Music playback control method and system based on smartwatch posture recognition
By using a smartwatch's posture recognition system to collect and analyze wrist movement data, and combining time-domain and frequency-domain feature extraction, the system distinguishes between movement and control actions, solving the problem of high misjudgment rate of smartwatches during exercise. This enables highly accurate music control and adaptive volume compensation, improving the user experience.
Patent Information
- Application Number
- CN202611435054.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing smartwatch music playback control solutions are prone to misjudgment and misoperation during exercise due to continuous wrist shaking. The misjudgment rate is high, the volume output is unstable, and it is difficult to meet users' needs for efficient, accurate and user-friendly music control.
The smartwatch's posture recognition system collects motion data from the user's wrist, extracts time and frequency domain features, and combines motion cycle features, motion amplitude features, motion abrupt change features, and motion duration features to distinguish between motion actions and control actions. It also generates playback control commands through a two-confirmation mechanism to achieve adaptive volume compensation.
It effectively distinguishes between motion actions and control actions, reduces the misjudgment rate, ensures high accuracy and stability of music playback operation, improves user experience, and achieves adaptive and constant volume output.
Smart Images

Figure CN123114264A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart wearable technology, and in particular to a music playback control method and system based on smartwatch posture recognition. Background Technology
[0002] With the deep integration of smart wearable technology and mobile multimedia applications, smartwatches have become the mainstream portable terminal for users to control music playback in scenarios such as exercise and commuting. Currently, most music playback control solutions based on wearable devices rely on the device's collected information about body movement to achieve music matching and playback control.
[0003] For example, Chinese patent CN109032384A discloses a music playback control method that acquires the shaking state information of a user's body parts while wearing a wearable device, matches the target music to be played based on the shaking state information, and acquires the shaking direction and amplitude of each shake action within the sampling period. Based on the shaking direction and amplitude, the target playback type is determined. The playback type is then determined to be rock, pop, folk, or classical based on the magnitude and variance of the horizontal, vertical, and forward / backward shaking beats. This optimizes music playback control operations, enriches the functionality of wearable devices, and increases user engagement. However, during periodic exercises such as running and fitness, users experience continuous and regular periodic wrist shaking. This shaking is a natural movement without intentional control. When users perform music control operations such as skipping tracks, pausing, and adjusting volume, their wrists make instantaneous and abrupt step movements. If the data is collected and matched based on the shaking direction and amplitude, it can easily lead to frequent misjudgments and misoperations triggered by the movement, significantly reducing the accuracy and stability of music control. In summary, current music playback control suffers from high error rates and unstable volume output, making it difficult to meet users' needs for efficient, accurate, and user-friendly music control on smartwatches.
[0004] Therefore, there is an urgent need for a music playback control method and system based on smartwatch posture recognition that can effectively distinguish between motion actions and control actions and achieve adaptive volume compensation. Summary of the Invention
[0005] This invention provides a music playback control method and system based on smartwatch posture recognition, which can effectively distinguish between motion actions and control actions and achieve precise control.
[0006] To solve the above-mentioned technical problems, this application provides the following technical solution: A music playback control system based on smartwatch gesture recognition includes: a data acquisition module, a playback control module, and a command execution module; The data acquisition module is used to collect raw motion data of the user's wrist in real time and send it to the playback control module; The playback control module receives raw motion data and performs time-domain and frequency-domain feature extraction to obtain motion classification features including motion cycle features, motion amplitude features, motion abrupt change features, and motion continuity features. Based on the motion abrupt change features and motion cycle features, it distinguishes between abrupt and continuous motions to obtain step characteristics; based on the motion continuity features and motion amplitude features, it distinguishes between effective and ineffective trigger durations to obtain continuity features; based on the motion cycle features and motion amplitude features, it distinguishes between continuous behavior with control intent and continuous behavior without control intent to obtain correlation features; based on the motion abrupt change features and motion continuity features, it eliminates unstable and invalid interference actions to obtain stability features; and then, based on the step characteristics, continuity features, correlation features, and stability features, it performs further analysis. The system first confirms the motion actions without control intent and the control actions with control intent. Based on the motion actions without control intent, it constructs the current motion state benchmark and uses the current motion state benchmark to calibrate the step characteristics, continuous characteristics, correlation characteristics, and stability characteristics. Based on the calibrated step characteristics, continuous characteristics, correlation characteristics, and stability characteristics, it performs a second confirmation to obtain the motion actions without control intent and the control actions with control intent. It then compares the motion actions without control intent and the control actions with control intent obtained from the two confirmations to determine whether both meet the preset control requirements. If yes, it generates and sends a playback control command to the command execution module based on the control action with control intent. If no, it returns to perform time-domain feature extraction and frequency-domain feature extraction. The instruction execution module is used to receive and execute playback control instructions.
[0007] The basic principle and beneficial effects of the solution are as follows: Temporal and frequency domain features are extracted from the collected raw motion data to obtain motion cycle features, motion amplitude features, motion abrupt change features, and motion continuity features. This comprehensively characterizes the user's wrist movements from the dimensions of motion regularity, motion intensity, change trend, and time, avoiding recognition bias caused by insufficient information in a single dimension. The combination of motion abrupt change features and motion cycle features yields step characteristics, which accurately identify the differences between abrupt and periodic movements, effectively distinguishing between abrupt control actions and continuous motion actions. The combination of motion continuity features and motion amplitude features yields continuity features, which accurately determine whether the action trigger duration is valid, filtering out short-term jitter and invalid micro-movements. The combination of motion cycle features and motion amplitude features yields correlation features, which can distinguish between actions with control intent and continuous actions without control intent generated by natural movement, eliminating continuous interference caused by movement. The combination of motion abrupt change features and motion continuity features yields stability features, which can eliminate unstable and irregular interference actions. This pairwise combination is complementary and non-overlapping, comprehensively characterizing motion attributes from four levels: step, continuity, correlation, and stability, improving the accuracy and reliability of motion differentiation.
[0008] In addition, the identified movements without intention to control are used as a reference benchmark to dynamically reflect the user's current movement intensity and patterns. This allows for adaptive calibration of four features: step, continuity, correlation, and stability. The action judgment threshold is adjusted in real time according to the user's movement state, enabling adaptive suppression of movement interference in different movement scenarios such as walking, running, and fitness. This fundamentally avoids misjudgments and omissions caused by changes in movement amplitude and cycle, improving the stability of action recognition in complex scenarios. Simultaneously, an initial action confirmation is performed using four mid-level features, followed by a second confirmation after motion benchmark calibration. The consistency of the two confirmations is compared and verified. Playback control commands are generated and executed only when both determined movements and control actions meet preset control requirements. Otherwise, features are re-extracted to form a closed-loop feedback correction. This mechanism filters out misidentifications, interference actions, and abnormal data to the greatest extent possible, ensuring that only genuine, effective, and clear control actions trigger control. This achieves zero-touch, high-precision, and fast-response control for music playback, pause, song skipping, and volume adjustment.
[0009] Because wrist movements during music control are often instantaneous and abrupt step movements, simply matching the direction and amplitude of the wrist movement can easily misinterpret natural movements like walking and running as control actions, leading to frequent false triggers. This solution extracts four action classification features from both the time and frequency domains, combining them in pairs to generate intermediate features such as step and continuous characteristics. This accurately characterizes the difference between the instantaneous abruptness of control actions and the continuous regularity of movement. A closed-loop logic of initial confirmation, motion benchmark calibration, secondary confirmation, and result comparison further distinguishes between the two types of actions, effectively filtering out interference from natural movements. This avoids the limitations of single-direction and amplitude matching, significantly reducing the false trigger rate of movement actions. It ensures that only explicitly defined control actions by the user generate commands, guaranteeing both ease of control and eliminating the inconvenience caused by misoperation. This improves the reliability and user experience of smartwatch music control, adapting to various usage scenarios.
[0010] In summary, this solution can effectively distinguish between periodic motions without control intent and control actions with control intent, and achieve highly robust, low-false-touch, and highly precise music playback control.
[0011] Furthermore, the data acquisition module is also used to acquire user terminal audio data and watch terminal audio data and send them to the playback control module; the playback control module is also used to perform time synchronization alignment processing, amplitude comparison analysis and phase difference calculation on the user terminal audio data and watch terminal audio data to obtain real-time distance change data between the watch terminal and the user terminal, perform correction calculation based on the real-time distance change data to obtain real-time volume compensation data to compensate for volume fluctuations caused by distance changes, and then generate and send volume adjustment commands to the command execution module based on the real-time volume compensation data; the command execution module is also used to receive and execute volume adjustment commands.
[0012] The beneficial effects are as follows: Since smartwatches mostly use speakers or bone conduction to output audio, the relative distance between the watch and the user's ear changes rapidly and irregularly when the user performs control actions such as raising or swinging their wrist. The attenuation of sound during propagation varies significantly with distance, directly causing the in-ear volume to fluctuate and making it difficult to maintain a stable volume output, severely impacting the listening experience. In this solution, sound from both the watch and the user's end is simultaneously acquired and time-synchronized to ensure strict temporal correspondence between the two sound data. Based on amplitude comparison and phase difference analysis of the two-end sound, the dynamic changes in the distance between the watch and the ear caused by wrist posture changes can be accurately reflected. By correcting the real-time distance change data, compensation data matching the distance fluctuation can be generated according to the sound propagation attenuation law, effectively offsetting the volume fluctuations caused by distance changes. In this way, through dual-end sound acquisition, time synchronization, amplitude and phase analysis, distance fitting, and volume attenuation compensation, adaptive and constant in-ear volume output is achieved, improving the music playback experience.
[0013] Furthermore, when the playback control module performs time-domain feature extraction and frequency-domain feature extraction, the time-domain feature extraction includes extracting the peak value, valley value, mean value, and variance of the original motion data. The frequency-domain feature extraction includes converting the original motion data into a frequency-domain signal through Fourier transform, extracting the main frequency, bandwidth, and spectral amplitude of the frequency-domain signal, and combining the time-domain features and frequency-domain features to obtain the motion classification features.
[0014] Furthermore, when the playback control module distinguishes between abrupt and continuous actions based on action mutation features and action cycle features to obtain step characteristics, it identifies the instantaneous change amplitude and rate of change of the action using action mutation features, and identifies the repetition pattern and cycle duration of the action using action cycle features. When the instantaneous change amplitude and rate of change of the action exceed the preset mutation threshold and there is no obvious repetition cycle, it is determined to be abrupt action; otherwise, it is determined to be a continuous action, thereby generating step characteristics that characterize the action mutation attribute. When distinguishing between effective and ineffective trigger durations based on action duration and action amplitude characteristics to obtain the duration feature, the actual duration of the action is obtained using the action duration feature, and the actual swing amplitude of the action is obtained using the action amplitude feature. When the action duration is not less than a preset duration threshold and the swing amplitude is not less than a preset amplitude threshold, it is determined to be an effective trigger duration; otherwise, it is determined to be an ineffective trigger duration, thus generating a duration feature that characterizes the effectiveness of the action trigger. When distinguishing between continuous behaviors with and without control intent based on the characteristics of the action cycle and the characteristics of the action amplitude to obtain correlation characteristics, the regularity of the action cycle and the stability of the amplitude are combined. When the continuous action cycle has no fixed regularity and the amplitude fluctuation is within the preset range of the control action, it is judged as a continuous behavior with control intent; otherwise, it is judged as a continuous behavior without control intent, thus generating correlation characteristics that characterize the correlation of action intent. When unstable and invalid interference actions are eliminated based on the action mutation features and action continuity features to obtain stability features, irregular instantaneous jitter is identified by action mutation features, and short-term discontinuous actions are identified by action continuity features. Instantaneous jitter and short-term discontinuous actions are judged as unstable and invalid interference actions and eliminated to generate stability features that characterize the stability of the action posture.
[0015] The beneficial effects are: it can accurately identify sudden and continuous actions, effective and ineffective trigger durations, accurately distinguish between controlled and uncontrolled behaviors, and efficiently eliminate interfering actions.
[0016] Furthermore, the playback control module performs action confirmation and calibration by: based on the extracted action classification features combined with step features, continuous features, correlation features and stability features, confirming the motion actions without control intention and the control actions with control intention, and obtaining the initially confirmed motion actions without control intention and control actions with control intention. Based on the initially confirmed motion actions without control intent, a current motion state benchmark is constructed. The step characteristics, persistence characteristics, correlation characteristics, and stability characteristics are dynamically calibrated in combination with the current motion state benchmark, and the judgment thresholds of the step characteristics, persistence characteristics, correlation characteristics, and stability characteristics are adjusted. Based on the calibrated step characteristics, persistence characteristics, correlation characteristics, and stability characteristics, the motion actions without control intent and the controlled actions with control intent are reconfirmed to obtain the reconfirmation results. The system compares the uncontrolled and controlled movements obtained from the first and second confirmations to determine whether both confirmations meet the preset control requirements. If both meet the requirements, the confirmation results are considered valid; otherwise, feature extraction and confirmation are performed again.
[0017] The beneficial effects are as follows: First, an initial action is confirmed, and then a state baseline is constructed using motion actions without control intent. The four mid-level features are dynamically calibrated to ensure that the feature judgment fits the current motion state and avoids recognition deviations caused by fixed thresholds. At the same time, by comparing the results of the two confirmations, real and valid control actions are screened out, and interfering and invalid action recognition results are eliminated, ensuring that only control actions that meet the requirements can trigger command generation. This not only makes full use of the reference value of motion actions without control intent, but also reduces the probability of misjudgment through a closed-loop verification mechanism, improves stability and reliability, and optimizes the user experience.
[0018] Furthermore, the playback control module generates and sends playback control instructions based on control actions with control intent, including: pre-determining the correspondence between preset control actions and playback control instructions; performing action type matching on control actions with control intent that have been confirmed twice and meet preset control requirements, and extracting the action amplitude, action direction, and action duration of the control actions with control intent and comparing them with preset control actions to determine the corresponding playback control instructions.
[0019] The beneficial effects are as follows: by setting the correspondence between preset control actions and playback control commands, the control intention of different control actions is clarified, avoiding confusion in command generation; by extracting key parameters of control actions for comparison and matching, it is ensured that the generated commands are highly consistent with the user's actual control intentions, reducing the probability of command misgeneration.
[0020] Furthermore, the playback control module performs time synchronization alignment, amplitude comparison analysis, and phase difference calculation on the user-end audio data and the watch-end audio data to obtain real-time distance change data, and performs correction calculations to obtain real-time volume compensation data. This includes: extracting the timestamps of the user-end audio data and the watch-end audio data, and eliminating the timing deviation between the two audio data through a timestamp alignment algorithm; performing amplitude comparison analysis on the two audio data to calculate the amplitude difference, and simultaneously extracting the phase information of the two audio data through Fourier transform to calculate the phase difference, and combining the sound propagation speed to fit the initial distance data between the watch-end and the user-end; removing outliers and fluctuating data in the initial distance data, and using a moving average algorithm to smooth and correct the initial distance data to obtain real-time distance change data; and generating real-time volume compensation data based on the real-time distance change data, combined with the sound propagation attenuation law and the correspondence between the distance change and the volume attenuation.
[0021] The beneficial effects are: eliminating the timing deviation of the two audio data and avoiding distance calculation errors caused by timing misalignment; based on the correction of distance change data and sound attenuation law, the generated volume compensation data can accurately match distance fluctuations, effectively offset the attenuation during sound propagation, and ensure that the in-ear volume remains constant.
[0022] Furthermore, the playback control module generates and sends volume adjustment commands based on real-time volume compensation data, including: pre-setting a volume adjustment gradient threshold, calculating the required volume adjustment amount based on real-time volume compensation data; verifying the rationality of the volume adjustment amount, eliminating abnormal adjustment amounts that exceed the watch's volume adjustment range, and generating a volume adjustment command.
[0023] The benefits are: it avoids auditory discomfort caused by adjusting the volume too high or too low, ensures that the adjustment operation meets the user's needs, and improves the user experience. Attached Figure Description
[0024] Figure 1 This is a system structure block diagram of Embodiment 1 of a music playback control system based on smartwatch gesture recognition; Figure 2 This is a flowchart of Embodiment 2 of a music playback control method based on smartwatch gesture recognition. Detailed Implementation
[0025] The following detailed description illustrates the specific implementation method: Example 1 This invention relates to a music playback control system based on smartwatch gesture recognition, as shown in the attached figure. Figure 1 As shown, the smartwatch, equipped with a three-axis accelerometer, built-in microphone, speaker, and main control processing unit, consists of a data acquisition module, a playback control module, and an instruction execution module interconnected by hardware. The data acquisition module communicates with the playback control module via serial port, and the playback control module is connected to the instruction execution module via bus. The specific implementation process is as follows: Step 1: Data Collection The data acquisition module has a built-in three-axis accelerometer with a preset sampling frequency of 50Hz. It continuously collects raw motion acceleration data generated by the user's wrist swing, hand raising, and daily walking and running in real time. The continuous time-series raw motion data is then encapsulated in real time and sent to the playback control module's buffer unit. At the same time, the data acquisition module has a built-in dual microphone. One microphone collects the sound data from the watch's speaker, and the other microphone collects the sound data from the user's ear. The two audio sampling frequencies are the same, and the data is sent to the playback control module synchronously.
[0026] This step is completed collaboratively by the motion data acquisition subunit and the audio data acquisition subunit in the data acquisition module. The two subunits share the same system clock inside the smartwatch as the time reference, ensuring that the timestamps of the motion data and audio data are aligned from the same source, providing a unified timing basis for subsequent feature extraction and distance calculation.
[0027] The data acquisition module incorporates a MEMS triaxial accelerometer, which is soldered to the smartwatch's main control PCB board and communicates with the playback control module's main control chip via an I2C serial bus. The sensor's range is preset to ±8g, and its resolution is configured to 16-bit, accurately capturing the acceleration changes of the wrist along the X, Y, and Z spatial axes. In this embodiment, the sensor's sampling frequency is preset to 50Hz. This sampling frequency fully covers the effective frequency band of human wrist movements (0.5–20Hz), accurately capturing the detailed features of sudden movements such as raising and swinging the wrist, while avoiding unnecessary power consumption due to excessive sampling. The sensor incorporates a first-order hardware low-pass filter with a cutoff frequency set to 20Hz, directly filtering out high-frequency noise introduced by the watch's mechanical vibration and electromagnetic interference at the acquisition end, improving the signal-to-noise ratio of the raw data.
[0028] During data acquisition, the triaxial accelerometer continuously samples raw motion acceleration data generated in real time across various scenarios, including wrist swings, natural arm raises, daily walking, jogging arm swings, and deliberate controlled movements. Each set of sampled data is accompanied by a high-precision timestamp generated by the system clock, forming a sequence of raw triaxial acceleration data with a time-series label. The acquired raw motion data is standardized and frame-encapsulated according to the sampling order. Each frame contains a 1-byte frame header, a 4-byte timestamp field, a 6-byte triaxial acceleration data field, and a 1-byte checksum field, with a total frame length of 12 bytes. The encapsulated data is then sent in real time to the ring buffer unit of the playback control module via the SPI serial peripheral interface. This ring buffer unit is designed to hold at least 5 seconds of continuous motion data and employs a read / write pointer separation mechanism. The write pointer cyclically writes new data in the sampling order, while the read pointer allows the playback control module to retrieve data as needed. This avoids data loss due to fluctuations in data transmission rate and ensures that subsequent feature extraction stages can obtain continuous and complete temporal motion data.
[0029] The data acquisition module has two independent sound pickup channels built-in, using a dual-microphone architecture to collect sound data from the watch and the user respectively: The first channel is the watch reference sound pickup channel, which uses an analog MEMS microphone and is installed inside the speaker outlet hole on the bottom of the watch, facing the speaker outlet surface. It is used to directly collect the raw driving sound signal output by the watch speaker as a sound source reference. The second channel is the user ear sound pickup channel, which uses a digital MEMS microphone and is installed on the upper edge of the watch body near the user's body, facing the user's ear. It is used to collect the actual in-ear sound signal around the user's ear, reflecting the sound state that the user actually hears.
[0030] Both microphones are uniformly configured with a 16kHz sampling frequency, 16-bit quantization precision, and mono acquisition mode to ensure complete consistency in sampling dimensions and quantization precision, eliminating system errors introduced by differences in sampling specifications. Audio acquisition startup and shutdown are linked to music playback status: when the smartwatch's music playback function is activated, both microphones are simultaneously awakened and enter acquisition mode; after music playback stops, the microphones automatically enter low-power sleep mode. During acquisition, both audio data streams are appended with corresponding channel identifiers and shared system timestamps, with the watch data marked as channel 0 and the user data marked as channel 1. After marking, the two audio data streams are synchronously transmitted to the dual-channel audio buffer of the playback control module via the I2S audio bus. The audio buffer is independently partitioned by channel number to avoid data confusion. Simultaneously, the playback control module can directly perform preliminary timing matching of the two audio streams through the timestamp field, providing a preliminary data foundation for subsequent time synchronization alignment, amplitude comparison, and phase difference calculation.
[0031] The data acquisition module is equipped with a synchronization control unit that uses the system's global clock as a reference to send synchronous sampling trigger signals to the accelerometer and dual microphones. This ensures that motion data and two audio data streams are acquired at the same time, guaranteeing complete consistency in the time reference of the three types of data. Simultaneously, the module incorporates a low-power management unit. In music playback standby mode, the accelerometer automatically switches to a 10Hz low-power sampling mode, and the microphone channels are completely shut down, maintaining basic motion sensing capabilities while reducing standby power consumption. Upon detecting a music playback start command, each acquisition unit switches to normal operating mode within 100ms, balancing device battery life and real-time functional response.
[0032] Step 2: Time-domain and frequency-domain feature extraction After receiving the raw motion data, the playback control module performs bidirectional feature extraction: during temporal feature extraction, the peak value of the raw motion data is calculated window by window. Valley value mean ,variance During frequency domain feature extraction, the time-domain motion signal is converted into a frequency-domain signal using the Discrete Fourier Transform. Extracting the dominant frequency in the frequency domain Effective bandwidth Spectral amplitude ; By integrating time-domain and frequency-domain parameters, four types of action classification features are generated and stored: action cycle features, action amplitude features, action mutation features, and action duration features.
[0033] This step is executed by the feature extraction and computation unit built into the playback control module. It takes the raw motion data collected and cached in step 1 as input, completes bidirectional feature extraction in the time domain and frequency domain through the sliding analysis window mechanism, and generates four types of action classification features through multi-dimensional parameter fusion, providing standardized feature basis for subsequent action differentiation, benchmark calibration and double confirmation.
[0034] The playback control module first performs vector synthesis on the cached raw three-axis motion acceleration data to obtain the three-axis synthesized acceleration timing data. The calculation formula is as follows: in , , The first The acceleration values at each sampling point along the X, Y, and Z spatial axes, in units of... ; The composite acceleration value for the corresponding sampling point, in units of .
[0035] Subsequently, a sliding analysis window was used to segment the synthesized acceleration time-series data. The length of the sliding analysis window was set to 100 sampling points, corresponding to a duration of 2 seconds at a sampling frequency of 50 Hz; the window sliding step size was set to 20 sampling points, corresponding to a duration of 0.4 seconds. These parameters were set based on the following: the typical duration of a single human music control action is 0.5 seconds to 1.5 seconds; a 2-second window length can completely cover the entire process of a single control action; and a 0.4-second sliding step size ensures 80% overlap between adjacent windows, preventing the omission of action features during window switching, while also balancing device computational efficiency and data continuity.
[0036] Before formal feature extraction, the synthetic acceleration data within a single window is preprocessed to remove outliers using the 3σ criterion: the mean and standard deviation of the synthetic acceleration within the single window are calculated, and sampling points whose values exceed the mean ± 3 times the standard deviation are identified as abnormal interference points and replaced with the linear interpolation results of adjacent sampling points. This processing can effectively filter out abnormal data generated by sudden electromagnetic interference from the sensor, instantaneous wrist collisions, etc., improving the reliability of subsequent feature extraction results.
[0037] For the preprocessed single-window synthetic acceleration data, four time-domain statistical characteristics are calculated window by window. The calculation method and physical meaning of each characteristic are as follows: Peak motion : Retrieves the maximum value of the composite acceleration within the window, in units of This is used to characterize the maximum acceleration impact intensity during an action, corresponding to the maximum force exerted by the wrist swing. For example, when a user quickly raises their hand to perform a song-changing operation, the peak value of the action within the window can reach 15. The above is true, while the peak arm swing motion during normal walking is typically 8-12. .
[0038] Motion trough : Take the minimum value of the composite acceleration within the window, in units of It is used to characterize the ultimate acceleration during the fall-off process of an action, and to help divide the complete swing range of an action, distinguishing between complete control actions and local micro-movements.
[0039] Mean of movement : Take the arithmetic mean of all synthetic acceleration sampling points within the window. The calculation formula is: in The total number of sampling points within the window, with the average action value in units of... It is used to characterize the overall acceleration benchmark level of the action within the window, and can help distinguish different motion intensities such as standing still, walking slowly, and running fast.
[0040] Action variance The formula for calculating the dispersion of the composite acceleration relative to the mean within the window is as follows: The unit of action variance is Variance is used to characterize the intensity and fluctuation of an action; periodic steady motion has a lower variance and smaller fluctuation, while abrupt control motion has a higher variance and faster change.
[0041] For the synthesized acceleration data within the same window, a mean-removing process is first performed to eliminate the interference of the DC component on the spectrum analysis, resulting in a zero-mean time-domain signal. Then, the time-domain motion signal is converted to the frequency domain using a Discrete Fourier Transform (DFT) to obtain the corresponding frequency-domain complex signal. The formula for calculating the DFT is: in The first one after removing the mean Each time-domain sampling point, in units of ; This represents the total number of sampling points within the window. The frequency domain spectral line number has a value range of 100. ; For the first The frequency domain complex values corresponding to each spectral line.
[0042] Based on the frequency domain signal obtained by transformation, three core frequency domain features are extracted: Frequency domain dominant frequency Take the spectral amplitude The frequency value corresponding to the largest spectral line, in units of The calculation formula is: ,in This is the spectral line number corresponding to the maximum amplitude. The original data sampling frequency is 50Hz in this embodiment. The dominant frequency reflects the dominant repetition frequency of the action. For example, the dominant frequency is stable at 1-2Hz when walking at a constant speed, and at 2-3Hz when jogging. However, the control action with a single sudden change has no stable dominant frequency, and the spectral energy distribution is scattered.
[0043] Effective bandwidth Take the frequency spectrum amplitude as the value decreases to the main frequency amplitude. When, the difference between the corresponding high-frequency cutoff frequency and the low-frequency cutoff frequency, in units of Effective bandwidth reflects the concentration of frequency components in an action; periodic motions have narrower bandwidths, while abrupt control actions have wider bandwidths.
[0044] Spectral amplitude : for frequency domain complex signals The modulus, in units of The frequency spectrum amplitude represents the energy intensity of the corresponding frequency component; the amplitude of the spectrum corresponding to the dominant frequency can reflect the energy magnitude of the dominant motion component and help determine the intensity level of the action.
[0045] The playback control module fuses the aforementioned time-domain and frequency-domain features in multiple dimensions to generate four types of action classification features: action period features, action amplitude features, action abrupt change features, and action duration features. These features are then written to the feature cache for subsequent modules to access. Action cycle characteristics: based on the dominant frequency in the frequency domain The core calculation basis is the period value. Simultaneously, cross-validation is performed using the time interval between adjacent peaks in the time domain to characterize the repetitive periodicity of the action; when there is no stable main frequency within the window, it is marked as a non-periodic action.
[0046] Movement amplitude characteristics: The difference between the peak and trough values in the time domain is used as the core, and the total energy of the spectrum is combined for weighted correction to characterize the overall swing amplitude of the wrist movement.
[0047] Action mutation characteristics: The core features are the rate of change of the mean and variance of actions in adjacent windows, combined with the instantaneous rate of change of acceleration in the time domain. Calculations are performed to characterize the degree of instantaneous change in the action.
[0048] Action duration characteristic: The actual duration of the action is obtained by counting the number of windows that continuously meet the preset amplitude threshold, multiplying it by the duration corresponding to the window sliding step size, which represents the time duration attribute of the action.
[0049] This step fully discloses the entire feature extraction process by clearly defining the sliding window parameters, preprocessing rules, and the specific calculation methods and fusion logic for time-domain and frequency-domain features. Those skilled in the art can unambiguously reproduce the entire computation process based on this content. Joint feature extraction in the time and frequency domains can comprehensively characterize action attributes from both time and frequency dimensions, effectively compensating for the blind spots of single-dimensional features. The four types of action classification features generated by fusion have clear boundaries and explicit physical meanings, providing a precise and reliable data foundation for subsequent mid-level feature generation, action differentiation, and dual verification, significantly improving the accuracy and anti-interference capability of subsequent action recognition.
[0050] Step 3: Refined generation of four types of mid-level features When generating step features, instantaneous change rates are identified using action abrupt change features. Identify cycle duration based on action cycle characteristics ;like If a predetermined mutation threshold is set and there is no fixed repetition period, the action is identified as a sudden change action like raising a hand to switch songs; otherwise, it is identified as a continuous walking or swaying action, and a quantified step feature score is output. When generating persistent features, if the duration of the action is... Preset duration threshold, swing amplitude A preset amplitude threshold is used to determine the effective trigger duration, generating a continuous feature; otherwise, it is judged as micro-motion interference and marked as invalid duration. When generating correlation features, daily running cycles with fixed patterns and stable amplitudes are judged as continuous behaviors without control intent; single, irregular hand raises or slight wrist shakes are judged as continuous behaviors with control intent, and correlation features are generated accordingly. When generating stability features, instantaneous jerks and discontinuous micro-motions with a duration of less than 0.2 seconds are identified and uniformly judged as invalid interference actions and discarded; the remaining valid data are used to generate stability features.
[0051] This step is executed by the built-in mid-level feature generation unit of the playback control module. It takes the four basic classification features output from step 2—motion cycle features, motion amplitude features, motion abrupt change features, and motion continuity features—as input. Through feature cross-combination, threshold judgment, and quantization mapping, it generates four types of quantized mid-level features: step feature, continuity feature, correlation feature, and stability feature. This provides multi-dimensional judgment criteria for subsequent double confirmation of actions. All mid-level features are normalized to the [0,1] value range; higher scores indicate stronger corresponding attributes.
[0052] Step characteristics are used to characterize the degree of instantaneous change and periodicity of an action, and are calculated by cross-calculation of action change characteristics and action periodicity characteristics. First, based on the action amplitude characteristics of adjacent sliding windows, the instantaneous rate of change of the action is calculated using the following formula: In the formula, , The first The, the The feature value of the motion amplitude corresponding to each sliding window, in units of ; , These are the center timestamps of the corresponding windows, in units of ; The instantaneous rate of change of the action, in units of .
[0053] The preset mutation threshold is set to 12. Simultaneously, the action attribute is determined by combining periodic stability: the period duration of three consecutive sliding windows is statistically analyzed. If the relative deviation between adjacent cycles does not exceed 15%, it is judged as a continuous action with a fixed repetition cycle; if the relative deviation exceeds 30% and there is no stable main frequency, it is judged as a sudden action without a fixed repetition cycle. When the instantaneous rate of change of the action is greater than the preset sudden change threshold and the action has no fixed repetition cycle, it is judged as a sudden action like raising your hand to change songs, and the corresponding step feature output is high; otherwise, it is judged as a continuous action like walking or swinging, and the corresponding step feature output is low. For example, when a user quickly raises their hand to change songs, the instantaneous rate of change can reach 18. Furthermore, it lacks a stable repetition period, with a step characteristic score of approximately 0.85; when the user walks at a constant speed while swinging the arm, the instantaneous rate of change is approximately 6. The period is stable at around 1.2s, and the step characteristic score is about 0.2.
[0054] Persistence features are used to characterize the effectiveness of action triggering. They are determined by cross-judgment of action duration and amplitude features, and are used to distinguish between effectively controlled actions and unconscious micro-movement interference. Action duration... It is directly obtained from the action duration characteristics, and is the cumulative duration of actions that continuously meet the minimum amplitude requirement, in units of... ; swing amplitude Take the average value of the amplitude characteristics of the movement within the corresponding time period, in units of... .
[0055] The preset duration threshold is set to 0.3 seconds, and the preset amplitude threshold is set to 5 seconds. When the duration of the action ≥Preset duration threshold, and swing amplitude When the duration is greater than or equal to the preset amplitude threshold, it is determined to be a valid trigger duration, and the continuous feature outputs a high score that is positively correlated with both duration and amplitude. If either of these values fails to meet the threshold requirement, it is determined to be an invalid trigger duration, and the continuous feature outputs a low score. For example, when a user deliberately shakes their wrist to perform a pause operation, the action lasts for 0.6 seconds, and the average amplitude is 9. The duration was determined to be valid, with a persistence score of approximately 0.75; during involuntary wrist muscle twitching, the duration was 0.12 seconds and the amplitude was 2.5 seconds. The duration was determined to be an invalid trigger duration, with a persistence feature score of approximately 0.1.
[0056] The correlation feature, used to characterize the degree of correlation between continuous actions and control intentions, is calculated jointly by action periodicity and action amplitude features. It distinguishes between continuous behaviors with control intentions and natural movements without control intentions. This step uses the coefficient of variation to quantify the stability of periodicity and amplitude. The formula for calculating the periodicity coefficient of variation is: In the formula, The standard deviation of the periodic characteristics of multiple consecutive window actions. This is the average value for the corresponding period. This is the periodic variation coefficient, which is dimensionless. Similarly, the amplitude variation coefficient can be calculated. .
[0057] The preset thresholds for the periodic coefficient of variation (PCV) and amplitude PCV are 15% and 20%, respectively. When both the PCV and amplitude PCV values for continuous actions are less than or equal to the thresholds, the behavior is considered non-controlled (e.g., daily running, walking, arm swinging), and the correlation feature score is low. Conversely, if the PCV has no fixed pattern and the amplitude fluctuations exceed the threshold range, the behavior is considered controlled (e.g., continuously shaking the wrist to adjust volume), and the correlation feature score is high. For example, when a user is jogging at a steady pace, the PCV value is approximately 10% and the amplitude PCV value is approximately 13%, indicating non-controlled behavior, and the correlation feature score is approximately 0.15. When a user shakes their wrist horizontally twice to switch songs, the PCV has no stable pattern, the amplitude fluctuates significantly, and the PCV value exceeds 40%, indicating controlled behavior, and the correlation feature score is approximately 0.82.
[0058] Stability features are used to eliminate unstable and invalid interference actions. They are determined jointly by action abrupt change features and action duration features, filtering out invalid data such as momentary jitter and short-term interference. A short-term interference duration threshold of 0.2s is set. If an action has an extremely high degree of momentary abrupt change but a duration of less than 0.2s, and there is no continuous trend between the preceding and following actions, it is judged as an unstable and invalid interference action of momentary jitter type and directly removed from the valid action sequence. For the remaining continuous valid actions, the smoothness of the action amplitude is further calculated; the higher the smoothness, the higher the stability feature score. For example, the momentary jitter caused by a watch accidentally touching a table, lasting 0.08s, has an extremely high degree of abrupt change and is judged as an invalid interference action and removed. A control action completed by a user smoothly raising their hand, lasting 0.5s with a gradual change in amplitude, has a stability feature score of approximately 0.9.
[0059] The four types of mid-level features generated in this step together form a feature vector, which is written to a dedicated cache for subsequent primary confirmation, benchmark calibration, and secondary confirmation stages. All judgment thresholds in this step are determined based on statistical analysis of action samples from 100 users of different age groups, covering typical scenarios such as walking, jogging, hand-raising operations, and unintentional shaking. The threshold selection balances recognition sensitivity and anti-interference capabilities, ensuring accurate recognition of normal control actions while filtering out most natural movements and unintentional interference, adapting to the complex scenarios of smartwatch wearable use, and conforming to the general laws of human kinematics.
[0060] This step fully discloses the entire mapping process from basic features to intermediate features by clarifying the quantification calculation methods, threshold determination logic, and parameter setting basis for the four types of intermediate features. Those skilled in the art can directly reproduce the feature generation logic based on this process. Through dual-feature cross-judgment, the shortcomings of insufficient discrimination of single features are effectively compensated for. It comprehensively characterizes action attributes from four dimensions: mutation attributes, trigger validity, intent relevance, and data stability. It accurately filters out various interferences such as micro-movements, jitters, and periodic movements, providing highly discriminative feature inputs for the subsequent dual confirmation mechanism, significantly reducing the probability of action misjudgment, and improving the overall recognition accuracy and anti-interference capability of the system.
[0061] Step 4: First Validation, Benchmark Building, Calibration and Second Validation The playback control module first integrates four types of features to complete an initial confirmation: distinguishing between uncontrolled movement actions such as walking and running and controlled actions such as skipping songs and pausing; using the daily walking movements identified in the initial confirmation as a sample set to construct a global motion state benchmark library; dynamically adjusting the judgment thresholds of the four types of features using the benchmark library, with the threshold increasing synchronously as the movement amplitude increases and decreasing adaptively as the movement amplitude decreases; and then completing a secondary confirmation based on the calibrated feature parameters to distinguish between the two types of actions again.
[0062] This step is executed by the action confirmation and calibration unit built into the playback control module. It takes the four types of quantified mid-level features generated in step 3—step features, continuous features, correlation features, and stability features—as input and executes according to the closed-loop logic of "first preliminary confirmation—building motion state benchmark—dynamic calibration of judgment threshold—second precise confirmation" to complete the two-level judgment of motion actions without control intention and control actions with control intention, thus solving the technical problem of poor adaptability of fixed thresholds in different motion scenarios.
[0063] Initially, a preliminary action classification based on an initial fixed threshold is established. A weighted feature fusion algorithm is then used to distinguish between the two types of actions. First, preset weights are assigned to the four mid-level features, with a total weight of 1. The weights for step features are 0.3, for persistence features 0.25, for correlation features 0.25, and for stability features 0.2. The weight allocation is based on the contribution of each feature to the recognition of control intent, with mutation attributes and trigger validity having a higher weight in intent determination.
[0064] The formula for calculating the overall confidence level of an action is: In the formula, The overall confidence level for the action is defined, and its value range is [value range missing]. Dimensionless; to The weight coefficients for the four types of features are listed in order. to The quantitative scores are, in order, for step characteristics, persistence characteristics, correlation characteristics, and stability characteristics.
[0065] The initial classification threshold is preset to 0.6, and the overall confidence level is... When the overall confidence level is high, it is determined to be a control action with control intent; When a user quickly raises their hand to switch songs, the scores for the four features are 0.85, 0.75, 0.82, and 0.9, respectively, resulting in a comprehensive confidence score of approximately 0.83, which is considered a controlled action with an intention to control. When a user walks at a constant speed and swings their arm, the scores for the four features are 0.2, 0.3, 0.15, and 0.85, respectively, resulting in a comprehensive confidence score of approximately 0.35, which is considered a movement without an intention to control.
[0066] Using the feature data corresponding to movements continuously determined to be without control intent in a single confirmation as a sample set, a current motion state benchmark library is constructed. The benchmark library storage unit has a built-in circular cache structure to store valid motion samples within the last 10 seconds. The sample data includes three core parameters for each set of movements: four types of mid-level feature scores, average movement amplitude, and average movement cycle.
[0067] The benchmark library is dynamically updated using a sliding window mechanism: whenever a new set of confirmed, uncontrolled movement samples is added, the oldest set of samples is automatically removed, and the mean and standard deviation of each feature of all samples in the benchmark library are recalculated to update the current movement intensity level. The movement intensity level is represented by the average value of the movement amplitude feature; the larger the amplitude, the higher the movement intensity. For example, when a user maintains a constant walking speed, the average movement amplitude in the benchmark library remains stable at [value missing]. Around 10%, the periodic characteristic fluctuation is less than 10%; when the user switches to jogging mode, the benchmark library completes parameter updates within 3-5 sliding window periods, and the average motion amplitude rises to The corresponding exercise intensity level will increase accordingly.
[0068] Based on the motion intensity parameters of the current motion state benchmark library, the initial judgment thresholds of four types of mid-level features are adaptively calibrated. The calibration rule is: the larger the motion amplitude, the feature judgment threshold is raised synchronously; the smaller the motion amplitude, the judgment threshold is adaptively lowered, so as to avoid the high amplitude and high abruptness of natural arm swing under violent motion being misjudged as control action.
[0069] The formula for calculating the threshold correction factor is: In the formula, This is the threshold correction coefficient, which is dimensionless. This is the intensity adjustment coefficient, with a value of 0.2. It is dimensionless and is used to control the rate at which the threshold increases with the intensity of the exercise, so as to avoid the control action being missed due to the threshold being too high. The average value of the current motion amplitude in the benchmark library, in units of ; The reference amplitude for low-intensity exercise is set to a value of [value missing]. This corresponds to the typical range of motion during a slow walk.
[0070] The formula for calculating the calibrated decision threshold is: In the formula, This is the initial threshold for judging the corresponding feature. This is the calibrated threshold. This calibration rule simultaneously applies to the mutation rate threshold of step characteristics, the duration and amplitude thresholds of persistence characteristics, the coefficient of variation threshold of correlation characteristics, and the duration threshold of stability characteristics, achieving synchronous adaptive adjustment of thresholds across all dimensions. For example, when a user is running fast, the average amplitude of their current movement is... The calculated correction coefficient is approximately 1.6, and the initial mutation threshold is... After calibration, approximately At this point, the mutation rate of the natural swing arm is approximately This will not trigger a judgment, while the mutation rate of deliberately raising one's hand can reach [a certain rate]. It can be accurately identified.
[0071] Threshold setting basis: The initial threshold and calibration coefficient in this step are determined based on statistical analysis of exercise samples from 120 different users, covering five typical wearable scenarios: sitting, walking slowly, walking briskly, jogging, and running briskly. The initial threshold is optimized for low-intensity scenarios to ensure that slight controlled movements can be effectively recognized; the calibration coefficient matches the natural growth pattern of movement amplitude and abrupt change when human exercise intensity increases, controlling the threshold increase within a reasonable range, balancing recognition sensitivity and anti-interference, and adapting to the full-scenario use needs of smartwatches.
[0072] After threshold calibration is completed, the action confirmation and calibration unit re-determines the action classification features of the current analysis window based on the calibrated judgment threshold, and generates calibrated step features, continuous features, correlation features and stability features; then, using the weighted fusion algorithm that is completely consistent with the first confirmation, the action comprehensive confidence is recalculated to complete the second confirmation, and the final classification results of motion actions without control intent and control actions with control intent are output.
[0073] The core function of secondary confirmation is to adapt the judgment threshold to the user's current actual movement state and eliminate the interference of changes in movement intensity on action classification. For example, when a user is jogging, primary confirmation may misjudge natural arm swing as a suspected controlled movement due to a large arm swing amplitude; after benchmark calibration to raise the judgment threshold, secondary confirmation can correctly judge it as a movement without controlled intent, effectively reducing the probability of misjudgment.
[0074] This step fully discloses the entire implementation logic of two-level confirmation and dynamic threshold calibration, clarifying the weighted fusion algorithm, benchmark library update mechanism, threshold calibration formula, and parameter values. Those skilled in the art can unambiguously reproduce the entire calculation process. Through a closed-loop mechanism of "initial judgment + benchmark calibration + secondary judgment," the technical problems of poor adaptability and high misjudgment rate of traditional fixed threshold schemes under different motion intensities are solved. This ensures the sensitivity of recognizing slight control actions in low-intensity scenarios while effectively filtering out interference and false triggers caused by vigorous motion, significantly improving the accuracy of action intent recognition and scene adaptability, and providing a reliable basis for the accurate generation of subsequent playback control commands.
[0075] Step 5: Comparison and Verification and Playback Command Generation The system compares the action classification results of the first and second confirmations. If the consistency between the two classifications reaches the preset similarity threshold, it is determined that the control requirements are met. The system has a pre-stored action mapping relationship: raising the wrist = play / pause, shaking the wrist to the left = previous track, and shaking the wrist to the right = next track. The system extracts the amplitude, direction, and duration parameters of the effective control action, matches them with the preset action library, generates the corresponding playback control command, and sends it to the command execution module. If the verification fails, the system automatically reverts to the feature extraction step for recalculation.
[0076] This step is executed by the instruction generation and verification unit built into the playback control module. It takes the first and second confirmation action classification results output from step 4 and completes the consistency comparison and verification of the two confirmation results, template matching of effective control actions, generation and issuance of playback control instructions in sequence. At the same time, a rollback mechanism is set up for verification failures, forming a complete instruction generation closed loop of "verification - matching - generation - rollback". It is the core link from action intent recognition to terminal control execution.
[0077] The instruction generation and verification unit first retrieves the first and second confirmation results corresponding to the same analysis window, and compares them from two dimensions: category consistency and confidence bias, calculating the overall similarity between the two confirmations. Category consistency refers to whether the classification labels for "movements without control intent" and "controlled actions with control intent" are consistent between the two confirmations; confidence bias refers to the relative deviation of the overall confidence of the actions output by the two confirmations, calculated using the following formula: In the formula, The relative deviation of the confidence levels between two confirmations is dimensionless. The overall confidence level of the action obtained from a single confirmation. The combined confidence level of the action obtained from the secondary confirmation, and the range of values for both is [missing information]. .
[0078] The preset similarity threshold is set as follows: the classification labels are consistent and the relative confidence deviation is ≤15%. When both of the above conditions are met in two confirmations, the preset control requirements are met, the confirmation result is valid, and the process can proceed to the subsequent action matching and instruction generation stage. If the classification labels are inconsistent, or the relative confidence deviation exceeds 15%, the control requirements are not met, and the confirmation result is invalid. For example: if a window action is confirmed as a control action with control intent in the first confirmation, with a comprehensive confidence of 0.83, and is also confirmed as a control action with control intent in the second confirmation, with a comprehensive confidence of 0.80, the calculated relative confidence deviation is approximately 3.6%, which meets the threshold requirement, and the verification is deemed successful. If the first confirmation is a control action and the second confirmation is a motion action, the classification labels are inconsistent, and the verification is directly deemed unsuccessful.
[0079] The playback control module's built-in non-volatile storage unit pre-stores a one-to-one mapping relationship between standardized control actions and playback control commands, forming a preset action template library. This library is generated based on standard control action samples from 100 users of different genders and age groups. Each action template includes four core parameters: three-axis acceleration waveform characteristics, typical amplitude range, dominant motion direction, and standard duration range. The specific mapping relationships are as follows: Slight wrist lift: The dominant direction is a sudden change in acceleration along the positive Y-axis, with a range of motion of 6-12. The duration of the action is between 0.4 and 0.8 seconds, corresponding to the play / pause control command; Left wrist swing motion: The dominant direction is the peak acceleration in the negative X-axis direction, and the range of motion is 8-15. The duration of the action is between 0.3 and 0.6 seconds, corresponding to the previous track switching control command; Right wrist swing motion: The dominant direction is the peak acceleration along the positive X-axis, and the range of motion is 8-15. The duration of the action is between 0.3 and 0.6 seconds, corresponding to the control command for switching to the next track.
[0080] Each action template has an independent matching threshold, which is used to determine the similarity in the subsequent action matching process, ensuring the distinction between different action types.
[0081] Once the verification passes and the action is determined to be a control action with intent, the instruction generation and verification unit extracts three core parameters of the control action from the original motion data and action classification features: the first is the action amplitude parameter, which is the peak value of the triaxial composite acceleration during the action period, in units of... Second, the motion direction parameter, which determines the dominant motion direction based on the extreme directions and component proportions of the X, Y, and Z axis accelerations; third, the motion duration parameter, which is the time difference between the starting trigger point and the ending fall point of the motion, in seconds.
[0082] The three extracted parameters are then matched one by one with the three types of templates in the preset action template library. The overall matching degree between the action to be matched and each type of template is calculated using the following formula: In the formula, For overall matching degree, the value range is: Dimensionless; The amplitude of the action to be matched. This is the standard amplitude for the template; all units are... ; The direction matching degree is 1 when the dominant directions are consistent and 0 when they are inconsistent; The duration of the action to be matched. The template standard duration is in seconds. , , These are weighting coefficients, with values of 0.4, 0.35, and 0.25 respectively, and a total weight of 1.
[0083] The result with the highest overall matching degree among the three template types is selected. If the highest matching degree is ≥0.7, the match is considered successful, and the playback control command type corresponding to the control action is determined. If the highest matching degree is <0.7, the match is considered unsuccessful, no command is generated, and the process returns to the feature extraction stage. For example: the peak value of the positive X-axis acceleration for a certain control action is 12. The duration is 0.5 seconds. The matching degree calculated with the right wrist sway template is 0.88, which exceeds the matching threshold of 0.7. Therefore, it is determined to be the next track switching instruction.
[0084] After successful matching and determination of the instruction type, the instruction generation and verification unit generates standardized playback control instruction frames. These frames use a fixed-length format, containing a 1-byte frame header identifier, a 1-byte instruction type code, a 2-byte instruction sequence number, and a 1-byte cyclic redundancy check code, with a total frame length of 5 bytes. The instruction type code is defined as follows: 0x01 corresponds to a play / pause instruction, 0x02 to a previous track instruction, and 0x03 to a next track instruction. The instruction sequence number increments sequentially according to the generation order and is used for deduplication and verification by the instruction execution module.
[0085] The generated instruction frame is sent to the instruction execution module in real time via the serial peripheral interface bus inside the smartwatch. Simultaneously, the instruction generation and verification unit records the type, timestamp, and sequence number of this instruction. The module has a built-in instruction debouncing mechanism: the same type of playback control instruction is sent only once within a 1-second interval, preventing multiple triggering of instructions due to window swiping and ensuring the stability of the operation.
[0086] If the comparison of the two confirmation results does not meet the preset control requirements, or if the control action fails to match the preset template, the recognition is deemed invalid and no playback control command is generated. At the same time, the data in the current analysis window is marked as data to be reviewed, and the system automatically reverts to the time domain feature extraction and frequency domain feature extraction steps in step 2. The system then re-executes the full-process recognition operation on the original motion data of the next set of sliding windows, forming a closed-loop fault tolerance mechanism to avoid misoperation caused by erroneous command issuance.
[0087] Threshold setting basis: The similarity threshold and matching threshold in this step are determined based on the test statistics of a large number of human action samples. The 15% confidence deviation threshold takes into account both the reliability of the judgment and the response efficiency. It filters out the occasional error of a single judgment through double verification and avoids operation delay caused by over-verification. The 0.7 action matching threshold can ensure that more than 95% of standard control actions are accurately identified, while filtering out the vast majority of similar interference actions, which is suitable for the actual needs of smartwatch wearable interaction scenarios.
[0088] This step fully discloses the complete process details of two result comparisons, action template matching, instruction generation and issuance, and rollback fault tolerance. It clarifies the calculation methods of various parameters, instruction frame structure, and threshold setting basis. Those skilled in the art can unambiguously reproduce all aspects of instruction generation based on this content. Through consistency verification of double confirmation results, errors that may occur in a single judgment are further filtered out, improving the accuracy of instruction generation. The standardized action mapping library and multi-parameter weighted matching mechanism ensure accurate differentiation of different control actions, effectively avoiding instruction confusion. The rollback mechanism for failed verification and the anti-jitter mechanism form a complete fault tolerance protection, significantly reducing the probability of misoperation and improving the reliability of the entire music control system and the user interaction experience.
[0089] Step 6: Audio Acquisition, Distance Calculation, and Volume Compensation The playback control module extracts the timestamps of the two audio streams, aligns them to eliminate time delay deviations, calculates the amplitude and phase differences between the two audio streams, and fits the initial distance using the speed of sound. A moving average algorithm is used to remove outlier distance points, smoothing the real-time distance change data. Based on an airborne acoustic attenuation model, a volume compensation coefficient is calculated according to the distance attenuation law, and this coefficient is used to generate compensation data to offset volume fluctuations. Multiple volume adjustment gradients are preset, and the compensation amount is verified to not exceed the watch's 0-100 volume range. Once compliant, a volume adjustment command is generated and issued. The command execution module simultaneously receives the playback control command and the volume adjustment command, executing music switching playback and real-time adaptive volume adjustment operations respectively.
[0090] This step is executed by the audio processing and volume compensation unit built into the playback control module. Taking the audio data from the watch and the user terminal collected in step 1 as input, it sequentially completes timing synchronization alignment, distance calculation, data smoothing, acoustic compensation calculation, and instruction verification and generation. Finally, the instruction execution module synchronously completes the playback control and volume adjustment operations, realizing adaptive and stable in-ear volume under changes in wrist position, and solving the problem of fluctuating volume caused by changes in the distance between the speaker and the ear due to the user raising their hand or swinging their arm.
[0091] The audio processing and volume compensation unit first extracts the timestamps of the same source system attached to the two audio data to initially align the time reference. Since the hardware conversion delay of the two microphones and the difference in acoustic propagation paths will introduce fixed and time-varying timing deviations, a cross-correlation algorithm is used to achieve accurate time synchronization alignment.
[0092] The formula for calculating the cross-correlation function of two discrete audio signals is: In the formula, This is the audio signal sequence from the watch. The user-side audio signal sequence. For relative time delay, This represents the total number of sampling points in the analyzed frame. When taking the maximum value, the corresponding This is the optimal time delay difference between the two signals. The unit is s.
[0093] After calculating the time delay difference, the user's audio signal is shifted along the time axis. This ensures that the characteristic waveforms of the two signals correspond perfectly in the time dimension, eliminating timing deviations. For example, if the propagation delay of the user-end signal relative to the watch-end signal is found to be 0.6ms after cross-correlation calculation, the user-end signal is shifted forward by 0.6ms to complete timing alignment and ensure that the time base for subsequent amplitude and phase calculations is consistent.
[0094] After timing alignment, amplitude analysis and phase analysis are performed on the two audio data in the same analysis frame, and the initial distance from the watch speaker to the user's ear is jointly fitted.
[0095] Amplitude difference calculation: Calculate the sound pressure level of the two signals separately. The formula for calculating the sound pressure level is: In the formula, This is the effective value of the signal sound pressure, in Pa. For reference sound pressure level, the value is taken as follows. ; Sound pressure level, measured in dB.
[0096] The sound pressure level difference between the two signals is It represents the amplitude attenuation during sound propagation.
[0097] Phase difference calculation: Perform Discrete Fourier Transform on the two signals respectively to extract the main frequency of the music signal. Corresponding phase value (Watch end) (User end) Calculate phase difference The unit is rad.
[0098] Initial distance fitting: combining the air speed of sound under normal temperature and pressure The distance is calculated based on the relationship between phase difference and propagation time, using the following formula: Simultaneously, the distance verification value is calculated based on the spherical wave propagation law of amplitude attenuation. The phase method distance and amplitude method distance are then weighted and fused with a 7:3 weight to obtain the final initial distance data, in meters. For example: taking the main frequency of music... The calculated phase difference is approximately 3.7 rad. Substituting this into the formula, the initial distance is approximately 0.2 m, which is consistent with the result calculated using the amplitude method, and is therefore determined to be a valid initial distance value. The initial distance data is calculated and output every 10 ms, forming a continuous initial distance sequence.
[0099] The initial distance sequence is susceptible to abnormal jumps caused by ambient reflected sound and background noise. Therefore, outlier removal and sliding smoothing are performed sequentially to obtain stable real-time distance change data.
[0100] First, the 3σ criterion is used to identify outliers: 10 consecutive initial distance points are taken as a statistical window, and the mean distance within the window is calculated. with standard deviation The value exceeds Points within the specified range are identified as abnormal interference points, and their replacements are made using the linear interpolation results of the two effective distance points before and after them.
[0101] The corrected distance sequence was then smoothed using a moving average algorithm, with a sliding window length of 5 distance points corresponding to a duration of 50ms. The formula for calculating the smoothed real-time distance is as follows: In the formula, For the first A corrected initial distance value, For the first The smoothed real-time distance value of a point. For example, if the initial distance sequence within a window is 0.21m, 0.22m, 0.78m (abnormal), 0.20m, and 0.23m, the abnormal point of 0.78m is removed by the 3σ criterion and interpolated to 0.21m. Then, the smoothed distance is calculated by moving average and is approximately 0.214m, effectively eliminating distance jumps caused by sudden interference.
[0102] This step relies on an airborne acoustic spherical wave attenuation model to calculate volume compensation based on real-time distance changes, thus offsetting the fluctuations in in-ear volume caused by these distance variations. The typical distance for a user wearing a watch is used. The reference distance corresponds to the reference in-ear volume; when the real-time distance is... At this time, the additional attenuation caused by sound propagation is the volume gain that needs to be compensated, and the calculation formula is: In the formula, The required volume gain is expressed in dB; the greater the distance, the greater the required gain.
[0103] Taking into account the volume level characteristics of smartwatches, each volume level corresponds to a gain change of approximately 0.3dB. Therefore, the gain is converted into a level compensation amount: In the formula, This represents the volume level compensation, measured in levels. A positive value increases the volume, while a negative value decreases it. For example, if the real-time distance between the watch and the user's ear becomes 0.3m after the user raises their wrist, the calculated gain is approximately 3.52dB. This corresponds to a compensation level of approximately 12 levels, meaning the volume needs to be increased by 12 levels to compensate for propagation attenuation and ensure the in-ear volume matches the baseline. The final output of real-time volume compensation data includes the target compensation level and adjustment direction, used for subsequent command generation.
[0104] To avoid auditory discomfort caused by sudden volume changes, the system presets multiple volume adjustment levels and performs reasonableness verification and smoothing on the calculated level compensation amounts: Range calibration: Verify the target volume level after compensation to ensure it is within the 0-100 range supported by the watch; if the calculated result is lower than 0, treat it as 0; if it is higher than 100, treat it as 100 to avoid exceeding the hardware adjustment range.
[0105] Gradient verification: Set the maximum single-step adjustment gradient to 5 levels. If the single compensation amount exceeds 5 levels, it will be split into multiple steps for gradual adjustment, with each step adjusting 5 levels and an interval of 200ms between steps, to achieve a smooth volume transition.
[0106] After verification, a standardized volume adjustment command frame is generated. The command frame contains a 1-byte frame header identifier, a 1-byte command type code, a 1-byte target volume level, and a 1-byte cyclic redundancy check code, for a total length of 4 bytes. The generated command is sent to the command execution module in real time via the smartwatch's internal serial peripheral interface bus. For example, if a 12-level volume increase is needed, and the current level is 40, it is broken down into three steps: 45→50→52, with each step spaced 200ms apart, and the three command frames are sent sequentially.
[0107] The instruction execution module has built-in independent playback control instruction channels and volume adjustment instruction channels. These two types of instructions are received in parallel and processed independently without blocking each other. Playback control instructions have the highest response priority, immediately executing playback, pause, and track skipping operations upon receipt. Volume adjustment instructions gradually adjust the speaker's output gain according to preset steps and intervals, achieving seamless adaptive volume adjustment. Simultaneously, the instruction execution module sends the actual volume level and playback status back to the playback control module, forming a feedback loop for subsequent fine-tuning and optimization of the compensation model parameters.
[0108] Threshold setting basis: The parameters such as the sliding window length, single-step adjustment gradient, and reference distance in this step are all determined based on wearable ergonomics and hearing tests. The 5-point sliding window balances the response speed and smoothness of distance tracking, avoiding excessive smoothing that could lead to compensation lag; the 5-level single-step adjustment gradient matches the human ear's hearing comfort threshold, preventing abrupt changes in volume; the 0.2m reference distance corresponds to the typical distance between the watch and the ear during daily wear, ensuring that the compensation reference closely matches actual usage conditions, and the parameter settings balance compensation accuracy and user experience.
[0109] This step fully discloses the entire process of dual-channel audio time synchronization, joint distance calculation, data smoothing filtering, acoustic attenuation compensation, command verification and issuance, and dual command execution. It clarifies the calculation formulas, parameter values, and verification rules for each stage, allowing those skilled in the art to unambiguously reproduce the entire volume adaptive compensation function. By using a distance calculation method combining phase difference and amplitude, the accuracy of distance detection in close-range scenarios is improved. Outlier removal and moving average processing ensure the stability of distance data, avoiding frequent volume fluctuations. The compensation mechanism based on the spherical wave acoustic attenuation model accurately offsets the volume attenuation caused by distance changes. Multi-level gradient verification achieves smooth and imperceptible volume adjustment. Working in conjunction with playback control functions, it significantly improves the music listening experience in scenarios involving movement and raising the hand, enhancing the system's practicality and user-friendliness.
[0110] Example 2 Based on Embodiment 1, this embodiment discloses a music playback control method based on smartwatch gesture recognition, as shown in the attached figure. Figure 2 As shown, the specific implementation process is as follows: Step 1: Real-time data acquisition. The smartwatch's sensor unit continuously acquires raw three-axis motion data of the user's wrist, with a fixed sampling period; simultaneously, it also acquires audio data from the watch's speaker and the user's ear. All acquired data is cached in real time and awaits algorithm calls.
[0111] Step 2, joint time-domain and frequency-domain feature extraction. Divide the cached raw motion data into a fixed-length sliding analysis window, and uniformly calculate the time-domain peak value, mean, and variance parameters within the window; convert the signal to the frequency domain using Fourier transform, and extract the main frequency and spectral amplitude parameters; integrate the calculation results, and standardize the output to output four basic motion classification features: motion period, motion amplitude, motion abrupt change, and motion duration.
[0112] Step 3: Generate four types of mid-level features in a layered manner. Multiple sets of judgment threshold matrices are preset: combining the rate of abrupt change with periodic patterns, abrupt / continuous actions are distinguished, and step characteristics are quantified and output; combining the duration of the action with the swing amplitude threshold, effective / ineffective trigger durations are divided, and continuous characteristics are output; combining periodic regularity with amplitude fluctuation range, human-controlled behavior and natural movement behavior are distinguished, and correlation characteristics are output; short-term, irregular jitter interference data are filtered out, and effective stable action data is retained to generate stability characteristics.
[0113] Step 4: Two-level confirmation and adaptive calibration calculation. The first round of action judgment is performed using four types of mid-level features to distinguish between uncontrolled and intentionally controlled movements. A dynamic motion baseline model is constructed using the natural movements identified in the first round, such as daily walking and jogging. The model stores a table of feature correction coefficients corresponding to different movement intensities. The baseline model is then used to globally calibrate and optimize the thresholds of the four types of features. Based on the calibrated feature parameters, the second round of action recognition and judgment is performed again to update the classification results for both types of actions.
[0114] Step 5: Result Comparison and Verification, and Dual Instruction Generation. The system automatically compares all classification labels and confidence scores from the first and second rounds. If the consistency of the two sets of recognition results meets the preset algorithm threshold, the verification is deemed successful. On the action control side, it matches the preset gesture action library parameters to generate playback control instructions for music playback, pause, and song skipping. On the audio compensation side, it performs time alignment, amplitude difference, and phase difference calculations on the two audio streams, and substitutes these into the acoustic distance formula to fit the real-time distance between the watch and the human ear. After anomaly smoothing correction, it calculates the volume compensation amount based on the acoustic attenuation relationship, and generates a volume adjustment instruction after verifying compliance with the gear selection criteria. If the two recognitions do not meet the preset requirements, the system automatically jumps back to Step 2 and re-executes the feature extraction process.
[0115] Step 6: Command Reception and Execution. The watch system in the background synchronously receives playback control commands and volume adjustment commands, prioritizing the execution of music playback switching operations corresponding to the gestures; it also synchronously reads the volume compensation command parameters and adjusts the speaker output volume in real time to offset the acoustic distance attenuation caused by wrist raising and swinging, ensuring that the volume heard by the human ear remains stable and consistent throughout.
[0116] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A music play control system based on smart watch gesture recognition, characterized in that, A music playback control system based on smartwatch gesture recognition includes: a data acquisition module, a playback control module, and a command execution module; The data acquisition module is used to collect raw motion data of the user's wrist in real time and send it to the playback control module; The playback control module receives raw motion data and performs time-domain and frequency-domain feature extraction to obtain motion classification features including motion cycle features, motion amplitude features, motion abrupt change features, and motion continuity features. Based on the motion abrupt change features and motion cycle features, it distinguishes between abrupt and continuous motions to obtain step characteristics; based on the motion continuity features and motion amplitude features, it distinguishes between effective and ineffective trigger durations to obtain continuity features; based on the motion cycle features and motion amplitude features, it distinguishes between continuous behavior with control intent and continuous behavior without control intent to obtain correlation features; based on the motion abrupt change features and motion continuity features, it eliminates unstable and invalid interference actions to obtain stability features; and then, based on the step characteristics, continuity features, correlation features, and stability features, it performs further analysis. The system first confirms the motion actions without control intent and the control actions with control intent. Based on the motion actions without control intent, it constructs the current motion state benchmark and uses the current motion state benchmark to calibrate the step characteristics, continuous characteristics, correlation characteristics, and stability characteristics. Based on the calibrated step characteristics, continuous characteristics, correlation characteristics, and stability characteristics, it performs a second confirmation to obtain the motion actions without control intent and the control actions with control intent. It then compares the motion actions without control intent and the control actions with control intent obtained from the two confirmations to determine whether both meet the preset control requirements. If yes, it generates and sends a playback control command to the command execution module based on the control action with control intent. If no, it returns to perform time-domain feature extraction and frequency-domain feature extraction. The instruction execution module is used to receive and execute playback control instructions. 2.The smart watch gesture recognition based music play control system of claim 1, wherein, The data acquisition module is also used to acquire user terminal audio data and watch terminal audio data and send them to the playback control module; the playback control module is also used to perform time synchronization alignment processing, amplitude comparison analysis and phase difference calculation on the user terminal audio data and watch terminal audio data to obtain real-time distance change data between the watch terminal and the user terminal, perform correction calculation based on the real-time distance change data to obtain real-time volume compensation data to compensate for volume fluctuations caused by distance changes, and then generate and send volume adjustment commands to the command execution module based on the real-time volume compensation data; the command execution module is also used to receive and execute volume adjustment commands. 3.The smart watch gesture recognition based music play control system of claim 1, wherein, When the playback control module performs time-domain feature extraction and frequency-domain feature extraction, the time-domain feature extraction includes extracting the motion peak, motion valley, motion mean, and motion variance of the original motion data. The frequency-domain feature extraction includes converting the original motion data into a frequency-domain signal through Fourier transform, extracting the main frequency, bandwidth, and spectral amplitude of the frequency-domain signal, and combining the time-domain features and frequency-domain features to obtain motion classification features. 4.The smart watch gesture recognition based music play control system of claim 1, wherein, When the playback control module distinguishes between abrupt and continuous actions based on the abrupt action characteristics and the periodic action characteristics to obtain step characteristics, it identifies the instantaneous change amplitude and rate of change of the action using the abrupt action characteristics and the repetition pattern and periodic duration of the action using the periodic action characteristics. When the instantaneous change amplitude and rate of change of the action exceed the preset abrupt action threshold and there is no obvious repetition period, it is determined to be an abrupt action; otherwise, it is determined to be a continuous action, thereby generating step characteristics that characterize the abrupt action attributes. When distinguishing between effective and ineffective trigger durations based on action duration and action amplitude characteristics to obtain the duration feature, the actual duration of the action is obtained using the action duration feature, and the actual swing amplitude of the action is obtained using the action amplitude feature. When the action duration is not less than a preset duration threshold and the swing amplitude is not less than a preset amplitude threshold, it is determined to be an effective trigger duration; otherwise, it is determined to be an ineffective trigger duration, thus generating a duration feature that characterizes the effectiveness of the action trigger. When distinguishing between continuous behaviors with and without control intent based on the characteristics of the action cycle and the characteristics of the action amplitude to obtain correlation characteristics, the regularity of the action cycle and the stability of the amplitude are combined. When the continuous action cycle has no fixed regularity and the amplitude fluctuation is within the preset range of the control action, it is judged as a continuous behavior with control intent; otherwise, it is judged as a continuous behavior without control intent, thus generating correlation characteristics that characterize the correlation of action intent. When unstable and invalid interference actions are eliminated based on the action mutation features and action continuity features to obtain stability features, irregular instantaneous jitter is identified by action mutation features, and short-term discontinuous actions are identified by action continuity features. Instantaneous jitter and short-term discontinuous actions are judged as unstable and invalid interference actions and eliminated to generate stability features that characterize the stability of the action posture. 5.The smart watch gesture recognition based music play control system of claim 1, wherein, The playback control module performs action confirmation and calibration, including: based on the extracted action classification features combined with step features, continuous features, correlation features and stability features, confirming the motion actions without control intention and the control actions with control intention, and obtaining the initially confirmed motion actions without control intention and control actions with control intention. Based on the initially confirmed motion actions without control intent, a current motion state benchmark is constructed. The step characteristics, persistence characteristics, correlation characteristics, and stability characteristics are dynamically calibrated in combination with the current motion state benchmark, and the judgment thresholds of the step characteristics, persistence characteristics, correlation characteristics, and stability characteristics are adjusted. Based on the calibrated step characteristics, persistence characteristics, correlation characteristics, and stability characteristics, the motion actions without control intent and the controlled actions with control intent are reconfirmed to obtain the reconfirmation results. The system compares the uncontrolled and controlled movements obtained from the first and second confirmations to determine whether both confirmations meet the preset control requirements. If both meet the requirements, the confirmation results are considered valid; otherwise, feature extraction and confirmation are performed again.
6. The music playback control system based on smartwatch posture recognition according to claim 1, characterized in that, The playback control module generates and sends playback control instructions based on control actions with control intent, including: pre-determining the correspondence between preset control actions and playback control instructions; performing action type matching on control actions with control intent that have been confirmed twice and meet preset control requirements, and extracting the action amplitude, action direction, and action duration of the control actions with control intent and comparing them with preset control actions to determine the corresponding playback control instructions.
7. The music playback control system based on smartwatch posture recognition according to claim 2, characterized in that, The playback control module performs time synchronization alignment, amplitude comparison analysis, and phase difference calculation on the user-end audio data and the watch-end audio data to obtain real-time distance change data, and performs correction calculations to obtain real-time volume compensation data. This includes: extracting the timestamps of the user-end audio data and the watch-end audio data, and eliminating timing deviations between the two audio data streams using a timestamp alignment algorithm; performing amplitude comparison analysis on the two audio data streams to calculate the amplitude difference, and simultaneously extracting the phase information of the two audio data streams using Fourier transform to calculate the phase difference, and fitting this with the sound propagation speed to obtain initial distance data between the watch-end and the user-end; removing outliers and fluctuations from the initial distance data, and using a moving average algorithm to smooth and correct the initial distance data to obtain real-time distance change data; and generating real-time volume compensation data based on the sound propagation attenuation law and the correspondence between distance change and volume attenuation.
8. The music playback control system based on smartwatch posture recognition according to claim 2, characterized in that, The playback control module generates and sends volume adjustment commands based on real-time volume compensation data, including: pre-setting a volume adjustment gradient threshold, calculating the required volume adjustment amount based on real-time volume compensation data, verifying the rationality of the volume adjustment amount, eliminating abnormal adjustment amounts that exceed the watch's volume adjustment range, and generating a volume adjustment command.
9. A music playback control method based on smartwatch gesture recognition, characterized in that, include: Step 1: Collect raw motion data of the user's wrist in real time; Step 2: Extract time-domain features and frequency-domain features from the original motion data to obtain motion classification features that include motion cycle features, motion amplitude features, motion abrupt change features, and motion duration features. Step 3: Based on the characteristics of abrupt action change and the characteristics of action cycle, distinguish between abrupt action and continuous action to obtain step characteristics; based on the characteristics of action duration and the characteristics of action amplitude, distinguish between effective trigger duration and ineffective trigger duration to obtain duration characteristics; based on the characteristics of action cycle and the characteristics of action amplitude, distinguish between continuous behavior with control intention and continuous behavior without control intention to obtain correlation characteristics. Based on the characteristics of sudden action change and the characteristics of continuous action, unstable and invalid interfering actions are eliminated to obtain the stability characteristics; Step 4: Based on the step characteristics, persistence characteristics, correlation characteristics, and stability characteristics, a first confirmation is performed to obtain the motion actions without control intention and the control actions with control intention; a current motion state benchmark is constructed based on the motion actions without control intention, and the step characteristics, persistence characteristics, correlation characteristics, and stability characteristics are calibrated using the current motion state benchmark; a second confirmation is performed based on the calibrated four characteristics to obtain the motion actions without control intention and the control actions with control intention. Step 5: Compare the motion actions without control intent and the control actions with control intent obtained from the two confirmations, and determine whether both meet the preset control requirements: If yes, generate and send playback control instructions based on the control actions with control intent; if no, return to step 2 to perform time domain feature extraction and frequency domain feature extraction again. Step 6: Receive and execute playback control commands.
10. The music playback control method based on smartwatch posture recognition according to claim 9, characterized in that, In step 1, audio data from both the user's device and the watch is also collected. In step 5, the user terminal audio data and the watch terminal audio data are also time-synchronized and aligned, amplitude comparison analysis and phase difference calculation are performed to obtain real-time distance change data between the watch terminal and the user terminal. Based on the real-time distance change data, correction calculation is performed to obtain real-time volume compensation data to offset the volume fluctuation caused by the distance change. Then, a volume adjustment command is generated and sent based on the real-time volume compensation data. In step 6, a volume adjustment command is also received and executed.
Citation Information
Patent Citations
Music playing control method, device, storage medium and wearable device
CN109032384A
