A smart mirror fitness training system based on motion recognition

By employing adaptive lighting adjustment, key point recognition, occlusion detection, and posture compensation technologies, the problem of motion recognition errors caused by occlusion in home fitness has been solved, enabling accurate training feedback and user guidance in complex environments.

CN121868834BActive Publication Date: 2026-11-10ZHEJIANG SU PRILIGY TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610027760.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-01-09
Publication Date
2026-11-10
Estimated Expiration
2046-01-09

AI Technical Summary

Technical Problem

Existing motion recognition-based smart fitness systems struggle to accurately identify key movements when users are obstructed by furniture or other objects during home workouts, leading to incorrect recognition results and impacting user experience and training effectiveness.

Method used

The image acquisition module performs adaptive lighting adjustment and generates a clear training image sequence by combining environmental depth information; the pose analysis module performs key point recognition and skeleton mapping, the occlusion judgment module detects occluded areas and calculates confidence; the pose compensation module dynamically compensates for missing key points to generate complete pose data; the action evaluation module performs action evaluation, and the feedback presentation module provides real-time guidance.

Benefits of technology

Maintaining the stability of action recognition in complex home scenarios, providing accurate training feedback, improving user experience and training effectiveness, and achieving complete pose reconstruction and action evaluation under occlusion conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121868834B_ABST
    Figure CN121868834B_ABST
Patent Text Reader

Abstract

This invention provides an intelligent mirror fitness training system based on motion recognition, belonging to the field of data processing technology. The system includes: an image acquisition module; a posture analysis module; an occlusion determination module, used to perform continuous analysis of the displacement changes of continuous key points to determine the occlusion area; and calculate the occlusion confidence of each occlusion area according to the temporal and spatial changes; a posture compensation module, used to dynamically compensate for missing key points, and perform spatial interpolation and temporal prediction based on the geometric constraints of unoccluded key points and the displacement trend curve of historical posture data; a motion evaluation module, used to decompose the current motion into multiple local joint motion segments, and compare the angle change curve of each segment with a preset motion template segment by segment to generate motion evaluation results; and a feedback presentation module, used to generate comprehensive training feedback data based on the motion evaluation results. This invention improves the autonomy and accuracy of the intelligent mirror fitness training system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent mirror fitness training system based on motion recognition. Background Technology

[0002] Currently, motion recognition-based smart fitness systems primarily utilize cameras installed in gyms or homes, combined with deep learning-based human posture recognition algorithms, to achieve real-time recognition and evaluation of user training movements. For example, some smart fitness mirror products use RGB or depth cameras to capture key points on the user's body, generating skeletal coordinates through convolutional neural networks (CNNs) or posture estimation algorithms (such as OpenPose and MediaPipe Pose), and comparing these coordinates with a standard motion database to determine if the user's movements are correct. Furthermore, these systems typically incorporate voice prompts or on-screen feedback to provide users with movement correction suggestions and record data such as training repetitions and angle deviations, enabling personalized exercise guidance.

[0003] However, in specific application scenarios, such as home fitness, existing systems often struggle to accurately identify key movement points when the user's body is partially obscured by furniture or other objects. For instance, when a user is performing squats or lunges, if their lower limbs are obscured by objects such as a yoga mat or chair legs in front of a mirror, the system's camera may capture incomplete or misjudged skeletal data, leading to incorrect recognition results or even incorrect corrective suggestions. In such cases, the system might determine that the user "did not squat" or that their "posture deviation was too large," distorting the training feedback and severely impacting the user experience and training effectiveness. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent mirror fitness training system based on motion recognition, which aims to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A smart mirror fitness training system based on motion recognition, the system comprising: The image acquisition module is used to acquire multiple frames of user pose image data in real time through the camera installed on the smart mirror, and combine environmental depth information and light intensity parameters to adaptively adjust the image sharpness and generate a training image sequence. The pose parsing module is used to identify key points and map skeletons of the human body based on the training image sequence, including temporal smoothing and noise removal of key point trajectories in adjacent frames, and generating a pose data sequence containing spatial coordinates and motion trends. The occlusion determination module is used to perform continuous analysis on the displacement changes of continuous key points based on the posture data sequence. When the trajectory interruption or displacement change of a specific key point is detected, it determines that there is an occlusion area in the corresponding body part. Based on the temporal and spatial changes of the occlusion features, it calculates the occlusion confidence of each occlusion area and generates occlusion feature label data containing occlusion location information and confidence information. The pose compensation module is used to dynamically compensate for missing key points based on occlusion feature labeling data. Based on the geometric constraints of unoccluded key points and the displacement trend curve of historical pose data, it performs spatial interpolation and temporal prediction to generate complete pose reconstruction data. The motion assessment module is used to decompose the user's current motion into multiple local joint motion segments based on the complete posture reconstruction data, and compare the angle change curve of each segment with the preset motion template segment by segment to generate motion assessment results. The feedback presentation module is used to generate comprehensive training feedback data based on the action evaluation results, including local action heatmaps, correction direction prompts, and rhythm guidance information. It also generates visual training guidance information for users through the smart mirror display interface and voice module.

[0006] The above-described solution of the present invention has at least the following beneficial effects: First, this invention introduces an environmental modeling and adaptive adjustment mechanism during the image acquisition stage. By acquiring information on illumination intensity, reflection distribution, and depth, an environmental feature model is established. This allows the system to dynamically adjust the camera's exposure time and gain threshold based on the actual brightness and reflection conditions of the training scene, thus obtaining clear and stable pose images even under uneven indoor lighting or reflective background conditions. Compared to existing technologies that rely solely on fixed parameters for image acquisition, this approach effectively reduces the impact of background interference on keypoint detection, ensuring stable and consistent image data quality and providing a reliable input foundation for subsequent action recognition and pose analysis.

[0007] Secondly, this invention uses a pose analysis module to perform keypoint recognition and skeleton mapping on the human body in the training image sequence. It then combines this with temporal smoothing and noise removal processing based on keypoint trajectory changes between adjacent frames, thereby obtaining a pose data sequence with continuous spatial coordinates and stable dynamic trends. This module not only improves the spatial accuracy of keypoint detection but also effectively filters out slight jitter and recognition noise during movement, making the action recognition process closer to the user's actual movement trajectory and improving the temporal coherence of pose recognition.

[0008] Furthermore, this invention addresses the low recognition rate of existing systems when the user's body is partially occluded by incorporating an occlusion detection and posture compensation mechanism. The occlusion detection module automatically identifies the occluded parts of the user's body by detecting the displacement continuity and trajectory abrupt changes of key points, and calculates the occlusion confidence level, enabling the system to determine the extent and stability of the occlusion. The posture compensation module, based on occlusion feature marker data and the geometric constraints of unoccluded key points, combined with displacement trend curves from historical frames, performs spatial interpolation and temporal prediction for missing key points, thus reconstructing complete human posture data even when the user's body is partially occluded. This technology not only solves the problem of missing key points caused by occlusion but also reduces the interference of incorrect postures on training feedback, enabling the system to maintain stable recognition performance in complex home scenarios.

[0009] Furthermore, the motion evaluation module of this invention, through structured decomposition of complete posture data, divides user movements into multiple local joint motion segments and compares them segment by segment with the standard angle change curve of a preset motion template, thereby achieving quantitative analysis of motion angle accuracy, time synchronization, and motion stability. This mechanism can output evaluation results containing multi-dimensional scoring information, enabling the system not only to determine whether the movement is completed, but also to identify the deviation points and durations in the user's movement, providing a data foundation for personalized correction and training optimization.

[0010] Finally, this invention generates visualized training feedback data from the evaluation results through a feedback presentation module, including local motion heatmaps, posture correction direction prompts, and rhythm guidance information, which are presented synchronously through the smart mirror's display interface and the voice module. In this way, users can obtain information such as posture deviation location, correction direction, and rhythm beat in real time during training, forming a visual and auditory dual-channel feedback loop. This method not only improves the interactive experience but also achieves dynamic closed-loop control throughout the entire process from motion acquisition, recognition, compensation, evaluation to guidance. For example, when a user performs squat training in a home environment, even if the lower limbs are partially obscured by a yoga mat, the system can still predict and recover leg joint data through historical posture trends and accurately determine the squat depth and knee joint angle deviation. Simultaneously, it displays the posture correction direction in real time on the mirror display interface and provides a voice prompt to "keep your back straight," thus achieving a complete, continuous, and intelligent training guidance effect. Attached Figure Description

[0011] Figure 1 This is an architecture diagram of an intelligent mirror fitness training system based on motion recognition, provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention proposes an intelligent mirror fitness training system based on motion recognition, the system comprising: The image acquisition module is used to acquire multiple frames of user pose image data in real time through the camera installed on the smart mirror, and combine environmental depth information and light intensity parameters to adaptively adjust the image sharpness and generate a training image sequence. The pose parsing module is used to identify key points and map skeletons of the human body based on the training image sequence, including temporal smoothing and noise removal of key point trajectories in adjacent frames, and generating a pose data sequence containing spatial coordinates and motion trends. The occlusion determination module is used to perform continuous analysis on the displacement changes of continuous key points based on the posture data sequence. When the trajectory interruption or displacement change of a specific key point is detected, it determines that there is an occlusion area in the corresponding body part. Based on the temporal and spatial changes of the occlusion features, it calculates the occlusion confidence of each occlusion area and generates occlusion feature label data containing occlusion location information and confidence information. The pose compensation module is used to dynamically compensate for missing key points based on occlusion feature labeling data. Based on the geometric constraints of unoccluded key points and the displacement trend curve of historical pose data, it performs spatial interpolation and temporal prediction to generate complete pose reconstruction data. The motion assessment module is used to decompose the user's current motion into multiple local joint motion segments based on the complete posture reconstruction data, and compare the angle change curve of each segment with the preset motion template segment by segment to generate motion assessment results. The feedback presentation module is used to generate comprehensive training feedback data based on the action evaluation results, including local action heatmaps, correction direction prompts, and rhythm guidance information. It also generates visual training guidance information for users through the smart mirror display interface and voice module.

[0014] In this embodiment of the invention, the image acquisition module can automatically adapt to different lighting and reflection conditions according to different training scenarios, dynamically adjusting the camera's exposure and gain. This ensures that the acquired multi-frame pose image data maintains high clarity even under varying brightness, reflective interference, or complex backgrounds. The resulting training image sequence provides a stable data foundation for subsequent pose recognition, reducing recognition errors caused by changes in lighting or reflections.

[0015] The pose parsing module extracts human keypoints and maps skeletons from training image sequences. Combining this with the continuity of inter-frame trajectories, it smooths out jitter and noise in motion, outputting a pose data sequence containing spatial location and motion trends. This process ensures accurate capture of dynamic changes in various parts of the human body, providing reliable input for subsequent occlusion analysis and pose compensation.

[0016] When pose data is discontinuous, the occlusion detection module can analyze the temporal changes in keypoint displacements to determine whether body parts are occluded. It then combines temporal persistence and spatial distribution information to calculate the confidence level of each occluded region, generating occlusion feature data with confidence weights. This mechanism helps distinguish between occasional and persistent occlusion, improving the accuracy of subsequent compensation processes.

[0017] The pose compensation module dynamically reconstructs missing keypoints based on occlusion feature data. This module uses the geometric constraints between unoccluded keypoints and historical pose change trends to perform spatial interpolation and temporal prediction on occluded areas, forming continuous and complete pose data. When abnormal jumps are detected during pose compensation, the system can trigger inter-frame smoothing correction and update constraint parameters to ensure the continuity and coordination of the human pose sequence.

[0018] The motion assessment module, after obtaining complete posture data, breaks down the user's movements into multiple joint motion segments and compares them against a preset standard template. The system quantifies the deviations in each joint's movement by calculating angle changes, time distribution, and motion stability, generating motion assessment results that include dimensions of angle accuracy, time synchronization, and stability, providing a data-driven reference for individual training levels.

[0019] The feedback presentation module generates training feedback data based on the evaluation results, including motion heatmaps, correction direction prompts, and rhythm guidance information. This data is presented in real time through mirror display and voice output, enabling users to receive immediate posture correction and rhythm guidance during training, thus creating a dynamic and interactive training experience.

[0020] In a preferred embodiment of the present invention, the method for setting a preset action template specifically includes: First, the system establishes a movement classification system based on different training types, using common training movements such as squats, lunges, push-ups, planks, and shoulder presses as basic movement categories. The standard movements for each category are performed by professional trainers or sports coaches according to proper form and recorded multiple times to ensure the accuracy of the movements.

[0021] Next, the system uses a pose recognition model to extract key points from the acquired standard motion videos, generating a standard motion skeleton sequence from the coordinates of multiple key points. The system further calculates information such as joint angle change curves, angular velocity change sequences, core force exertion points, and the time range for motion completion at each stage of the motion. To improve adaptability, the system averages multiple standard sampling results and forms a unified template by fitting common trends at key moments.

[0022] Subsequently, the system generates different difficulty versions of the same action based on different user skill levels (such as beginner, intermediate, and advanced). The differences between the difficulty versions mainly lie in the range of motion, rhythm requirements, and stability thresholds, thus making the template more universal.

[0023] Finally, the system stores the processed standard angle curves, stage division information, and time series parameters as template reference data for subsequent action comparison, temporal matching, and score generation. Those skilled in the art can use the above steps to generate action templates with unified standards, stable structures, and applicability to different training types.

[0024] In this embodiment of the invention, the attitude resolution module specifically includes: First, a pre-defined human pose recognition model is used to detect key points frame by frame in the training image sequence received from the image acquisition module. This model can identify multiple skeletal nodes of the human body, such as the head, shoulders, elbows, hips, knees, and ankles, and output their two-dimensional coordinate data in the image coordinate system. Subsequently, the system uses depth information or camera calibration parameters to map the two-dimensional coordinates to three-dimensional spatial coordinates, thereby obtaining the position data of each key point in three-dimensional space.

[0025] Next, to reduce the impact of image jitter, lighting variations, and sensor noise on the recognition results, the system performs temporal smoothing on the keypoint trajectories of consecutive frames. This process smooths out short-term random fluctuations by calculating the average displacement trend of each keypoint across multiple time frames, preserving the continuity of natural human movement. For abnormal keypoint data that significantly deviates from the motion trend, the system removes noise through interpolation of adjacent frames and correction of historical pose data trends, ensuring the stability and accuracy of the pose data.

[0026] Finally, based on the time-smoothed keypoint data, the system performs skeletal mapping on the human body structure, connecting each keypoint according to the human body's topology to form a skeletal model reflecting the relationships between the human joints. Simultaneously, the system extracts the direction of displacement change and the rate of velocity change between adjacent frames, generating motion trend information reflecting the user's dynamic posture changes. This step yields a complete posture data sequence, providing an accurate input basis for occlusion detection and motion evaluation.

[0027] In this embodiment of the invention, the feedback presentation module specifically includes: First, the system receives motion assessment data, including the angular deviation amplitude of each joint, time synchronization score, and motion stability score. Based on this data, corresponding visual feedback is generated. By mapping the angular deviation values ​​to color gradients, the system generates a local motion heatmap on the human skeleton map, highlighting areas with larger deviations as warm-colored regions, thus visually displaying the user's main motion deviation locations.

[0028] Subsequently, based on the angular deviation direction parameters in the evaluation results, the system calculates the correction direction for each joint in space. This direction is determined by the average trend of the deviation vector and is indicated by an arrow on the mirrored display interface, providing the user with intuitive posture correction prompts. For example, when insufficient knee flexion angle is detected, the system generates a downward arrow in the leg area, instructing the user to squat further.

[0029] In terms of rhythm guidance, the system calculates the stability of the user's movement rhythm based on the movement time distribution data and generates rhythm guidance signals. When the system detects that the movement rhythm is too fast or too slow, it broadcasts prompts such as "slow down the rhythm" or "maintain the speed" through the voice module and displays the current rhythm progress in the form of a beat bar on the display interface to help users establish a stable training rhythm.

[0030] Finally, the system integrates heat distribution, correction direction, and rhythm information into a comprehensive feedback interface, which is output synchronously with voice broadcast, forming a multimodal interactive experience combining images, text, and voice. Through this process, users can understand their own movement deviations and rhythm control in real time during training, thereby obtaining continuous and accurate movement correction feedback without the need for external coaching, achieving the goal of self-training optimization.

[0031] In a preferred embodiment of the present invention, the image acquisition module includes: The environment modeling unit is used to extract information on light intensity, reflection distribution and background depth based on multiple frames of user pose image data, establish an environment feature model and generate environment parameter data; The adaptive adjustment unit is used to dynamically adjust the exposure time and gain threshold of the camera based on environmental parameter data. When the rate of change of light exceeds the preset light change threshold, it performs fast compensation processing to generate exposure correction image data. The reflection correction unit is used to detect high reflectivity areas in the image based on the exposure correction image data, and to perform brightness equalization correction on the reflectivity areas by interpolation of neighboring frame pixels to generate correction frame data. The multi-frame fusion unit is used to perform motion vector alignment and temporal weighted fusion between adjacent frames based on the correction frame data to generate a training image sequence.

[0032] In this embodiment of the invention, the quality and adaptability of the training image sequence can be effectively improved through multi-level data acquisition and optimization processing of the image acquisition module. The environment modeling unit can extract illumination, reflection, and depth information from multiple frames of images to establish an environment parameter model, enabling the system to have the ability to self-recognize scene features. When the user is in a training scene with rapidly changing light or strong reflection interference, the adaptive adjustment unit dynamically adjusts the exposure time and gain threshold according to the environmental parameters, enabling the camera to quickly complete brightness compensation when the light changes abruptly, thereby ensuring the continuity and stability of image details. The reflection correction unit further eliminates visual noise caused by factors such as background specular reflection and sweat reflection by detecting high-reflection areas and performing interpolation correction of adjacent frame pixels, ensuring that the key point recognition area maintains true contrast. The multi-frame fusion unit utilizes the temporal consistency of motion vectors in adjacent frames to perform weighted superposition and temporal smoothing of multiple frames of images, thereby eliminating jitter and blur in dynamic actions and obtaining a clearer and more stable training image sequence. The collaborative processing of this module enables subsequent pose recognition and motion analysis to be based on high-quality image input, reducing recognition errors and pose loss rate, making it particularly suitable for user fitness training processes in complex multi-light source scenarios.

[0033] In a preferred embodiment of the present invention, the method for setting the gain threshold specifically includes: First, the system collects raw image data under different lighting conditions in the training environment. By measuring the image brightness distribution, noise intensity, and exposure saturation, it determines the typical response range of the camera under low light, high light, and normal lighting conditions.

[0034] Subsequently, the system determined the allowable gain adjustment range while ensuring image recognizability by statistically analyzing the changing trends of image noise and brightness under different lighting scenarios. Following the principle of appropriately increasing gain in low light and decreasing gain in high light, the system determined the most reasonable gain range through correlation analysis between image sharpness and noise performance.

[0035] Next, the system sets the gain threshold in stages based on the degree of brightness deviation. For example, when the brightness is below a certain level, the gain is increased to a medium level; when the brightness decreases further, it is increased to the maximum safe gain; if the brightness is too high, the gain is reduced to avoid overexposure. This setting process uses a step-by-step adjustment method to make the gain adjustment more continuous and stable.

[0036] Finally, the system solidifies the gain threshold list into environmental adaptation parameters, enabling the image acquisition module to quickly select the corresponding threshold for exposure adjustment based on real-time lighting conditions.

[0037] In a preferred embodiment of the present invention, the method for setting a preset illumination change threshold specifically includes: First, the system calculates the average brightness value of multiple consecutive frames of images to obtain a brightness change sequence, which is used to characterize the stability of ambient lighting. By detecting the difference in brightness at multiple moments, the system can determine whether there are rapid changes in lighting.

[0038] Subsequently, the system analyzes illumination fluctuation samples from typical training environments (such as homes, gyms, and outdoor shaded areas) to determine the rate of brightness change in different scenarios and establishes a reasonable threshold based on the user's acceptable level of image stability. For example, if the brightness changes by more than a certain margin within a short period, it will be considered a sudden change in illumination, requiring the triggering of a rapid compensation mechanism.

[0039] Next, the system divides the rate of change of illumination into multiple levels and determines the threshold corresponding to different levels through gradient evaluation, so that the system can execute different levels of exposure compensation strategies according to the actual degree of change.

[0040] Ultimately, the system solidifies the selected illumination change threshold into an environmental perception parameter, which is used to determine whether rapid exposure correction is needed, thereby ensuring stable image quality.

[0041] In this embodiment of the invention, the adaptive adjustment unit specifically includes: First, the system receives environmental parameter data generated by the environmental modeling unit, which includes information such as light intensity, background reflectivity, and depth changes. Based on these parameters, the system builds a light change trend model to determine whether the current scene lighting is stable. When a rapid change in light intensity over time is detected, the system initiates an adaptive adjustment mechanism.

[0042] Subsequently, the system dynamically adjusts the camera's exposure time and gain threshold. Specifically, when the ambient brightness decreases, the system gradually extends the exposure time and appropriately increases the gain value to ensure that the image brightness remains within a recognizable range; conversely, when the ambient brightness is too high, the system shortens the exposure time and reduces the gain to prevent overexposure of the image.

[0043] When the rate of change in illumination exceeds a preset threshold, the system executes a rapid compensation strategy. This strategy calculates a compensation ratio based on the rate of change in illumination and adjusts the camera parameters in real time, making exposure adjustment proactive rather than reactive.

[0044] Finally, the system uses a continuous frame brightness equalization detection mechanism to evaluate the adjustment effect in real time. Once the image brightness is detected to be stable, the system locks the current parameters and outputs the exposure-corrected image data. Through this process, image brightness equalization can be maintained under conditions of sudden changes in illumination or partial occlusion, thus providing a stable and reliable visual input for subsequent pose recognition.

[0045] In this embodiment of the invention, the multi-frame fusion unit specifically includes: First, the system performs motion estimation processing on consecutive correction frames. By detecting the direction and magnitude of pixel block displacement between adjacent frames, a motion vector field is calculated to characterize the user's body motion trajectory in adjacent frames. This motion vector reflects the displacement trend of various parts of the human body and is used to achieve temporal alignment in multi-frame data.

[0046] Next, the system performs inter-frame alignment based on the motion vector field. The position of each pixel block is spatially relocated to ensure that pixels of the same human body part in adjacent frames correspond to the same spatial location as much as possible. To prevent blurring or misalignment caused by rapid local motion, the system uses an adaptive weighting method to allocate the inter-frame fusion ratio, with lower weights for faster-moving areas and higher weights for stationary or slowly moving areas, thus ensuring the clarity and coherence of the fused image.

[0047] Subsequently, the system performs temporal weighted fusion processing. By superimposing multiple frames of images according to a weighted ratio along the time dimension, a training image sequence is generated. This method can smooth sudden noise, stabilize illumination changes, and improve the image signal-to-noise ratio while preserving motion details.

[0048] Ultimately, the fused training image sequence is more spatially stable and more temporally continuous, which can improve the robustness of subsequent pose recognition and key point tracking, enabling the system to maintain high recognition accuracy under conditions of rapid motion or fluctuating lighting.

[0049] In a preferred embodiment of the present invention, the occlusion determination module includes: The trajectory analysis unit is used to calculate the spatial displacement curves of each key point based on the attitude data sequence, extract the motion continuity index of adjacent frames, and generate trajectory continuity data. The occlusion recognition unit is used to detect interruptions or abrupt changes in the spatial displacement curve based on trajectory continuity data. When the continuity is lower than a preset continuity threshold, it generates preliminary occlusion marker data. The occlusion classification unit is used to determine whether the occlusion is self-occlusion or external occlusion based on the preliminary occlusion marking data and the human skeleton topology, and to classify self-occlusion into cross-type or overlapping type, and generate classified occlusion data. The confidence calculation unit is used to calculate the occlusion confidence matrix and generate occlusion feature label data based on the classified occlusion data, combined with the occlusion duration, occlusion area and the visibility ratio of surrounding key points.

[0050] In this embodiment of the invention, the occlusion determination module dynamically analyzes human posture data to achieve automatic occlusion detection and confidence labeling during training. The trajectory analysis unit calculates the displacement curves of each key point in the posture data sequence and extracts motion consistency indices between consecutive frames to determine whether the trajectory of the key points is continuous. The occlusion recognition unit detects locations of trajectory interruptions or abrupt changes based on trajectory continuity data, thereby identifying potential occlusion events. When a specific part is occluded by an object, clothing, or the body itself, the system can promptly generate preliminary labeling data. The occlusion classification unit introduces human skeletal topology information to determine whether the occlusion type is external or self-occlusion, and further subdivides self-occlusion types, such as limb crossing or body overlap, forming a more detailed description of the occlusion type. The confidence calculation unit combines the occlusion duration, the area of ​​the occluded region, and the proportion of surrounding visible key points to calculate the confidence of each occluded region, generating a quantifiable occlusion confidence matrix. This result not only accurately reflects the stability and severity of occlusion but also provides a basis for subsequent posture compensation, enabling the system to distinguish between occluded regions with high confidence and unstable occlusion. This module enables the system to detect and reliably assess occlusion in real time during complex movements, thereby avoiding posture loss and motion misjudgment caused by occlusion.

[0051] In a preferred embodiment of the present invention, the method for setting the preset continuity threshold specifically includes: First, the system analyzes the displacement trends of keypoints during normal motion based on a large number of standard training examples, and statistically analyzes the magnitude of the displacement direction changes and the rate of velocity change of keypoints in consecutive frames. These indicators usually remain stable during normal motion, and therefore can be used to define the typical range of continuous motion.

[0052] Subsequently, the system collects sample data including occlusion, misidentification, or rapid motion, and analyzes the changes in the continuity index of adjacent frames. During occlusion and misidentification, the continuity decreases significantly; the decrease is greater during sudden motion changes than during moderate motion.

[0053] Next, the system compares the data distribution of normal and abnormal samples and defines the dividing point between them as the continuity threshold. To enhance the applicability of the threshold, the system uses a median offset method to correct the threshold, making it applicable to multiple scenarios and the differences in body shape movements of different users.

[0054] Ultimately, the system uses this continuity threshold as the basis for the occlusion recognition module to determine whether the trajectory is interrupted or abruptly changed, in order to identify abnormal key point behaviors.

[0055] In this embodiment of the invention, the occlusion classification unit specifically includes: First, the system receives preliminary occlusion marker data output by the occlusion recognition unit and extracts the corresponding key point location information. Based on the human skeletal topology model, the system calculates the spatial distance and directional relationship between the occluded key point and its adjacent joints. When the spatial angle between the occluded point and the adjacent joint is less than a set range and their movement directions tend to be the same, it indicates that the occlusion may be caused by the body's own movement, i.e., self-occlusion.

[0056] Next, the system further analyzes the types of self-occlusion. If two body parts (such as the arm and chest, or the legs) intersect or overlap briefly on the 3D spatial trajectory, and the voxel distribution of the occluded area shows obvious intersection, it is judged as intersecting self-occlusion; if the intersecting area maintains a high overlap ratio in multiple consecutive frames, it is judged as overlapping self-occlusion.

[0057] For cases where the self-occlusion condition is not met, the system identifies the occlusion as external occlusion. This type of occlusion is usually characterized by abrupt changes in the trajectory of key points but continuous movement of adjacent key points, and a relatively long duration of occlusion, which may be caused by obstructions from furniture, walls, or external objects.

[0058] Finally, the system generates categorized occlusion data, including occlusion type, occlusion location number, and duration. This data will be passed to the subsequent confidence calculation module to determine the confidence level of the occluded area and the pose compensation strategy. Through this classification process, the system can distinguish different occlusion sources, ensuring the relevance and stability of the compensation algorithm, and making human pose reconstruction more accurate and reliable.

[0059] In a preferred embodiment of the present invention, the attitude compensation module includes: The occlusion resolution unit is used to extract the spatial location missing region of the occluded key points based on the occlusion feature label data, and to identify the corresponding associated skeletal structure of the unoccluded key points to generate occlusion association data. The constraint generation unit is used to construct a local skeleton topology constraint model based on the spatial relative position of unoccluded key points and the proportion of bone length in the occlusion association data. When the occlusion duration exceeds the preset first duration threshold, the angle constraint interval is dynamically adjusted to generate spatial constraint data. The attitude trend prediction unit is used to calculate the motion direction and velocity change rate of the occluded key points based on the displacement trend information of key points in spatial constraint data and historical attitude data. When the motion direction remains unchanged, it performs time extrapolation prediction, and when the direction changes frequently, it performs neighboring frame interpolation prediction to generate key point prediction data. The continuous correction unit is used to perform consistency detection on the displacement changes of consecutive frames based on the preliminary compensation coordinate data. When an abnormal jump point is detected, it triggers the smooth correction of adjacent frames and feeds the correction result back to the constraint generation unit for constraint correction to generate complete attitude reconstruction data.

[0060] In this embodiment of the invention, the posture compensation module achieves complete restoration of human posture under occlusion by fusing occlusion data parsing, geometric constraint reconstruction, and dynamic prediction. The occlusion parsing unit processes the occlusion feature marker data, identifies spatial missing regions of occluded key points, and generates occlusion-related data based on the correspondence between unoccluded key points and skeletal structures, thereby determining the constraint boundaries for posture reconstruction. The constraint generation unit utilizes the spatial proportional relationship of unoccluded key points and skeletal length constraints to establish a local topological model, ensuring structural consistency in the prediction compensation process. When the occlusion duration is too long, this unit can also adaptively adjust the angle constraint range to prevent posture distortion caused by miscompensation. The posture trend prediction unit calculates the motion direction and velocity change trend of key points based on spatial constraint data and historical posture trajectories, and dynamically selects time extrapolation or adjacent frame interpolation methods based on stability to balance action continuity and prediction accuracy. The confidence fusion unit combines occlusion confidence information to perform weighted fusion of predicted coordinates, giving greater weight to data from highly confident regions in the reconstruction, thus improving overall posture consistency. The continuous correction unit further performs inter-frame consistency detection on the generated pose sequence. If anomalies are detected, smooth correction is performed through adjacent frames and feedback constraint models are applied to form a closed-loop optimization of the pose compensation process. The synergistic effect of this module enables the system to reconstruct a coherent, stable, and physiologically consistent complete pose even when body parts are occluded, providing a reliable foundation for action recognition and evaluation.

[0061] In a preferred embodiment of the present invention, the method for setting a preset first duration threshold specifically includes: First, the system statistically analyzes multiple actual training videos to determine the duration characteristics of occlusion for different body parts such as arms, legs, and torso in common movements. For example, self-occlusion (such as an arm covering the torso) is usually short-lived, while external occlusion (such as furniture occlusion) is longer-lasting.

[0062] Subsequently, the system sets different thresholds based on the importance of different key points. For example, key joints (knee, hip) have a greater impact on posture compensation, so their occlusion duration threshold is relatively short, while end joints (wrist, toe) can be set with higher thresholds.

[0063] Next, to adapt to various movement rhythms, the system dynamically adjusts based on the user's movement speed, for example, appropriately increasing the duration threshold when the user's movements are slow.

[0064] Finally, based on the principles of statistical distribution and action rationality, an optimal interval value is determined and fixed as a preset first duration threshold, which is used to determine whether the constraint model needs to be adjusted before attitude compensation.

[0065] In this embodiment of the invention, the constraint generation unit specifically includes: First, the system receives occlusion association data from the occlusion resolution unit and extracts the 3D coordinates of unoccluded keypoints and their spatial topological relationships with occluded points. Based on a length ratio model of the human skeletal structure, the system calculates distance parameters between adjacent keypoints to maintain the relative constancy of bone length. For example, when the knee joint is occluded, the system keeps the length ratio of the thigh to the lower leg unchanged, ensuring geometric consistency in pose reconstruction.

[0066] Next, the system generates angle constraint parameters based on the relative positions of unoccluded key points. Specifically, the system calculates the range of angle changes between joints and limits it within the physiologically mobile range of the human body to prevent abnormal postures during compensation calculations. When the occlusion duration is short, the system uses a default constraint range to maintain model stability; however, when the occlusion duration exceeds a preset first duration threshold, the system activates a dynamic adjustment mechanism, automatically expanding or contracting the constraint range based on the angle change trends of joint movements in historical frames, making the reconstruction results closer to the user's actual movements.

[0067] Subsequently, the system constructs a local skeletal topological constraint model by fusing spatial geometric constraints and temporal constraints. This model not only maintains the consistency of human body structural proportions but also adjusts the constraint strength according to the direction of movement. For example, during lower limb movement, if the occlusion point is located at a high-dynamic joint (such as the knee), the system will reduce the weight of the angle constraint to enhance prediction flexibility; if the occlusion point is located in a low-dynamic area (such as the shoulder), the constraint strength will be increased to maintain stability.

[0068] Finally, the system integrates the spatial position constraints, angle range constraints, and dynamic temporal constraints of all joints to generate spatial constraint data. This data serves as the input parameter for the subsequent trend prediction unit, limiting the prediction range in compensation calculations and ensuring that posture reconstruction conforms to the laws of human anatomy while also possessing dynamic adaptability.

[0069] In a preferred embodiment of the present invention, the motion evaluation module includes: The template calling unit is used to call the preset action template corresponding to the current training type from the system database, and extract the standard angle change curves and time series parameters of each stage in the template to generate template reference data; The phase division unit is used to calculate the angle change rate based on the complete attitude reconstruction data, and divide it into the initial phase, transition phase and termination phase according to the rate change trend to generate motion phase data. The deviation calculation unit is used to compare the user's joint angle curve with the standard angle curve of the corresponding stage in the template reference data segment by segment based on the action stage data, calculate the angle deviation amplitude and duration of each joint, and generate local deviation data. The timing matching unit is used to extract the angular velocity change sequence of each stage based on the local deviation data, and synchronously compare it with the time series parameters in the template reference data to calculate the difference in motion time distribution and generate timing matching degree data. The comprehensive scoring unit is used to calculate the comprehensive training score based on the temporal matching degree data and local deviation data, and generate action evaluation results including angle accuracy score, time synchronization score and action stability score.

[0070] In this embodiment of the invention, the motion evaluation module achieves quantitative analysis and dynamic scoring of user training motions through multi-dimensional motion analysis and comparison with standard templates. The template invocation unit can automatically retrieve preset motion templates matching the training type from the system database and extract standard angle change curves and time series parameters for each stage of the template, enabling the system to adapt to different motions. The stage division unit divides the motion process into start, transition, and termination stages based on the time change rate of each joint angle in the complete posture data, ensuring the temporal logical continuity of the evaluation process. The deviation calculation unit calculates the angle deviation value and duration of each joint by comparing the user's angle changes with the template standard curve stage by stage, generating local deviation data so that motion errors can be quantified both spatially and temporally. The timing matching unit further analyzes the consistency of motion rhythm by comparing the user's angular velocity changes with the time series distribution of the template to detect motion speed and rhythm deviations. The comprehensive scoring unit weighted and fused the above multiple indicators to generate three results: angle accuracy score, timing synchronization score, and motion stability score, providing trainees with quantitative motion performance data. This module enables the system to achieve accurate motion assessment without relying on external sensors, allowing users to intuitively understand the gap between their own movements and standard movements after training, providing data support for personalized training adjustments.

[0071] In this embodiment of the invention, the stage division unit specifically includes: First, the system receives complete posture reconstruction data output by the posture compensation module and extracts the angle change curves of major joints (such as hip, knee, elbow, shoulder, etc.) over time. The system calculates the angle change rate between adjacent time frames and identifies the key stages of the action based on the rising or falling trend of the change rate.

[0072] When the rate of angle change approaches zero and remains stable, the system determines it to be in the initial stage; when the rate of angle change rises or falls rapidly, it indicates that the action has entered the transition stage; when the rate of angle change falls back and tends to stabilize, it is determined to be in the termination stage.

[0073] During this process, the system combines the overall trend of changes in multiple joint angles to avoid misjudgment of stages due to sudden changes in a single joint. For example, in a squatting movement, the synchronous changes in the knee and hip angles are detected simultaneously, and the system only recognizes the "standing-up termination stage" when both angles have stabilized and recovered.

[0074] Finally, the system records the time range, keyframe number, and average angular velocity information of each stage as motion stage data, providing structured staged input for the deviation calculation unit, making the evaluation process more consistent with the actual rhythmic changes of human movement.

[0075] In this embodiment of the invention, the deviation calculation unit specifically includes: First, the system determines the temporal correspondence between user actions and template standard actions based on the action phase data. For each phase, the system extracts the joint angle curves from the user posture data and compares them point by point with the standard angle curves in the template reference data.

[0076] During the comparison process, the system calculates the angle difference for each frame and takes its absolute value as the deviation amplitude. When the angle deviation exceeds a preset angle threshold for multiple consecutive frames, the system records this stage as the "deviation duration segment" and calculates the duration of the deviation.

[0077] The system then analyzes the spatial characteristics of the deviation. For example, if the deviation is concentrated in the lower limb joints and the angle is too small, the system determines it as "insufficient squatting"; if the change in the angle of the upper limb joints is lagging, it may be "rhythm delay".

[0078] Finally, the system summarizes the deviation magnitude, duration, and type of deviation at each stage to form local deviation data. This data reflects the distribution of differences between user actions and the standard template, providing a quantitative basis for subsequent time-series matching and comprehensive scoring.

[0079] In a preferred embodiment of the present invention, the confidence calculation unit includes: The occlusion time series analysis subunit is used to extract the occlusion duration of each key point and calculate the occlusion time stability parameter based on the classified occlusion data. When the occlusion time exceeds the preset second duration threshold, the corresponding position confidence is reduced and time-weighted data is generated. The spatial feature evaluation subunit is used to calculate the spatial integrity index of the occluded area based on time-weighted data, combined with the proportion of visible points in the neighborhood of each key point and the spatial distribution density, and to improve the confidence of the area when adjacent visible points are dense and in the same direction, and generate spatial correction data. The neighborhood weight diffusion subunit is used to establish neighborhood transit relationships within the topological structure of key points based on spatially corrected data, perform hierarchical diffusion and smoothing of confidence, and generate neighborhood fusion data. The dynamic confidence update subunit is used to calculate the inter-frame confidence change rate based on the neighborhood fusion data and historical frame confidence records. When the change rate exceeds the preset first change rate threshold, confidence correction is performed and an occlusion confidence matrix is ​​generated.

[0080] In this embodiment of the invention, the confidence calculation unit achieves dynamic evaluation and updating of the confidence level of the occlusion region through comprehensive analysis of the temporal characteristics, spatial characteristics, and neighborhood topology of the occlusion state. The occlusion time-series analysis subunit calculates the occlusion duration of each key point and determines the reliability of the occlusion through a time stability parameter. When a key point is continuously occluded for more than a set second duration threshold, its confidence value automatically decreases to reflect the uncertainty of the information in that area. The spatial feature evaluation subunit calculates the spatial integrity of the occluded region by combining the visibility ratio and spatial distribution density of neighboring key points. When surrounding key points are dense and move in the same direction, the system automatically increases the confidence value of that region to ensure the continuity of local area data. The neighborhood weight diffusion subunit establishes a neighborhood transmission relationship through the topological structure of human body key points, diffusing the influence of high-confidence regions to surrounding key points, achieving spatial smoothing and transition of confidence information. The dynamic confidence update subunit uses historical frame confidence records to detect the inter-frame confidence change rate. When fluctuations exceed a stability threshold, corrections are performed to generate an updated occlusion confidence matrix. This module enables the system to track the visibility and reliability of each part in real time during the action, ensuring that the weighted decision during posture compensation is more reasonable, reducing posture reconstruction deviations caused by incorrect confidence assessments, thereby improving overall recognition accuracy and action continuity.

[0081] In a preferred embodiment of the present invention, the method for setting the preset second duration threshold specifically includes: First, the system statistically analyzes the duration differences of different occlusion situations based on historical occlusion samples, paying particular attention to the typical duration distribution of long-term and occasional occlusion.

[0082] Subsequently, by analyzing the impact of occlusion on the accuracy of attitude prediction, the system determined that after a certain duration, the impact of occlusion on prediction increases significantly, and the confidence value of that region needs to be reduced.

[0083] Next, the system segments the duration of occlusion, distinguishing between short-term occlusion, transitional occlusion, and long-term occlusion, and increases the adaptive correction ratio based on the stability of the occlusion.

[0084] Ultimately, the system determines the second duration threshold based on statistical analysis, making the confidence update process more consistent with actual motion occlusion.

[0085] In a preferred embodiment of the present invention, the method for setting the preset first rate of change threshold specifically includes: First, based on multiple sets of training motion data, the system statistically analyzes the displacement changes of key points on the human body during normal movement. The system organizes the displacement differences between consecutive frames, calculates the average level and typical fluctuation range of the change amplitude, and thus forms the basic change rate reference interval for each key point.

[0086] The system then categorizes key points based on their characteristics in different training movements. For example, for joints with large variations, such as the knee and hip joints, the system appropriately raises the baseline reference range; while for joints with higher stability, such as the shoulder and head, the reference range is appropriately lowered. This method makes the rate of change threshold more consistent with the movement patterns of different key points.

[0087] Next, the system adjusts the threshold a second time based on the type of training movement. For fast-paced or explosive movements, the system increases the threshold to tolerate greater variations; conversely, for slow-paced or flexible movements, the system decreases the threshold to sensitively detect subtle changes.

[0088] Finally, the system performs personalized fine-tuning of the threshold based on the user's historical training records. When the user's behavior is consistently stable over a long period, the system appropriately increases the threshold; if the user frequently exhibits jittery or unstable movements, the threshold is appropriately decreased to enhance the system's ability to identify abnormal fluctuations. The result determined through this process is set as the preset first rate of change threshold and used for subsequent inter-frame confidence change judgments.

[0089] In this embodiment of the invention, the occlusion timing analysis subunit specifically includes: First, the system receives occlusion classification data output by the occlusion classification unit. This data includes the start frame, end frame, and occlusion duration for each occluded keypoint. The system performs time statistics on the occlusion interval of each keypoint and calculates the total duration for which it remains occluded in consecutive frames.

[0090] Next, the system calculates the occlusion time stability parameter based on the changing trend of the occlusion duration. When the occlusion duration is short and occurs intermittently, it indicates that the occlusion has random characteristics, and the system determines it to be of low stability; when the occlusion duration is long and changes slowly, it is determined to be of high stability.

[0091] To prevent misidentification caused by high-stability occlusion, the system uses time weighting when assigning confidence scores. When the duration of occlusion of a key point exceeds a preset second duration threshold, the system automatically reduces its confidence score weight, thereby decreasing the reference value of subsequent compensation algorithms in that area and reducing error propagation.

[0092] Finally, the system outputs time-weighted data, which records the occlusion duration, stability parameters, and adjusted time weight values ​​for each key point, providing a temporal basis for spatial feature evaluation.

[0093] In this embodiment of the invention, the spatial feature evaluation subunit specifically includes: First, the system determines the neighborhood range of each occluded keypoint based on time-weighted data. The neighborhood typically consists of surrounding keypoints that are topologically connected to the target keypoint, such as the shoulder and elbow, or the hip and knee. The system calculates the percentage of visible keypoints within this neighborhood as a visible point ratio indicator.

[0094] Subsequently, the system analyzes the spatial distribution density of each visible key point in the neighborhood. If adjacent visible points are evenly distributed in space and their directional changes have the same trend, it indicates that the structure of the region is continuous, and the system will determine that the occlusion location is highly inferable; if the visible points in the neighborhood are sparse or their directions change drastically, it is determined to be a structurally unstable region.

[0095] The system calculates the spatial integrity index of the occluded area based on the above results. When the integrity is high, the confidence level of the area is increased; when the integrity is low, the time-weighted low confidence level is maintained.

[0096] Finally, the system outputs spatial correction data, including the spatial integrity score of each occluded region and the updated confidence assignment, for subsequent neighborhood weight diffusion processing.

[0097] In this embodiment of the invention, the neighborhood weight diffusion subunit specifically includes: First, based on the topology of the human skeleton, the system constructs a graph network structure of all key points according to their joint connections. The system reads the confidence value of each node from the spatially corrected data and uses it as the initial weight distribution.

[0098] Subsequently, the system performs hierarchical diffusion operations on the topology network. Specifically, for nodes with high confidence, the system propagates weights to neighboring nodes along the connection path; for nodes with low confidence, it receives compensating weights from neighboring high-confidence nodes. This diffusion process is repeated across multiple levels until the weight distribution converges to a smooth state.

[0099] During the diffusion process, the system introduces a smoothing constraint parameter to prevent over-propagation. That is, when the spatial distance or movement direction difference between adjacent nodes exceeds the stability threshold, the weight transfer stops to ensure that confidence diffusion only occurs in structurally similar regions.

[0100] Finally, the system outputs neighborhood fusion data, forming a confidence matrix with global consistency and smooth distribution, providing basic input for dynamic confidence update sub-units.

[0101] In this embodiment of the invention, the dynamic confidence update subunit specifically includes: First, the system reads the neighborhood fusion data of the current frame and the confidence records in historical frames, and performs inter-frame comparisons on the confidence values ​​of each keypoint. The system calculates the confidence change rate of each keypoint to measure the degree of difference in confidence distribution between the current frame and the previous frame.

[0102] When the confidence change rate of a key point exceeds a preset first change rate threshold, the system considers the confidence change in that region to be abnormal. At this time, the system performs corrections based on the direction of change: if the confidence of the current frame suddenly increases while there is no significant structural improvement in the neighborhood, confidence weakening is performed; if the confidence suddenly decreases while the neighborhood remains stable, confidence enhancement is performed.

[0103] Subsequently, the system performs global normalization on the corrected confidence distribution to keep the overall confidence matrix within a uniform numerical range, ensuring stable input for subsequent compensation calculations.

[0104] Finally, the system outputs an occlusion confidence matrix, which records the final confidence value and spatial distribution of each keypoint. This matrix will be used as input to the confidence fusion unit in the pose compensation module to perform weighted fusion of the prediction results, thereby achieving reliability control during pose reconstruction.

[0105] In a preferred embodiment of the present invention, the attitude trend prediction unit includes: The motion pattern recognition subunit is used to analyze the continuous displacement characteristics of key points and determine the motion pattern based on spatial constraint data and historical posture data. When the frequency of displacement direction change is lower than the preset second rate of change threshold, it is identified as stable; otherwise, it is identified as abrupt and generates pattern identification data. The short-term trend generation subunit is used to extract the displacement direction and velocity change rate of the most recent multi-frames from the stable key point sequence in the pattern identification data, perform time extrapolation prediction, and generate short-term prediction data. The long-term trend generation subunit is used to extract global displacement change trends and acceleration distribution characteristics based on the key point sequence of abrupt or nonlinear motion in the pattern identification data, and generate long-term prediction data through neighboring frame interpolation and orientation smoothing calculation. The adaptive trend fusion subunit is used to calculate the dynamic fusion weight of short-term and long-term forecast data based on the directional consistency parameter. When the consistency is high, the proportion of short-term forecast is increased; otherwise, the proportion of long-term forecast is increased, thus generating key point trend forecast data.

[0106] In this embodiment of the invention, the attitude trend prediction unit achieves dynamic position prediction and trajectory continuation of occluded key points through multi-timescale motion modeling and adaptive fusion mechanisms. The motion pattern recognition subunit analyzes the displacement characteristics of historical attitude data to determine the frequency of change of the current action pattern and classifies it into two types: stable and abrupt. This distinction allows the system to make targeted adjustments to the prediction strategy, thereby avoiding error accumulation caused by a single prediction pattern. The short-term trend generation subunit, for stable motion, extracts the motion direction and velocity change rate of key points in the most recent multiple frames and uses time extrapolation for prediction, making the short-term motion trajectory more consistent. The long-term trend generation subunit, for abrupt or nonlinear actions, extracts global displacement trends and acceleration distribution features and combines them with directional smoothing calculations in adjacent frames to generate long-term prediction results, enabling the system to maintain stable trajectory inference even under highly dynamic changes. The adaptive trend fusion subunit performs weighted fusion of short-term and long-term prediction results, automatically adjusting the ratio of the two according to the direction consistency parameter. When the motion direction is consistent, the weight of short-term prediction is increased, and when the direction changes frequently, the proportion of long-term prediction is increased, thereby achieving dynamic adaptation under different action types. Through this module, the system can achieve temporal extension and reasonable prediction of key point positions when posture data is lost or partially occluded, ensuring the spatial continuity and dynamic reproducibility of human movements, and providing highly reliable input for subsequent posture compensation and evaluation.

[0107] In a preferred embodiment of the present invention, the method for setting the preset second rate of change threshold specifically includes: First, the system analyzes the frequency of displacement direction changes during normal actions and statistically analyzes the difference in the rate of change between stable and abrupt actions from a large number of samples.

[0108] Subsequently, by establishing a distribution map of the rate of change of direction, the system is able to find the boundary point between stable motion and abrupt motion.

[0109] Next, the system modifies the threshold by sampling the movement habits of different users multiple times, making it more consistent with the general performance of multiple users in different training movements.

[0110] Ultimately, the system solidifies the rate of change threshold for action pattern classification, enabling the prediction module to automatically select short-term or long-term trend models.

[0111] In this embodiment of the invention, the short-term trend generation subunit specifically includes: First, the system filters out keypoint sequences marked as "stable" from the pattern identification data. These keypoints exhibit minimal changes in their direction of motion, making them suitable for short-term prediction using time extrapolation.

[0112] Subsequently, the system extracts the displacement direction and velocity change rate of these key points in the most recent frames, and calculates the predicted displacement path for the next few frames by fitting their temporal variation trends. The system employs a weighted moving average mechanism during the calculation process, giving higher weight to the displacement features of the most recent frames to reflect the current motion trend.

[0113] To improve the reliability of predictions, the system also limits the prediction range of key points based on spatial constraint data when generating prediction results, so as to avoid generating postures that exceed the physiological range of human movement.

[0114] Finally, the system outputs short-term prediction data, which includes the future location coordinates and displacement trend information of each stable key point. This data is then fused with long-term trend data to ensure the accuracy and stability of the prediction results on a short timescale.

[0115] In this embodiment of the invention, the long-term trend generation subunit specifically includes: First, the average acceleration and direction change distribution are calculated. Then, abrupt motion prediction is performed based on neighboring frame interpolation and direction smoothing calculations. Specifically, several frames are taken before and after the abrupt change, and the displacement direction is smoothed using interpolation methods. Based on this, time extrapolation is performed to generate a long-term trend curve. Finally, the long-term trend prediction results are output in sequence form, including the estimated coordinates and velocity change directions of key points in several future frames, providing a long-term dynamic feature reference for subsequent adaptive fusion.

[0116] In this embodiment of the invention, the adaptive trend fusion subunit specifically includes: First, the system receives short-term and long-term prediction data and performs a consistency analysis on the displacement direction of the same key point within the same time frame. The system calculates the difference in direction and velocity between the two data sets, using this as the basis for determining the direction consistency parameters.

[0117] When the directional consistency parameter is high, it indicates that the short-term forecast is basically consistent with the long-term trend. At this time, the system increases the weight of the short-term forecast results to enhance the immediacy of the response. When the consistency is low, it indicates that the short-term forecast is greatly affected by noise or sudden movements. The system then increases the weight of the long-term forecast to enhance the overall stability.

[0118] Subsequently, the system weights and merges the short-term and long-term prediction results based on the calculated dynamic weights to generate new trend prediction data. The system also introduces a confidence smoothing mechanism to balance the changes in the fusion results and avoid prediction jumps when weights switch.

[0119] Finally, the system outputs key point trend prediction data, including the fused displacement direction, velocity change trend and predicted stability information, providing continuous and reliable motion trend input for subsequent confidence fusion and attitude compensation.

[0120] In a preferred embodiment of the present invention, the confidence fusion unit includes: The prediction confidence matching subunit is used to establish a mapping relationship between predicted coordinates and occlusion confidence based on key point trend prediction data and occlusion feature labeling data, and to calculate the confidence weight value of each key point to generate weighted prediction data. The error detection and correction subunit is used to calculate the displacement difference of key points in adjacent frames based on the weighted prediction data. When the displacement difference exceeds the preset stable threshold, reverse correction is performed according to the historical trend direction to generate error correction data. The multi-frame smoothing fusion subunit is used to perform time-series smoothing processing on the key point coordinates of multiple consecutive frames based on error correction data, and replace the predicted coordinates with the latest visible frame coordinates when the occlusion is removed, generating time-series fusion data; The confidence strategy update subunit is used to calculate the fusion accuracy index based on the time-series fusion data, update the confidence allocation strategy for subsequent frames, and generate preliminary compensation coordinate data.

[0121] In this embodiment of the invention, the confidence fusion unit achieves adaptive fusion of predicted coordinates and actual observation data by introducing a confidence weighting mechanism and multi-frame temporal correction, thereby improving the reliability and stability of attitude compensation results. The prediction confidence matching subunit establishes a mapping relationship between predicted coordinates and occlusion confidence based on keypoint trend prediction data and occlusion feature labeling data, applying confidence weight to each keypoint, giving greater weight to parts with higher confidence in the compensation calculation. This makes the compensation results more dependent on reliable data, while for low-confidence areas, a smooth transition is achieved through neighborhood interpolation, avoiding error propagation in occluded areas. The error detection and correction subunit dynamically detects the weighted prediction results, identifies prediction errors by calculating the displacement differences of keypoints in adjacent frames, and performs reverse correction based on historical trend directions, making the keypoint trajectory more consistent with the actual movement trend. The multi-frame smoothing fusion subunit, based on error correction, performs time-weighted fusion of coordinates from multiple consecutive frames, preserving short-term motion changes while weakening the impact of local abnormal fluctuations; when occlusion is removed, the system can automatically replace the predicted coordinates with the latest visible frame data, achieving real-time calibration. The confidence strategy update subunit adjusts the confidence allocation weights of subsequent frames based on the accuracy evaluation index of the fusion result, enabling the fusion strategy to possess self-learning and convergence characteristics. Through this module, the pose compensation process forms a complete closed loop from prediction to correction to optimization, making the output preliminary compensation coordinate data smoother, more continuous, and more in line with the actual human motion patterns, providing a stable input foundation for the final pose reconstruction.

[0122] In a preferred embodiment of the present invention, the method for setting the preset stability threshold specifically includes: First, the system statistically analyzes the displacement difference range of key points during normal motion and forms a basic stable range by detecting the natural fluctuations in displacement between adjacent frames.

[0123] Subsequently, the system collected samples containing abnormal jumps and misidentifications due to occlusion, and statistically analyzed their displacement difference characteristics, finding that they were usually much higher than the normal range of motion.

[0124] Next, the system uses the boundary between the upper limit of the normal range and the lower limit of the abnormal range as a stabilization threshold, and makes fine adjustments according to the joint type, such as allowing greater variation for end joints.

[0125] Ultimately, the system uses a stable threshold for error detection to ensure that trajectory correction is performed within a reasonable range.

[0126] In this embodiment of the invention, the prediction confidence matching subunit specifically includes: First, the system receives keypoint trend prediction data output by the attitude trend prediction unit and occlusion feature labeling data generated by the occlusion determination module. The former includes the predicted position and motion direction of each keypoint in the next few frames, while the latter includes the occlusion confidence level and spatial occlusion range of each part.

[0127] The system performs spatial matching on two types of data. Specifically, based on the 3D coordinates of key points, the system determines whether a point is within an occlusion region and establishes a correspondence between the predicted point and the corresponding occlusion confidence level. When a predicted point is within an occlusion region, its corresponding confidence weight is automatically reduced according to the occlusion confidence level; conversely, if it is within an unoccluded region, it is assigned a higher confidence weight.

[0128] Subsequently, the system makes a secondary correction based on the motion stability of the predicted point and historical confidence records. For example, when a key point maintains a high confidence state in historical frames and the current prediction result is consistent with the historical trend direction, the system appropriately increases the confidence weighting value of that point to enhance the continuity of the prediction.

[0129] Finally, the system combines the corrected weight values ​​with the predicted coordinates to generate weighted prediction data. This data can provide a dynamic confidence distribution in subsequent fusion calculations, enabling the attitude compensation process to achieve adaptive reliability control under occlusion interference, thereby improving compensation accuracy and action continuity.

[0130] In this embodiment of the invention, the confidence policy update subunit specifically includes: First, the system receives the temporal fusion data output by the multi-frame smooth fusion subunit and extracts the fusion coordinates and confidence residual values ​​of each key point. The fusion residual reflects the degree of deviation between the predicted coordinates and the actual visible frame coordinates, and is an important parameter for measuring the fusion accuracy.

[0131] The system calculates the fusion accuracy index based on these residual values. If the fusion residual of a key point is small across multiple frames, it indicates that the prediction result is reliable, and the system increases the confidence of that point in subsequent frames. If the residual is large and changes frequently, it indicates that the point is strongly affected by occlusion or sudden motion changes, and the system correspondingly lowers the confidence value of that point and introduces stronger spatial constraint correction in the next frame.

[0132] Next, the system dynamically adjusts the overall confidence allocation strategy. Specifically, the system reallocates the confidence weights of different regions based on the global balance of the keypoint confidence distribution to ensure that keypoints throughout the body maintain a reasonable proportion in subsequent predictions. For example, when the upper limbs are severely occluded, the system reduces the predictive confidence influence of that region and increases the confidence weight of the lower limb region to maintain the overall balance of the posture.

[0133] Finally, the system combines the updated confidence strategy with the fused coordinates to generate preliminary compensation coordinate data. This data will be used in the continuous correction phase of the attitude compensation module, enabling the system to achieve adaptive optimization of confidence and error feedback adjustment in the next reconstruction process.

[0134] In a preferred embodiment of the present invention, the continuous correction unit includes: The anomaly detection subunit is used to calculate the displacement change rate of key points in adjacent frames based on the preliminary compensation coordinate data. When the displacement change of any key point exceeds the preset stability threshold, the point is identified as an abnormal node and anomaly detection data is generated. The neighboring frame interpolation subunit is used to extract the position sequence of the abnormal node in the neighboring frame based on the abnormal detection data, perform temporal interpolation correction based on the displacement direction and velocity change rate of the neighboring frame, and generate interpolation correction data. The trajectory smoothing correction subunit is used to smooth the motion trajectory of key points in consecutive frames based on interpolation correction data. It generates smooth trajectory data by calculating the directional change amplitude of adjacent frames and limiting over-correction. The constraint feedback subunit is used to calculate the attitude consistency parameters based on the smooth trajectory data. When there is a deviation between the trajectory direction and the skeletal topology constraint model, it automatically adjusts the angle constraint range of the constraint generation unit, updates the spatial constraint data, and generates complete attitude reconstruction data.

[0135] In this embodiment of the invention, the continuous correction unit achieves temporal consistency correction and adaptive adjustment of topological constraints during the attitude reconstruction process through a linkage mechanism of anomaly detection, interpolation smoothing, and constraint feedback. The anomaly detection subunit calculates the displacement rate of key points in adjacent frames based on preliminary compensation coordinate data. When abnormal jumps or movements exceeding the normal range are detected at certain key points, these nodes are marked as anomalies, effectively preventing further amplification of attitude prediction errors. The adjacent frame interpolation subunit calculates the displacement direction and velocity change rate of adjacent frames based on the position sequence of the anomaly nodes in the preceding and following frames, and reconstructs the anomalies using temporal interpolation, ensuring that the corrected trajectory naturally connects with the trajectory of adjacent frames in space. The trajectory smoothing correction subunit further performs continuous frame direction difference analysis on the corrected motion trajectory. By limiting the magnitude of over-correction, the motion trajectory of the human body key points maintains the naturalness and dynamic consistency of the movement rhythm after correction. The constraint feedback subunit calculates attitude consistency parameters based on the smoothed trajectory and provides feedback correction for parts deviating from the skeletal topological constraints, thereby adjusting the angle limitation range of the constraint generation unit and preventing local movements from violating the body's structural logic. This module not only suppresses local pose abrupt changes, but also maintains the coherence of the skeletal structure at the global level, so that the pose reconstruction results are both in line with temporal dynamics and meet the constraints of human body structure, providing reliable and smooth three-dimensional pose data support for action recognition and evaluation.

[0136] In this embodiment of the invention, the constraint feedback subunit specifically includes: First, the system receives smoothed trajectory data output from the trajectory smoothing correction subunit. This data contains information on the motion trajectory and orientation changes of key points in consecutive frames. The system then uses a human skeletal topology model to perform consistency analysis on the trajectory direction and joint angles of each key point, calculating a posture consistency parameter. This parameter measures whether the actual motion trajectory conforms to the constraints of the human body structure.

[0137] When a deviation is detected between the trajectory direction of a key point and the topological model, the system activates a constraint feedback mechanism. The system compares the angular change at the deviation point with the preset range of the constraint model to determine if it exceeds the range of human movement. If it exceeds a threshold, the system automatically adjusts the angular constraint range in the constraint generation unit, ensuring that subsequent compensation calculations are performed within a spatial range that better conforms to the laws of human movement.

[0138] Simultaneously, the system performs coordinated corrections on multiple related joints. For example, when abnormal knee joint movement leads to hip posture imbalance, the system will simultaneously adjust the constraint parameters of the hip and knee to restore their angular relationship to the characteristics of the human kinetic chain.

[0139] Finally, the system re-inputs the updated spatial constraint data into the attitude compensation module to generate new, complete attitude reconstruction data. Through this closed-loop feedback method, the system can continuously correct reconstruction errors during dynamic training, maintaining a high degree of consistency between attitude calculation and actual human movement.

[0140] In a preferred embodiment of the present invention, the comprehensive scoring unit includes: The index extraction subunit is used to extract the joint angle error, time delay difference and stability fluctuation amplitude from the temporal matching degree data and local deviation data, and convert them into angle accuracy index, time consistency index and motion stability index respectively to generate raw scoring data. The indicator weight allocation subunit is used to determine the dynamic weight parameters of the angle accuracy indicator, time consistency indicator, and action stability indicator based on the training type and the user's historical training records, and generate weight allocation data. The comprehensive calculation subunit is used to perform weighted average calculation on each indicator based on the original scoring data and weight allocation data. By accumulating the scores of the three dimensions of angle accuracy, time consistency and motion stability according to the weight ratio, a single frame comprehensive score data is generated. The scoring time-series fusion subunit is used to perform time-weighted fusion processing on the comprehensive scoring data of multiple consecutive frames to generate stage comprehensive scoring data; The scoring trend analysis subunit is used to calculate the scoring trend based on the stage comprehensive scoring data and historical training results. When the overall scoring change rate is detected to exceed the preset scoring change threshold, the weight allocation data is reversed and the action evaluation results including angle accuracy score, time synchronization score, action stability score and stage comprehensive score are generated.

[0141] In this embodiment of the invention, the comprehensive scoring unit achieves quantitative evaluation of user actions in terms of angle accuracy, time consistency, and stability through multi-indicator extraction, weight allocation, and temporal fusion, and possesses adaptive scoring capabilities that dynamically adjust according to training performance. The indicator extraction subunit extracts basic features such as joint angle error, time delay difference, and stability fluctuation amplitude from temporal matching data and local deviation data, transforming different action dimensions into comparable quantitative indicators to form the original scoring data. The indicator weight allocation subunit dynamically determines the weight ratio of each indicator based on the training type and the user's historical performance, making the scoring more personalized and targeted for different users and action scenarios. The comprehensive calculation subunit performs a weighted average calculation based on the indicators and weight allocation results to generate single-frame comprehensive scoring data, thereby achieving real-time evaluation of action performance in each time slice. The scoring temporal fusion subunit further performs time-weighted fusion on the scoring results of multiple consecutive frames, eliminating the interference of local fluctuations on the overall score, making the distribution results more stable and continuous. The scoring trend analysis subunit tracks scoring changes at the stage level. When the rate of change in scores exceeds a set threshold, it automatically adjusts the weight allocation ratio, enabling the system to develop an adaptive scoring mechanism during continuous training. Through this module, the system can generate comprehensive evaluation results including angle accuracy, time synchronization, and movement stability. This allows trainees to not only understand the current quality of their movements but also to grasp the direction for training improvement through trend changes, thus building a sustainable movement performance monitoring and optimization system.

[0142] In a preferred embodiment of the present invention, the method for setting the preset score change threshold specifically includes: First, the system analyzes the score change trends across multiple training cycles, statistically analyzing the differences in amplitude between normal and abnormal fluctuations. For example, beginners typically experience larger fluctuations, while intermediate and advanced users show more stable changes.

[0143] Subsequently, the system establishes a score distribution model based on the score composition structure (angle, time, stability) to identify the score change magnitude that usually reflects training anomalies.

[0144] Next, the system uses statistical methods to determine the boundary interval of the rate of change in scores, and determines the final threshold by multi-dimensional weighting based on the stability of score fluctuations, the difficulty of training actions, and the user's skill level.

[0145] Finally, the system solidifies the threshold, which is used to trigger a weight adjustment mechanism when the score changes beyond the threshold, giving the scoring system an adaptive capability.

[0146] In this embodiment of the invention, the indicator extraction subunit specifically includes: First, the system receives data from the timing matching unit and the deviation calculation unit. The timing matching data reflects the rhythmic consistency of the user's movements over time, while the local deviation data includes information on the degree and duration of joint angle deviation. The system synchronizes and aligns the two types of data to ensure that the angle curve of each joint corresponds one-to-one with the time series.

[0147] Next, the system performs statistical analysis on the angle error of each joint. By calculating the average deviation and peak deviation of the joint angle, an angle accuracy index for that joint is obtained, which is used to measure the degree of spatial angular alignment between the user's actions and the standard template. Simultaneously, the system calculates the time delay difference between adjacent stages to generate a time consistency index, which reflects the degree of synchronization between the action completion sequence and the standard template.

[0148] The system then extracts the stability features of the motion by detecting the fluctuation amplitude and rate of change of the angle curve. If the curve is smooth and there is no obvious jitter, the stability index is high; if there are frequent fluctuations in the angle change, the stability score is low.

[0149] Finally, the system integrates the three types of indicators (angle accuracy, time consistency, and movement stability) into raw scoring data and represents them in a standardized manner to make the scores of different movement types comparable. This step provides a quantitative input basis for subsequent weight allocation and comprehensive scoring.

[0150] In this embodiment of the invention, the indicator weight allocation subunit specifically includes: First, the system reads the user's training type and movement characteristic classification information. For example, for strength training (such as squats and push-ups), the system sets the angle accuracy weight as the main indicator; while for rhythmic or flexibility training (such as yoga and stretching), the system increases the weight ratio of time consistency and stability indicators.

[0151] Next, the system retrieves the user's historical training records and analyzes the score fluctuation trends and common deviation patterns in past training. If a user consistently performs poorly on a certain indicator, such as unstable angle control, the system automatically increases the weight of that indicator to enhance the targeted nature of the training feedback.

[0152] Subsequently, the system normalizes the weights of the three types of indicators to ensure that the total weight remains constant. The system also introduces a dynamic weight adjustment mechanism, fine-tuning the weights based on score fluctuations during real-time user training, enabling the system to dynamically optimize evaluation weights according to actual performance.

[0153] Finally, the system outputs weight allocation data, recording the weight parameters and adjustment strategies for each scoring dimension, which are used in the comprehensive calculation phase. Through this process, the system can adaptively adjust the scoring logic according to different training objectives and user characteristics, realizing a personalized training evaluation mechanism.

[0154] In this embodiment of the invention, the scoring trend analysis subunit specifically includes: First, the system reads the stage-specific comprehensive score data output by the comprehensive calculation subunit and compares it with the user's historical training results stored in the database. The system compares the score change curves of each stage to calculate the overall score change rate and the direction of score fluctuation, in order to determine the improvement or decline trend of the user's training performance.

[0155] Next, the system performs adaptive adjustments based on the trend of score changes. When the system detects that a certain score dimension (such as time synchronization score) continues to decline during consecutive training cycles while other dimensions remain stable, it indicates that the user has a rhythm control deviation. The system then reverses this by increasing the weight of that dimension, making subsequent training pay more attention to time coordination feedback. If the score fluctuation is too large, the system reduces the sensitivity to change to prevent short-term fluctuations from affecting the overall evaluation stability.

[0156] Subsequently, the system reapplies the trend-corrected weights to the scores of each dimension in the current training cycle and recalculates the comprehensive score. This result includes four dimensions: angle accuracy score, time synchronization score, motion stability score, and stage comprehensive score, which respectively reflect the user's comprehensive performance in space, time, and rhythm.

[0157] Finally, the system stores the evaluation results and correlates them with historical data to generate long-term training trend reports. Through this trend analysis mechanism, the system can not only achieve dynamic scoring optimization but also continuously track user training progress, providing quantitative evidence for subsequent personalized training plans.

[0158] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A smart mirror fitness training system based on motion recognition, characterized in that, The system includes: The image acquisition module is used to acquire multiple frames of user pose image data in real time through the camera installed on the smart mirror, and combine environmental depth information and light intensity parameters to adaptively adjust the image sharpness and generate a training image sequence. The pose parsing module is used to identify key points and map skeletons of the human body based on the training image sequence, including temporal smoothing and noise removal of key point trajectories in adjacent frames, and generating a pose data sequence containing spatial coordinates and motion trends. The occlusion determination module is used to perform continuous analysis on the displacement changes of continuous key points based on the posture data sequence. When the trajectory interruption or displacement change of a specific key point is detected, it determines that there is an occlusion area in the corresponding body part. Based on the temporal and spatial changes of the occlusion features, it calculates the occlusion confidence of each occlusion area and generates occlusion feature label data containing occlusion location information and confidence information. The pose compensation module is used to dynamically compensate for missing key points based on occlusion feature labeling data. Based on the geometric constraints of unoccluded key points and the displacement trend curve of historical pose data, it performs spatial interpolation and temporal prediction to generate complete pose reconstruction data. The motion assessment module is used to decompose the user's current motion into multiple local joint motion segments based on the complete posture reconstruction data, and compare the angle change curve of each segment with the preset motion template segment by segment to generate motion assessment results. The feedback presentation module is used to generate comprehensive training feedback data based on the action evaluation results, including local action heat map, correction direction prompts and rhythm guidance information, and generate visual training guidance information to users through the smart mirror display interface and voice module. The occlusion determination module includes: The trajectory analysis unit is used to calculate the spatial displacement curves of each key point based on the attitude data sequence, extract the motion continuity index of adjacent frames, and generate trajectory continuity data. The occlusion recognition unit is used to detect interruptions or abrupt changes in the spatial displacement curve based on trajectory continuity data. When the continuity is lower than a preset continuity threshold, it generates preliminary occlusion marker data. The occlusion classification unit is used to determine whether the occlusion is self-occlusion or external occlusion based on the preliminary occlusion marking data and the human skeleton topology, and to classify self-occlusion into cross-type or overlapping type, and generate classified occlusion data. The confidence calculation unit is used to calculate the occlusion confidence matrix and generate occlusion feature label data based on the classified occlusion data, combined with the occlusion duration, occlusion area and the visibility ratio of surrounding key points.

2. The intelligent mirror fitness training system based on motion recognition according to claim 1, characterized in that, The image acquisition module includes: The environment modeling unit is used to extract information on light intensity, reflection distribution and background depth based on multiple frames of user pose image data, establish an environment feature model, and generate environment parameter data. The adaptive adjustment unit is used to dynamically adjust the exposure time and gain threshold of the camera based on environmental parameter data. When the rate of change of light exceeds the preset light change threshold, it performs fast compensation processing to generate exposure correction image data. The reflection correction unit is used to detect high reflectivity areas in the image based on the exposure correction image data, and to perform brightness equalization correction on the reflectivity areas by interpolation of neighboring frame pixels to generate correction frame data. The multi-frame fusion unit is used to perform motion vector alignment and temporal weighted fusion between adjacent frames based on the correction frame data to generate a training image sequence.

3. The intelligent mirror fitness training system based on motion recognition according to claim 1, characterized in that, The attitude compensation module includes: The occlusion resolution unit is used to extract the spatial location missing region of the occluded key points based on the occlusion feature label data, and to identify the corresponding associated skeletal structure of the unoccluded key points to generate occlusion association data. The constraint generation unit is used to construct a local skeleton topology constraint model based on the spatial relative position of unoccluded key points and the proportion of bone length in the occlusion association data. When the occlusion duration exceeds the preset first duration threshold, the angle constraint interval is dynamically adjusted to generate spatial constraint data. The attitude trend prediction unit is used to calculate the motion direction and velocity change rate of the occluded key points based on the displacement trend information of key points in spatial constraint data and historical attitude data. When the motion direction remains unchanged, it performs time extrapolation prediction, and when the direction changes frequently, it performs neighboring frame interpolation prediction to generate key point trend prediction data. The confidence fusion unit is used to perform confidence-weighted fusion on the predicted coordinates based on the key point trend prediction data and the occlusion confidence of each part in the occlusion feature label data, to generate preliminary compensation coordinate data. The continuous correction unit is used to perform consistency detection on the displacement changes of consecutive frames based on the preliminary compensation coordinate data. When an abnormal jump point is detected, it triggers the smooth correction of adjacent frames and feeds the correction result back to the constraint generation unit for constraint correction to generate complete attitude reconstruction data.

4. The intelligent mirror fitness training system based on motion recognition according to claim 1, characterized in that, The motion assessment module includes: The template calling unit is used to call the preset action template corresponding to the current training type from the system database, and extract the standard angle change curves and time series parameters of each stage in the template to generate template reference data; The phase division unit is used to calculate the angle change rate based on the complete attitude reconstruction data, and divide it into the initial phase, transition phase and termination phase according to the rate change trend to generate motion phase data. The deviation calculation unit is used to compare the user's joint angle curve with the standard angle curve of the corresponding stage in the template reference data segment by segment based on the action stage data, calculate the angle deviation amplitude and duration of each joint, and generate local deviation data. The timing matching unit is used to extract the angular velocity change sequence of each stage based on the local deviation data, and synchronously compare it with the time series parameters in the template reference data to calculate the difference in motion time distribution and generate timing matching degree data. The comprehensive scoring unit is used to calculate the comprehensive training score based on the temporal matching degree data and local deviation data, and generate action evaluation results including angle accuracy score, time synchronization score and action stability score.

5. The intelligent mirror fitness training system based on motion recognition according to claim 1, characterized in that, The confidence calculation unit includes: The occlusion time series analysis subunit is used to extract the occlusion duration of each key point and calculate the occlusion time stability parameter based on the classified occlusion data. When the occlusion time exceeds the preset second duration threshold, the corresponding position confidence is reduced and time-weighted data is generated. The spatial feature evaluation subunit is used to calculate the spatial integrity index of the occluded area based on time-weighted data, combined with the proportion of visible points in the neighborhood of each key point and the spatial distribution density, and to improve the confidence of the area when adjacent visible points are dense and in the same direction, and generate spatial correction data. The neighborhood weight diffusion subunit is used to establish neighborhood transit relationships within the topological structure of key points based on spatially corrected data, perform hierarchical diffusion and smoothing of confidence, and generate neighborhood fusion data. The dynamic confidence update subunit is used to calculate the inter-frame confidence change rate based on the neighborhood fusion data and historical frame confidence records. When the change rate exceeds the preset first change rate threshold, confidence correction is performed and an occlusion confidence matrix is ​​generated.

6. The intelligent mirror fitness training system based on motion recognition according to claim 3, characterized in that, The attitude trend prediction unit includes: The motion pattern recognition subunit is used to analyze the continuous displacement characteristics of key points and determine the motion pattern based on spatial constraint data and historical posture data. When the frequency of displacement direction change is lower than the preset second rate of change threshold, it is identified as stable; otherwise, it is identified as abrupt and generates pattern identification data. The short-term trend generation subunit is used to extract the displacement direction and velocity change rate of the most recent multi-frames from the stable key point sequence in the pattern identification data, perform time extrapolation prediction, and generate short-term prediction data. The long-term trend generation subunit is used to extract global displacement change trends and acceleration distribution characteristics based on the key point sequence of abrupt or nonlinear motion in the pattern identification data, and generate long-term prediction data through neighboring frame interpolation and orientation smoothing calculation. The adaptive trend fusion subunit is used to calculate the dynamic fusion weight of short-term and long-term forecast data based on the directional consistency parameter. When the consistency is high, the proportion of short-term forecast is increased; otherwise, the proportion of long-term forecast is increased, thus generating key point trend forecast data.

7. The intelligent mirror fitness training system based on motion recognition according to claim 3, characterized in that, The confidence fusion unit includes: The prediction confidence matching subunit is used to establish a mapping relationship between predicted coordinates and occlusion confidence based on key point trend prediction data and occlusion feature label data, and to calculate the confidence weight value of each key point to generate weighted prediction data. The error detection and correction subunit is used to calculate the displacement difference of key points in adjacent frames based on the weighted prediction data. When the detected displacement difference exceeds the preset stable threshold, reverse correction is performed according to the historical trend direction to generate error correction data. The multi-frame smoothing fusion subunit is used to perform time-series smoothing processing on the key point coordinates of multiple consecutive frames based on error correction data, and replace the predicted coordinates with the latest visible frame coordinates when the occlusion is removed, generating time-series fusion data; The confidence strategy update subunit is used to calculate the fusion accuracy index based on the time-series fusion data, update the confidence allocation strategy for subsequent frames, and generate preliminary compensation coordinate data.

8. The intelligent mirror fitness training system based on motion recognition according to claim 3, characterized in that, The continuous correction unit includes: The anomaly detection subunit is used to calculate the displacement change rate of key points in adjacent frames based on the preliminary compensation coordinate data. When the displacement change of any key point exceeds the preset stability threshold, the point is identified as an abnormal node and anomaly detection data is generated. The neighboring frame interpolation subunit is used to extract the position sequence of the abnormal node in the neighboring frame based on the abnormal detection data, perform temporal interpolation correction based on the displacement direction and velocity change rate of the neighboring frame, and generate interpolation correction data. The trajectory smoothing correction subunit is used to smooth the motion trajectory of key points in consecutive frames based on interpolation correction data. It generates smooth trajectory data by calculating the directional change amplitude of adjacent frames and limiting over-correction. The constraint feedback subunit is used to calculate the attitude consistency parameters based on the smooth trajectory data. When there is a deviation between the trajectory direction and the skeletal topology constraint model, it automatically adjusts the angle constraint range of the constraint generation unit, updates the spatial constraint data, and generates complete attitude reconstruction data.

9. The intelligent mirror fitness training system based on motion recognition according to claim 4, characterized in that, The comprehensive scoring unit includes: The index extraction subunit is used to extract the joint angle error, time delay difference and stability fluctuation amplitude from the temporal matching degree data and local deviation data, and convert them into angle accuracy index, time consistency index and motion stability index respectively to generate raw scoring data. The indicator weight allocation subunit is used to determine the dynamic weight parameters of the angle accuracy indicator, time consistency indicator, and action stability indicator based on the training type and the user's historical training records, and generate weight allocation data. The comprehensive calculation subunit is used to perform weighted average calculation on each indicator based on the original scoring data and weight allocation data. By accumulating the scores of the three dimensions of angle accuracy, time consistency and motion stability according to the weight ratio, a single frame comprehensive score data is generated. The scoring time-series fusion subunit is used to perform time-weighted fusion processing on the comprehensive scoring data of multiple consecutive frames to generate stage comprehensive scoring data; The scoring trend analysis subunit is used to calculate the scoring trend based on the stage comprehensive scoring data and historical training results. When the overall scoring change rate is detected to exceed the preset scoring change threshold, the weight allocation data is reversed and the action evaluation results including angle accuracy score, time synchronization score, action stability score and stage comprehensive score are generated.

Citation Information

Patent Citations

  • Intelligent mirror design method based on automatic image recognition and motion model comparison

    CN109389054A

  • Running posture continuous capturing and analyzing method for training

    CN120977007A