimination processing for real-time video streams. The input is the real-time video stream and the probe reliability heatmap, and the output is a valid action marker sequence. The valid action marker sequence consists of frame number, timestamp, valid action state marker and motion energy field. The motion energy field is synthesized by global displacement amplitude statistics and local texture change amplitude statistics of human body regions in adjacent frames. The precision measurement and orchestration module performs downsampling and low-resolution human body region localization on the real-time video stream. The downsampling beat is derived from the frame rate level in the probe geometric relationship registration table. The low-resolution human body region localization outputs a human body bounding box sequence and writes it into the human body proportion field. Then, the motion energy field is calculated within the sliding time window of the human body bounding box sequence, and the motion energy field is fused with the average thermal weight of the probe reliability heatmap within the human body bounding box coverage area to obtain a coarse screening score. When the coarse screening score crosses the corresponding threshold in the threshold table, the action valid state is marked as valid; otherwise, the action valid state is marked as invalid, and this mark is written into the action valid mark sequence in sequence with the frame number. The threshold table is obtained by indexing the configuration version number carried in the probe registration package. The threshold table entries include the motion energy threshold, the human body proportion threshold, and the thermal weight threshold, which together constitute the minimum set constraint for coarse screening judgment.

[0074] The precise measurement trigger is a gated triggering process based on the effective action marker sequence. The inputs are the effective action marker sequence and the joint visibility state machine, and the outputs are a gated marker sequence and a precise measurement trigger sequence. The gated marker sequence consists of a frame number, a gated state marker, and a key joint state summary field. The key joint state summary field is generated by combining the set of joint numbers in the visible state of the joint visibility state machine with the statistics of their consecutive visible frames. Within the frame segment where the effective action state is marked as effective, the precise measurement orchestration module reads the state transition record of the joint visibility state machine, writes the joint numbers in missing or abrupt states into the masked joint set, and writes the joint numbers in the visible state with a consecutive visible frame count exceeding a threshold into the available joint set. Subsequently, the masked joint set and the available joint set are mapped to the joint dimension weights of the probe reliability heatmap to generate a gated score and write it into the gated state marker. The precision measurement trigger sequence consists of a frame number, a trigger state marker, a trigger reason code, and a trigger quota index field. The precision measurement orchestration module generates a candidate trigger frame set based on the gating marker sequence and sorts the candidate trigger frames according to the gating score to form a trigger queue. The trigger queue is truncated within the time window corresponding to the output refresh cycle threshold. The truncated length is limited by the precision measurement frame quota. The truncated frame number set is written into the precision measurement trigger sequence and simultaneously written into the trigger quota index field. The trigger reason code is obtained by jointly encoding the action valid state marker, the gating state marker, and the size of the shielded joint set, so that the subsequent key point map module has traceable information about the trigger reason when reading the precision measurement trigger sequence.

[0075] The key-stage sampling orchestration involves clock reshaping of the precision measurement trigger sequence. The inputs are the precision measurement trigger sequence, the probe reliability heatmap, and the stage sampling rate table; the output is a sampling plan. The stage sampling rate table is written into the probe registration package and bound to the training project. The stage sampling rate table fields include the stage number, rate value, and the length of the effective frame window. The precision measurement orchestration module calculates the peak-valley structure of the motion energy field within the continuous frame window corresponding to the precision measurement trigger sequence and generates key-stage candidate markers based on the thermal weight stability of the torso region grid cells in conjunction with the probe reliability heatmap. The key-stage candidate markers consist of a frame number, a candidate state marker, and a stability summary field. Subsequently, the precision measurement orchestration module maps the key-stage candidate markers to rate values ​​according to the stage sampling rate table and performs trigger interval compression or trigger interval stretching on the precision measurement trigger sequence to obtain the sampling plan. The sampling plan fields include a frame number sequence, a sampling rate sequence, and a sampling plan version number. The sampling plan version number and configuration version number are jointly written into the sampling plan header field for consistency verification with the stage credibility of subsequent stage slice packages. For the frame segment marked as closed by the gated state, the fine measurement and orchestration module writes the empty trigger segment mark and keeps the coarse screening and judgment link running continuously. At the same time, the empty trigger segment mark and the trigger reason code are written into the abnormal record area of ​​the fine measurement and orchestration packet.

[0076] In the engineering implementation, after the trainee starts a session in the basketball dribbling training scenario, the probe registration package output by the probe registration module enters the precision measurement and orchestration module. When the trainee enters continuous dribbling action, the motion energy field shows a periodic peak-valley structure within the sliding time window. The action valid state marker is continuously written into the valid state, and the hip and knee joints remain visible in the joint visibility state machine, with the gating state marker written into the open state. The precision measurement and orchestration module forms a trigger queue based on the precision measurement frame quota within the output refresh cycle threshold time window and writes it into the precision measurement trigger sequence. When screen jitter causes the ankle joint state to migrate to the jump state, the size of the shielded joint set increases, the gating score decreases, the trigger cause code changes synchronously, and is written into the precision measurement trigger sequence. When the candidate marker of the key stage points to the acceleration and change of direction segment, the corresponding multiplier value of the stage sampling multiplier table is written into the sampling multiplier sequence, and the sampling plan performs compression processing on the trigger interval, thereby reorganizing the trigger density within the frame window into the sampling plan.

[0077] The precision measurement orchestration package is generated by encapsulating the precision measurement trigger sequence and the sampling plan. The session number, configuration version number and generation timestamp are written in the package header. At the same time, the gating mark sequence summary, the action valid mark sequence summary and the abnormal record area index are also written. The precision measurement orchestration module transmits the precision measurement orchestration package to the key point map module as the input for "execute key point precision measurement based on precision measurement orchestration package" in S400, so that the key point map module extracts precision measurement frames according to the precision measurement trigger sequence and organizes the time order of precision measurement frames according to the sampling plan.

[0078] Summary of the technical effects of this step: This step transforms the reliability and visibility information in the probe registration package into a precision measurement trigger sequence and sampling plan, enabling the precision measurement of key points to obtain traceable triggering basis and cycle constraints under the same session number and configuration version number link.

[0079] S400: Based on the precision measurement and arrangement package, perform precision measurement of key points, construct virtual probe maps of key points, perform consistency verification and repair marking, and generate key point map packages;

[0080] Specifically, the key point mapping module receives the precision measurement orchestration package output by S300 as input, parses the precision measurement trigger sequence and sampling plan in the precision measurement orchestration package, and reads the session number and configuration version number in the precision measurement orchestration package header to establish the processing context for this step. Among them, the precision measurement trigger sequence records the trigger frame number, trigger state flag, trigger reason code and trigger quota index field, and the sampling plan records the frame number sequence, sampling rate sequence and sampling plan version number. The key point mapping module binds the local circular cache index according to the session number, and binds the model version record item and log partition according to the configuration version number, so that the key point mapping package generated in this step and the subsequent stage slicing module can complete cross-step connection under the same version link. After receiving the real-time video stream, the keypoint mapping module performs sequential loading processing on the frame number sequence according to the sampling plan. The sequential loading processing includes frame number deduplication, frame sequence continuity verification, and timestamp alignment registration. The frame sequence continuity verification focuses on the count of dropped frames, the count of duplicate frames, and the count of out-of-order frames. The timestamp alignment registration focuses on the consistency verification between the trigger quota index field in the precision measurement trigger sequence and the timestamp source mark of the video stream. When dropped frames or out-of-order frames occur, the keypoint mapping module writes a precision measurement frame anomaly record and registers the abnormal frame number segment in the anomaly index field of the keypoint mapping package for segment skipping processing during subsequent kinematic event sequence extraction.

[0081] During the keypoint precision measurement phase, the keypoint mapping module extracts corresponding precision measurement frames according to the precision measurement trigger sequence and performs preprocessing such as human region cropping, scale normalization, and noise suppression on the precision measurement frames to obtain the precision measurement input tensor. Human region cropping is driven by the human bounding box coordinates output by low-resolution human region localization. Scale normalization maps the human region cropping result to a fixed input size and writes it into the scale normalization parameter field. Noise suppression includes two types of operations: brightness normalization and edge-preserving filtering, and writes them into the preprocessing label field. The keypoint mapping module inputs the precision measurement input tensor into a lightweight convolutional neural network (CNN) to perform joint keypoint inference, outputting joint coordinates and a joint confidence sequence. The lightweight convolutional neural network is loaded into the model version record and obtained by indexing the configuration version number. In each precision measurement frame, the joint keypoint inference generates a joint number, two-dimensional coordinates, and a node confidence field. The node confidence field is obtained by mapping the joint confidence sequence to the gating level of the trigger reason code. The gating level mapping is written into the gating mapping table of the keypoint mapping module and archived with the configuration version number. Subsequently, the keypoint mapping module generates a set of joint nodes based on the joint number and two-dimensional coordinates of each precision measurement frame, and embeds the node confidence field within the joint node set. The joint node set, bone segment edge set, node confidence field, and edge constraint field constitute the minimum set of the keypoint virtual probe map. The bone segment edge set is generated from the human skeleton connection relationship table and written with bone segment endpoint number pairs. The edge constraint field is generated by combining bone segment length constraint items and joint angle constraint items and written with constraint type code and constraint threshold fields. To enhance temporal consistency, the keypoint mapping module synchronously registers the temporal connection relationships of joints with the same name in adjacent precision measurement frames within the keypoint virtual probe map, and writes the temporal connection relationships into the extended edge field. The extended edge field and the minimum set are stored separately, and an extended field presence marker is registered in the keypoint map packet header, facilitating subsequent modules to execute parsing branches according to the packet header marker.

[0082] During the consistency verification phase, the keypoint mapping module performs three types of verifications on the keypoint virtual probe map: length interval constraints, angle interval constraints, and node velocity smoothing constraints, and calculates consistency residuals. The length interval constraints revolve around the bone segment edge set. The keypoint mapping module calculates the bone segment length sequence for each bone segment edge and generates length baseline statistics within a sliding time window. The length baseline statistics include median length and dispersion length fields. Subsequently, the median length and dispersion length are mapped to length interval thresholds and written into the edge constraint field. The extent to which the bone segment length sequence exceeds the length interval threshold is recorded as a length residual. The angle interval constraints revolve around adjacent edge pairs of the bone segment edge set. The keypoint mapping module generates joint angle sequences according to joint numbers and maps the joint angle sequences to angle interval thresholds. The extent to which the joint angle sequences cross the angle interval threshold is recorded as an angle residual. The node velocity smoothing constraints revolve around the temporal connection relationship of the joint node set. The keypoint mapping module calculates node velocity sequences for adjacent precision measurement frames and compares the node velocity sequences with velocity smoothing thresholds. The extent to which the node velocity sequences cross the velocity smoothing threshold is recorded as a velocity residual. The keypoint map module aggregates length residuals, angle residuals, and velocity residuals by joint number and frame number to generate consistency residuals and writes them into the consistency residual field. At the same time, it writes the residual source code to distinguish the three types of sources: length residual, angle residual, and velocity residual. The consistency residual field serves as the trigger for repair marking and is solidified in this step. It also maintains the same index alignment relationship with the joint node set within the keypoint map package, allowing the subsequent alignment residual module to trace the source of abnormal observations.

[0083] During the repair marking phase, the key point map module generates a repair mask matrix for nodes whose consistency residual exceeds the threshold and writes it into the key point map package. The repair mask matrix establishes a two-dimensional mask bitmap according to the frame number and joint number. The mask bit value of the mask bitmap is determined by the consistency residual field and the residual source code. The generation timestamp of the mask bit and the trigger threshold version number are written into the header field of the repair mask matrix. For nodes with anomalous mask bits, the keypoint mapping module performs constraint repair processing, which includes two links: bone segment length backprojection and short-window smoothing reestimation. Bone segment length backprojection corrects the coordinates of the anomalous node along the constraint direction of the bone segment edge set to within the length interval threshold range. Short-window smoothing reestimation performs weighted smoothing on the same-named joint nodes in adjacent precision measurement frames, with the weights given by the node confidence field and the gating mapping table. The corrected node coordinates are written into the repair coordinate field of the joint node set. The repair coordinate field is included in the extended field partition within the keypoint mapping package, and a repair field existence marker is registered in the package header. This allows the subsequent slicing module to select either the original coordinates or the repair coordinates based on the package header marker when extracting kinematic event sequences. For nodes with missing mask bits, the keypoint mapping module writes a missing marker while keeping the edge constraint field unchanged. Simultaneously, it registers the missing node segment in the anomalous index field for reference during subsequent confidence synthesis.

[0084] In the engineering implementation, the trainee performs a squatting motion in a physical training scenario. The precision measurement and arrangement package output by the precision measurement and arrangement module enters the keypoint map module. The sampling plan provides a sparse frame number sequence within the starting and ending frame windows, and a high-magnification frame number sequence within the squatting turning frame window. The keypoint map module extracts precision measurement frames within the squatting turning frame window according to the precision measurement trigger sequence and completes the precision measurement of keypoints, generating a set of joint nodes including hip, knee, and ankle joint nodes and a corresponding set of bone segment edges. Due to the trainee's forearm occlusion causing abrupt changes in the coordinates of elbow joint nodes in some frames, the length residual and velocity residual aggregate in the consistency residual field and cross the threshold. The repair mask matrix writes the abnormal state mask bit at the corresponding frame number and joint number position, and triggers bone segment length backprojection and short window smoothing reestimation. The repair coordinate field is simultaneously written into the joint node set. When the stage slicing module reads the keypoint map package later, it selects the repair coordinates according to the existence mark in the repair field of the package header to extract the extreme value event of the joint angle. The abnormal index field is simultaneously used for stage credibility calculation, thereby completing the input connection with S500.

[0085] After generating the keypoint virtual probe map, consistency residuals, repair mask matrix, and anomaly index field, this step encapsulates the above products into a keypoint map package. The session number, configuration version number, sampling plan version number, frame number range, and verification summary fields are written into the header of the keypoint map package. The keypoint map package is written into the circular cache index and transmitted to the stage slicing module as the input source for "extracting kinematic event sequences and matching stage boundary rules based on the keypoint map package" in S500, so as to maintain terminology consistency and version traceability across steps.

[0086] Summary of the technical effects of this step: This step organizes the precise measurement of key points around the triggering beat of the precision measurement and arrangement package, and solidifies the key point observations into a virtual probe map of key points. At the same time, through consistency verification and repair marking, traceable abnormal observation marks and repair records are formed, providing stable input for subsequent stage slicing.

[0087] S500: Based on the key point map package, extract kinematic event sequences and match stage boundary rules to generate a stage slice package containing a stage slice set and stage credibility.

[0088] Specifically, the stage slicing module receives the keypoint map package output by S400 as input, parses the header fields of the keypoint map package, and establishes a processing context binding the session number and configuration version number. Simultaneously, it reads the joint node set, bone segment edge set, node confidence field, edge constraint field, consistency residual field, repair mask matrix, and anomaly index field from the keypoint map package. Before starting event extraction, the stage slicing module performs frame number range verification and field completeness verification on the keypoint map package. The frame number range verification focuses on the matching relationship between the sampling plan version number and the frame number range, while the field completeness verification focuses on the continuity of the joint node set index and the missing node confidence field marker. When the anomaly index field contains lost frame segments or missing node segments, the stage slicing module writes a slice anomaly record and maps the anomaly segment to a skip segment marker. The skip segment marker and the frame number range are jointly written to the event extraction log partition of this step. Understandably, this step involves constructing a kinematic event sequence based on joint timing observations. The kinematic event sequence is a set of events that have discriminative significance for the boundaries of training action phases. In this embodiment, the event set consists of joint angle extreme value events, angular velocity zero-crossing events, support switching events, and trunk stability events. During the event extraction process, the event code, occurrence frame number, event confidence, and evidence index fields are kept in the same source registration.

[0089] Further, the stage slicing module selects the joint coordinate source for event calculation from the keypoint map package. When the keypoint map package header has a repair field presence marker, the stage slicing module reads the repair coordinate field from the joint node set and simultaneously reads the repair mask matrix, writing the node frames marked as abnormal by the repair mask matrix into the culling mask. When the keypoint map package header lacks a repair field presence marker, the stage slicing module reads the two-dimensional coordinate field from the joint node set and writes the node frames with consistency residual fields exceeding a threshold into the culling mask. Subsequently, the stage slicing module constructs a joint angle calculation link based on the bone segment edge set and edge constraint field. The joint angle calculation link maps adjacent bone segment edges to joint angle sequences and writes angle missing markers within the frames covered by the skip segment markers. The stage slicing module performs short-window smoothing on the joint angle sequences in frames outside the angle missing markers and generates angular velocity sequences. The angular velocity sequences are obtained by the angle difference between adjacent frames and written into the angular velocity field. For each joint number, the stage slicing module maps the node confidence field and the removal mask together to the event calculation weight. The event calculation weight is called in the event confidence calculation stage to form a traceable weight link consistent with the key point map package.

[0090] During the extraction of joint angle extreme events, the stage slicing module performs extreme candidate detection on the joint angle sequence within a sliding time window. Extreme candidate detection includes local maxima detection and local minima detection, and performs consistency verification on the stable angular velocity sign segments on both sides of the extreme point. When the consistency verification meets the continuous frame number threshold, the stage slicing module generates a joint angle extreme event and registers the event code, occurrence frame number, and event confidence score. The event confidence score is synthesized from the extreme value amplitude, the stability of the angular velocity signs on both sides, and the event calculation weight. The extraction of angular velocity zero-crossing events revolves around the angular velocity sequence. The stage slicing module performs zero-crossing candidate detection on the angular velocity sequence and performs joint verification on the difference in angular velocity amplitude before and after the zero-crossing point and the node confidence score field. When the verification is satisfied, the angular velocity zero-crossing event is registered and written into the evidence index field. The evidence index field points to the node index of the corresponding frame number and joint number within the keypoint map package, forming a traceable link between the event and the keypoint map. The support switching event extraction revolves around the contact discrimination features between the lower limb joint node set and the ground. The stage slicing module generates support candidate markers from the vertical displacement sequence, horizontal velocity sequence, and stable duration frame count of the ankle and knee joints, and verifies the mutual exclusion relationship between the left and right foot support candidate markers. When the verification is successful, the support switching event is registered. When the abnormal index field covers the support candidate frame segment, the stage slicing module writes a support switching missing marker and includes the missing marker in the event completeness score. The trunk stability event extraction revolves around the trunk principal axis direction and the approximate trajectory of the center of mass. The stage slicing module generates a trunk principal axis direction sequence from the line vector connecting the shoulder and hip joints, and generates an approximate trajectory sequence of the center of mass from the weighted coordinates of the hip and knee joints. When the directional jitter amplitude of the trunk principal axis direction sequence is lower than the stability threshold and the drift amplitude of the approximate trajectory sequence of the center of mass within the sliding time window is lower than the drift threshold, the trunk stability event is registered and written to the stability summary field. The stability summary field and the event confidence are jointly written to the event record area of ​​the kinematic event sequence.

[0091] During the stage boundary rule matching process, the stage slicing module reads the action stage template skeleton from the session configuration package bound to the same session number as the keypoint map package, parses the stage number table and the stage boundary event rule table, and solidifies the stage boundary rules as rule matching input. The stage boundary rules consist of an event code sequence pattern, a duration frame threshold, and an event integrity threshold. The event code sequence pattern describes the ordered combination relationship between the stage start event code and the stage end event code. The duration frame threshold describes the minimum duration segment of the stage boundary event, and the event integrity threshold describes the coverage ratio constraint of the stage boundary event within the target frame window. The stage slicing module sorts the kinematic event sequence by occurrence frame number to generate an event sequence index and performs pattern matching on the event sequence index. Pattern matching includes event code sequence pattern comparison, time interval verification, and duration frame verification. When the pattern matching passes, the stage slicing module registers the stage start and end frame numbers and generates a stage slice entry. The stage slice entry writes the stage number, stage start and end frame numbers, stage boundary event index, and stage intra-event index set. The stage intra-event index set points to the event record area index of the kinematic event sequence, thereby providing event anchors for subsequent stage feature token generation. For frame windows with conflicting pattern matching, the stage slicing module generates conflict markers using conflict resolution rules. The conflict resolution rules revolve around event confidence, event calculation weight, and continuous frame verification results. The conflict markers are written into the exception field of the stage slice entry and into the slice exception record, so that the subsequent rule engine can trace the source of stage division.

[0092] During the stage credibility generation process, the stage slicing module calculates the event completeness score for each stage slice entry and combines it with the mean weight of the probe reliability heatmap to generate stage credibility. The event completeness score is synthesized from the number of events covered by the stage boundary event index, the count of missing support switching markers, the count of missing angle markers, and the coverage ratio of jump markers. The mean weight of the probe reliability heatmap is obtained by statistically analyzing the heatmap weights corresponding to the human body bounding box region within the frame window covered by the stage slice entry, and is written into the heatmap weight summary field within the stage slice entry. The stage credibility is synthesized from the event completeness score and the heatmap weight summary field and written into the stage credibility field. The stage credibility field, along with the stage number and the stage start and end frame numbers, serves as the input gating field for the subsequent S600 "generating stage feature tokens based on stage slice packages". After generating the stage credibility field, the stage slicing module assembles the stage slice entries into a stage slice set according to the frame number order, and writes the session number, configuration version number, sampling plan version number and generation timestamp into the header of the stage slice set. Then, the stage slice set and the stage credibility are encapsulated to form a stage slice package, and the kinematic event sequence summary, event sequence index summary and slice anomaly record index are written into the stage slice package, so that the stage slice package has cross-step traceability capability after being transmitted to the alignment residual module. After being read by S600, the stage slice package is used to generate stage feature tokens and generate an alignment index table.

[0093] Summary of the technical effects of this step: This step transforms the key point map package into a kinematic event sequence and completes the stage boundary rule matching. The stage slice set and stage credibility form a traceable stage division result under the same session number and configuration version number link, providing stage anchors and credibility inputs for subsequent weighted alignment and residual organization.

[0094] S600: Based on the stage slice package, generate stage feature tokens and align them according to the probe reliability heatmap to generate an alignment residual package containing the alignment residual tensor and the missing test label set;

[0095] Specifically, the alignment residual module receives the stage slice package output by S500 as input, reads the stage slice set, stage confidence, kinematic event sequence summary, event sequence index summary and slice anomaly record index within the stage slice package, and establishes a processing context based on the session number, configuration version number and sampling plan version number in the stage slice package header. Under this processing context, the alignment residual module reads the session configuration package bound to the same session number to obtain the action stage template skeleton, and extracts the state transition records of the joint visibility state machine from the session log area of ​​the probe registration module, or reads the preset state transition record index field in the stage slice package to locate the partition where the state transition record is located, so that the generation of the missing test label set is consistent with the joint visibility state machine formed by the S200 stage. Before token generation, the alignment residual module performs consistency verification. The consistency verification process revolves around the matching relationship between the configuration version number and the threshold table version number, the continuity of the frame number range of the stage slice set, and the coverage ratio of the jump segment marker in the slice abnormal record index. When there is a stage slice entry whose jump segment marker coverage ratio crosses the threshold, the entry is registered as a low-confidence slice and written into the token gating flag, so that the entry enters the deweighted branch when the alignment index table is generated in the future. At the same time, the frame number range of the abnormal segment is written into the abnormal record area of ​​the alignment residual module and archived with the session number.

[0096] In the stage feature token generation process, the alignment residual module uses each stage slice entry in the stage slice set as the basic processing unit, reading the stage number, stage start and end frame number, event index set within the stage, and stage confidence field one by one. Within the frame window defined by the stage start and end frame number, it extracts the corresponding joint node set and node confidence field from the keypoint map packet buffer to form the token calculation input packet. In this embodiment, the stage feature token is defined as a data carrier that structurally expresses the kinematic morphology of the stage. Its minimum set consists of joint angle vectors, angular velocity vectors, trajectory curvature vectors, left-right symmetry coefficients, and rhythm counts. The joint angle vectors are generated by reshaping the joint angle sequence obtained by mapping the bone segment edge set within the stage frame window using fixed sampling points. The angular velocity vectors... The process involves reshaping adjacent frame differences of the joint angle sequence at the same sampling point within the stage frame window. The trajectory curvature vector is jointly generated by the trajectory arc length sequence and turning change sequence of the joint node set within the stage frame window. The left-right symmetry coefficient is statistically synthesized from the differences between the left and right corresponding joints in the joint angle vector and trajectory curvature vector. The rhythm count is obtained by statistically analyzing the event intervals of the joint angle extreme events and angular velocity zero-crossing events pointed to by the event index set within the stage and written into the rhythm field. The alignment residual module writes the above fields into the stage feature token record and writes the stage number, stage start and end frame number, sampling plan version number, and token digest check code into the record header field. The token digest check code is generated by performing a secure hash algorithm on the field sequence of the stage feature token record for subsequent auditing and traceability. The alignment residual module also generates a token quality marker, which is synthesized from the mean of the stage confidence and node confidence fields and the missing marker count in the slice abnormal record index, and written into the quality field of the stage feature token record, so that the subsequent alignment residual tensor has a weight entry based on the quality field during calculation.

[0097] In the probe reliability heatmap weighted alignment processing, the alignment residual module first generates an alignment index table based on the stage boundary event rule table and stage slice set in the action stage template skeleton. The alignment index table is used to register the correspondence between the target stage feature token and the standard action template token. The standard action template token is located by the feature token dictionary within the stage in the session configuration package and matched with the stage number according to the training item. The alignment index table writes the stage number, token sampling point number, standard template sampling point number and alignment window frame number range field, and binds the alignment window frame number range field with the stage start and end frame numbers, so that the window boundary when connecting across stages has a traceable caliber. Subsequently, the alignment residual module generates a weight vector based on the probe reliability heatmap. In this embodiment, the weight vector is defined as a set of weighted coefficients that map the mean thermal weight, thermal weight stability, and joint dimension weight of the human body bounding box within the stage frame window to the same dimension. The minimum set of the weight vector includes the stage weight coefficient and the joint weight coefficient. The stage weight coefficient is synthesized from the mean thermal weight and the token quality flag, and the joint weight coefficient is synthesized from the joint dimension weight of the heatmap and the node confidence field. When the coverage ratio of the jump segment flag corresponding to a certain stage slice entry in the stage slice set crosses the threshold, the stage weight coefficient of that stage is written into the deweighting flag and simultaneously written into the abnormal record area. The alignment residual module then calculates the alignment residual tensor according to the alignment index table and weight vector. In this embodiment, the alignment residual tensor is defined as a structured stack of the difference results of the stage feature token and the standard action template token in the sampling point dimension, feature dimension, and joint dimension. It is written in a partitioned storage method, and the joint angle vector residual, angular velocity vector residual, trajectory curvature vector residual, left and right symmetry coefficient residual, and rhythm count residual are written into the residual partition respectively. The stage number, alignment window frame number interval field, weight vector summary, and residual source code are registered in the tensor header field. The residual source code is used to distinguish the residual triggered by the weight reduction of the joint weight coefficient and the residual triggered by the weight reduction of the stage weight coefficient, so as to allow the S700 deviation decomposition to perform a traceable decomposition of the residual source.

[0098] In the missing test tag set generation process, the alignment residual module performs state transition encoding on the state transition records within the stage frame window. The state transition encoding process reorganizes the visible state, suspected occlusion state, missing state, and jump state of the joint visibility state machine into a state sequence according to the joint number and frame number, and slices the state sequence according to the alignment window frame number interval field of the alignment index table to obtain the stage missing test fragments. The alignment residual module further maps the stage missing test fragments into a missing test tag set. The missing test tag set is written with the joint number, missing test type code, missing test start and end frame number, and alignment window index field. The missing test type code is obtained by the state type encoding of the suspected occlusion state, missing state, and jump state. The alignment window index field points to the entry index of the alignment index table, so that the missing test tag set has a cross-packet reference path when the S700 generates the evidence index. After generating the missing test label set, the alignment residual module encapsulates the missing test label set and the alignment residual tensor together to form an alignment residual package. The header of the alignment residual package contains the session number, configuration version number, sampling plan version number, generation timestamp, and verification summary field. The verification summary field is synthesized from the alignment index table summary, weight vector summary, residual partition summary, and missing test label set summary. The alignment residual package is written to the circular cache index and transmitted to the decision strategy module as the direct input for S700 to "perform bias deconstruction based on the alignment residual package and generate anomaly decision records and evidence indexes according to the rule engine".

[0099] In the engineering implementation, after the trainee starts a session in the sprint training scenario, the phase slice package output by the phase slice module enters the alignment residual module under the same session number. When the ankle joint shows alternating transitions between suspected occlusion state and jump state due to backlighting during the sprint phase, the state transition record is mapped to a missing test label set by the state transition encoding process and the missing test type code and missing test start and end frame number are written in. The alignment residual module generates joint weight coefficients from the probe reliability heatmap and registers weight reduction marks for ankle joint-related residual partitions. Then, according to the alignment index table, it completes the weighted alignment of the phase feature tokens and standard action template tokens of the sprint phase, and obtains the alignment residual tensor containing the joint angle vector residual and angular velocity vector residual. Together with the missing test label set, it is encapsulated into an alignment residual package, so that the subsequent deviation deconstruction has a consistent index basis when reading the residual source code and the missing test type code.

[0100] Summary of the technical effects of this step: This step transforms the set of stage slices within the stage slice package into stage feature tokens, and introduces the probe reliability heatmap and joint visibility state machine into the alignment calculation link, forming a unified encapsulation of the alignment residual tensor and the missing test label set, so that the deviation deconstruction and evidence index generation of S700 have traceable alignment basis and missing test mark entry.

[0101] S700, based on the aligned residual package, performs deviation deconstruction and generates anomaly judgment records and evidence indexes according to the rule engine, generates error correction instruction packages containing priority, timing, channel, and verification window fields, and updates individualized rule parameter groups.

[0102] Specifically, the decision strategy module receives the alignment residual packet output by S600 as input, reads the alignment residual tensor and missing test label set in the alignment residual packet, and establishes a decision context based on the session number, configuration version number, and sampling plan version number written in the header of the alignment residual packet. At the same time, it reads the threshold table, priority mapping table, and channel mapping table from the session configuration packet bound to the session number to form the current snapshot of the individualized rule parameter group. The alignment residual tensor is a residual stacking structure organized by stage number, alignment window index field, feature partition mark, and joint number. The missing test label set is a missing test registration structure organized by joint number, missing test type code, missing test start and end frame number, and alignment window index field. When accessing, the decision strategy module performs consistency verification on the alignment window index field, registers missing index or index conflict as residual index abnormal records, and writes the abnormal record to the decision log partition associated with the configuration version number, so that the subsequent evidence index generation has an auditable entry point. The deviation deconstruction is a processing chain that splits and aggregates the alignment residual tensor according to a preset dimension. Its inputs are the alignment residual tensor and the alignment window index field, and its outputs are a time-series deviation vector, an amplitude deviation vector, a cooperative deviation vector, and a stability deviation vector. During execution, the decision strategy module first locates the residual partitions according to the stage number and slices the residual fragments according to the alignment window index field. Then, it performs time-series reorganization and partition aggregation on the residual fragments. Time-series reorganization is carried out around the sampling point sequence corresponding to the alignment window index field, and partition aggregation is carried out around the joint angle vector residual, angular velocity vector residual, trajectory curvature vector residual, left-right symmetry coefficient residual, and rhythm count residual. The time-series deviation vector... The sampling point index field is generated by the ordered concatenation of residual fragments along the sampling point sequence dimension. The amplitude deviation vector is generated by the amplitude statistics of the residual fragments and written into the amplitude summary field. The cooperative deviation vector is generated by the correlation aggregation of multi-joint residuals and written into the cooperative joint pair field. The stability deviation vector is generated by the aggregation of high-frequency jitter residuals of the residual fragments and written into the jitter summary field. When generating the four types of deviation vectors, the judgment strategy module simultaneously introduces the missing measurement label set, writes a missing measurement mask mark to the residual fragments covered by the missing measurement type code, and writes the missing measurement mask mark and the amplitude summary field together into the deviation deconstruction intermediate record, so that the subsequent rule engine can distinguish between observation missing and motion anomaly during judgment.

[0103] The decision strategy module inputs the four types of deviation vectors into the rule engine to perform anomaly determination. The rule engine is a rule execution component consisting of a rule entry set, a threshold table, a severity mapping table, a priority mapping table, and a channel mapping table. The rule entry set defines the mapping relationship between deviation type codes and anomaly type codes and the association relationship between stage numbers and rule application domains. The threshold table defines the threshold caliber for each deviation type code under different stage numbers. The severity mapping table defines the mapping caliber between the deviation amplitude summary field and the severity code. The priority mapping table defines the mapping caliber between the anomaly type code and the priority field. The channel mapping table defines the mapping caliber between the anomaly type code and the output channel field. The decision strategy module iterates through the deviation deconstruction results one by one according to the stage number, and inputs the deviation type code, stage number, amplitude summary field, and other parameters into the rule engine. The joint pair field and missing test mask marker are sent to the rule engine. The rule engine first performs threshold matching to obtain the threshold matching result, then performs severity mapping to generate a severity code, and performs rule entry set matching to generate an anomaly type code, finally generating an anomaly determination record. The anomaly determination record includes an anomaly type code, severity code, stage number, and evidence index set. The evidence index set is an index set that associates the anomaly determination record with the upstream data structure. Its fields consist of alignment window index field, residual partition marker, joint number, missing test type code reference, and stage slice entry index. When the evidence index set is generated, the session number and configuration version number are written at the same time, so that the anomaly determination record can return to the node index of the key point map package and the stage start and end frame number of the stage slice set when tracing across main steps.

[0104] During the generation of the error correction instruction package, the judgment strategy module arranges and processes the input instructions from the anomaly judgment record and outputs a set of instruction fragments, which constitute the error correction instruction package. Specifically, the priority field of the error correction instruction package is mapped from the priority mapping table along the anomaly type code dimension and written into the priority field; the timing field is mapped from the stage start and end frame numbers corresponding to the stage number and the output refresh cycle threshold and written into the timing field; the channel field is mapped from the channel mapping table and written into the channel field; and the verification window field is combined from the stage start and end frame numbers, the alignment window index field, and the preset verification window length and written into the verification window field. The judgment strategy module, during arrangement, performs... Multiple anomaly judgment records with the same stage number undergo conflict resolution. Conflict resolution is carried out based on the overlap ratio of the priority field, severity code, and evidence index set. Records with an overlap ratio exceeding the threshold are merged into the same instruction fragment and written with a merge mark. Records with an overlap ratio below the threshold are retained as independent instruction fragments and written with a parallel mark. When the error correction instruction package is encapsulated, the session number, configuration version number, generation timestamp, and instruction digest check code are written into it and transmitted to the instruction output module. The instruction output module outputs the corresponding instruction fragment in the display channel, voice channel, or vibration channel according to the channel field. At the same time, the verification window field is written into the verification scheduling table and sent back to the judgment strategy module as the rule parameter update trigger condition.

[0105] During the update of the individualized rule parameter group, the parameter update module reads the alignment window index field pointed to by the verification window field when the verification schedule table triggers the corresponding verification window ends, and reads back the new round of alignment residual tensor fragments in the window to perform residual fallback amplitude statistics. The residual fallback amplitude statistics are the fallback summary field obtained by statistically analyzing the fallback amplitude of the amplitude summary field associated with the same anomaly type code in the verification window. The parameter update module recalibrates the fallback summary field and the anomaly type code into the threshold table, writes the recalibrated threshold entries into the new version of the threshold table, and writes the joint distribution of severity code and priority field into the new version of the priority mapping table. The channel mapping table is remapped when the channel congestion flag is triggered. The channel congestion flag is generated by the channel queuing delay summary returned by the instruction output module and archived with the session number. The updated individualized rule parameter group is written with the parameter group version number, change reason code and parameter digest verification code, and written back to the session configuration module. This enables the subsequent session configuration package generated by S100 to read the new version threshold table, priority mapping table and channel mapping table under the same training project, forming a closed-loop evolution record across sessions.

[0106] In the engineering implementation, when a trainee enters the sprint phase of a sprint training session, the phase number in the phase slice set corresponds to the sprint phase. The judgment strategy module reads the alignment window index field corresponding to the phase number and generates an amplitude deviation vector and a stability deviation vector for the ankle-related residual partition. It also detects the suspected occlusion state missing test type code associated with the alignment window index field in the missing test label set. Based on this, the rule engine generates an anomaly type code and a severity code and forms an evidence index set. The instruction orchestration process maps the anomaly judgment record to the instruction fragment of the display channel corresponding to the channel field and writes it into the verification window field. When the verification window ends, the parameter update module reads back the new residual fragment corresponding to the same alignment window index field to generate residual falloff amplitude statistics and writes them into the new version of the threshold table and the new version of the priority mapping table. The updated individualized rule parameter group is written back to the session configuration module for reference when the next session configuration package is generated.

[0107] In summary, the technical effects of this step are as follows: This step transforms the alignment residual tensor and missing label set in the alignment residual package into a traceable set of anomaly determination records and evidence indexes, organizes the anomaly determination results into error correction instruction packages, and forms an evolution chain of version-auditable individualized rule parameter groups through verification window-driven rule parameter updates.

[0108] Example 2: Figure 2 A structural block diagram of a motion capture system based on sports training according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0109] Session configuration module 01 generates a session configuration package and transmits it to the probe registration module. Specifically, the session configuration module receives the individualized rule parameter group written back by the parameter update module and loads it into the configuration storage area of ​​the session configuration module. It reads the version record of the individualized rule parameter group associated with the current session trigger in the configuration storage area, performs field caliber fixing, version field registration, verification summary generation and timestamp alignment registration processing to form a session configuration package. The session configuration package writes the summary field, version field and output channel declaration field of the individualized rule parameter group and writes them into the cache index area of ​​the session configuration module. The session configuration module transmits the session configuration package to the probe registration module through the module interface, and at the same time registers the generation record and transmission record of the session configuration package in the configuration storage area and associates them with the version field.

[0110] The probe registration module 02 generates a probe registration package based on the session configuration package and transmits it to the precision measurement and orchestration module. Specifically, the probe registration module receives the session configuration package from the session configuration module, parses the version field, verification summary field, and output channel declaration field in the session configuration package, completes the acquisition channel binding, acquisition parameter loading, and channel identifier registration processing, and performs timestamp source verification, frame sequence continuity verification, image quality marking, and occlusion marking processing after the preview frame sequence enters the probe registration module, generating a quality record associated with the frame number range. Based on the quality record, the probe registration module performs geometric relationship registration and state transition registration processing to form a probe registration package. The probe registration package writes the geometric relationship registration table field, reliability heatmap field, joint visibility state machine field, and anomaly index field, and registers the version field and verification summary field consistent with the session configuration package in the packet header. The probe registration module transmits the probe registration package to the precision measurement and orchestration module, and simultaneously writes the state transition registration record into the session log area of ​​the probe registration module and associates it with the version field.

[0111] The precision measurement and orchestration module 03 generates a precision measurement and orchestration package based on the probe registration package and transmits it to the key point map module. Specifically, the precision measurement and orchestration module receives the probe registration package from the probe registration module, reads the geometric relationship registration table field, reliability heatmap field, joint visibility state machine field, and anomaly index field from the probe registration package, performs real-time frame segment coarse screening, gating mark generation, trigger queue assembly, and quota constraint pruning, and registers the pruning results and time window slicing results as a scheduling field set. The precision measurement and orchestration module performs version field alignment registration and index reorganization on the scheduling field set to generate a precision measurement and orchestration package. The precision measurement and orchestration package writes the trigger index field, frame number sequence field, time window field, and anomaly record field, and registers the version field and verification summary field consistent with the probe registration package in the package header. The precision measurement and orchestration module transmits the precision measurement and orchestration package to the key point map module, and at the same time registers the correspondence between the trigger index field and the frame number sequence field in the scheduling buffer of the precision measurement and orchestration module for back-reading and verification by the key point map module.

[0112] The keypoint mapping module 04 generates a keypoint mapping package based on the precision measurement and orchestration package and transmits it to the stage slicing module. Specifically, the keypoint mapping module receives the precision measurement and orchestration package from the precision measurement and orchestration module, parses the trigger index field, frame number sequence field, and time window field in the precision measurement and orchestration package, calls the acquisition channel according to the trigger index field to load the corresponding frame segment and performs preprocessing mark registration processing, and then performs keypoint inference, node set construction, edge set assembly, confidence field registration, and constraint field registration processing to form the basic record of the keypoint mapping. The keypoint mapping module performs consistency verification and repair mark processing on the basic record of the keypoint mapping, writes the consistency residual field, repair mask field, and abnormal index field into the keypoint mapping package, and registers the version field and verification summary field consistent with the precision measurement and orchestration package in the package header. The keypoint mapping module transmits the keypoint mapping package to the stage slicing module, and at the same time writes the keypoint mapping package into the circular cache index area of ​​the keypoint mapping module and registers the frame number range field, so that the stage slicing module can read back and locate according to the frame number range field.

[0113] The stage slicing module 05 generates a stage slice package based on the keypoint map package and transmits it to the alignment residual module. Specifically, the stage slicing module receives the keypoint map package from the keypoint map module, reads the node set field, edge set field, confidence field, constraint field, consistency residual field, and anomaly index field from the keypoint map package, performs event extraction, event sequence index assembly, and slice anomaly registration processing, and calls the session configuration package stored in the session configuration module to read the stage boundary rule field to complete rule matching. The stage slicing module registers the stage number field and the stage start and end frame number field according to the rule matching result, generates the stage slice set field, and registers the stage confidence field and evidence index field on the stage slice set field. The stage slicing module encapsulates the stage slice set field and the stage confidence field into a stage slice package, and registers the version field and verification summary field consistent with the keypoint map package in the package header. The stage slicing module transmits the stage slice package to the alignment residual module, and at the same time registers the associated records of the stage number field, the stage start and end frame number field, and the anomaly index field in the slice log area of ​​the stage slicing module.

[0114] The alignment residual module 06 generates an alignment residual package based on the stage slice package and transmits it to the decision strategy module. Specifically, the alignment residual module receives the stage slice package from the stage slice module, parses the stage slice set field, stage credibility field, and evidence index field in the stage slice package, and calls the session configuration package stored in the session configuration module to read the standard template field set to complete the index binding. Within the stage start and end frame number field range defined by the stage slice set field, the alignment residual module reads back the key point map package cache record of the key point map module, performs token reorganization, alignment index assembly, weight digest generation, and weighted alignment processing to form residual words. The segment set is processed and the residual source code field is registered in the residual field set; the alignment residual module synchronously reads the joint visibility state machine field or its log index record from the probe registration module, performs state transition encoding processing to generate a missing test field set and registers the alignment window index field in the missing test field set; the alignment residual module encapsulates the residual field set and the missing test field set into an alignment residual package and registers the version field and verification digest field consistent with the stage slice package in the package header; the alignment residual module transmits the alignment residual package to the decision strategy module, and at the same time registers the residual index exception record and the missing test field set index record in the exception record area of ​​the alignment residual module.

[0115] The judgment strategy module 07 generates an error correction instruction package based on the aligned residual package and transmits it to the instruction output module. Simultaneously, it outputs an individualized rule parameter group to the parameter update module. Specifically, the judgment strategy module receives the aligned residual package from the aligned residual module, reads the residual field set, missing test field set, and residual source code field from the aligned residual package, and performs deviation deconstruction, rule entry matching, threshold caliber comparison, severity code generation, priority arbitration, and evidence index set registration processing to form an anomaly judgment record and write it into the anomaly type code field, severity code field, and stage number field. The judgment strategy module in... Based on the anomaly determination record, instruction orchestration processing is performed, writing the priority field, timing field, channel field, and verification window field into the error correction instruction package, and registering the version field and verification summary field consistent with the alignment residual package in the package header; the determination strategy module transmits the error correction instruction package to the instruction output module, and at the same time encapsulates the recalibration input field and version change record of the anomaly determination record into an individualized rule parameter group and outputs it to the parameter update module. The individualized rule parameter group is written with the version field, change reason code field, and parameter summary field, and is indexed and associated with the verification window field of the error correction instruction package.

[0116] The parameter update module 08 updates the individualized rule parameter group and writes it back to the session configuration module. Specifically, the parameter update module receives the individualized rule parameter group from the decision strategy module, parses the version field, change reason code field, and parameter summary field in the individualized rule parameter group, performs parameter loading, version chain registration, conflict entry arbitration, and recalibration writing processing, and reads back the residual field set fragment of the corresponding window when the verification window field trigger condition returned by the instruction output module is reached, completes the generation of the fallback summary field, and writes it to the update record area of ​​the individualized rule parameter group; the parameter update module writes the updated individualized rule parameter group back to the session configuration module and registers the write-back timestamp field and write-back confirmation field. After receiving the write-back, the session configuration module loads the individualized rule parameter group into the configuration storage area and updates the generation caliber of the session configuration package, so that the session configuration package subsequently received by the probe registration module is consistent with the version field.

Claims

1. A motion capture method based on sports training, characterized in that, include: S100: Obtain training items, computing power indicators, and camera parameters, and generate a session configuration package containing resource budget parameter groups and action phase template skeletons; S200, based on session configuration package and preview video stream, performs deployment guidance, self-calibration, and reliability modeling, and generates a probe registration package containing probe geometry relationship registration table, probe reliability heat map, and joint visibility state machine; S300, based on the probe registration package, performs coarse screening judgment, fine measurement triggering, and key stage sampling arrangement, and generates a fine measurement arrangement package containing fine measurement triggering sequence and sampling plan; S400: Based on the precision measurement and arrangement package, perform precision measurement of key points, construct virtual probe maps of key points, perform consistency verification and repair marking, and generate key point map packages; S500: Based on the key point map package, extract kinematic event sequences and match stage boundary rules to generate a stage slice package containing a stage slice set and stage credibility. S600: Based on the stage slice package, generate stage feature tokens and align them weighted according to the probe reliability heatmap to generate an alignment residual package containing the alignment residual tensor and the missing test label set; S700, based on the aligned residual package, performs deviation deconstruction and generates anomaly judgment records and evidence indexes according to the rule engine, generates error correction instruction packages containing priority, timing, channel, and verification window fields, and updates individualized rule parameter groups.

2. The method according to claim 1, characterized in that, The resource budget parameter set includes a single-frame inference latency threshold, a precision frame quota, an output refresh cycle threshold, a stage sampling rate table, and a cache occupancy threshold; the action stage template skeleton includes a stage number table, a stage boundary event rule table, and a stage feature token dictionary.

3. The method according to claim 1, characterized in that, The deployment guidance includes: extracting the character proportion score, image jitter score, backlight probability score, and occlusion ratio score from the preview video stream, and generating a deployment prompt record; The self-calibration includes: estimating the field of view coverage boundary, human scale consistency parameters, and key joint visibility distribution within the guided action segment, and updating the probe geometric relationship registration table; The joint visibility state machine includes a visible state, a suspected occlusion state, a missing state, and a jump state.

4. The method according to claim 1, characterized in that, The coarse screening determination includes: The real-time video stream is downsampled and low-resolution human body region localization is performed, motion energy sequences are calculated, and effective action marker sequences are generated. The precision measurement trigger is based on the effective action marker sequence and the joint visibility state machine to generate a gated marker sequence, and the precision measurement trigger sequence is generated according to the gated marker sequence; The key-stage sampling orchestration is based on the stage sampling multiplier table to generate a key-stage sampling plan.

5. The method according to claim 1, characterized in that, The keypoint virtual probe map includes a set of joint nodes, a set of bone segments and edges, a node confidence field, and an edge constraint field. The consistency check includes: calculating the consistency residual based on the length interval constraint, angle interval constraint and node velocity smoothing constraint of the bone segment edge set; The repair markers include: generating a repair mask matrix for nodes whose consistency residuals exceed a threshold and writing it into the key point map package.

6. The method according to claim 1, characterized in that, The kinematic event sequence consists of joint angle extreme value events, angular velocity zero crossing events, support switching events, and trunk stability events; The stage boundary rules consist of the event code sequence pattern, the duration frame threshold, and the event integrity threshold. The stage credibility is synthesized by the mean weight of the probe reliability heatmap in the stage slice set and the event completeness score.

7. The method according to claim 1, characterized in that, The stage feature token consists of joint angle vector, angular velocity vector, trajectory curvature vector, left-right symmetry coefficient, and rhythm count; The weighted alignment includes: generating an alignment index table based on stage boundary rules, generating a weight vector based on probe reliability heatmap, and calculating the alignment residual tensor according to the alignment index table and the weight vector; The missing label set is generated by the state transition encoding of the joint visibility state machine.

8. The method according to claim 1, characterized in that, The deviation deconstruction includes: generating a temporal deviation vector based on the alignment index table, generating an amplitude deviation vector based on the amplitude statistics of the alignment residual tensor, generating a collaborative deviation vector based on the multi-joint related residuals, and generating a stability deviation vector based on the high-frequency jitter residuals; the anomaly determination record includes an anomaly type code, severity code, stage number, and evidence index set.

9. The method according to claim 1, characterized in that, The error correction instruction package consists of a set of instruction fragments, each of which includes a target stage number, a target deviation type code, a priority field, an output timing field, an output channel field, and a verification window field. The individualized rule parameter group includes a threshold table, a priority mapping table, and a channel mapping table. The update recalibrates the threshold table and the priority mapping table by statistically analyzing the residual fall-off magnitude within the verification window.

10. A motion capture system based on sports training, applied to the method of any one of claims 1-9, characterized in that, include: The session configuration module generates a session configuration package and sends it to the probe registration module; The probe registration module generates a probe registration package based on the session configuration package and transmits it to the precision measurement orchestration module. The precision measurement and arrangement module generates a precision measurement and arrangement package based on the probe registration package and transmits it to the key point map module; The key point mapping module generates a key point mapping package based on the precision measurement and arrangement package and transmits it to the stage slicing module; The stage slicing module generates a stage slicing package based on the key point map package and sends it to the alignment residual module; The alignment residual module generates an alignment residual package based on the stage slice package and sends it to the decision strategy module; The decision strategy module generates an error correction instruction package based on the alignment residual package and sends it to the instruction output module, while simultaneously outputting an individualized rule parameter group to the parameter update module. The parameter update module updates the individualized rule parameter group and writes it back to the session configuration module.

Images