Exercise guidance method and device and electronic equipment

Abstract

The invention discloses an exercise guidance method and device and electronic equipment. The method comprises the steps that motion process data corresponding to target motion are acquired, and the motion process data comprise process data of target motion executed by a target part; according to the motion process data, action prompt parameters corresponding to the target action are determined, and the action prompt parameters comprise an action type corresponding to the target action and a track feature corresponding to the target action; determining a reconstruction action sequence corresponding to the target action according to the action prompt parameter; performing simulation variation adjustment on original action parameters corresponding to the reconstructed action sequence to obtain a deduction action sequence; and obtaining guidance parameters corresponding to the target action according to the reconstructed action sequence and the deduced action sequence. According to the invention, the technical problem that the guidance cost and the guidance accuracy are difficult to balance during motion guidance in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of sports instruction, and more specifically, to a sports instruction method, apparatus, and electronic device. Background Technology

[0002] With the increasing awareness of fitness among the general public, the market demand for intelligent sports assistive technologies that can provide personalized and precise guidance is becoming increasingly urgent. Currently, the technological implementation paths in this field are mainly divided into two categories, but both involve an irreconcilable contradiction between guidance cost and guidance accuracy.

[0003] One type is the solution based on environmental sensing devices. While this type of solution can acquire relatively complete motion information, its equipment procurement and deployment costs are high, and it is sensitive to and has stringent requirements for environmental conditions, making it difficult to achieve large-scale widespread application in the daily exercise scenarios of ordinary users.

[0004] Another type is the solution based on low-cost, single-point wearable sensors. Its advantages lie in convenience and universality, but it is limited by the inherent information dimension of the sensor, and can only collect local motion data of the body. It cannot effectively reconstruct the user's full-body coordinated posture when performing a target action, and it cannot perceive the interaction information between the moving subject and the external environment.

[0005] Therefore, most existing products only provide basic statistical indicators such as the number of repetitions and local peak speed, and cannot provide in-depth and actionable technical guidance on core issues such as where the movement is not standardized and how to adjust it to optimize the exercise effect.

[0006] There is currently no effective solution to the above problems. Summary of the Invention

[0007] This invention provides a method, apparatus, and electronic device for providing exercise guidance, thereby addressing the technical problem in the related art of balancing guidance cost and accuracy during exercise guidance.

[0008] According to one aspect of the present invention, a motion guidance method is provided, comprising: acquiring motion process data corresponding to a target motion, wherein the motion process data includes process data of a target body part performing a target action; determining motion prompt parameters corresponding to the target action based on the motion process data, wherein the motion prompt parameters include: the action type corresponding to the target action and the trajectory features corresponding to the target action; determining a reconstructed motion sequence corresponding to the target action based on the motion prompt parameters; performing simulation variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence to obtain a deduced motion sequence; and obtaining guidance parameters corresponding to the target action based on the reconstructed motion sequence and the deduced motion sequence.

[0009] Optionally, based on the motion process data, determining the motion prompt parameters corresponding to the target action includes: determining the linear acceleration corresponding to the target part in the motion process data; determining the acceleration modulus corresponding to the linear acceleration; marking the corresponding time as the action time when the acceleration modulus exceeds a predetermined modulus threshold; extracting data segments of fixed duration before and after the action time from the motion process data to obtain the motion process data corresponding to the target action; and generating the motion prompt parameters corresponding to the target action based on the motion process data.

[0010] Optionally, generating motion prompt parameters corresponding to the target action based on the motion process data includes: when the motion process data is collected by a motion device, converting the motion process data in a local coordinate system based on the motion device to obtain converted motion data in a global coordinate system based on physical space; determining a motion feature vector based on the converted motion data, wherein the motion feature vector includes a temporal feature vector and a spatial feature vector characterizing the target action; and determining the action type corresponding to the target action based on the motion feature vector.

[0011] Optionally, determining the reconstructed action sequence corresponding to the target action based on the action prompt parameters includes: determining an initial action sequence based on the action prompt parameters; and correcting the initial action sequence to obtain the reconstructed action sequence.

[0012] Optionally, determining the initial action sequence based on the action prompt parameters includes: determining the trajectory dynamics features corresponding to the target part based on the action prompt parameters; and determining the initial action sequence matching the target action from a professional action prior library based on the trajectory dynamics features. The professional action prior library includes multiple entries stored as key-value pairs. The value in each key-value pair represents an index feature, and the key represents an action sequence. The corresponding key is determined based on the corresponding value. The index feature includes at least one of the following: instantaneous kinematic features, energy envelope features, and spatial geometric features. The instantaneous kinematic features represent the relative displacement of the target part in physical space. The energy envelope features include energy peak temporal features and energy decay features. The spatial geometric features include trajectory curvature features and trajectory torsion features corresponding to the target part.

[0013] Optionally, based on the trajectory dynamics features, an initial action sequence matching the target action is determined from a professional action prior library, including: determining the cosine similarity between the trajectory dynamics features and multiple values ​​in the professional action prior library; determining a target value from the multiple values ​​based on the cosine similarity between the features and the values; and determining the action sequence represented by the key corresponding to the target value as the initial action sequence.

[0014] Optionally, correcting the initial action sequence to obtain the reconstructed action sequence includes: determining the theoretical position sequence corresponding to the initial action sequence for the target part; comparing the theoretical position sequence with the actual position sequence of the target part to determine error data corresponding to the initial action sequence; backpropagating the error data to correct the initial action sequence and obtain the reconstructed action sequence.

[0015] Optionally, backpropagating the error data to correct the initial action sequence and obtain a reconstructed action sequence includes: determining the biological constraints corresponding to the skeletal key points; and under the biological constraints, backpropagating the error data to correct the initial action sequence and obtain a reconstructed action sequence.

[0016] Optionally, performing simulation variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence to obtain the deduced motion sequence includes: when the target motion is a ball motion, determining the motion parameters and motion vector parameters corresponding to the motion moment based on the reconstructed motion sequence, wherein the motion parameters include at least one of the following: motion velocity, peak acceleration, vibration spectrum; determining the incoming ball velocity vector parameters based on the motion parameters and the collision physics network; determining the ball's motion trajectory based on the motion vector parameters and the incoming ball velocity vector parameters; and performing simulation variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence based on the ball's motion trajectory to obtain the deduced motion sequence.

[0017] Optionally, performing simulated variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence to obtain a deduced motion sequence includes: when the target motion is a ball-like motion, performing simulated variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence to obtain multiple candidate deduced motions; retrieving a benefit function, wherein the benefit function includes a distance benefit term, a ball flight speed benefit term, and a motion adjustment amplitude term, the distance benefit term representing the benefit term brought about by the distance between the ball's landing point and the boundary line; determining the benefit index corresponding to each of the multiple candidate deduced motions based on the benefit function; and determining the deduced motion sequence from the multiple candidate deduced motions based on the corresponding benefit index.

[0018] Optionally, the original motion parameters corresponding to the reconstructed motion sequence are subjected to simulated variational adjustment to obtain multiple candidate deduced motions, including: determining motion parameters whose influence index on the hitting result is greater than a predetermined influence index to obtain the original motion parameters; constructing a multidimensional probability distribution centered on the original motion parameters; sampling multiple variational parameters from the multidimensional probability distribution; and simulating the hitting process corresponding to the adjusted motion based on the multiple variational parameters to generate the multiple candidate deduced motions.

[0019] Optionally, determining the revenue index corresponding to each of the plurality of candidate deduced actions based on the revenue function includes: when the corresponding candidate deduced actions include deduced action parameters, determining the state of the target object at the next moment based on the deduced action parameters corresponding to the current moment, thereby obtaining the corresponding object moment state, wherein the deduced action parameters include force parameters and speed parameters; determining the landing point corresponding to each of the plurality of candidate deduced actions based on the object moment states corresponding to each of the plurality of candidate deduced actions; and determining the revenue index corresponding to each of the plurality of candidate deduced actions and the distance revenue term based on the landing points corresponding to each of the plurality of candidate deduced actions.

[0020] Optionally, based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained, including: when the guidance parameters include guidance voice text, determining the action difference parameters between the reconstructed action sequence and the deduced action sequence on key parameters, wherein the key parameters are parameters that result in an action deformation index greater than a predetermined deformation threshold; converting the action difference parameters into the guidance voice text, and playing the guidance voice text.

[0021] Optionally, based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained, including: when the guidance parameters include a guidance video, determining a first video corresponding to the reconstructed action sequence and a second video corresponding to the deduced action sequence; superimposing the first video and the second video to obtain the guidance video.

[0022] According to one aspect of the present invention, a motion guidance device is provided, comprising: an acquisition module, configured to acquire motion process data corresponding to a target motion, wherein the motion process data includes process data of a target body part performing a target action; a first determination module, configured to determine motion prompt parameters corresponding to the target action based on the motion process data, wherein the motion prompt parameters include: a motion type corresponding to the target action and a trajectory feature corresponding to the target action; a second determination module, configured to determine a reconstructed motion sequence corresponding to the target action based on the motion prompt parameters; a simulation variational module, configured to perform simulation variational adjustments on the original motion parameters corresponding to the reconstructed motion sequence to obtain a deduced motion sequence; and a third determination module, configured to obtain guidance parameters corresponding to the target action based on the reconstructed motion sequence and the deduced motion sequence.

[0023] According to one aspect of the present invention, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the motion guidance method as described above.

[0024] According to one aspect of the present invention, a computer-readable storage medium is provided, wherein when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the motion guidance method as described above.

[0025] In this embodiment of the invention, motion process data corresponding to the target motion is acquired, including process data of the target body part performing the target action; based on the motion process data, motion prompt parameters corresponding to the target action are determined, including: the action type corresponding to the target action and the trajectory features corresponding to the target action; based on the motion prompt parameters, a reconstructed action sequence corresponding to the target action is determined; the original action parameters corresponding to the reconstructed action sequence are subjected to simulated variational adjustment to obtain a deduced action sequence; based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained. This method combines data-driven motion reconstruction with counterfactual deduction. Motion prompt parameters are extracted from local motion process data to generate a precise reconstructed action sequence matching the target action. Then, simulated variational adjustment of the original action parameters under physiological constraints is performed to obtain the deduced action sequence. The difference between the reconstructed action sequence and the deduced action sequence is compared to quantify the impact of motion adjustment on the exercise effect. This achieves the goal of high-precision motion analysis and personalized optimization guidance using low-cost local motion process data, thus achieving a dual optimization effect of significantly reducing guidance costs and significantly improving guidance accuracy. This solves the technical problem in related technologies where it is difficult to balance guidance costs and guidance accuracy during exercise guidance. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0027] Figure 1 This is a flowchart of a motion guidance method according to an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of a motion guidance method provided by an optional embodiment of the present invention;

[0029] Figure 3 This is a structural block diagram of a motion guidance device according to an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] Example 1

[0033] According to an embodiment of the present invention, an embodiment of a motion guidance method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] Figure 1 This is a flowchart of a motion guidance method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0035] Step S102: Obtain motion process data corresponding to the target motion, wherein the motion process data includes process data of the target part performing the target action;

[0036] In step S102 of this application, the raw sensor data sequence generated by the user during the relevant sports activities is obtained. This data fully records the dynamic changes of specific parts of the user's body when performing a specific technical action.

[0037] This involves the target sport, which can be of various types and is not limited here. It can be adapted to the actual application and scenario. For example, it can refer to a certain type of sports, such as ball sports, frisbee sports, etc. Ball sports can include badminton, tennis, table tennis, etc.

[0038] This involves motion process data, which can be collected by a user-worn smart motion sensing device. It can be a time-series signal collected by the inertial measurement unit built into the motion sensing device, and can include triaxial linear acceleration, triaxial angular velocity, and triaxial magnetometer data. This data is continuously recorded at a high sampling frequency, forming a multidimensional time series characterizing the user's motion state. Therefore, this application can be achieved by collecting motion process data from specific parts of the body.

[0039] This involves the target body part, which can be a part of the user's body where the motion sensing device is fixedly worn. For example, if the target sport is badminton, when the user wears the motion sensing device on their wrist, the target body part can be the wrist. During badminton, this is to facilitate data collection without affecting the movement, and the data must reflect clear actions.

[0040] This involves target actions, which are single action units performed by a user during a target movement, possessing a clear start and end point and specific technical semantics. Specifically, in badminton, forehand clears, backhand drop shots, and jump smashes can all be analyzed as independent target actions.

[0041] This step establishes a low-cost data acquisition entry point, requiring only process data of the target body part performing the target action. For example, users only need to wear ordinary smart devices to obtain core motion data in natural motion scenarios without being bothered. This eliminates the need for expensive and complex multi-camera arrays or optical motion capture equipment, significantly reducing the hardware cost and environmental deployment complexity of motion data acquisition from the source, and laying the data foundation for the realization of popularized intelligent motion guidance.

[0042] Step S104: Based on the motion process data, determine the motion prompt parameters corresponding to the target action, wherein the motion prompt parameters include: the action type corresponding to the target action and the trajectory features corresponding to the target action;

[0043] In step S104 of this application, the original motion process data obtained in step S102 is processed and analyzed to identify specific data segments containing the target action. Then, the data segments are processed to generate a set of structured and semantic descriptive parameters to accurately characterize the core attributes and kinematic features of the target action.

[0044] This involves motion cue parameters, which are structured outputs obtained by the system after performing high-level abstraction on the raw sensor data. They serve as a bridge to transform low-dimensional time-series signals into high-level semantic descriptions and can include two core components: motion type and trajectory features. Together, they constitute the precise conditions or guiding signals for performing full-body posture reasoning.

[0045] This involves action type, which is a semantic classification label for the technical category to which the target action belongs. Specifically, in badminton, this could be a forehand high clear, a backhand drop shot, or a jump smash. This label is obtained through pattern recognition of action data segments, providing a clear prior knowledge of the action intent for subsequent models.

[0046] This involves trajectory features, which are the quantification and encoding representation of the spatial motion path of the target part during the execution of the target action. They are generated by denoising the raw sensor data, transforming it from the device coordinate system to the world coordinate system, and extracting temporal kinematic information. They can usually be represented as a high-dimensional feature vector or embedding tensor to accurately describe the spatiotemporal pattern of the target part's motion.

[0047] This step enables feature extraction and information condensation from raw sensor signals to high-level semantic parameters. High-frequency, multi-dimensional time-series data is transformed into structured conditional inputs with higher information density that can be directly utilized by subsequent generative models. This reduces the complexity and computational cost of subsequent model processing. By providing precise action types and trajectory features as strong guiding signals, it provides a reliable and efficient conditional input foundation for the core innovation of inferring whole-body posture from local trajectories. This is a crucial preprocessing step to ensure the accuracy and rationality of whole-body motion reconstruction.

[0048] Step S106: Based on the action cue parameters, determine the reconstructed action sequence corresponding to the target action;

[0049] In step S106 of this application, the action prompt parameters generated in step S104 are used as precise guiding conditions to generate a matching reconstructed action sequence.

[0050] This involves reconstructing motion sequences, which are serialized data generated by the system based on local data inference or generation. These sequences fully describe the changes in the user's whole-body posture over time when performing a target action. They are typically represented as a parametric sequence of a standard human body model and may include rotational and displacement information of key joints in three-dimensional space, thus reproducing the user's coherent whole-body movements from the initial posture to the completion of the shot and the recovery phase. Optionally, the reconstructed motion sequence can be a three-dimensional, multi-joint motion sequence conforming to human kinematics and biomechanical constraints.

[0051] This step overcomes the fundamental technical limitation that relying solely on single-point sensor data from moving parts of the target body cannot perceive the entire body's posture, achieving a leap from local observation to global reconstruction. By leveraging the powerful prior knowledge learning and condition generation capabilities of generative artificial intelligence models, the system can supplement the most reasonable full-body movements. This allows for the digital reproduction of full-body 3D movements—previously requiring expensive multi-camera motion capture systems—at extremely low hardware costs. This provides an indispensable and high-dimensional motion data foundation for subsequent precise physical deduction and movement comparison guidance.

[0052] Step S108: Perform simulation variational adjustment on the original action parameters corresponding to the reconstructed action sequence to obtain the deduced action sequence;

[0053] In step S108 provided in this application, parameter changes are generated based on the user's original action parameters, and then the complete motion process and physical results corresponding to the adjusted action are simulated and executed.

[0054] This involves raw motion parameters, which specifically refer to several quantifiable kinematic parameters extracted from the reconstructed motion sequence that have a key impact on the motion effect. Specifically, in badminton swing motion, these may include wrist pronation / external rotation angle, wrist flexion / extension angle, timing of impact, and swing speed, which are the direct objects of subsequent optimization and adjustment.

[0055] This involves the deduced action sequence, which is a sequence of actions derived through simulation variational adjustment and physical deduction. In the presence of multiple action sequences, the deduced action sequence can be identified as the optimized sequence that yields the highest benefit for a specific objective. This sequence represents the better solution or suggested action calculated by the system based on physical laws for the current situation, serving as the benchmark for subsequent comparison with the user's original actions and the generation of guidance suggestions.

[0056] This step provides tactical-level guidance capabilities that traditional data-driven equipment lacks, enabling a qualitative leap from movement reproduction to movement optimization. Linking movement adjustments to the final athletic outcome allows for the provision of physically based and actionable optimization suggestions, similar to those offered by professional coaches. This is the core technological guarantee for delivering high-value guidance with low-cost equipment.

[0057] Step S110: Based on the reconstructed action sequence and the deduced action sequence, obtain the guidance parameters corresponding to the target action.

[0058] In step S110 of this application, a multi-dimensional comparative analysis is performed between the reconstructed action sequence generated in step S106 and the deduced action sequence obtained in step S108. This includes calculating the differences between the two in key kinematic parameters and converting these quantified differences into feedback formats that are easy for users to understand and execute, ultimately generating specific output parameters to guide users in improving their actions.

[0059] This involves guidance parameters, which are the final output of this method and the result data used to guide users to improve their target actions. They are the transformation of the system's analysis conclusions into user-oriented interactive parameters.

[0060] This step enables the final transformation and delivery of intelligent analysis results into personalized and actionable guidance. It translates complex internal calculations into intuitive and actionable instructions that users can readily understand and execute, transforming the system into a visually appealing and semantically clear interactive coach. This significantly enhances the user experience and the acceptability of the guidance, making professional-grade sports guidance from low-cost devices truly accessible.

[0061] Through the above steps S102-S110, motion process data corresponding to the target motion is obtained, wherein the motion process data includes the process data of the target part performing the target action; based on the motion process data, motion prompt parameters corresponding to the target action are determined, wherein the motion prompt parameters include: the action type corresponding to the target action and the trajectory features corresponding to the target action; based on the motion prompt parameters, the reconstructed action sequence corresponding to the target action is determined; the original action parameters corresponding to the reconstructed action sequence are simulated and varied to obtain the deduced action sequence; based on the reconstructed action sequence and the deduced action sequence, the guidance parameters corresponding to the target action are obtained. This method combines data-driven motion reconstruction with counterfactual inference. By extracting motion cue parameters from local motion process data, it generates a precise reconstructed motion sequence that matches the target motion. Then, it performs simulated variational adjustments on the original motion parameters under physiological constraints to obtain the inferred motion sequence. By comparing the differences between the reconstructed and inferred motion sequences, the impact of motion adjustments on the motion effect is quantified. This achieves the goal of high-precision motion analysis and personalized optimization guidance using low-cost local motion process data. As a result, it achieves the dual optimization effect of significantly reducing guidance costs and significantly improving guidance accuracy. This solves the technical problem in related technologies where it is difficult to balance guidance costs and guidance accuracy when providing motion guidance.

[0062] As an optional embodiment, based on motion process data, determining motion prompt parameters corresponding to the target action includes: determining the linear acceleration corresponding to the target part in the motion process data; determining the acceleration magnitude corresponding to the linear acceleration; marking the corresponding time as the action time when the acceleration magnitude exceeds a predetermined magnitude threshold; extracting data segments of fixed duration before and after the action time from the motion process data to obtain motion process data corresponding to the target action; and generating motion prompt parameters corresponding to the target action based on the motion process data.

[0063] This embodiment illustrates the process of accurately locating and extracting effective data fragments directly related to the target action by detecting specific dynamic events, and then generating action cue parameters.

[0064] Among them, linear acceleration is involved. Linear acceleration is a physical quantity that is directly measured by the accelerometer of the relevant motion sensing device and describes the change in the speed of motion of the target part along each coordinate axis. It is one of the core raw data that characterizes the intensity of motion.

[0065] This involves the acceleration modulus, a scalar value obtained by calculating the combined magnitude of the three-axis linear acceleration vectors. It eliminates the influence of direction and can comprehensively reflect the overall intensity of the target part's motion at any instant.

[0066] This involves a predetermined modulus threshold, which is a pre-set critical value for acceleration modulus used to distinguish between normal motion and target motion. It is derived from the analysis of a large amount of motion data. When the measured modulus exceeds this threshold, it is determined that a target motion worth analyzing has occurred.

[0067] This involves the moment of motion, which is the precise point in time when the acceleration modulus first exceeds a predetermined modulus threshold. Specifically, in badminton, this usually corresponds to the instant the racket collides with the shuttlecock during the hitting motion, and is the core reference point for segmenting and locating the entire motion sequence.

[0068] This involves data segments, which are subsequences formed by extracting fixed-length segments of raw motion data both forward and backward from the moment of the action. A data segment can include the preparation, execution, and conclusion phases of the target action, serving as a complete data unit for subsequent feature extraction and motion analysis.

[0069] In this step, the system first calculates the instantaneous motion intensity of the wearing site from the continuous sensor data stream; when the intensity exceeds a defined threshold, the system marks this instant as a critical event point. Subsequently, the system extracts a fixed-length data window around this event point, containing complete motion context information. Finally, feature extraction and parameter generation are performed based on this clean, targeted data segment.

[0070] This method enables precise and automatic event detection and data segmentation of target actions. By avoiding the processing of lengthy continuous data containing a large amount of invalid motion, it can efficiently and accurately capture every valuable motion instance from the continuous motion data stream, significantly reducing the amount of data that subsequent calculation modules need to process and improving the overall efficiency of the system. Reliable physical indicators ensure that the extracted data segments are highly correlated with the target action to be analyzed, laying an accurate data foundation for generating high-quality motion cue parameters, thereby ensuring the reliability of subsequent motion reconstruction and guided analysis.

[0071] Specifically, in badminton applications, taking the forehand smash as an example, the threshold setting is based on motion data statistics. The predetermined modulus threshold for hit detection is set to 3g, and the data extraction window is 1.0 second before and 0.5 seconds after the hit. Event detection involves real-time calculation of the wristband's acceleration modulus. When the detected modulus first exceeds 3g, that moment is marked as the hit action. Data extraction extracts 1.5 seconds of sensor data centered on the action, forming a data segment containing the complete backswing, hit, and follow-through. After filtering and coordinate system transformation of the segment data, it is input into a trained badminton action classification model, outputting action type labels and extracting trajectory feature vectors, ultimately generating structured action prompts.

[0072] As an optional embodiment, generating motion prompt parameters corresponding to the target action based on motion process data includes: when the motion process data is collected by a motion device, converting the motion process data in a local coordinate system based on the motion device to obtain converted motion data in a global coordinate system based on the physical space; determining motion feature vectors based on the converted motion data, wherein the motion feature vectors include temporal feature vectors and spatial feature vectors representing the target action; and determining the action type corresponding to the target action based on the motion feature vectors.

[0073] This embodiment illustrates the process of accurately identifying action semantic categories and generating high-quality action cue parameters by performing coordinate system 1 and high-level feature extraction on raw sensor data.

[0074] This includes sports equipment, which refers to smart wearable devices worn on a user's target body part to collect data during exercise. These devices can include smart bracelets, smartwatches, etc., and have an integrated inertial measurement unit.

[0075] This involves a local coordinate system, which is established with the motion device itself as the origin and can move and rotate along with the motion device worn by the user. The acceleration, angular velocity, and other data directly measured by the sensors are all based on this coordinate system, and its direction is independent of the user's actual body orientation.

[0076] This involves a global coordinate system, which is a reference coordinate system with constant orientation established based on a fixed physical space. Transforming data to this coordinate system can eliminate interference caused by changes in device wearing orientation and posture, giving the description of the motion trajectory absolute spatial directional meaning.

[0077] This involves transforming motion data, which is motion process data after coordinate transformation. By using sensor fusion algorithms, the original data based on the device's local coordinate system is solved and transformed to a global coordinate system to obtain data describing the actual motion direction and trajectory of the target part in world space.

[0078] This involves motion feature vectors, which are high-dimensional vectors obtained by further abstracting and encoding transformed motion data. They are used to comprehensively and compactly represent a motion segment, including at least temporal feature vectors extracted from the time dimension and spatial feature vectors extracted from the spatial dimension.

[0079] This involves temporal feature vectors, which mainly describe the evolution pattern of an action on the time axis. These vectors can include the rhythm of the action, the timing of explosive power, the frequency domain characteristics of speed changes, etc., reflecting the dynamic characteristics of the action.

[0080] This involves spatial feature vectors, which mainly describe the geometric shape of the motion trajectory in three-dimensional space and can reflect the spatial structural characteristics of the motion.

[0081] This step first addresses the orientation dependency issue of sensor data by using a coordinate transformation algorithm to unify the raw data from the device coordinate system to a fixed world coordinate system, resulting in consistent and comparable transformed motion data. Then, features that characterize the temporal evolution and spatial morphological structure of the motion are extracted from this standardized data, forming a robust motion feature vector. Finally, using this feature vector as input, a pre-trained classification model outputs the specific type label corresponding to the motion.

[0082] This approach enables deep purification and high-order semantic understanding of raw sensor data. Specifically, coordinate transformation eliminates noise introduced by the randomness of device wearing, ensuring the consistency of data benchmarks across different actions and users, providing standardized input for feature extraction. Simultaneously, the fusion of temporal and spatial multi-dimensional feature extraction captures the dynamics and kinematic patterns of actions more comprehensively and fundamentally, enhancing the discriminative and representational capabilities of action features. Using high-quality feature vectors for action type recognition improves the accuracy and robustness of action semantic classification, thereby ensuring the reliability of the generated action type—a crucial cue parameter.

[0083] Specifically, in badminton applications, a smart bracelet worn on the wrist continuously collects data. Its sensor data is based on a local coordinate system (the origin is at the center of the bracelet, and the axis changes with wrist rotation). To eliminate the influence of the wearing direction, the system uses the Madgwick algorithm for sensor fusion, transforming the raw data into a global coordinate system based on gravity and geomagnetic north, resulting in consistent motion data. Subsequently, motion feature vectors are extracted from the transformed wrist trajectory. The root mean square velocity, zero-crossing rate, and spectral energy of acceleration in a specific frequency band (e.g., 3-8Hz, reflecting the force characteristics of the shot) are calculated to quantify the rhythm and explosive power of the movement, yielding a temporal feature vector. Principal component analysis (PCA) eigenvalues, mean curvature, and variance of torsion are calculated to describe whether the trajectory is a straight drive or a high clear, yielding a spatial feature vector. Finally, the feature vectors that integrate temporal and spatial information are input into a pre-trained lightweight 1D-CNN classifier, which outputs specific action types, such as "Forehand_Clear" or "Backhand_Drop", thereby generating high-quality action cue parameters.

[0084] As an optional embodiment, determining the reconstructed action sequence corresponding to the target action based on the action prompt parameters includes: determining the initial action sequence based on the action prompt parameters; and correcting the initial action sequence to obtain the reconstructed action sequence.

[0085] This embodiment illustrates the specific process of determining the final reconstructed action sequence using a two-step strategy of first matching and then correcting.

[0086] This involves an initial action sequence, which can be a pre-determined action sequence or, optionally, a complete full-body 3D action data sequence that is most likely to correspond to the action based on action prompt parameters and retrieved from a pre-built professional action prior library. It can represent a candidate action prototype that is reasonable in terms of movement mode but may have subtle deviations from the user's actual body posture.

[0087] In this step, the system first uses motion cue parameters as search criteria to find the most similar full-body motion pattern from a massive professional motion database, using this as the starting point for generation. Subsequently, the system compares the initial motion sequence with the actual wrist trajectory measured by the sensor, calculates the error between the two, and uses optimization algorithms such as backpropagation to iteratively adjust the full-body posture while maintaining the naturalness and rationality of the movements, until the wrist trajectory of the generated motion coincides with the measured trajectory, finally obtaining the corrected reconstructed motion sequence.

[0088] This method achieves an effective balance and dual guarantee between the rationality and accuracy of generated movements. By retrieving the initial sequence from a professional prior library, the professionalism and rationality of the generated movements in terms of motion patterns are ensured. This avoids generating bizarre postures that violate common sense about movement from scratch, providing a high-quality and high-starting-point initial value for subsequent corrections. Through constraint correction based on real sensor data, the generated movements are forced to precisely match the individual user's execution details. This solves the problem of superficial resemblance but lack of true resemblance that may occur with simple prior matching, significantly improving the accuracy of the reconstructed movements in restoring the user's actual posture.

[0089] Specifically, in badminton applications, taking the forehand smash as an example, the system matches motion prompt parameters, including the "Forehand_Smash" tag and wrist trajectory features, with a priori library of professional badminton motions. This library stores the wrist dynamics feature vectors of professional athletes during smashes ("keys") and the corresponding standard SMPL (Smart Smash Proportional) full-body motion sequences ("values"). By calculating cosine similarity, the most matching professional smash motion is retrieved as the initial motion sequence. The system then imports this initial sequence into a conditional diffusion model. During the iterative denoising process of the model, the wrist endpoint position of the generated posture is strictly constrained to align with the trajectory data actually recorded by the smart bracelet. By calculating the error between the theoretical and measured wrist positions, and using backpropagation to adjust the posture parameters of all joints while maintaining biomechanical constraints such as bone length and joint range of motion, a reconstructed motion sequence is ultimately generated that conforms to the professional smash pattern while accurately reproducing the user's individual force application habits and angles.

[0090] As an optional embodiment, determining the initial action sequence based on action cue parameters includes: determining the trajectory dynamics characteristics corresponding to the target part based on the action cue parameters; and determining the initial action sequence matching the target action from a professional action prior library based on the trajectory dynamics characteristics. The professional action prior library includes multiple entries stored as key-value pairs. The value in each key-value pair represents an index feature, and the key represents the action sequence. The corresponding key is determined based on the corresponding value. The index features include at least one of the following: instantaneous kinematic features, energy envelope features, and spatial geometric features. The instantaneous kinematic features represent the relative displacement of the target part in physical space. The energy envelope features include energy peak temporal features and energy decay features. The spatial geometric features include the trajectory curvature features and trajectory torsion features corresponding to the target part.

[0091] In this embodiment, the process of determining the initial sequence of actions is described.

[0092] This involves trajectory dynamics features, which are a set of quantitative feature vectors calculated and refined from motion cue parameters to characterize the essence of the motion pattern of the target part, and describe the characteristics of motion from different physical dimensions.

[0093] This involves a professional movement prior library, which is a data warehouse that pre-builds and stores a large amount of standard movement data of professional athletes. Each movement added to the library can be cleaned, labeled, and parameterized to form structured knowledge that can be efficiently retrieved and used by computers.

[0094] This involves entries, which are the basic units for storing individual action data in the professional action prior library. Each entry can be organized using an efficient key-value data structure.

[0095] This involves key-value pairs, a data storage structure where the value stores index features used for fast retrieval and comparison, and the key stores the complete action sequence data corresponding to that index feature. By calculating the similarity between the input feature value and the values ​​in the database, the corresponding key can be found and returned.

[0096] This involves index features, which are feature vectors extracted from professional motion data and stored at the median position of key-value pairs. These features serve as the basis for retrieval and can include features that describe wrist trajectories from different perspectives. Specifically, they are divided into three categories: instantaneous kinematic features, energy envelope features, and spatial geometric features.

[0097] This involves instantaneous kinematic features, which are used to describe the motion state of the target part at each moment. The core is its relative displacement in physical space, that is, the sequence of wrist position changes over time, which can reflect the path and amplitude of the movement.

[0098] This involves energy envelope features, which are used to describe the temporal pattern of energy release during an action.

[0099] This involves spatial geometric features, which are used to describe the bending and twisting morphology of the target part's motion trajectory in three-dimensional space.

[0100] In this step, the system first performs in-depth processing of the motion cue parameters to calculate multi-dimensional trajectory dynamics features that comprehensively reflect the movement patterns of the target body part. Then, these features are used as query vectors and compared with the pre-stored index feature vectors of all entries in the professional motion prior library. Finally, the key corresponding to the entry with the highest similarity—that is, the best-matching professional full-body motion sequence—is used as the initial motion sequence generated for this operation.

[0101] This approach enables precise and efficient action retrieval based on multi-dimensional motion. By integrating features from kinematic, energetic, and geometric dimensions, a highly robust action retrieval method is constructed, improving the accuracy of finding the most relevant and reasonable action prototypes from a vast prior database. Furthermore, the use of a key-value pair index structure transforms complex action sequence similarity calculations into high-dimensional feature vector similarity calculations, significantly enhancing the speed and efficiency of retrieval and matching, meeting the system's real-time or near-real-time response requirements.

[0102] Specifically, in badminton applications, taking the backhand drop shot as an example, the system first extracts multi-dimensional trajectory dynamics features from the action prompt parameters: instantaneous kinematic features (such as the world coordinate system displacement sequence of the wrist within 0.5 seconds before and after the shot), energy envelope features (such as the relative timing of the energy peak and the decay rate), and spatial geometric features (such as the average curvature and torsion changes of the trajectory). Next, the system concatenates these features into a query vector and performs a cosine similarity comparison with the index feature vectors (values) stored in each entry of a professional badminton action prior library. This library is organized in key-value pairs, where the "value" is the vector composed of the three types of features mentioned above, and the "key" is the corresponding parameterized professional full-body action sequence. Finally, the system selects the entry with the highest similarity and determines the professional action sequence represented by its "key" as the initial action sequence for this task, serving as a high-quality starting point for generating personalized user actions.

[0103] As an optional embodiment, based on trajectory dynamics characteristics, an initial action sequence matching the target action is determined from a professional action prior library, including: determining the cosine similarity between the trajectory dynamics characteristics and multiple values ​​in the professional action prior library; determining the target value from the multiple values ​​based on the cosine similarity with the multiple values; and determining the action sequence represented by the key corresponding to the target value as the initial action sequence.

[0104] This embodiment illustrates the process of performing efficient and accurate similarity retrieval in a professional action prior database by calculating a specific quantitative index of cosine similarity, thereby accurately matching and obtaining the initial action sequence.

[0105] This involves cosine similarity, a mathematical measure used to measure the consistency of two vectors in direction. Its value range is [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors are, that is, the more similar the feature patterns are. It can be used to quantify the similarity between the trajectory dynamic feature vector of the user's current action and the index feature vector (value) of each entry in the prior library.

[0106] This involves a target value, which is the index feature vector that is most similar to the user's current trajectory dynamics after calculating all cosine similarities. It represents the professional action feature found in the library that has the highest matching degree with the user's current action pattern.

[0107] This involves keys, which are the parts of the data that store the complete whole-body movement sequence and are paired with the target value in the same entry. Once the target value is determined, the movement sequence represented by the key associated with it can be directly retrieved based on the mapping relationship between the key and the value pair.

[0108] In this step, the user's current action trajectory dynamic feature vector is first treated as the query vector. The system iterates through the index feature vectors (values) of all entries in the professional action prior library, calculating the cosine similarity between each value and the query vector. Next, the system sorts or compares all calculated similarities, selecting the index feature vector with the highest similarity and marking it as the target value. Finally, based on the key-value pair storage structure, the system retrieves the key paired with the target value; the complete action sequence stored in this key is the initial action sequence matched in this retrieval.

[0109] This method enables fast and stable action matching based on a vector space model. Cosine similarity, as a metric, focuses on comparing the direction of feature vectors rather than their absolute magnitude, making it insensitive to changes in feature amplitude and thus improving the robustness of the matching. It effectively addresses the differences in feature vector norms caused by varying force exerted by different users. Furthermore, the calculation of cosine similarity is computationally efficient, suitable for large-scale, fast retrieval in high-dimensional vector spaces, meeting the performance requirements of real-time or near-real-time responses. Through quantified and automated matching mechanisms, the most relevant action prototypes can be accurately located from massive amounts of professional prior knowledge, ensuring that the initial action sequences are not only highly matched in kinematic patterns.

[0110] Specifically, in badminton applications, taking the forehand high clear as an example, the system extracts the trajectory dynamics feature vector from the user's action as the query vector and calculates the cosine similarity between it and each of the "values" (i.e., pre-stored index feature vectors) in the professional action prior library. After calculation, the system selects the index feature vector with the highest similarity as the target value. Then, based on the key-value pair mapping relationship, it retrieves the "key" paired with the target value, i.e., a parameterized professional forehand high clear full-body SMPL action sequence, and determines it as the initial action sequence for this application. This process efficiently achieves accurate matching of the best action prototype from a massive amount of professional priors through cosine similarity, laying an accurate foundation for the subsequent generation of personalized actions.

[0111] As an optional embodiment, correcting the initial action sequence to obtain the reconstructed action sequence includes: determining the theoretical position sequence corresponding to the initial action sequence for the target part; comparing the theoretical position sequence with the actual position sequence of the target part to determine the error data corresponding to the initial action sequence; backpropagating the error data to correct the initial action sequence and obtain the reconstructed action sequence.

[0112] This embodiment illustrates the process of iteratively optimizing the whole-body posture by establishing and minimizing the difference between the theoretical and actual positions of the observation points, thereby correcting the initial action sequence based on prior matching into a reconstructed action sequence that precisely matches the individual execution details of the user.

[0113] This involves a theoretical position sequence, which is a sequence of theoretical motion trajectory points of the target part in three-dimensional space, calculated by forward kinematics based on the human body model and joint parameters in the initial motion sequence.

[0114] This involves the real position sequence, which is the sequence of real spatial position points of the target part recorded in the global coordinate system during actual movement, obtained by the motion sensing device and processed by coordinate transformation and other methods.

[0115] This involves error data, which is the difference information calculated by comparing the theoretical position sequence with the actual position sequence step by step. It is usually represented as the spatial distance vector between two position points at each moment, which quantifies the inaccuracy of the initial action sequence in reproducing the user's actual wrist trajectory.

[0116] In this step, firstly, a human kinematics model is used to calculate the theoretical trajectory of the corresponding motion target part based on the initial whole-body posture parameters. Then, this theoretical trajectory is compared point-by-point with the actual motion target part trajectory obtained from sensor measurements and conversion, and the trajectory error is calculated. Finally, optimization techniques such as backpropagation are used to iteratively adjust the joint parameters in the initial whole-body posture with the goal of minimizing this trajectory error, until the generated motion sequence's motion target part trajectory highly matches the actual trajectory. The optimized sequence obtained at this point is the final reconstructed motion sequence.

[0117] This method enables the generation of actions that are forcibly aligned and precisely fitted to the user's actual performance details. By iteratively optimizing and forcing the generated full-body posture to conform to this observable fact, it solves the problem of superficial resemblance that may arise from relying solely on prior action matching—that is, the action pattern may appear reasonable, but it doesn't match the user's actual force application habits and individual characteristics such as flexibility. Correction through backpropagation errors improves the accuracy and fidelity of the reconstructed action in restoring the user's personalized posture, ensuring that the full-body movements inferred from local data are consistent with reality.

[0118] Specifically, in badminton applications, taking the forehand smash as an example, the system first calculates the theoretical position sequence of the wrist joint using forward kinematics based on the matched initial motion sequence (professional smash posture). Simultaneously, a smart bracelet collects and transforms coordinates to obtain the user's actual wrist position sequence during the smash. The system compares these two sequences frame by frame, calculating the spatial position deviation at each moment to form error data. Subsequently, the system uses optimization algorithms such as gradient descent to backpropagate this error. While maintaining biomechanical constraints such as the length of the human skeleton, it iteratively adjusts the parameters of various joints in the body, such as the shoulder, elbow, and waist, in the initial motion sequence until the generated wrist trajectory highly overlaps with the measured trajectory, ultimately obtaining a reconstructed motion sequence that accurately reproduces the user's personalized smash posture.

[0119] As an optional embodiment, backpropagating error data and correcting the initial action sequence to obtain a reconstructed action sequence includes: determining the biological constraints corresponding to the skeletal key points; and under the biological constraints, backpropagating error data and correcting the initial action sequence to obtain a reconstructed action sequence.

[0120] In this embodiment, the process of obtaining the reconstructed action sequence is described.

[0121] This involves skeletal key points, which are core joints defined in human motion models. These can include the shoulder, elbow, wrist, hip, knee, and ankle. These points connect to the bones, and their movement determines the posture of the entire human body. Correcting the motion sequence is essentially adjusting the joint parameters corresponding to these key points.

[0122] This involves biological constraints, which are a series of physical limitations defined based on human anatomy and kinesiology. In motion correction, these constraints primarily manifest as two core types: constant bone length constraints and joint range of motion constraints.

[0123] In this step, when performing backpropagation optimization to minimize the trajectory error of the target moving part, the system incorporates biological constraints as hard boundary conditions or penalty terms into the optimization objective. When the algorithm attempts to adjust joint angles to reduce the positional error of the target moving part, any adjustments that might lead to changes in bone length or joints exceeding their normal range of motion are suppressed or prohibited.

[0124] This approach ensures the physiological rationality and naturalness of the reconstructed movement sequences. By introducing and strictly adhering to biological constraints, the system forces the optimization process to search within a solution space that conforms to the actual movement capabilities of the human body. This aligns with the laws of human movement, generating realistic and usable movements while avoiding the generation of distorted or impossible actions. This makes the comparative analysis and movement guidance suggestions more valuable for practical reference.

[0125] Specifically, in badminton applications, taking the correction of a jump smash as an example, the system applies strict biological constraints simultaneously while optimizing wrist trajectory errors through backpropagation. Specifically, the system ensures that the bone length between key skeletal points remains constant during optimization, while limiting the rotation angles of each joint to within the range of human physiological activity. Under these dual constraints, the system iteratively adjusts the whole-body posture parameters, ensuring that the generated reconstructed motion sequence accurately matches the actual wrist trajectory recorded by the wristband while maintaining a coordinated and natural body posture. This results in an output of a reasonable three-dimensional motion that conforms to both the user's individual force characteristics and the laws of human kinematics.

[0126] As an optional embodiment, the original motion parameters corresponding to the reconstructed motion sequence are subjected to simulated variational adjustment to obtain the deduced motion sequence, including: when the target motion is a ball motion, determining the motion parameters and motion vector parameters corresponding to the motion moment based on the reconstructed motion sequence, wherein the motion parameters include at least one of the following: motion velocity, peak acceleration, vibration spectrum; determining the incoming ball velocity vector parameters based on the motion parameters and the collision physical network; determining the ball motion trajectory corresponding to the ball based on the motion vector parameters and the incoming ball velocity vector parameters; and performing simulated variational adjustment on the original motion parameters corresponding to the reconstructed motion sequence based on the ball motion trajectory to obtain the deduced motion sequence.

[0127] This embodiment illustrates the process of inferring the ball's condition and forward deduce its trajectory by comprehensively analyzing the human body's movement characteristics and collision physics information at the moment of impact in specific application scenarios of ball sports.

[0128] This involves motion parameters, which are a series of dynamic characteristics describing the user's hitting action itself at the moment of the action.

[0129] This involves motion vector parameters, which specifically refer to the direction and velocity vector of the racket face at the moment of the action, calculated from the reconstructed motion sequence. These parameters are used to quantify the direction and magnitude of the initial force applied to the ball by the user when hitting it.

[0130] This involves a collision physics network, a pre-trained neural network model capable of learning and modeling the complex nonlinear physical relationship between the motion characteristics at the moment of impact in ball sports and the collision outcome. Input motion parameters can probabilistically infer the most likely incoming ball velocity vector parameters.

[0131] This involves the incoming ball velocity vector parameter, which is the direction and speed of the ball's flight toward the user just before the collision, inferred by the collision physics network based on the current ball-hitting characteristics.

[0132] This involves the trajectory of the ball's motion, which is based on the inferred velocity vector of the incoming ball and the user's motion vector parameters. It can be combined with physical parameters such as the ball's mass, air resistance coefficient, and gravitational acceleration to calculate the complete three-dimensional flight path of the ball from being hit to landing in a virtual physical environment through numerical integration.

[0133] In this step, the system first extracts the dynamic features and geometric direction at the moment of impact from the reconstructed hitting motion. Then, using a collision physics network, it infers a reasonable incoming ball velocity based on the hitting features. Next, it performs collision calculations between the user's swing vector and the inferred incoming ball vector in the physics engine to obtain the ball's velocity and simulate its complete flight trajectory. Finally, using this inferred original hitting result as a benchmark and optimization target, it initiates a subsequent simulation variational adjustment process to find motion parameters that can produce a better flight trajectory.

[0134] This approach enables intelligent completion of missing environmental information and physical closed-loop deduction. Through physics-heuristic reverse inference and forward simulation, the system constructs a complete virtual ball-hitting interaction scenario that conforms to physical laws. This allows subsequent simulation variational adjustments to be a directed and interpretable exploration within a context with clear physical meaning and optimization objectives.

[0135] Specifically, in badminton applications, taking a net smash as an example, the system first extracts motion parameters (such as racket head peak velocity and high-frequency vibration spectrum) and motion vector parameters (racket face direction and velocity) at the moment of impact from the reconstructed smashing motion. Then, these parameters are input into a pre-trained badminton collision physics network, which, based on collision physics principles, infers the possible velocity vector parameters of the incoming ball (such as a slower incoming ball velocity and a flatter incident angle). Next, in a differentiable physics simulation environment, combining the inferred incoming ball velocity, the user's swing vector, and the badminton aerodynamic model, the system calculates the complete flight trajectory of the ball (ball motion trajectory) through numerical integration. Finally, using this trajectory as a benchmark (if the simulation results show the ball hitting the net), the system initiates a simulated variational adjustment of the original smashing motion parameters (such as racket face angle and hitting timing) to find an optimized motion sequence (deduced motion sequence) that leads to the ball crossing the net and landing accurately.

[0136] As an optional embodiment, the original motion parameters corresponding to the reconstructed motion sequence are simulated and varied to obtain the deduced motion sequence, including: when the target motion is a ball motion, simulating and varying the original motion parameters corresponding to the reconstructed motion sequence to obtain multiple candidate deduced motions; retrieving a benefit function, wherein the benefit function includes a distance benefit term, a ball flight speed benefit term, and a motion adjustment amplitude term, and the distance benefit term represents the benefit term brought about by the distance between the ball's landing point and the boundary line; determining the benefit index corresponding to each of the multiple candidate deduced motions based on the benefit function; and determining the deduced motion sequence from the multiple candidate deduced motions based on the corresponding benefit index.

[0137] This embodiment illustrates the process of scientifically and automatically determining the optimal improvement action by defining a comprehensive quantitative evaluation criterion to evaluate and select the best among multiple adjusted candidate actions in parallel.

[0138] This involves candidate actions, which are multiple different and possible variations of the action sequence generated by performing a series of systematic small variations on the original action parameters of the reconstructed action sequence. Each candidate action can represent a possibility of adjustment in that way, and is an alternative to be evaluated and compared.

[0139] This involves a payoff function, a mathematical function used to quantitatively evaluate the overall quality or detriment of the shot result corresponding to each candidate deduced action, transforming the complex effects and costs of the shot into a comparable scalar value. A higher function value indicates a better overall payoff for the action.

[0140] This involves a distance benefit term, which is a component of the benefit function. It can be used to evaluate the quality of the ball's landing point. The benefit is calculated based on the distance between the ball's landing point and the opponent's boundary line. Generally, it is encouraged to land closer to the boundary line but ensure that the ball is within the boundary and avoids going out of bounds.

[0141] This includes a ball flight speed benefit term, which is another component of the benefit function and is used to evaluate the ball's flight speed after it is hit. Generally, assuming accuracy is maintained, a faster ball speed means a greater threat, and therefore a higher ball speed will bring a higher benefit. This term encourages seeking a balance between speed and control in motion adjustments.

[0142] This includes the action adjustment magnitude term, which is a term in the profit function used to penalize excessive action adjustments and measures the degree of change of the candidate inferred action relative to the original action parameters.

[0143] This includes a benefit index, which is the final scalar score obtained after each candidate deduction action is calculated using a benefit function. It reflects the overall performance of the action in three dimensions: accuracy of landing point, threat level of ball speed, and cost of action modification. It is the direct basis for the system to select the best.

[0144] In this step, the system first generates multiple candidate actions that fluctuate slightly around the original action parameters. Then, for each candidate action, the hitting process is simulated in a virtual physical environment to obtain its corresponding ball trajectory, landing point, and velocity. Next, a predefined reward function is called, and the reward index for each action is calculated by combining its landing point, ball velocity, and differences from the original action. Finally, the system compares the reward indices of all candidate actions and selects the candidate action with the highest reward index as the final sequence of deduced actions.

[0145] This approach enables intelligent action optimization, moving from blind trial-and-error to goal-oriented selection. By designing a multi-objective and trade-off benefit function, coaching experience is encoded into computable mathematical rules. This allows the system to automatically and objectively evaluate a vast number of candidate actions and scientifically select the best improvement plan under multiple constraints. This overcomes the limitations of traditional methods that rely on human experience or single indicators for evaluation, ensuring that the guidance provided is not only effective but also efficient and easy to implement. This significantly enhances the scientific rigor, practicality, and personalization of virtual coaching.

[0146] Specifically, in badminton applications, taking the backhand high clear as an example, the system performs minor random perturbations on key parameters in the reconstructed motion (such as wrist pronation angle and hitting point height) to generate dozens of candidate deduced motions. In the physical simulation, the hitting effect corresponding to each motion is calculated in parallel, and a reward function is called for scoring: the distance reward is given a high score based on the distance of the landing point from the opponent's baseline (the closer to the baseline, the higher the reward); the ball flight speed reward encourages higher ball speed; and the motion adjustment range reward penalizes adjustments that deviate too much from the original motion. The reward index for each motion is obtained by combining these three calculations. Finally, the system selects the motion variant with the highest reward index as the deduced motion sequence, such as an optimized motion that, while maintaining reasonable force, allows the landing point to be closer to the baseline and has a moderate ball speed, thus providing users with clear and actionable improvement suggestions.

[0147] As an optional embodiment, the original motion parameters corresponding to the reconstructed motion sequence are simulated and varied to obtain multiple candidate deduced motions, including: determining motion parameters whose influence index on the hitting result is greater than a predetermined influence index to obtain original motion parameters; constructing a multidimensional probability distribution centered on the original motion parameters; sampling multiple variational parameters from the multidimensional probability distribution; and simulating the hitting process corresponding to the adjusted motion based on the multiple variational parameters to generate multiple candidate deduced motions.

[0148] This embodiment illustrates how to efficiently generate a batch of meaningful and diverse action variants by focusing on key parameters, constructing a probabilistic model, and systematically sampling.

[0149] This includes the impact index of the shot result, which is a quantitative indicator used to measure the degree of influence of changes in a specific action parameter in the reconstructed action sequence on the final shot effect. It can be obtained through theoretical analysis, pre-experimental data, or sensitivity analysis based on physical models.

[0150] This involves a predetermined impact index, which is a pre-set threshold used to filter out key action parameters that have a significant impact on the results. Only parameters with an impact index exceeding this threshold are selected for subsequent variational adjustment to ensure that the optimization search focuses on the most effective dimensions.

[0151] This involves multidimensional probability distributions, which are mathematical models defined in a multidimensional parameter space centered on the selected original action parameter values. These distributions describe the range and probability of reasonable variations in these parameters. One such distribution is a multidimensional Gaussian distribution, with its mean set to the original parameter values ​​and its variance controlling the magnitude of the adjustment.

[0152] This involves variational parameters, which are a set of specific values ​​randomly sampled from the aforementioned multidimensional probability distribution. These variational parameters represent a small shift or change in a specific combination based on the original motion parameters. Each set of variational parameters corresponds to a possible motion adjustment scheme.

[0153] In this step, firstly, based on physics knowledge or data analysis, several core motion parameters that have the greatest impact on the quality of the shot are identified and used as the original motion parameters to be optimized. Next, a multidimensional probability distribution is constructed around the original values ​​of these parameters, defining the fluctuation of each parameter within its reasonable range. Then, multiple random samples are taken from this distribution to obtain multiple sets of different combinations of variational parameters. Finally, each set of variational parameters is applied to the original reconstructed motion to generate the adjusted motion, and the corresponding complete shot process is simulated in a physics simulation engine, thereby generating a series of candidate deduced motions.

[0154] This approach enables efficient and purposeful exploration within a vast space of action parameters. By pre-screening key parameters, inefficient or even ineffective searches across countless unimportant parameter dimensions are avoided, significantly narrowing the optimization search space and improving computational efficiency and optimization focus. Sampling is performed using a probability distribution centered on the original action, which systematically generates a series of action variants that maintain a certain degree of similarity while incorporating reasonable variations, near the user's original habits. This ensures that the explored candidate actions are diverse yet remain within the user's learnable and executable range, avoiding the generation of completely unrealistic and bizarre actions.

[0155] Specifically, in badminton application scenarios, in optimizing the forehand slice drop shot, the system first uses a physical model to determine that wrist flexion / extension angle and hitting timing are the two parameters with the highest impact on the drop shot's landing point and net clearance height, and uses these as the original motion parameters to be optimized. Then, a two-dimensional Gaussian distribution (multidimensional probability distribution) is constructed using the mean of these two parameters, with its standard deviation set according to the variable range of human movement. Next, 128 sets of variational parameters are randomly sampled from this distribution (e.g., one set is {wrist angle +2 degrees, hitting timing -0.03 seconds}). Finally, each set of parameters is substituted into the reconstructed slice drop shot motion, and the hitting is simulated in parallel within the physics engine, generating 128 candidate deduced motions. These motions all fluctuate reasonably around the user's original habits, providing a high-quality and diverse candidate set for subsequent optimization.

[0156] As an optional embodiment, the revenue index corresponding to each of the multiple candidate deduced actions is determined based on the revenue function, including: when the corresponding candidate deduced actions include deduced action parameters, the state of the target object at the next moment is determined based on the deduced action parameters corresponding to the current moment, thereby obtaining the corresponding object moment state, wherein the deduced action parameters include force parameters and velocity parameters; the landing point corresponding to each of the multiple candidate deduced actions is determined based on the object moment states corresponding to each of the multiple candidate deduced actions; and the revenue index corresponding to each of the multiple candidate deduced actions and the distance revenue term is determined based on the landing points corresponding to each of the multiple candidate deduced actions.

[0157] This embodiment illustrates how to use physical simulation to gradually evolve the dynamic parameters inherent in each candidate deduced action into specific and quantifiable hitting results, and thereby accurately calculate its gains in the dimension of landing accuracy, thus providing objective and quantitative basic data for the comprehensive evaluation of the gain function.

[0158] This involves deduced motion parameters, which are specific values ​​describing the mechanical characteristics of the ball at the moment of impact in the candidate deduced motion, and may include force parameters and velocity parameters.

[0159] This involves target objects, which are the main objects affected by user actions in a physical simulation environment.

[0160] This involves the object's instantaneous state, which is a complete physical state description of the target object at a specific moment, and may include the object's spatial position coordinates, velocity vector, and acceleration vector, etc.

[0161] This involves the landing point, which refers to the spatial coordinates of the target object when it finally collides with the ground after being hit by the user, following a complete flight trajectory simulation in the virtual physical environment.

[0162] In this step, for each candidate deduced action, its deduced action parameters at the instant of impact are first extracted. These parameters are used as initial conditions and input into a differentiable physical simulation environment. Starting from the instant of impact, the simulator integrates step by step according to the laws of physics, calculating the state of the ball at each tiny next moment. This process is iterated until the ball hits the ground, thus determining its final landing coordinates. Finally, the calculated landing coordinates are substituted into the formula for the distance gain term to obtain the gain index of the candidate action in terms of landing accuracy.

[0163] This method enables precise and repeatable physical prediction and quantitative evaluation of shot results. By directly linking motion parameters to the final effect through rigorous physical simulation, the calculation of distance gains is based entirely on objective physical laws rather than subjective experience or rough estimates, thus enhancing the scientific rigor and credibility of the evaluation. High-precision numerical integral simulation accurately predicts changes in landing point caused by subtle adjustments in motion, making the feedback of the gain function extremely sensitive and accurate. This allows for the identification of subtle but crucial improvements from a large number of similar candidate motions.

[0164] Specifically, in badminton application scenarios, taking the optimization of the badminton smash as an example, for each candidate deduced action, the system extracts the deduced action parameters at the moment of impact (such as the racket head linear velocity vector and impact force estimation). In a differentiable physics simulation environment, taking this moment as the initial state (current time), the badminton shuttlecock is regarded as the target object. Combining air resistance and gravity models, the system progressively calculates the object's state (position and velocity) at each time step through numerical integration until the shuttlecock lands and its precise landing point coordinates are obtained. Subsequently, based on the distance between this landing point and the opponent's baseline, a distance gain term function is called to calculate the gain index (e.g., a high score is obtained if the landing point is less than 10 centimeters from the baseline). This process provides a quantified landing point accuracy score based on rigorous physics simulation for each smash variation.

[0165] As an optional embodiment, guidance parameters corresponding to the target action are obtained based on the reconstructed action sequence and the deduced action sequence, including: when the guidance parameters include guidance voice text, determining the action difference parameters between the reconstructed action sequence and the deduced action sequence on key parameters, wherein the key parameters are parameters whose resulting action deformation index is greater than a predetermined deformation threshold; converting the action difference parameters into guidance voice text, and playing the guidance voice text.

[0166] In this embodiment, it is illustrated that by identifying the most significant pose differences and quantifying them into natural language commands, clear and actionable voice guidance is provided to the user, thus completing the process from analysis results to user-perceptible suggestions.

[0167] This includes guidance voice text, which refers to natural language statements generated by the system to guide users in improving their actions. These statements can be composed of predefined semantic templates and specific action difference parameter values.

[0168] This involves key parameters, which are specific kinematic parameters that are determined to have a substantial impact on the action form and hitting effect in the comparison between the reconstructed action sequence and the deduced action sequence.

[0169] This involves the motion deformation index, which is an indicator used to quantify the degree of difference between two motion sequences on a specific parameter. It can be calculated as the absolute value of the difference, the relative rate of change, or a normalized value, and is used to objectively measure the magnitude of change of that parameter.

[0170] This involves a predetermined deformation threshold, which is a pre-set critical value used to determine whether the change of a certain parameter is significant enough to warrant making suggestions to the user. When the deformation index of a certain parameter exceeds this threshold, the parameter will be identified as a key parameter and thus enter the subsequent guidance information generation process.

[0171] This involves motion difference parameters, which are quantitative descriptions of the specific differences corresponding to key parameters. They can include the direction of the difference and the specific value, and are the specific data for filling in the guidance voice text template.

[0172] In this step, the reconstructed and deduced actions are first compared item by item across various kinematic parameters, and the deformation index for each parameter is calculated. Next, parameters whose deformation indices exceed a predetermined deformation threshold are identified as key parameters, and their specific action difference parameters are recorded. Then, the system inputs these difference parameters into a pre-set natural language generation template to form a complete guidance speech text. Finally, the device's speech synthesis and playback functions are invoked to convert the text into speech and deliver it to the user.

[0173] This approach enables the transformation of intelligent analysis results into the most natural and convenient form of user interaction. By setting deformation thresholds, unimportant minor differences are automatically filtered out, ensuring that the generated voice suggestions focus on the core issues that most need improvement, avoiding information overload and allowing users to grasp the key points. Converting quantified parameter differences into natural language makes professional and complex kinematic analysis conclusions easy to understand and immediately implement, significantly lowering the user's comprehension threshold. Real-time or post-event feedback via voice provides users with guidance without needing to look at a screen, especially suitable for listening during exercise breaks or debriefing, improving the convenience of guidance and user experience, making the virtual coach interaction more human-centered and efficient.

[0174] Specifically, in badminton applications, taking the forehand flick as an example, the system compares the user's reconstructed motion with the optimized deduced motion, calculating the motion deformation index of each parameter. It finds a 15-degree difference in the wrist upward angle (exceeding the predetermined deformation threshold of 10 degrees), thus identifying it as a key parameter, with the corresponding motion difference parameter being an increase of 15 degrees. The system fills this difference into a template and generates a guiding voice text: "It is recommended that you increase the wrist upward angle by another 15 degrees when flicking the shuttlecock." This voice text is then played through the device's speaker.

[0175] As an optional embodiment, guidance parameters corresponding to the target action are obtained based on the reconstructed action sequence and the deduced action sequence, including: when the guidance parameters include a guidance video, determining a first video corresponding to the reconstructed action sequence and a second video corresponding to the deduced action sequence; superimposing the first video and the second video to obtain a guidance video.

[0176] This embodiment illustrates the specific process of generating intuitive and immersive visual teaching materials by rendering the user's actual actions and the optimized actions suggested by the system in three dimensions, and then using overlay technology for synchronous comparison and display.

[0177] This includes instructional videos, which are a form of visual guidance that is the final output. They are dynamic video sequences that can use graphics, animations, and other means to intuitively present the differences in the user's actions that need improvement. They can usually be played on the user's mobile phone, tablet, or AR / VR device for action review and learning.

[0178] Among them, the first video is a 3D animation video rendered based on the reconstructed action sequence. The virtual human model in the video strictly reproduces the user's posture and movement process when performing the target action.

[0179] This includes a second video, a 3D animation rendered based on a deduced sequence of movements. The virtual human model in this video demonstrates physically optimized movements and postures.

[0180] This involves overlay, a video or graphics processing technique that refers to synchronizing two video streams or image sequences in time, aligning them in space, and then merging them into the same screen using methods such as transparency, color differentiation, or side-by-side arrangement.

[0181] In this step, a 3D graphics rendering engine is first used to drive the virtual character model to generate a coherent 3D animation by using the human model parameters corresponding to the reconstructed and deduced action sequences. This animation is then recorded or generated in real-time as the first and second videos. Subsequently, the system aligns the two videos along their timelines and places the two virtual human models within the same virtual sports field 3D scene. Finally, using graphics overlay technology, the two models are displayed simultaneously with different visual styles, generating a single instructional video for direct viewing and comparison by the user.

[0182] This method enables the efficient transformation of sports guidance information from abstract data to a concrete visual experience. By creating a shadow coach-like visual effect, users can intuitively see every subtle difference between their actual movements and ideal movements in terms of spatial posture, trajectory, and timing of force application. This visual comparison is far easier to understand and remember than verbal or textual descriptions, greatly reducing the cognitive load on users to comprehend technical points and providing strong visual feedback. Simultaneously, the generated visual videos are easy to save, replay, and analyze in slow motion, providing users with an efficient learning tool similar to how professional athletes use video recordings for technical review, enhancing the immersion, enjoyment, and learning effectiveness of the guidance.

[0183] Specifically, in badminton applications, taking the backhand flat drive as an example, the system first renders the reconstructed sequence of the user's drive into a first video, where the virtual human model is presented in semi-transparent red. Simultaneously, the optimized deduced sequence of actions is rendered into a second video, where the model is presented as a highlighted green solid line. Then, the system synchronizes the two videos on the timeline, using the moment of impact as a reference, and places them in the same virtual court scene. Finally, using graphics overlay technology, the red phantom and green solid models are played synchronously to generate a tutorial video. Users can intuitively compare the differences in backswing amplitude, impact point position, and follow-through trajectory using their mobile phone screen or AR glasses, achieving efficient visual review and learning of the movements.

[0184] Based on the above embodiments and optional embodiments, an optional implementation method is provided, which is described in detail below.

[0185] Among related technologies, mainstream intelligent motion assistance technologies mainly fall into two categories, but both have significant limitations:

[0186] Computer vision-based solutions rely on deploying multiple cameras on sports fields or using high-precision optical motion capture equipment. Their drawbacks include high hardware costs, complex system setup, and susceptibility to environmental factors such as changes in lighting and object obstruction. Therefore, this approach is difficult to implement in everyday environments like sports halls or homes where ordinary users lack dedicated cameras.

[0187] Wearable sensor-based solutions utilize inertial sensors in smart bracelets or watches worn on the wrist to record motion data. Their core drawback lies in the severe lack of information dimension; relying solely on motion data from a single wrist node cannot reconstruct coordinated body posture, nor can it perceive external environmental information. Therefore, existing products typically only provide basic statistical data such as the number of swings and maximum speed, failing to offer in-depth, tactical, or detailed guidance on where mistakes were made or how to hit the next ball.

[0188] In view of this, an optional embodiment of the present invention provides a motion guidance method that does not attempt to accurately calculate missing information using traditional algorithms, but instead uses generative AI (VLA model) to generate the most reasonable human movements based on massive prior motion data, and uses a physical world model to deduce the motion results in virtual space, thereby realizing the leap from data recording to virtual coaching. Figure 2 This is a schematic diagram of the motion guidance method provided by an optional embodiment of the present invention, such as... Figure 2 As shown, it will be introduced below.

[0189] (a) Multimodal perception and feature extraction;

[0190] First, the user wears a smart wearable device with an integrated inertial measurement unit (IMU), such as a smart bracelet or watch. This device synchronously collects raw data from multimodal sensors during motion at a high sampling rate, mainly including triaxial linear acceleration, triaxial angular velocity, and triaxial magnetometer data, which together constitute the underlying signal stream describing the motion state of the target part.

[0191] Subsequently, the system processes the aforementioned raw data stream in real time. The core task is to identify key motion events and extract representative features. The processing follows these sub-steps:

[0192] 1. Event detection and sequence segmentation employ an adaptive sliding window algorithm to continuously calculate the synthetic modulus of linear acceleration at the wrist. When this modulus exceeds a threshold for a hitting event (e.g., 3g) set based on prior knowledge, the system marks it as the "hitting moment." Using this moment as a reference, a fixed duration (e.g., 1.0 second before and 0.5 seconds after) of sensor data is extracted before and after it to form an independent action sequence segment. This segment fully covers the preparation, execution, and subsequent stages of a single hitting action.

[0193] 2. Data normalization and spatial alignment: Preprocessing is performed on the segmented motion sequence fragments. First, high-frequency noise is removed by low-pass filtering. Then, a sensor fusion algorithm (such as the Madgwick algorithm) is used to transform the raw data in the device's local coordinate system to a global world coordinate system based on gravity and geomagnetic directions, eliminating the uncertainty caused by the device's wearing orientation, thereby obtaining a standardized motion trajectory with clear physical spatial meaning.

[0194] 3. Feature Encoding and Semantic Recognition: The preprocessed spatiotemporal trajectory data is constructed into a regular action tensor and input into a lightweight neural network model (such as a one-dimensional convolutional neural network 1D-CNN or a temporal convolutional network TCN) deployed on the terminal. This model extracts and classifies features from the input tensor and outputs corresponding action semantic labels, such as forehand smash, backhand drop shot, etc.

[0195] Finally, the system integrates action type labels with deep feature vectors extracted from trajectory data to generate a structured action prompt. An example of this prompt is shown below:

[0196] [ACTION:"Forehand_Smash"]+[INTENSITY:"High"]+[TRAJECTORY_EMBEDDING:<Tensor_Data> This cue word compactly encapsulates the core semantics and kinematic properties of the current action, serving as a key conditional input to drive subsequent generative models to infer full-body pose.

[0197] (ii) Full-body pose generation based on VLA model;

[0198] The system inputs the structured action cues generated in step (1) into a pre-trained VLA (Vision-Language-Action) generation model. This model uses a conditional diffusion model as its core architecture, and its workflow mainly includes three hierarchical processing stages: a conditional encoding layer, a denoising generation layer, and a pose decoding layer. The conditional encoding layer receives the action cues and, using a Transformer encoder structure, encodes the wrist trajectory temporal features and action semantic labels into a high-dimensional conditional embedding vector. This vector serves as a high-level guiding signal controlling the entire generation process. The denoising generation layer is the core of the model. In each iterative denoising step, this layer receives a noisy full-body pose sequence estimate and the aforementioned conditional embedding vector. Through a multi-layer self-attention mechanism, the model progressively predicts and removes noise. This layer is designed with integrated biomechanical constraint strategies to ensure that the length of the human skeleton remains constant during generation, and that the rotation angles of all joints are limited to a physiologically permissible range. The pose decoding layer is responsible for decoding the features in the latent space output by the denoising generation layer back into specific physical human body model parameters. The output is usually in the parameterized form of standard human body models such as SMPL, such as a 6D rotation representation of 24 joints and a three-dimensional displacement of the root node, thus fully defining the full-body three-dimensional pose at each moment.

[0199] The model abandons the traditional inverse kinematics solution method and instead adopts a condition generation mechanism based on massive prior knowledge. This mechanism mainly includes two key processes: retrieval and matching, and constraint generation. In the retrieval and matching process, the model does not generate from scratch, but retrieves the most relevant full-body movement patterns from a pre-built prior library containing massive amounts of professional athlete movement data. This library is organized in key-value pairs: "keys" are wrist trajectory dynamic feature vectors extracted from professional movements; "values" are the corresponding, parameterized standard full-body movement sequences. The system maps the user's current wrist features to the same latent space, calculates the similarity with all "keys" in the library, activates and retrieves the most matching full-body movement "value" as a high-quality generation starting point.

[0200] Specifically, during badminton, instantaneous kinematic features include the linear velocity vector, angular velocity vector, and relative wrist displacement in the world coordinate system for each time step. Energy envelope features include the peak timing and energy decay rate of the action sequence, used to characterize the rhythm and explosiveness of the movement. Spatial geometric features include the curvature and torsion of the wrist trajectory, used to distinguish between linear movements (such as flat drives) and arc-shaped movements (such as high clears). A prior library containing massive amounts of professional motion data is pre-built. Each entry in the library is a key-value pair, categorized into several semantic clusters such as force application mode and shot height: Value (motion representation): Represents the whole-body movement using the SMPL (Skinned Multi-Person Linear Model) parameter sequence, including 6D rotation vectors of 24 joints and root node displacement, accurately describing the 3D posture changes of the human body. Key (index feature): The wrist trajectory dynamic feature vector corresponding to the action (i.e., the three types of features mentioned above). The wrist trajectory dynamic features are concatenated into a one-dimensional vector, mapped to the Latent Space through a conditional encoder, and the cosine similarity between it and the Key vector in the pre-trained action library is calculated, thereby activating the most relevant whole-body posture pattern (Value).

[0201] The constraint generation is based on the retrieved full-body movements as priors. The model performs multi-step iterative denoising generation. After generating an intermediate full-body pose in each step, the model calculates the error between the theoretical position of the wrist joint in that pose and the actual trajectory position recorded by the sensor. By backpropagating this error, the model can dynamically adjust the full-body pose parameters. Under the premise of keeping other parts of the body coordinated and natural (following biomechanical constraints), the model forces the wrist trajectory of the generated movement to fit the measured data very closely.

[0202] Specifically, during badminton, the "whole-body motion prior" serves as the initial state (or guiding signal) for the diffusion model. The model performs 50 steps of denoising, predicting a temporary whole-body posture after each step. Within this posture, there exists a theoretical position for the wrist, and the error between this theoretical position and the sensor's measured position is calculated. This error is backpropagated to adjust the whole-body posture parameters, forcibly "pulling" the wrist back to the actual trajectory recorded by the sensor, while simultaneously utilizing biomechanical constraints (such as constant bone length) to maintain the natural coordination of other body parts.

[0203] Through the above mechanism, the system can reconstruct a complete, reasonable and personalized three-dimensional sequence of the user's whole body movements with high fidelity based solely on data from a single part collected by the smart bracelet, without requiring any visual input devices.

[0204] (iii) Physical scenario deduction based on world model;

[0205] This step aims to place the reconstructed full-body movements within a virtual simulation environment that conforms to real physical laws, thereby enabling scientific prediction of motion outcomes and intelligent search for motion optimization solutions. This environment is a differentiable physical simulation space, and its core components and working mechanisms are as follows:

[0206] 1. Construction of a differentiable physics simulation environment:

[0207] The system imports the 3D full-body motion sequence generated in step two into a virtual world model built on physical laws. The core layer of this environment includes a physical property layer, a dynamics solver, and differentiable properties. The physical property layer defines all physical constants and object properties for the simulation, including environmental parameters such as court geometry, net height, gravitational acceleration, air drag coefficient, and ground friction coefficient, as well as equipment parameters such as racket elasticity and badminton aerodynamics (e.g., considering the nonlinear relationship between drag and the square of velocity). The dynamics solver integrates rigid body dynamics and collision detection / response models. It can accurately calculate the motion state of objects (such as the human body, racket, and shuttlecock) in the virtual scene at every moment using numerical integration methods (such as the fourth-order Runge-Kutta method) based on the input forces, velocities, and collision conditions. Differentiability is a key innovation of this simulation environment, supporting automatic differentiation. This means that the system can not only perform forward simulation to calculate the "ball's trajectory and landing point" from the "actions", but also calculate the gradient of the "error between the expected landing point and the actual landing point" with respect to the "original action parameters" through backpropagation. This provides a mathematical basis for directly and efficiently optimizing action parameters.

[0208] 2. Contextual completion:

[0209] To address the issue of single-point sensors being unable to directly observe incoming ball information, the system uses physical reasoning to probabilistically complete the missing environmental context. Collision Feature Extraction: The system accurately identifies the moment of impact from sensor data and extracts features such as peak acceleration and high-frequency vibration spectrum at that moment. These features encode the intensity and nature of the collision. Inverse Dynamics Inference: A pre-trained collision physics network is used. The extracted collision features and the racket face velocity vector calculated from the reconstructed motion are input together. This network learns and models collision physics, outputting the most probable probability distribution of the incoming ball velocity vector under given impact conditions. Forward Trajectory Generation: Using the inferred incoming ball vector distribution and the user's swing vector as input, and combining parameters such as the collision recovery coefficient, the initial launch state of the ball is calculated. Subsequently, multiple sampling points are simulated in parallel in a differentiable physics environment. Numerical integration is used to calculate the complete flight trajectory of the ball under the combined effects of air resistance and gravity, ultimately generating a "probabilistic trajectory cloud" representing uncertainty and its confidence interval.

[0210] 3. Counterfactual deduction and action optimization:

[0211] Based on a complete virtual physics scenario, the system performs "if-then" counterfactual reasoning to search for a better action strategy. The system first identifies the key action parameters (e.g., wrist pronation / external rotation angle, wrist flexion / extension angle, and timing of the shot) that have the most significant impact on the shot's effectiveness (e.g., landing point, speed). Centering on these original user parameter values, a multidimensional probability distribution (e.g., Gaussian distribution) is constructed, and a large number (e.g., N=128 groups) of variational parameters are sampled from it. Each group of parameters represents a small, reasonable adjustment to the original action. Using a GPU-accelerated physics engine, the system simulates the shot process corresponding to all parameter samples in batches and in parallel. For the result of each simulation, a preset reward function is called for quantitative evaluation. This function comprehensively considers factors such as landing point accuracy (e.g., distance from the opponent's boundary line), ball flight speed, and the magnitude of action adjustments. An example of the calculation formula is shown below:

[0212]

[0213] Among them, Accuracy is the distance between the landing point and the opponent's baseline / sideline (the closer the better), Speed ​​is the ball's flight speed, Effort is the range of motion adjustment, and w1, w2, w3 are the corresponding weight coefficients, which can be adjusted adaptively according to the actual application and scenario.

[0214] The system compares the payoff values ​​of all candidate motion variants and selects the motion parameter sequence with the highest payoff, marking it as the "optimal motion solution." This solution represents the system's recommended direction for motion improvement; it is the optimal or near-optimal suggestion obtained by balancing the hitting effect with the cost of motion modification under the constraints of physical laws.

[0215] The system achieves a leap from motion reproduction to motion optimization and strategy generation, providing users with precise technical guidance that has clear physical basis and can be quantified and evaluated.

[0216] (iv) Comparison of virtual and real elements and intuitive feedback;

[0217] The system's intelligent analysis results are transformed into multimodal guidance information that users can intuitively perceive and easily understand, completing a closed loop from data analysis to user experience.

[0218] First, the system precisely compares the reconstructed action sequence generated in step two (i.e., a digital reproduction of the user's actual actions) with the "optimal action solution" obtained in step three. By calculating the quantitative differences between the two in key kinematic parameters (such as joint angles, motion trajectory, and force application timing), the system identifies the core dimensions that need improvement. Based on this difference analysis, the system generates two complementary forms of feedback: Visual action comparison feedback: The system uses a 3D graphics engine to spatiotemporally overlay and render a virtual human model representing the user's actual actions (e.g., rendered as a semi-transparent red phantom) and a virtual human model representing the system's recommended actions (e.g., rendered as a bright green solid image) in the same virtual scene. This animation can be played on the user's mobile terminal (such as a smartphone or tablet) or augmented reality (AR) device. This intuitive comparison of "virtual and real in the same frame" can clearly show the subtle differences in posture, trajectory, and rhythm, providing the user with an immersive visual review experience. Semantic language guidance feedback: The system further transforms the above-mentioned quantitative parameter differences into concise, specific, and operable natural language instructions through predefined semantic mapping rules. For example, the system may generate guidance statements such as "It is recommended to move the hitting point forward by about 10 centimeters" or "The wrist pronation angle can be increased by 5 degrees at the moment of impact." This text information can be read aloud through the device's speech synthesis function or displayed directly on the interactive interface in text form.

[0219] Through the multimodal feedback mechanism that combines "visual contrast" and "verbal guidance", the system can efficiently and user-friendly convey complex motion analysis conclusions to users, thereby significantly reducing the understanding threshold and improving the efficiency of training guidance and user experience.

[0220] The above optional implementation methods can achieve at least the following beneficial effects:

[0221] (1) Users only need to wear ordinary smart bracelets to infer complete three-dimensional movements of the whole body through generative AI models, which eliminates the dependence on expensive multi-camera arrays or professional optical motion capture equipment, and greatly reduces the threshold and popularization cost of professional motion analysis.

[0222] (2) It innovatively utilizes pre-trained large models and physical world models to generate reasonable full-body postures from local wrist data and can probabilistically infer missing environmental information, thus solving the fundamental problem of insufficient data dimensions in wearable devices;

[0223] (3) By superimposing and comparing the user's real actions with the system's recommended actions in three dimensions and generating voice guidance, an intuitive virtual shadow coach experience is created, which greatly improves the user's understanding efficiency and training immersion.

[0224] (4) Only inertial sensor data is processed throughout the process, without collecting or processing any image or video information. This avoids the complex privacy compliance issues of visual solutions and protects the privacy of users and sports venues.

[0225] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0226] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0227] Example 2

[0228] According to an embodiment of the present invention, an apparatus for implementing the above-described motion guidance method is also provided. Figure 3 This is a structural block diagram of a motion guidance device according to an embodiment of the present invention, such as... Figure 3As shown, the device includes: an acquisition module 302, a first determination module 304, a second determination module 306, a simulation variation module 308, and a third determination module 310. The device will be described in detail below.

[0229] The acquisition module 302 is used to acquire motion process data corresponding to the target motion, wherein the motion process data includes process data of the target part performing the target action; the first determination module 304, connected to the acquisition module 302, is used to determine the action prompt parameters corresponding to the target action based on the motion process data, wherein the action prompt parameters include: the action type corresponding to the target action and the trajectory features corresponding to the target action; the second determination module 306, connected to the first determination module 304, is used to determine the reconstructed action sequence corresponding to the target action based on the action prompt parameters; the simulation variation module 308, connected to the second determination module 306, is used to perform simulation variation adjustment on the original action parameters corresponding to the reconstructed action sequence to obtain the deduced action sequence; the third determination module 310, connected to the simulation variation module 308, is used to obtain the guidance parameters corresponding to the target action based on the reconstructed action sequence and the deduced action sequence.

[0230] It should be noted here that the above-mentioned acquisition module 302, first determination module 304, second determination module 306, simulation variation module 308 and third determination module 310 correspond to steps S102 to S110 in the motion guidance method. The multiple modules and the corresponding steps are the same in terms of implementation instances and application scenarios, but are not limited to the content disclosed in the above embodiment 1.

[0231] Example 3

[0232] According to another aspect of the present invention, an electronic device is also provided, comprising: a processor; and a memory for storing processor-executable instructions, wherein the processor is configured to execute instructions to implement the motion guidance method of any of the above embodiments.

[0233] Example 4

[0234] According to another aspect of the present invention, a computer-readable storage medium is also provided, which, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform any of the motion guidance methods described above.

[0235] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0236] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0237] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0238] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0239] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0240] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0241] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle 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 method for guiding exercise, characterized in that, include: Acquire motion process data corresponding to the target motion, wherein the motion process data includes process data of the target part performing the target action; Based on the motion process data, motion prompt parameters corresponding to the target motion are determined, wherein the motion prompt parameters include: the motion type corresponding to the target motion, and the trajectory features corresponding to the target motion; Based on the action prompt parameters, determine the reconstructed action sequence corresponding to the target action; The original action parameters corresponding to the reconstructed action sequence are simulated and varied to obtain the deduced action sequence; Based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained.

2. The method according to claim 1, characterized in that, Based on the motion process data, determine the motion cue parameters corresponding to the target motion, including: Determine the linear acceleration corresponding to the target part in the motion process data; Determine the acceleration modulus corresponding to the linear acceleration; When the acceleration modulus exceeds a predetermined modulus threshold, the corresponding moment is marked as the action moment; From the motion process data, extract data segments of fixed duration before and after the moment of the action to obtain motion process data corresponding to the target action; Based on the motion process data, motion prompt parameters corresponding to the target motion are generated.

3. The method according to claim 1, characterized in that, Based on the motion process data, motion prompt parameters corresponding to the target motion are generated, including: When the motion process data is collected by a motion device, the motion process data in a local coordinate system based on the motion device is transformed to obtain transformed motion data in a global coordinate system based on the physical space. Based on the transformed motion data, a motion feature vector is determined, wherein the motion feature vector includes a temporal feature vector and a spatial feature vector characterizing the target action; Based on the motion feature vector, the action type corresponding to the target action is determined.

4. The method according to claim 1, characterized in that, Based on the action cue parameters, determine the reconstructed action sequence corresponding to the target action, including: Based on the action prompt parameters, determine the initial action sequence; The initial action sequence is corrected to obtain the reconstructed action sequence.

5. The method according to claim 4, characterized in that, Based on the action prompt parameters, determine the initial action sequence, including: Based on the motion prompt parameters, determine the trajectory dynamics characteristics corresponding to the target part; Based on the trajectory dynamics characteristics, an initial action sequence matching the target action is determined from a professional action prior library. This library includes multiple entries stored as key-value pairs. The value in each key-value pair represents an index feature, and the key represents the action sequence. The corresponding key is determined based on the corresponding value. The index features include at least one of the following: instantaneous kinematic features, energy envelope features, and spatial geometric features. The instantaneous kinematic features represent the relative displacement of the target part in physical space. The energy envelope features include energy peak temporal features and energy decay features. The spatial geometric features include the trajectory curvature features and trajectory torsion features corresponding to the target part.

6. The method according to claim 5, characterized in that, Based on the trajectory dynamics characteristics, an initial action sequence matching the target action is determined from a professional action prior library, including: Determine the cosine similarity between the trajectory dynamics features and multiple values ​​in the professional action prior library; The target value is determined from the plurality of values ​​based on the cosine similarity to each of the plurality of values. The action sequence represented by the key corresponding to the target value is determined as the initial action sequence.

7. The method according to claim 4, characterized in that, The reconstructed action sequence is obtained by correcting the initial action sequence, including: Determine the theoretical position sequence corresponding to the initial action sequence for the target part; By comparing the theoretical position sequence with the actual position sequence of the target part, error data corresponding to the initial action sequence is determined. The error data is backpropagated to correct the initial action sequence, resulting in a reconstructed action sequence.

8. The method according to claim 7, characterized in that, Backpropagation of the error data corrects the initial action sequence, resulting in a reconstructed action sequence, including: Determine the biological constraints corresponding to key skeletal points; Under the biological constraints, the error data is backpropagated to correct the initial action sequence, resulting in a reconstructed action sequence.

9. The method according to claim 1, characterized in that, The original action parameters corresponding to the reconstructed action sequence are subjected to simulated variational adjustment to obtain the deduced action sequence, including: When the target motion is a ball sport, the motion parameters and motion vector parameters corresponding to the moment of the motion are determined based on the reconstructed motion sequence. The motion parameters include at least one of the following: motion velocity, peak acceleration, and vibration spectrum. Based on the motion parameters and the collision physics network, the velocity vector parameters of the incoming ball are determined; Based on the motion vector parameters and the incoming ball velocity vector parameters, the ball's trajectory is determined. Based on the trajectory of the sphere, the original motion parameters corresponding to the reconstructed motion sequence are simulated and varied to obtain the deduced motion sequence.

10. The method according to claim 1, characterized in that, The original action parameters corresponding to the reconstructed action sequence are subjected to simulated variational adjustment to obtain the deduced action sequence, including: When the target motion is a ball game, the original motion parameters corresponding to the reconstructed motion sequence are simulated and varied to obtain multiple candidate deduced motions; The revenue function is retrieved, which includes a distance revenue term, a ball flight speed revenue term, and an action adjustment range term. The distance revenue term represents the revenue generated by the distance between the ball's landing point and the boundary line. Based on the aforementioned profit function, determine the profit index corresponding to each of the multiple candidate deduction actions; Based on the corresponding profit index, the sequence of deduced actions is determined from the plurality of candidate deduced actions.

11. The method according to claim 10, characterized in that, The original action parameters corresponding to the reconstructed action sequence are subjected to simulation variational adjustment to obtain multiple candidate inferred actions, including: The original motion parameters are obtained by determining the motion parameters whose influence on the shot result is greater than a predetermined influence index. A multidimensional probability distribution is constructed centered on the original action parameters; Multiple variational parameters are obtained by sampling from the multidimensional probability distribution; Based on the multiple variational parameters, the ball-hitting process corresponding to the adjusted action is simulated to generate the multiple candidate deduced actions.

12. The method according to claim 10, characterized in that, Based on the aforementioned payoff function, determine the payoff indices corresponding to the plurality of candidate deduction actions, including: When the corresponding candidate deduction action includes deduction action parameters, the state of the target object at the next moment is determined based on the corresponding deduction action parameters at the current moment, and the corresponding object time state is obtained. The deduction action parameters include force parameters and velocity parameters. Based on the object's current state corresponding to each of the multiple candidate deduced actions, determine the landing point corresponding to each of the multiple candidate deduced actions; Based on the landing points corresponding to the multiple candidate deduction actions, the profit indices corresponding to the multiple candidate deduction actions and the distance profit items are determined.

13. The method according to claim 1, characterized in that, Based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained, including: When the guidance parameters include guidance speech text, determine the action difference parameters between the reconstructed action sequence and the deduced action sequence in terms of key parameters, wherein the key parameters are parameters that result in an action deformation index greater than a predetermined deformation threshold; The action difference parameters are converted into the guidance voice text, and the guidance voice text is played.

14. The method according to any one of claims 1 to 13, characterized in that, Based on the reconstructed action sequence and the deduced action sequence, guidance parameters corresponding to the target action are obtained, including: When the guidance parameters include a guidance video, a first video corresponding to the reconstructed action sequence and a second video corresponding to the deduced action sequence are determined. The guidance video is obtained by overlaying the first video and the second video.

15. A motion guidance device, characterized in that, include: The acquisition module is used to acquire motion process data corresponding to the target motion, wherein the motion process data includes process data of the target part performing the target action; The first determining module is used to determine motion prompt parameters corresponding to the target motion based on the motion process data, wherein the motion prompt parameters include: the motion type corresponding to the target motion and the trajectory features corresponding to the target motion; The second determining module is used to determine the reconstructed action sequence corresponding to the target action based on the action prompt parameters; The simulation variational module is used to perform simulation variational adjustments on the original action parameters corresponding to the reconstructed action sequence to obtain the deduced action sequence. The third determining module is used to obtain guidance parameters corresponding to the target action based on the reconstructed action sequence and the deduced action sequence.

16. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the motion guidance method as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the motion guidance method as described in any one of claims 1 to 14.

Images