Motion analysis and feedback method and device and computer readable storage medium

Abstract

The invention discloses a motion analysis and feedback method and device and a computer readable storage medium, and the method comprises the steps: building a neutral position reference model based on the skeleton features of a user, and calculating the instantaneous deviation features of a real-time posture and the reference model; resolving a time domain accumulation feature representing the stubborn degree of the action error and a time domain change rate feature representing the deterioration trend of the action error in real time; retrieving the kinematic chain knowledge graph to obtain attribution information; inputting the features into an adaptive feedback control model, dynamically calculating abstract hierarchy parameters of a teaching strategy based on the magnitude of time domain accumulation features, respectively calculating mood specificity parameters and intervention opportunity parameters in combination with instantaneous deviation features and time domain change rate features, and generating prompt control parameters; and finally, system prompts are assembled based on the prompt control parameters, and self-adaptive feedback information is generated. According to the application, the crossing from mechanical error correction to adaptive closed-loop teaching can be realized.

Description

Technical Field

[0001] This invention relates to the field of sports science and technology, and in particular to a motion analysis and feedback method, device, and computer-readable storage medium. Background Technology

[0002] With the development of computer vision technology, camera-based motion assistance systems have been widely used in home fitness, rehabilitation training, and professional sports guidance. Current mainstream solutions typically utilize pose estimation techniques (such as extracting skeletal key points) to capture the user's real-time movements and geometrically compare them with preset standard movement templates. When the system detects that the user's movement parameters (such as joint angles) deviate from preset thresholds, it triggers corresponding error correction prompts.

[0003] However, existing motion analysis and feedback technologies still have significant limitations in practical applications, mainly because their feedback mechanisms are usually "static" and "instantaneous." Existing systems often make judgments based solely on the instantaneous deviation of the current frame, lacking the ability to perceive the evolution of user action errors over time. Specifically, existing technologies struggle to distinguish between "initial errors" and "persistent repetitive errors" (i.e., lacking consideration for the cumulative effect of errors), and also fail to capture fatigue signs of "a sharp decline in action quality" (i.e., lacking the ability to predict the trend of error deterioration).

[0004] This "open-loop" or "fixed-gain" feedback logic leads to the following technical challenges: First, rigid teaching strategies. Whether the user makes an accidental error due to a lack of understanding of the movement or a persistent error due to long-term habits, the system provides homogenized corrective instructions. For persistent errors, simple repetitive instructions are often ineffective, failing to dynamically switch to more easily understood "metaphorical" or "external focus" strategies like a live coach. Second, a lack of safety warning mechanisms. When a user's movement stability rapidly deteriorates due to physical exhaustion, existing systems often mechanically insist on "maintaining standard movements," failing to promptly detect risks and trigger immediate interruption or de-escalation instructions, thus increasing the risk of sports injuries.

[0005] Therefore, how to design an adaptive motion analysis method that dynamically adjusts the feedback strategy and intervention timing has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This application provides a motion analysis and feedback method, aiming to solve the technical problem that existing motion feedback mechanisms have rigid teaching strategies and cannot adapt to the user's real-time state.

[0007] To achieve the above objectives, embodiments of this application provide a motion analysis and feedback method, including:

[0008] Acquire user motion video stream, extract user skeletal features based on initial data of the user motion video stream, and construct user neutral position reference model based on the skeletal features;

[0009] The deviation between the real-time pose of the user's motion video stream and the neutral position reference model is calculated based on the neutral position reference model to obtain instantaneous deviation characteristics;

[0010] Based on the distribution of the instantaneous deviation characteristics in the time dimension, the time-domain cumulative characteristics that characterize the persistence of the action error and the time-domain rate of change characteristics that characterize the deterioration trend of the action error are calculated in real time.

[0011] Based on the instantaneous deviation features, a preset motion chain knowledge graph is retrieved to obtain an initial content payload containing attribution information;

[0012] The instantaneous deviation feature, the time-domain cumulative feature, and the time-domain rate of change feature are input into a preset adaptive feedback control model. The abstract level parameters of the teaching strategy are dynamically calculated based on the magnitude of the time-domain cumulative feature, the tone-specific parameters are calculated based on the instantaneous deviation feature, and the intervention timing parameters are calculated based on the time-domain rate of change feature.

[0013] The abstract level parameters, the tone-specific parameters, and the intervention timing parameters are combined to obtain the prompt control parameters;

[0014] The system prompts are assembled based on the initial content load and the prompt control parameters, and adaptive feedback information is generated based on the system prompts.

[0015] To achieve the above objectives, embodiments of this application also propose a motion analysis and feedback device, including a memory, a processor, and a motion analysis and feedback program stored in the memory and executable on the processor. When the processor executes the motion analysis and feedback program, it implements the motion analysis and feedback method as described in any of the above claims.

[0016] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a motion analysis and feedback program, which, when executed by a processor, implements the motion analysis and feedback method as described in any of the preceding claims.

[0017] The motion analysis and feedback method of this application has the following beneficial effects:

[0018] 1. Achieved dynamic adaptive adjustment of teaching strategies: This application employs an "adaptive feedback control model" to calculate in real time the "time-domain cumulative characteristics" representing the persistence of movement errors and the "time-domain rate of change characteristics" representing the worsening trend of movement errors. The system no longer provides mechanical feedback based solely on current instantaneous deviations, but instead senses the user's "historical performance inertia" and "current state trend," dynamically mapping these dynamic characteristics to the abstract level of the teaching strategy, the severity of the tone, and the timing of intervention. This achieves adaptive guidance "tailored to individual needs," much like a real coach, significantly improving user compliance and training experience.

[0019] 2. Breakthrough in Correcting Stubborn Errors: This application utilizes a dynamic switching logic for setting "abstract level parameters of teaching strategies," specifically automatically switching the strategy from "internal focus mode" to "external metaphor mode" when the "time-domain cumulative characteristics" exceed a saturation threshold. The system can automatically identify "cognitive deadlock" scenarios where conventional instructions (such as "contracting muscles") fail, and bypass user cognitive barriers through environmental analogies or vivid metaphors (such as "tearing the floor"), activating the body's self-organizing movement patterns. This effectively solves the problem of users "understanding but unable to perform" when facing stubborn errors, significantly improving the success rate of correcting complex actions or deep-seated errors.

[0020] 3. Achieved proactive defense against motion risks: This application introduces the monitoring of "time-domain rate of change characteristics" and compares them with "safety damping limits." When an excessively rapid rate of motion deterioration is detected, a "proactive blocking mode" is triggered. In this way, the system possesses the "predictive" perception capability for motion collapse (such as instantaneous instability due to exhaustion), and can immediately generate an interruption command the instant the error amplitude has not yet reached its maximum value but the deterioration trend is already out of control. This transforms traditional "hindsight" error correction into proactive "early warning" defense, effectively preventing motion injuries caused by user fatigue or sudden loss of control.

[0021] 4. Improved the scientific rigor and accuracy of attribution analysis: This application constructs a user-specific "neutral position reference model" and a pre-built "kinetic chain knowledge graph." On the one hand, it eliminates systematic misjudgments caused by individual differences in limb proportions (such as long femurs); on the other hand, it achieves deep indexing from "visual representation errors" (such as genu valgum) to "physiological roots" (such as gluteus medius inhibition). Furthermore, it ensures the biomechanical scientific rigor and accuracy of feedback information, helping users understand and improve their movement patterns from the root causes. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a modular structure diagram of an embodiment of the motion analysis and feedback device of the present invention;

[0024] Figure 2 This is a flowchart illustrating an embodiment of the motion analysis and feedback method of the present invention.

[0025] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0028] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The quantifier "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of "first," "second," and "third," etc., does not indicate any order and can be interpreted as names.

[0029] like Figure 1 As shown, Figure 1 This is a schematic diagram of the hardware operating environment server 1 (also called motion analysis and feedback device) involved in the embodiment of the present invention.

[0030] The server in this embodiment of the invention includes devices with display functions such as "Internet of Things devices", smart air conditioners, smart lights, smart power supplies with network connectivity, AR / VR devices with network connectivity, smart speakers, autonomous vehicles, PCs, smartphones, tablets, e-book readers, and portable computers.

[0031] like Figure 1 As shown, the server 1 includes: a memory 11, a processor 12, and a network interface 13.

[0032] The memory 11 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the server 1, such as the hard disk of the server 1. In other embodiments, the memory 11 can also be an external storage device of the server 1, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the server 1.

[0033] Furthermore, the memory 11 may include both internal storage units of the server 1 and external storage devices. The memory 11 can be used not only to store application software and various types of data installed on the server 1, such as the code of the motion analysis and feedback program 10, but also to temporarily store data that has been output or will be output.

[0034] In some embodiments, processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 11 or process data, such as executing motion analysis and feedback program 10.

[0035] The network interface 13 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface), which is typically used to establish communication connections between the server 1 and other electronic devices.

[0036] The network can be the Internet, a cloud network, a Wi-Fi network, a Personal Area Network (PAN), a Local Area Network (LAN), and / or a Metropolitan Area Network (MAN). Various devices in the network environment can be configured to connect to the communication network according to various wired and wireless communication protocols. Examples of such wired and wireless communication protocols may include, but are not limited to, at least one of the following: Transmission Control Protocol and Internet Protocol (TCP / IP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), File Transfer Protocol (FTP), ZigBee, EDGE, IEEE 802.11, Li-Fi, 802.16, IEEE 802.11s, IEEE 802.11g, multi-hop communication, wireless access points (APs), device-to-device communication, cellular communication protocols, and / or Bluetooth communication protocols, or combinations thereof.

[0037] Optionally, the server may also include a user interface, which may include a display, an input unit such as a keyboard, and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be referred to as a screen or display unit, used to display information processed in server 1 and to display a visual user interface.

[0038] Figure 1 Only server 1, which includes components 11-13 and motion analysis and feedback program 10, is shown. Those skilled in the art will understand that... Figure 1 The structure shown does not constitute a limitation on server 1 and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0039] In this embodiment, the processor 12 can be used to call the motion analysis and feedback program stored in the memory 11 and perform the following operations:

[0040] Acquire user motion video stream, extract user skeletal features based on initial data of the user motion video stream, and construct user neutral position reference model based on the skeletal features;

[0041] The deviation between the real-time pose of the user's motion video stream and the neutral position reference model is calculated based on the neutral position reference model to obtain instantaneous deviation characteristics;

[0042] Based on the distribution of the instantaneous deviation characteristics in the time dimension, the time-domain cumulative characteristics that characterize the persistence of the action error and the time-domain rate of change characteristics that characterize the deterioration trend of the action error are calculated in real time.

[0043] Based on the instantaneous deviation features, a preset motion chain knowledge graph is retrieved to obtain an initial content payload containing attribution information;

[0044] The instantaneous deviation feature, the time-domain cumulative feature, and the time-domain rate of change feature are input into a preset adaptive feedback control model. The abstract level parameters of the teaching strategy are dynamically calculated based on the magnitude of the time-domain cumulative feature, the tone-specific parameters are calculated based on the instantaneous deviation feature, and the intervention timing parameters are calculated based on the time-domain rate of change feature.

[0045] The abstract level parameters, the tone-specific parameters, and the intervention timing parameters are combined to obtain the prompt control parameters;

[0046] The system prompts are assembled based on the initial content load and the prompt control parameters, and adaptive feedback information is generated based on the system prompts.

[0047] Based on the hardware architecture of the aforementioned motion analysis and feedback device, embodiments of the motion analysis and feedback method of this invention are proposed. The motion analysis and feedback method of this invention aims to solve the technical problem that existing motion feedback mechanisms suffer from rigid teaching strategies and cannot adapt to the real-time state of users.

[0048] Reference Figure 2 , Figure 2 This is an embodiment of the motion analysis and feedback method of the present invention, which includes the following steps:

[0049] S10. Acquire the user's motion video stream, extract the user's skeletal features based on the initialization data of the user's motion video stream, and construct the user's neutral position reference model based on the skeletal features.

[0050] Specifically, this step aims to establish a "personalized" evaluation benchmark. In traditional motion analysis, systems typically use universal standard templates (e.g., the standard for squatting for everyone is that the knees should not extend beyond the toes). However, different users have significant differences in body structure (such as the length ratio of the femur to the tibia), and directly applying a universal template can easily lead to misjudgments. Therefore, this application constructs a "neutral position reference model" to provide a "zero point" or "ideal trajectory" that adapts to the individual physiological structure of the user for subsequent error calculation.

[0051] Specifically, the user's movement video stream can be acquired in real time through a camera or wearable device.

[0052] In some embodiments of this application, step S10 can be implemented by the following sub-steps S11 to S12:

[0053] S11. Extract the user's static limb length ratio and joint range of motion from the initialization data.

[0054] Initialization data refers to a sequence of video frames or a set of images collected when the system guides the user to perform specific actions (such as standing naturally, T-shaped movement, maximum hip flexion, etc.) before the formal training begins.

[0055] In its implementation, the system first uses a pose estimation algorithm (such as a skeletal point detection algorithm based on a convolutional neural network) to identify key points of the human body from video frames of the initialized data, including but not limited to the 3D coordinates of the shoulder, hip, knee and ankle joints.

[0056] Then, based on the aforementioned 3D coordinates, the system calculates the static limb length ratio. For example, the system calculates the length L of the femur (thigh bone). femur Length L of the tibia (shinbone) tibia The ratio R leg =L femur / L tibia ; and the ratio of trunk length to lower limb length.

[0057] At the same time, the system determines the user's "range of motion" (ROM) by analyzing the user's extreme positions in specific test movements, such as the maximum dorsiflexion angle of the ankle joint.

[0058] For example, assuming user A has a "long femur" body type, the system calculates their R by analyzing their initial standing frame. leg The value was 12 (significantly higher than the average of 10). Additionally, the initial test revealed that User A had limited ankle dorsiflexion, with a maximum angle of only 15 degrees. These data will be structured and stored as the user's "physiological fingerprint."

[0059] S12. Based on the static limb length ratio, the geometric parameters of the preset general biomechanical model are corrected to generate a neutral position reference model that adapts to the user's body structure.

[0060] Among these, the pre-defined general biomechanical model is a parameterized standard skeletal linkage model constructed based on average human statistical data. This model defines the ideal joint angle curves under standard movements (such as squats).

[0061] During this step, the system inputs the individualized data extracted in step S11 into the general model and performs geometric parameter correction. Specifically, the system stretches or shortens the corresponding link length in the general model according to the user's actual limb proportions, and recalculates the ideal angle that the user should present when performing standard movements based on the constraints of the kinematic chain.

[0062] Continuing the previous example, for User A, who has a long femur, the biomechanically correct posture during a squat requires the torso to lean further forward or the knees to extend slightly beyond the toes to maintain balance in the center of gravity. If an uncorrected generic model were used, User A's posture would be incorrectly flagged as "excessive forward torso lean" or "knees lurching forward." However, in this step, the system corrects the model based on User A's proportions, generating a "neutral position reference model" specific to User A. In this reference model, the "ideal torso angle" at the bottom of the squat is automatically adjusted to 45 degrees (instead of the generic 60 degrees). This means that as long as User A maintains an angle of approximately 45 degrees, the system considers it "correct" and no longer generates false alarms.

[0063] It is understood that by performing steps S11 and S12, this application is able to eliminate systematic biases caused by anthropometry differences. This ensures that the "instantaneous bias characteristics" calculated in subsequent steps truly reflect the user's motion technique issues, rather than natural phenomena caused by body structure, thus laying a data foundation for subsequent high-precision adaptive feedback.

[0064] S20. Calculate the deviation between the real-time pose of the user motion video stream and the neutral position reference model based on the neutral position reference model to obtain instantaneous deviation characteristics.

[0065] Specifically, this step aims to transform the user's real-time, unstructured motion performance into a quantifiable, single-dimensional numerical signal (i.e., an "error signal") for subsequent processing by a feedback control model (PID). This process is not a simple image comparison, but rather a weighted calculation process based on biomechanical key points.

[0066] In the specific implementation details, step S20 can be further expanded into the following processing logic S21-S23:

[0067] S21. Extract a real-time pose vector from the current frame of the user motion video stream, wherein the real-time pose vector contains real-time angle values ​​of multiple key joints.

[0068] Specifically, the system processes the input real-time video stream frame by frame, and uses pose estimation algorithms (such as the aforementioned skeletal keypoint detection technology) to extract the real-time pose vectors (Pose) of each key joint of the user in the current frame. user (t). This vector contains a set of angle values ​​for key joints (e.g., hip flexion angle, knee eversion / inversion angle, ankle dorsiflexion angle, etc.).

[0069] S22. Determine the user's current action phase and index the reference pose vector corresponding to the current action phase in the neutral position reference model.

[0070] Specifically, the system performs a phase alignment operation. Since the user's movement rate varies, the system needs to index the corresponding phase reference pose vector (Pose) in the neutral position reference model generated in step S12, based on the current completion level of the movement (e.g., squatting to the lowest point, or just beginning to squat). ref (t). This ensures that the comparison is performed at the “same moment of action”.

[0071] S23. Based on the preset joint weight vector, calculate the weighted distance between the real-time posture vector and the reference posture vector to obtain the instantaneous deviation feature.

[0072] Next, the system calculates the difference between the real-time posture vector and the reference posture vector. To more accurately reflect the severity of the movement error, this application introduces a weight vector w, assigning different levels of attention to deviations at different joints. For example, in a "squat" movement, the risk weight of knee valgus is much higher than that of elbow position deviation; therefore, the weight coefficient for the knee joint will be set to a larger value. In some embodiments of this application, the instantaneous deviation feature $e(t)$ can be calculated using the following weighted Euclidean distance formula or weighted absolute value sum formula:

[0073]

[0074] In the formula:

[0075] M represents the total number of key joints being monitored;

[0076] w j The preset weight coefficient (w) for the j-th joint j ∈[0,1]);

[0077] The real-time angle of the user's j-th joint at time t;

[0078] This is the ideal angle for the corresponding phase in the neutral position reference model.

[0079] Continuing with the case of user A mentioned above, suppose user A is performing a squat, and at the current time t, they are at the lowest point of the squat.

[0080] At this point, the neutral position reference model indicates that, for user A's body structure, the knee valgus angle is... It should be 0 degrees (i.e., the knee is aligned with the toes).

[0081] However, the system detected in real time that User A's left knee was collapsing inward, and the measured angle was... It is 15 degrees (inward).

[0082] Assume the system has a preset weight w for the knee joint. knee =0.8 (high risk), while the torso angle, although slightly deviating, has a lower weight. The system substitutes these values ​​into the above formula and finally outputs the instantaneous deviation characteristic e(t) at that moment as a large value (e.g., 12.0). This value quantifies the "severity" of user A's erroneous action at that instant and serves as the basic signal input to the subsequent PID controller.

[0083] It is understood that, through step S20, this application reduces the complex visual image differences into a single, time-varying error scalar stream. This instantaneous bias feature e(t) not only reflects "where the error is", but also reflects "the severity of the error" through weighted summation, providing the necessary input data foundation for calculating "error integral (stubbornness)" and "error derivative (deterioration trend)" in subsequent steps.

[0084] S30. Based on the distribution of the instantaneous deviation characteristics in the time dimension, calculate in real time the time-domain cumulative characteristics that characterize the persistence of the action error, and the time-domain rate of change characteristics that characterize the deterioration trend of the action error.

[0085] Specifically, this step aims to introduce an analysis of the "time dimension." The technical solution of this application argues that the same error magnitude has drastically different characteristics depending on whether it is a "first offense" versus "repeated offenses," or whether it occurs "slowly" versus "suddenly." Therefore, this step aims to calculate these two key time-domain dynamic indicators to provide input for the "I term (integral)" and "D term (derivative)" of the subsequent PID controller.

[0086] In some embodiments of this application, the calculation process for the "time-domain cumulative feature" in step S30 above can be implemented by the following sub-steps S31 to S34:

[0087] S31. Create a sliding time window with a preset window length.

[0088] The sliding time window refers to a first-in, first-out (FIFO) queue buffer allocated in memory. This window is used to store the instantaneous deviation characteristic sequence over a recent period.

[0089] The preset window length (denoted as W) can be defined based on time (e.g., the last 3 seconds) or on the number of frames (e.g., the last 90 frames). This window length determines how long the system "remembers" the user's historical behavior. If W is set too short, the system will be too sensitive and easily forget the user's past performance; if W is set too long, the system's feedback will have a significant lag.

[0090] In some preferred embodiments, the window length W is set to cover a complete repetition duration, for example, W = 50 frames.

[0091] S32. Calculate the weighted sum of the instantaneous deviation features within the sliding time window, as the time-domain cumulative feature.

[0092] This step essentially performs a "numerical integration" operation in discrete mathematics. The system accumulates all historical error data in the buffer to quantify the "stubbornness" of the action error.

[0093] Specifically, the time-domain cumulative feature (denoted as I) err (t) can be calculated using the following formula:

[0094]

[0095] In the formula:

[0096] e(tk) represents the instantaneous deviation characteristic at the k-th historical moment in the sliding window;

[0097] Δ is the sampling time interval (i.e., the frame interval);

[0098] λ k The attenuation weighting coefficient (usually λ) k ∈(0,1]). In some embodiments, to emphasize "recent performance", the closer to the current time t, the greater the weight (λ0>λ). W-1 In some simplified embodiments, it can be set to equal weight (λ). k =1).

[0099] Let's continue with the example of "User A doing squats".

[0100] Scenario 1 (Initial Mistake): User A's form is correct in the first 4 squats (e(t)≈0). On the 5th squat, a knee valgus occurs occasionally, causing the instantaneous error e(t) to spike to 10. At this point, since most of the historical data within the sliding window is 0, the calculated weighted sum I... err (t) is still relatively small. Based on this, the system judges that this is just an "accidental error", and the PID controller will output a milder prompt in the future.

[0101] Scenario 2 (Persistent Error): User A experiences knee valgus during 5 consecutive squats. At this point, the sliding window is filled with a high amount of error data e(t). The calculated I... err (t) will accumulate to a huge value (e.g., exceeding a preset saturation threshold). This value tells the system: "User A is not only wrong, but this error is very persistent, and ordinary reminders are ineffective." This will trigger the "dimension reduction" or "switching" of the teaching strategy in subsequent steps.

[0102] S33. Calculate the difference between the instantaneous deviation feature of the current frame and the instantaneous deviation feature of the previous frame;

[0103] S34. The ratio of the difference to the sampling time interval is determined as the time-domain rate of change feature.

[0104] Specifically, these two steps aim to capture the "trend of evolution" of user movement errors in real time through numerical calculations. In motion control theory, the error value alone (P term) only reflects "how wrong it is at the moment," while the rate of change of the error (D term) reveals "how quickly the error is worsening or improving." This is crucial for predicting motion risks (such as movement collapse due to exhaustion).

[0105] Specifically, the time-domain rate of change feature (denoted as D) err The calculation of (t) essentially involves performing a first-order difference operation in the discrete-time domain. Its calculation logic is as follows:

[0106] First, in step S33, the system extracts the instantaneous deviation feature e(t-1) of the previous frame (at time t-1) from the buffer queue, and subtracts it from the instantaneous deviation feature e(t) of the current frame (at time t) to obtain the error increment Δe = e(t) - e(t-1).

[0107] If Δe > 0, it indicates that the error is increasing (the action is deteriorating);

[0108] If Δe < 0, it means the error is decreasing (the action is being corrected);

[0109] If Δe < 0, it means the action remains stable (regardless of whether it is right or wrong).

[0110] Subsequently, in step S34, the system obtains the sampling time interval Δt of the video stream (e.g., for a 30FPS video stream, Δt ≈ 0.033 seconds). The system calculates the ratio of Δe to Δt, thus obtaining the time-domain rate of change characteristic:

[0111]

[0112] Continuing with the example of "User A doing squats", let's assume that he is doing his last set of high-intensity training and is in a state of extreme fatigue.

[0113] At frame t-1, user A's knee remains stable, with an instantaneous deviation e(t-1) of 2 degrees (within the acceptable range). However, at frame t (only 0.033 seconds later), due to sudden exhaustion of the gluteus medius, user A loses control of the femur, causing the left knee to momentarily valgus inward, and the instantaneous deviation e(t) surges to 8 degrees. At this point, although 8 degrees may not yet reach the static threshold of "serious error" (e.g., 15 degrees) in absolute terms, simple P-item control might only provide a mild warning.

[0114] However, after the system executes steps S33 and S34, the calculation reveals that: (degrees / second). Thus, the system captured a time-domain rate of change characteristic as high as 181.8.

[0115] This extremely high positive value is a clear "pre-collapse warning signal." It tells the system that the user's defenses are crumbling in an instant. Based on this signal, the subsequent PID controller (D-term channel) will bypass the normal logic and directly trigger an "advanced blocking" command (such as immediately playing a rapid alarm sound), thereby forcing the user to stop moving before the knee completely collapses and causes ligament damage.

[0116] In summary, steps S31 to S34 together constitute the dynamic feature extraction mechanism of this application, which aims to upgrade a single instantaneous error signal into a time-domain control signal that includes historical and trend dimensions.

[0117] Specifically, steps S31 and S32 establish a sliding window to perform numerical integration calculations, accurately quantifying the cumulative persistence of motion errors, enabling the system to distinguish between occasional errors and habitual errors; while steps S33 and S34 calculate the difference ratio of inter-frame errors to capture the rate of deterioration of motion quality in real time, giving the system a dynamic early warning capability for fatigue collapse or motion injury risks. Together, they provide the necessary time-varying input parameters for subsequent adaptive PID feedback control.

[0118] Meanwhile, by combining the above-mentioned simplified differential and first-level low-pass filtering, this application minimizes the computational load while ensuring the real-time performance and sensitivity of the time-domain rate of change characteristics, thereby achieving smooth deployment on edge devices such as mobile phones and tablets.

[0119] S40. Based on the instantaneous deviation features, retrieve the preset motion chain knowledge graph to obtain the initial content load containing attribution information.

[0120] Specifically, existing motion correction systems often remain at the "phenomenal level" of feedback (e.g., "Don't let your knees buckle inward"), which only tells users "what error has occurred" but cannot explain "why it happened." This step utilizes a pre-built kinetic chain knowledge graph to simulate the diagnostic logic of a professional physical therapist, deducing the physiological root cause from the apparent error, thereby generating an initial content load that can truly guide the user's improvement.

[0121] In some embodiments of this application, step S40 can be implemented by the following sub-steps S41 to S42:

[0122] S41. Match the instantaneous deviation feature to the corresponding error representation node.

[0123] First, the system needs to discretize or map the instantaneous deviation feature e(t) of continuous numerical values ​​to semantic entry points in a pre-defined motion chain knowledge graph. This pre-defined motion chain knowledge graph is a graph-based data structure containing two types of core nodes:

[0124] Symptom Node: Corresponds to geometric errors that can be visually detected (such as "knee valgus" or "anterior pelvic tilt").

[0125] Root Cause Node: Corresponds to the muscle or joint dysfunction that leads to the error (such as "insufficient gluteus medius muscle strength" or "tight iliopsoas muscle").

[0126] During this step, the system compares e(t) with a preset classification threshold. When e(t) falls within a specific interval, the corresponding error representation node is activated.

[0127] Continuing with the example of user A performing a squat, the previous steps calculated the instantaneous deviation feature of user A's left knee, e(t), to be 12 degrees (inward valgus direction). The system searched the map entry layer and found that this value matched the trigger condition of the appearance node

Knee Valgus

[0128] S42. Index along the graph path to the associated root cause node, and use the physiological cause described by the root cause node as the initial content load.

[0129] After locking onto the representation node, the system traverses downstream along predefined directed edges. These edges represent causal relationships in sports biomechanics and are typically predefined by sports science experts.

[0130] The system identifies one or more associated root cause attribution nodes based on the attribution map path and extracts the textual description data (i.e., "physiological causes") stored in these nodes. This data includes not only anatomical terms but may also contain specific descriptions of sensation cues. This extracted information constitutes the initial content payload C. load .

[0131] For the node

knee valgus

[0132] Path A (probability 0.7): points to the node

Gluteus Medius Inhibition

[0133] Path B (probability 0.3): points to the node

Dorsiflexion restricted

[0134] In this embodiment (assuming a maximum probability path-first strategy is used), the system indexes along path A to the [Gluteus medius inhibition] node. The attribute data stored in this node includes:

[0135] Anatomical cause: "Insufficient strength of the lateral gluteal muscles".

[0136] Description of the feeling of exertion: "Feel the outer side of your hips contracting."

[0137] Key point of force: "Gluteus medius".

[0138] The system packages the above information to generate the initial content payload C. load ={Focus:"Gluteus Medius",Explanation:"Weak hip abductor muscles",Cue:"Squeeze outer glutes"}.

[0139] Understandably, through steps S41 and S42, the system completes an "expert diagnosis." The subsequent PID controller (S50) will no longer simply read this content out, but will decide "how to read it" (whether to shout it out loud, guide it gently, or use a different metaphor), thereby transforming this standardized "prescription" into a personalized "treatment plan."

[0140] S50. Input the instantaneous deviation feature, the time-domain cumulative feature, and the time-domain rate of change feature into a preset adaptive feedback control model. Calculate the abstract level parameters of the teaching strategy based on the magnitude of the time-domain cumulative feature, calculate the tone-specific parameters based on the instantaneous deviation feature, and calculate the intervention timing parameters based on the time-domain rate of change feature.

[0141] Specifically, the adaptive feedback control model employs a variant architecture of the classic PID control algorithm. Unlike PID outputs of voltage or torque in industrial control, this model uses semantic mapping logic to map the three input features of the PID to three control dimensions of Natural Language Generation (NLG):

[0142] Integral term (I term) mapping: The time-domain cumulative feature corresponds to the integral term in the PID algorithm, which is used to characterize the amount of error accumulated by the user within the historical time window (i.e. the persistence of the error).

[0143] Differential term (D term) mapping: The time-domain rate of change feature corresponds to the differential term in the PID algorithm, which is used to characterize the rate of change of user action deviation (i.e. the trend of error deterioration).

[0144] Proportional term (P term) mapping: The instantaneous deviation feature corresponds to the proportional term in the PID algorithm and is used to characterize the current error magnitude.

[0145] In the specific implementation details, step S50 includes independent calculation processes for these three dimensions.

[0146] In one embodiment, the time-domain cumulative characteristics of the input (i.e., the integral term I in the PID) are processed using an adaptive feedback control model. err When this occurs, the system executes the following logic to dynamically calculate the abstract level parameters of the teaching strategy. This process is implemented through steps S51 to S53, aiming to solve the problem of correcting "persistent errors".

[0147] S51. Determine whether the time-domain cumulative feature exceeds a preset saturation threshold.

[0148] Specifically, the system first reads the preset saturation threshold (denoted as Threshold). sat This threshold represents the system's "tolerance limit" or "turning point" for repeated user errors.

[0149] In control theory, the integral term reflects the accumulation of steady-state error. If I err If the error continues to increase and exceeds the threshold, it means that the system's previous conventional control signals (i.e., ordinary error correction instructions) have failed to eliminate the error, and the system has entered an "integral saturation" state. In a sports teaching scenario, this corresponds to the user being stuck in a cognitive deadlock of "understanding but not being able to do it," and continuing to repeat the same instructions is often futile.

[0150] Continuing with the example of user A performing squats, let's assume the system sets a saturation threshold (Threshold). sat =200. The system monitors the time-domain cumulative characteristic I of user A in real time. err .

[0151] Phase 1 (Initial Stage): In the first few sets of training, User A occasionally experienced knee valgus, with a cumulative value of I. err It slowly increases to 50. At this point, 50 < 200, and the result is "No".

[0152] Phase Two (Persistent Stage): User A repeatedly exhibits the same internal deduction error in multiple consecutive sets, despite repeated system prompts. err It accumulates rapidly and reaches 210 in the Nth frame. At this point, 210 > 200, and the result is "yes".

[0153] S52. If so, the abstract level parameter is set to the external metaphor mode, wherein the external metaphor mode indicates the generation of guiding instructions that include environmental analogies or external images.

[0154] When the accumulated error is small, the system defaults to the Internal Focus strategy. This is the most basic and direct biomechanical correction method, designed to guide users to focus on their own limb movements and muscle contractions.

[0155] In this mode, the system-generated guidance instructions typically include specific anatomical terms, joint names, or muscle activation sensations.

[0156] Specifically, in Phase 1 (I) err =50), the system executes step S53. The system sets the abstract level parameter to Strategy="Internal_Focus". Combining the attribution information (gluteal medius inhibition) retrieved in the previous step S40, the system finally generates the feedback information as: "Please pay attention to your left knee, actively contract your lateral hip muscles, and open your knee outwards."

[0157] S53. If not, the abstract level parameter is set to internal focus mode, wherein the internal focus mode indicates the generation of guidance instructions containing control instructions for specific body parts.

[0158] When the accumulated error "saturates", the system forcibly triggers a strategy switch and enters the External Metaphor / Focus mode.

[0159] According to Motor Control Science, when learners focus excessively on bodily details (internal focus) but are unable to correct their movements, it is often because the coordination between major and antagonistic muscle groups is impaired ("paralyzed in analysis"). In this case, shifting attention to the impact of the movement on the environment (external focus) can utilize the body's self-organizing capabilities to more effectively activate the correct movement patterns.

[0160] In this mode, the system prohibits the use of muscle / joint terminology and instead generates figurative metaphors, environmental analogies, or goal-oriented imagery through LLM.

[0161] Specifically, in Phase Two (I) err =210), the system detects that the user is stuck in a persistent error and executes step S52. The system modifies the abstraction level parameter to Strategy="External_Metaphor". At this point, although the root cause is still "gluteus medius weakness", the system no longer mentions "hips" or "knees". Based on this strategy parameter, LLM generates a completely new guiding instruction: "Imagine your feet are stepping on a crack in the floor, try to tear the floor apart to the sides." This external metaphor of "tearing the floor" can subconsciously induce the user to perform hip abduction and external rotation movements, thereby correcting knee valgus without the user's awareness.

[0162] It is understood that through steps S51 to S53, this application implements a hierarchical diagnosis-and-treatment-style intelligent feedback mechanism: precise anatomical correction is used for minor errors; for stubborn errors, it automatically switches to holistic imagery guidance, thereby significantly improving the teaching success rate in complex scenarios.

[0163] In one embodiment, when processing the instantaneous deviation characteristics of the input (i.e., the proportional term e(t) in the PID) using an adaptive feedback control model, the system executes the following logic to calculate the tone-specific parameters. This process corresponds to the proportional (P) element in the PID control law, and aims to adjust the "emotional intensity" or "severity level" of the feedback rhetoric in real time according to the severity of the current error.

[0164] Specifically, calculating the tone-specific parameters based on the instantaneous deviation features includes: mapping the instantaneous deviation features to tone severity values ​​using a nonlinear mapping function, wherein the tone severity values ​​are positively correlated with the instantaneous deviation features.

[0165] In real-world sports instruction scenarios, the magnitude of an error and the required intensity of the warning are often not simply linearly related. For minor errors (which may be measurement noise or fluctuations within acceptable limits), the system should maintain "tolerance" to avoid frequently interrupting the user; however, for serious errors that exceed safe limits, the system must quickly reach "maximum severity" to ensure the warning effect. Therefore, this embodiment uses a nonlinear mapping function to implement this logic.

[0166] In some preferred embodiments of this application, the nonlinear mapping function employs a variant of the Sigmoid function (S-shaped growth curve). The system calculates the severity value (denoted as α) using the following formula:

[0167]

[0168] In the formula:

[0169] |e(t)| is the absolute value of the instantaneous deviation characteristic (representing the magnitude of the error);

[0170] k is the slope factor, used to control the "sensitivity" of tone changes. The larger the k value, the steeper the transition from mild to severe tone;

[0171] e mid This is the critical deviation threshold for tone switching (i.e., when the error reaches this value, the severity is 0.5);

[0172] α is the tone-specific parameter of the output, with a value range of (0,1). This value is positively correlated with |e(t)|.

[0173] The calculated α value will be used to define the tone constraint for generative AI (LLM):

[0174] Mild range (α < 0.3): The system instructs LLMs to use suggestive and encouraging language. For example, use modal verbs such as "try...", "can...", or "pay attention to feelings...".

[0175] Neutral range (0.3 ≤ α ≤ 0.7): The system instructs LLMs to use a declarative, objective tone. Directly describe facts and correct actions, such as "Please turn your knees outward."

[0176] Severity Range (α>0.7): The system instructs LLMs to use imperative, short, and abrupt language. Modifiers are omitted, and verb phrases such as "knees out!" or "tuck in!" are used directly to create a sense of urgency.

[0177] Let's continue with the example of user A performing squats. Assume the system sets a critical threshold e. mid=15 degrees, sensitivity k=0.5.

[0178] Scenario 1 (Slight Deviation): User A's knee valgus angle e(t) = 5 degrees. The system substitutes this into the formula and calculates α ≈ 0.006 (extremely low value). At this point, although an error is detected, the system judges it to be within an acceptable range. The tone parameter α instructs the LLM to generate an extremely mild reminder: "Well done, try paying a little more attention to the direction of your knees."

[0179] Scenario 2 (Significant Error): User A's knee valgus angle increases to e(t) = 15 degrees. Substituting into the formula, we get α = 0.5 (median). The system's tone changes to a standard instruction: "Please actively control your knee to align it with your toes."

[0180] Scenario 3 (Critical Error): User A's movements become uncontrollable, and the angle of knee valgus surges to e(t) = 25 degrees. Substituting into the formula, the exponent term -0.5 × (25 - 15) = -5, causing the denominator to rapidly approach 1, ultimately yielding α ≈ 0.99 (saturation value). At this point, the tone parameter indicates that the LLM enters "maximum alert" mode. The feedback message no longer includes any polite language, but directly outputs: "Open your knees!"

[0181] Through this nonlinear mapping, this application is able to simulate the emotional responses of a real coach: being lenient with minor mistakes to protect the user's confidence, and being strict with major mistakes to ensure the safety of the action, thereby achieving an extremely natural and efficient teaching interaction.

[0182] In one embodiment, the time-domain rate of change characteristic of the input (i.e., the derivative term D in the PID) is processed using an adaptive feedback control model. err When (t) occurs, the system executes the following logic (steps S54-S55) to calculate the intervention timing parameters. This process corresponds to the derivative (D) element in the PID control law, aiming to dynamically determine whether feedback should be given calmly after the action ends or should intervene urgently at the moment the risk occurs, based on the rate of error deterioration.

[0183] S54. Compare the time-domain rate of change characteristics with the preset safety damping limit.

[0184] The system first reads the preset "Safety Damping Limit" (denoted as Limit). safety This limit is not based on the absolute magnitude of the error (such as angle), but on the speed at which the error occurs (such as angular velocity). In sports biomechanics, slow movement deviations can usually be corrected by muscle adjustments, but high-speed deviations (such as instantaneous knee valgus or sudden lumbar instability) often mean a loss of muscle control (exhaustion) or a surge in the risk of ligament injury.

[0185] The system will calculate D in real time err (t) and Limit safety Perform numerical comparisons.

[0186] Let's continue with the example of user A performing squats. Assume the system sets a safety damping limit (Limit) based on user A's weight and training level. safety = 100 degrees / second (representing the safe upper limit of the knee valgus velocity).

[0187] Scenario 1 (Normal Fatigue): During the last few squats, User A's form begins to deteriorate, with the knee valgus angle slowly increasing from 0 degrees to 10 degrees within 1 second. At this point, the calculated rate of change D... err ≈10 degrees / second. Since 10 < 100, the system determines that the limit has not been exceeded. At this time, the system tends to maintain the conventional "post-action feedback mode", that is, wait for the user to complete the squat and stand up before giving the "watch your knees" prompt, so as not to disrupt the user's breathing rhythm.

[0188] S55. When the time-domain rate of change characteristic exceeds the safety damping limit, the intervention timing parameter is set to the advanced blocking mode, wherein the advanced blocking mode indicates that an immediate interruption command is triggered when feedback information is generated.

[0189] When D is detected err (t) surged instantly and broke through the limit. safety At this point, the system determines that the current state is "high-risk and out of control". The system then forcibly sets the intervention timing parameter to "Pre-emptive Blocking Mode".

[0190] This mode has the highest system priority, and its specific execution logic is as follows:

[0191] 1. Generation Interruption: Instructs the generation module to stop generating long sentences or explanatory text, and only generate the shortest immediate interruption command (such as "Stop!", "Danger!", or a simple beeping alarm).

[0192] 2. Output preemption: Instructs the audio output module to interrupt the current background music or audio stream and immediately play the above interrupt command.

[0193] Let's continue with the example of user A doing squats.

[0194] Scenario 2 (Sudden Risk of Exhaustion): User A attempts to lift a heavy weight while on the verge of exhaustion. At the bottom of the squat, the gluteus medius suddenly loses strength, and the left knee collapses inward by 15 degrees in 0.1 seconds.

[0195] At this point, the time-domain rate of change characteristic spikes: D err ≈15 / 0.1=150 degrees / second.

[0196] Since 150 > 100, the system determines that the safety damping limit has been exceeded. This extremely high rate of change indicates that the ligament is subjected to enormous shear force and is about to be damaged.

[0197] The system immediately activated the "advanced blocking mode." Before user A had even passed the most difficult stage at the bottom, the system immediately issued a rapid "Beep—! Stop!" alarm.

[0198] User A, upon hearing the auditory stimulus, instinctively released the force or stopped the action, thus avoiding a possible meniscus or ligament tear.

[0199] Through steps S54 and S55, this application utilizes the "predictive" characteristic of the PID differential term to intervene within the "golden window" when the error has not yet reached its maximum magnitude but the deterioration trend is already out of control, thus achieving a technological leap from "hindsight" to "early warning".

[0200] It is understandable that, through the mathematical mapping of the PID control law and the teaching semantics described above, this application realizes a fundamental leap from "static gain" to "dynamic adaptation" in the feedback strategy, enabling the system to make real-time decisions on the optimal teaching path based on the severity, persistence, and deterioration trend of errors, just like a senior coach.

[0201] S60. Combine the abstract level parameters, the tone-specific parameters, and the intervention timing parameters to obtain the prompt control parameters.

[0202] Specifically, this step aims to complete the structured encapsulation of the control signals. The parameters generated in the preceding steps are logically discrete. In order to drive the subsequent generative artificial intelligence model (LLM) to generate in a controlled manner, these scattered signals must be integrated into a unified, standardized prompt control parameter, usually in the form of a vector or JSON object.

[0203] Specifically, this step executes the following aggregation logic:

[0204] 1. The system constructs a predefined control vector template V. ctrl =[P tone ,P strategy ,P timing ].

[0205] 2. The system fills the specific values ​​of each dimension calculated in step S50 into the corresponding slots, and the final generated combination is the prompt control parameter:

[0206] P tone(Tone slot): Fill in the tone severity value obtained from the instantaneous deviation feature mapping or the corresponding enumeration label (such as "Tone:Imperative").

[0207] P strategy (Strategy slot): Enter the abstract mode label determined by the time-domain cumulative features (e.g., "Strategy:External_Metaphor").

[0208] P {timing} (Timing Slot): Enter the intervention mode label determined by the time-domain rate of change characteristics (e.g., "Timing:Pre-emptive").

[0209] Let's continue with the example of user A performing squats. Assume that in the preceding steps, the calculation results for each channel of the PID controller are as follows:

[0210] P channel (tone): Severe knee valgus (e(t) = 25 degrees) was detected, and the tone severity α≈0.99 was calculated, corresponding to the label "Severe / Commanding".

[0211] Channel I (Policy): Error stubbornness (integral saturation) detected, corresponding to the label "External_Metaphor".

[0212] Channel D (Timing): Extremely rapid deterioration detected (signs of impending exhaustion), corresponding to the tag "Immediate_Interrupt".

[0213] In step S60, the system combines the three independent signals to generate prompt control parameters with the following structure (using JSON format as an example):

[0214]

[0215] This encapsulated data packet will serve as an immutable "hard constraint," directly injected into the LLM's system prompt in the next step. This ensures that every piece of feedback generated by the AI ​​is not merely "explaining errors," but rather rigorously executing the teaching tactics planned by the PID control law—that is, using the harshest tone (P), the most vivid metaphors (I), and stopping the user at the most dangerous moment (D).

[0216] It is understandable that, through the above-mentioned standardized combination and automatic derivation of auxiliary decision fields, this application achieves a seamless bridge between the semantic PID three-way control signals and the structured instructions that can be directly executed by the large language model, thereby ensuring a high degree of coordination and precise controllability of the feedback speech in terms of tone, strategy and timing.

[0217] S70. Assemble system prompts based on the initial content load and the prompt control parameters, and generate adaptive feedback information based on the system prompts.

[0218] Specifically, in order to ensure that the output of the generative artificial intelligence model (LLM) strictly follows the planning of the PID control law (i.e., no illusions are produced, and the tone, strategy, and timing are completely accurate), the system cannot directly "feed" the raw data to the model, but must build a strongly constrained system prompt.

[0219] In some embodiments of this application, the specific assembly process of step S70 can be implemented by the following steps S71-S73:

[0220] S71. Construct a structured prompt template that includes role definition fields, context state fields, and instruction constraint fields.

[0221] The system first invokes a pre-defined "structured hint template." This template typically uses JSON (JavaScript Object Notation) format to ensure that the LLM can accurately parse the semantic weights of each field. The template contains three core fields:

[0222] Role Definition field: Sets the LLM's identity criteria (e.g., "You are a strict but professional biomechanical correction specialist").

[0223] Context State field: Used to input objective facts, i.e., "what happened".

[0224] Instruction Constraints: Used to input control parameters, i.e., "how to say it".

[0225] S72. Fill the initial content payload into the context state field.

[0226] The system will fill the data stream generated by the previous steps into the above template:

[0227] Fill in the content payload: Fill the context state field with the initial content payload (such as “gluteal medius inhibition”, “force sensation: tearing the floor”) extracted from the knowledge graph in step S40. This establishes the “core information content” of the feedback.

[0228] Enter control parameters: Enter the prompt control parameters (tone, strategy, timing) obtained in step S60 into the instruction constraint field. This establishes the "expression" of the feedback.

[0229] Let's continue with the example of user A doing squats. Suppose that in the previous steps, the system detected that user A is in a high-risk state of "persistent error" and "extreme fatigue" (integral saturation, differential overlimit).

[0230] Therefore, the initial content payload (from S40) is {Cause:"Glute Medius Weakness",Metaphor_Cue:"Tear the floor apart"}.

[0231] The prompt control parameters (from S60) are {Tone:"Severe / Commanding",Strategy:"External_Metaphor",Timing:"Pre-emptive_Block"}.

[0232] The final system prompt generated by the system assembly is shown below:

[0233]

[0234] S73 fills the abstract level parameter, the tone-specific parameter, and the intervention timing parameter into the instruction constraint field to form the system prompt.

[0235] Finally, the system inputs the constructed JSON Prompt into the pre-trained large language model. The model performs "controlled generation" of information in context_state based on the strong constraints in instruction_constraints.

[0236] Based on the strong constraints of the above case, LLM will not generate invalid feedback such as "Dear user, it seems your gluteus medius is a bit weak, please try contracting it" (this is the default output of a regular LLM). Instead, LLM will strictly follow the "command tone" + "external metaphor" + "immediate interruption" to generate the final adaptive feedback message: "Stop! Tear the floor apart immediately!" Upon hearing this, the user immediately stops and performs the activation action, and the error in the next frame drops rapidly. err As the praise begins to fade, the system switches back to gentle praise, forming a complete positive closed loop.

[0237] Meanwhile, if intervention_timing is detected as "advanced blocking mode", the system will also trigger non-verbal feedback (such as a beeping sound) synchronously, and output the text synthesized by TTS speech in sync, thereby completing the intervention on user A's high-risk actions within milliseconds.

[0238] It is understandable that, through the above-mentioned structured prompts and the use of a strictly constrained large language model, this application has achieved a closed-loop mapping of the entire chain from physical error to physiological attribution to optimal teaching techniques, making the feedback both accurate and professional as well as natural and humane, truly reaching the guidance level of a top-notch real coach.

[0239] In summary, the motion analysis and feedback method of this application has the following beneficial effects:

[0240] 1. Achieved dynamic adaptive adjustment of teaching strategies: This application employs an "adaptive feedback control model" to calculate in real time the "time-domain cumulative characteristics" representing the persistence of movement errors and the "time-domain rate of change characteristics" representing the worsening trend of movement errors. The system no longer provides mechanical feedback based solely on current instantaneous deviations, but instead senses the user's "historical performance inertia" and "current state trend," dynamically mapping these dynamic characteristics to the abstract level of the teaching strategy, the severity of the tone, and the timing of intervention. This achieves adaptive guidance "tailored to individual needs," much like a real coach, significantly improving user compliance and training experience.

[0241] 2. Breakthrough in Correcting Stubborn Errors: This application utilizes a dynamic switching logic for setting "abstract level parameters of teaching strategies," specifically automatically switching the strategy from "internal focus mode" to "external metaphor mode" when the "time-domain cumulative characteristics" exceed a saturation threshold. The system can automatically identify "cognitive deadlock" scenarios where conventional instructions (such as "contracting muscles") fail, and bypass user cognitive barriers through environmental analogies or vivid metaphors (such as "tearing the floor"), activating the body's self-organizing movement patterns. This effectively solves the problem of users "understanding but unable to perform" when facing stubborn errors, significantly improving the success rate of correcting complex actions or deep-seated errors.

[0242] 3. Achieved proactive defense against motion risks: This application introduces the monitoring of "time-domain rate of change characteristics" and compares them with "safety damping limits." When an excessively rapid rate of motion deterioration is detected, a "proactive blocking mode" is triggered. In this way, the system possesses the "predictive" perception capability for motion collapse (such as instantaneous instability due to exhaustion), and can immediately generate an interruption command the instant the error amplitude has not yet reached its maximum value but the deterioration trend is already out of control. This transforms traditional "hindsight" error correction into proactive "early warning" defense, effectively preventing motion injuries caused by user fatigue or sudden loss of control.

[0243] 4. Improved the scientific rigor and accuracy of attribution analysis: This application constructs a user-specific "neutral position reference model" and a pre-built "kinetic chain knowledge graph." On the one hand, it eliminates systematic misjudgments caused by individual differences in limb proportions (such as long femurs); on the other hand, it achieves deep indexing from "visual representation errors" (such as genu valgum) to "physiological roots" (such as gluteus medius inhibition). Furthermore, it ensures the biomechanical scientific rigor and accuracy of feedback information, helping users understand and improve their movement patterns from the root causes.

[0244] Furthermore, this invention also proposes a computer-readable storage medium, which can be any one or any combination of several of the following: hard disk, multimedia card, SD card, flash memory card, SMC, read-only memory (ROM), erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB memory, etc. The computer-readable storage medium includes a motion analysis and feedback program 10. The specific implementation of the computer-readable storage medium of this invention is largely the same as the motion analysis and feedback method and the specific implementation of server 1 described above, and will not be repeated here.

[0245] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0246] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0247] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0248] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0249] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0250] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A motion analysis and feedback method, characterized in that, include: Acquire user motion video stream, extract user skeletal features based on initial data of the user motion video stream, and construct user neutral position reference model based on the skeletal features; The deviation between the real-time pose of the user's motion video stream and the neutral position reference model is calculated based on the neutral position reference model to obtain instantaneous deviation characteristics; Based on the distribution of the instantaneous deviation characteristics in the time dimension, the time-domain cumulative characteristics that characterize the persistence of the action error and the time-domain rate of change characteristics that characterize the deterioration trend of the action error are calculated in real time. Based on the instantaneous deviation features, a preset motion chain knowledge graph is retrieved to obtain an initial content payload containing attribution information; The instantaneous deviation feature, the time-domain cumulative feature, and the time-domain rate of change feature are input into a preset adaptive feedback control model. The abstract level parameters of the teaching strategy are dynamically calculated based on the magnitude of the time-domain cumulative feature, the tone-specific parameters are calculated based on the instantaneous deviation feature, and the intervention timing parameters are calculated based on the time-domain rate of change feature. The abstract level parameters, the tone-specific parameters, and the intervention timing parameters are combined to obtain the prompt control parameters; The system prompts are assembled based on the initial content load and the prompt control parameters, and adaptive feedback information is generated based on the system prompts.

2. The motion analysis and feedback method as described in claim 1, characterized in that, Based on the initial data of the user's motion video stream, the user's skeletal features are extracted, and a neutral bit reference model of the user is constructed based on the skeletal features, including: Extract the user's static limb length ratio and joint range of motion from the initialization data; Based on the static limb length ratio, the geometric parameters of the preset general biomechanical model are corrected to generate a neutral position reference model that adapts to the user's body structure.

3. The motion analysis and feedback method as described in claim 1, characterized in that, Real-time calculation of the time-domain cumulative features characterizing the persistence of motion errors, and the time-domain rate of change features characterizing the worsening trend of motion errors, including: Create a sliding time window with a preset window length; Calculate the weighted sum of the instantaneous deviation features within the sliding time window, and use it as the time-domain cumulative feature; Calculate the difference between the instantaneous deviation feature of the current frame and the instantaneous deviation feature of the previous frame; The ratio of the difference to the sampling time interval is determined as the time-domain rate of change feature.

4. The motion analysis and feedback method as described in claim 1, characterized in that, The pre-built motion chain knowledge graph contains multiple error representation nodes and root cause attribution nodes associated with the error representation nodes; Based on the instantaneous deviation features, a pre-set motion chain knowledge graph is retrieved to obtain an initial content payload containing attribution information, including: The instantaneous deviation features are matched to the corresponding error manifestation nodes; Index along the graph path to the associated root cause node, and use the physiological cause described by the root cause node as the initial content load.

5. The motion analysis and feedback method as described in any one of claims 1 to 4, characterized in that, The adaptive feedback control model employs a PID control algorithm; wherein... The time-domain cumulative feature corresponds to the integral term in the PID control algorithm and is used to characterize the amount of error accumulated by the user within the historical time window. The time-domain rate of change characteristic corresponds to the derivative term in the PID control algorithm and is used to characterize the rate of change of the user's action deviation.

6. The motion analysis and feedback method as described in claim 5, characterized in that, The abstract hierarchical parameters of the magnitude dynamic computation teaching strategy based on the aforementioned time-domain cumulative characteristics include: Determine whether the time-domain cumulative feature exceeds a preset saturation threshold; If so, the abstract level parameter is set to the external metaphor mode, wherein the external metaphor mode indicates the generation of guiding instructions that include environmental analogies or external images; If not, the abstract level parameter is set to internal focus mode, wherein the internal focus mode indicates the generation of guidance instructions containing control instructions for specific body parts.

7. The motion analysis and feedback method as described in claim 5, characterized in that, Calculating tone specificity parameters based on the instantaneous deviation features includes: mapping the instantaneous deviation features to tone severity values ​​using a nonlinear mapping function, wherein the tone severity values ​​are positively correlated with the instantaneous deviation features; and / or The intervention timing parameters are calculated based on the aforementioned time-domain rate of change characteristics, including: Compare the time-domain rate of change characteristics with the preset safety damping limit; When the time-domain rate of change characteristic exceeds the safety damping limit, the intervention timing parameter is set to the advanced blocking mode, wherein the advanced blocking mode indicates that an immediate interruption command is triggered when feedback information is generated.

8. The motion analysis and feedback method as described in claim 1, characterized in that, Assemble system prompts based on the initial content load and the prompt control parameters, including: Construct a structured prompt template that includes role definition fields, context state fields, and instruction constraint fields; Fill the initial content payload into the context state field; The abstract level parameter, the tone-specific parameter, and the intervention timing parameter are filled into the instruction constraint field to form the system prompt.

9. A motion analysis and feedback device, characterized in that, It includes a memory, a processor, and a motion analysis and feedback program stored in the memory and executable on the processor, wherein the processor, when executing the motion analysis and feedback program, implements the motion analysis and feedback method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a motion analysis and feedback program, which, when executed by a processor, implements the motion analysis and feedback method as described in any one of claims 1-8.

Images