AR intelligent glasses control method and system and AR intelligent glasses

Abstract

The invention provides an AR intelligent glasses control method and system and AR intelligent glasses, and the method comprises the steps: obtaining original inclination angle data and accelerometer data from a built-in sensor of the AR intelligent glasses, and generating an original data sequence containing an inclination angle value, an acceleration value and a corresponding time stamp; performing de-noising processing on the inclination angle value in the original data sequence by applying a low-pass filtering algorithm to obtain a filtered inclination angle measurement value and a corresponding time stamp sequence; training an adaptive parameter adjustment model by using a high-sensitivity mode sign or a stability priority mode sign and combining stored user interaction record data in past 30 days, and generating personalized response curve parameters and sensitivity adjustment coefficients; and updating an attitude resolving matrix by using the inclination angle data after deviation adjustment, calculating head three-dimensional space positioning information through a quaternion rotation algorithm in combination with gyroscope data, and generating a final attitude angle data sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly relates to an AR smart glasses control method and system and AR smart glasses. BACKGROUND

[0002] As the core carrier of next-generation human-computer interaction, augmented reality (AR) smart glasses are reshaping the boundaries of digital experience. Their precise pose perception capability directly determines the quality of virtual information fusion with the real world. Tilt control technology, as the foundation of pose perception, plays a decisive role in achieving immersive interactive experience.

[0003] Currently, the tilt control schemes of AR smart glasses on the market generally use fixed parameter configurations and static threshold settings. Such schemes often struggle to make dynamic adjustments in the face of differences in user usage habits (such as head movement amplitude, speed preference) and changing application environments (such as sudden changes in light, motion state switching). Specifically, fixed parameters may lead to response lag in scenarios requiring high sensitivity (such as quickly turning the head to view information), while in scenarios requiring high stability (such as reading quietly), unnecessary interface jitter may occur due to minor disturbances. In addition, traditional tilt measurement systems often lack effective and adaptive error compensation mechanisms, and are easily affected by sensor drift, environmental noise, and cumulative errors under complex dynamic conditions in actual applications, leading to a decline in the reliability of tilt data and further restricting the stability of the entire pose control system and the consistency of user experience.

[0004] The fixed setting of parameter thresholds and response curves for tilt recognition results in the system's inability to adapt to individual differences and changes in user habits. This lack of adaptability further exacerbates the balance between measurement accuracy and response speed, as the system attempts to improve recognition sensitivity, it often amplifies measurement errors caused by various environmental noise and hardware drift. The accumulation and propagation of measurement errors not only affect the reliability of tilt data, but also restrict the stability of the entire pose control system and the consistency of user experience. These interrelated technical bottlenecks form a complex constraint network, making it difficult for traditional single optimization strategies to make breakthrough progress.

[0005] How to build a set of AR smart glasses control technology that can automatically evolve tilt control parameters according to user usage habits and environmental changes, while achieving real-time and adaptive compensation for tilt measurement errors, to eliminate measurement errors caused by various factors and improve response sensitivity and recognition accuracy, has become one of the key issues in promoting AR smart glasses technology breakthroughs. In particular at the algorithm level, designing a control strategy that can dynamically learn user behavior patterns, perceive subtle changes in the environment, and optimize internal model parameters accordingly, is a core challenge in improving the intelligence and practicality of AR glasses. SUMMARY

[0006] The purpose of the present application is to provide an AR smart glasses control method, system and AR smart glasses, aiming to solve the technical problems of fixed parameters in the existing AR smart glasses tilt angle control scheme, inability to adapt to user habits and environmental changes, and lack of effective error compensation mechanism leading to insufficient control accuracy and response sensitivity.

[0007] To achieve the above purpose, the present application provides an AR smart glasses control method, which comprises the following steps: Step S101: Obtain the original tilt angle data and accelerometer data from the sensors built-in in the augmented reality smart glasses, and generate an original data sequence containing the tilt angle value, acceleration value and corresponding time marker.

[0008] Step S102: Apply a low-pass filter algorithm to the tilt angle value in the original data sequence for denoising processing, to obtain the filtered tilt angle measurement value and the corresponding time marker sequence.

[0009] Step S103: Calculate the head motion frequency according to the filtered tilt angle measurement value and the time marker sequence, and if the head motion frequency exceeds the preset threshold of a preset number of times per second (for example, 5 times), generate a high sensitivity mode flag; if the tilt angle change rate is lower than a preset degree per second (for example, 0.5 degrees) and lasts for a preset time length (for example, 5 seconds), generate a stability priority mode flag.

[0010] Step S104: Use the high sensitivity mode flag or the stability priority mode flag, combined with the stored past preset number of days (for example, 30 days) of user interaction record data, to train an adaptive parameter adjustment model, to generate personalized response curve parameters and sensitivity adjustment coefficients.

[0011] Step S105: Fuse the light intensity change rate obtained by the ambient light sensor and the angular velocity change rate obtained by the gyroscope, and if the light intensity change rate exceeds a preset lux per second (for example, 50 lux per second) or the angular velocity change rate exceeds a preset degree per second (for example, 10 degrees per second), adjust the personalized response curve parameters and sensitivity adjustment coefficients to generate environment-adaptive corrected control parameters.

[0012] Step S106: According to the environment-adaptive corrected control parameters, use the Kalman filter algorithm to fuse and calibrate the filtered tilt angle measurement value and the accelerometer data, to generate a predicted tilt angle value, and if the deviation between the predicted tilt angle value and the actual measured value exceeds a preset degree (for example, 2 degrees), activate the error correction process.

[0013] Step S107: In the error correction process, a correction factor is generated, a multi-dimensional error analysis model is applied to the predicted inclination value for real-time deviation adjustment, and if the deviation direction and amplitude of a continuous predetermined number (for example, 5) of sampling points are greater than a preset degree (for example, 1 degree), a reference reset is performed, and inclination data after deviation adjustment is generated.

[0014] Step S108: The inclination data after deviation adjustment is used to update the attitude solution matrix, the three-dimensional space positioning information of the head is calculated by the quaternion rotation algorithm combined with the gyroscope data, and the final attitude angle data sequence is generated.

[0015] Step S109: According to the stability judgment of the final attitude angle data sequence, if the angle change variance is lower than a preset square degree (for example, 0.1 degree square), and the improvement amplitude of the user interaction frequency exceeds a preset percentage (for example, 10%), the learning weight of the adaptive parameter adjustment model is adjusted, the optimized inclination control parameter is generated, and the closed-loop adjustment process is completed.

[0016] On the other hand, the application also provides an AR smart glasses control system, which comprises a data acquisition module, a low-pass filter module, a mode flag generation module, an adaptive parameter adjustment module, an environmental adaptability correction module, a data fusion calibration module, an error correction module, an attitude solution module and a closed-loop adjustment module. Each module works together to execute the above method.

[0017] On the other hand, the application also provides an AR smart glasses, which comprises a processor, a memory, a display and the above-mentioned AR smart glasses control system.

[0018] In summary, the application proposes an AR smart glasses control method, system and AR smart glasses, which can automatically evolve the tilt angle control technology according to the user's usage habits and environmental changes. The core invention consists of a multi-level, adaptive, closed-loop optimization control process: by learning the user's long-term interaction data and recognizing the user's current head movement mode (high sensitivity or stability priority), the personalized response curve parameters and sensitivity adjustment coefficients are dynamically generated, so that the control strategy fits the individual user habits; real-time fusion of ambient light intensity change rate and gyroscope angular velocity change rate information, real-time adjustment of personalized parameters, enhances the adaptability and robustness of the system to external environmental changes; Kalman filtering is used for preliminary sensor data fusion and calibration, and a deep error correction process is designed, which includes multi-dimensional error analysis, real-time deviation adjustment and condition-triggered reference reset, significantly improving the accuracy and anti-interference ability of the tilt angle data; using high-precision tilt angle data after multiple corrections, combined with quaternion algorithm for accurate head three-dimensional pose solving; and through monitoring the long-term stability of the final pose output and the feedback of the user interaction efficiency, the learning weight of the adaptive parameter adjustment model is optimized in reverse, realizing the continuous evolution of the control system performance.

[0019] Compared with the prior art, the application mainly realizes the following technical effects: significantly improving the personalization and intelligence level of AR smart glasses tilt angle control, automatically learning and adapting to the usage habits of different users (such as head movement amplitude, speed preference), and intelligently switching to the appropriate response mode (high sensitivity or high stability) according to the user's current movement state (rapid rotation or relative stillness), avoiding the response lag or over-sensitivity problems caused by traditional fixed parameter strategy. Enhances the adaptability and robustness of AR smart glasses in complex and variable environments, dynamically adjusts the control parameters by real-time sensing of light mutations, device violent movements and other environmental factors, ensuring the stability and accuracy of pose perception in different lighting conditions and movement states, improving the usability of AR glasses in various real scenes; improves the measurement accuracy of tilt angle data and the accuracy of pose solving, combines low-pass filtering, Kalman filtering fusion, and innovative multi-dimensional error analysis and reference reset mechanism, effectively suppresses the errors introduced by sensor noise, drift and external interference, provides a high-quality data foundation for subsequent accurate pose solving and precise alignment of virtual information and the real world; realizes the self-optimization and continuous evolution ability of the control system, through the establishment of a closed-loop feedback regulation mechanism based on the final pose stability and user interaction efficiency, the tilt angle control performance of AR smart glasses can be continuously optimized and improved during use, and the optimal user experience is maintained in the long term. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 is a flow chart of an AR smart glasses control method provided by an embodiment of the present application. Figure 2 is a detailed flow chart of step S104 (adaptive parameter adjustment model training) in the embodiment of the present application. Figure 3 is a detailed flow chart of step S107 (error correction flow) in the embodiment of the present application. Figure 4 is a module structure schematic diagram of an AR smart glasses control system provided by an embodiment of the present application. Figure 5 is a network architecture schematic diagram of AR smart glasses in the embodiment of the present application. Figure 6 is a simulation effect comparison diagram of the tilt error changing with time in the embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0023] Embodiment one Please refer to Figure 1 is a flow chart of an AR smart glasses control method provided by an embodiment of the present application.

[0024] Step S101: Obtain original tilt angle data and accelerometer data from the sensors built-in in the augmented reality smart glasses, and generate original data sequence containing tilt angle value, acceleration value and corresponding time mark.

[0025] The original data is obtained from the sensor source to construct a data sequence containing the inclination data and acceleration values. If the time mark is missing or abnormal during data acquisition, a preset interpolation method is used to supplement the missing part to obtain a complete original data sequence. According to the complete original data sequence, the change trend of the inclination data and the acceleration value is analyzed to determine whether there is abnormal fluctuation. If the change trend of the inclination data or the acceleration value exceeds the preset threshold range, an abnormal mark is triggered, and the corresponding time mark and value record are obtained. By comparing the time mark and value record corresponding to the abnormal mark, the data is smoothed by using the Kalman filtering algorithm to obtain an optimized data sequence. According to the optimized data sequence, a dynamic data stream containing real-time monitoring results is generated to determine whether the data meets the preset stable condition. If the data in the dynamic data stream does not meet the stable condition, the data generation process is adjusted to obtain new value records to determine the final output result.

[0026] Specifically, the original inclination data and accelerometer data are obtained from the sensors built-in the augmented reality smart glasses, and an original data sequence containing inclination values, acceleration values and corresponding time marks is generated, which can be implemented in the following way. First, assuming that the gyroscope sensor and accelerometer sensor built-in the smart glasses collect data at a frequency of 100Hz, the sensor driver is called through the embedded system interface inside the device to read the original data stream.

[0027] Specifically, the inertial measurement unit (IMU) built-in the AR smart glasses usually contains a three-axis gyroscope, a three-axis accelerometer and an optional three-axis magnetometer. In this step, the original inclination data (such as Roll, Pitch, Yaw angles) obtained by integrating the angular velocity output by the gyroscope and the three-axis acceleration data directly output by the accelerometer are mainly used. The data acquisition frequency can be set to a high value (such as 100Hz or higher) to ensure the real-time nature of the data. Each set of inclination and acceleration data is attached with a high-precision time stamp to form an original data sequence. For example, at t1, the inclination {θx1, θy1, θz1} and acceleration {ax1, ay1, az1} are collected, and recorded as (t1, {θx1, θy1, θz1}, {ax1, ay1, az1}).

[0028] For example, at a certain time, the inclination data returned by the gyroscope sensor is a roll angle (Roll) of 15.2 degrees, a pitch angle (Pitch) of -8.3 degrees, and a yaw angle (Yaw) of 45.7 degrees, while the data returned by the accelerometer is an x-axis acceleration of 0.5 m / s 2 , a y-axis acceleration of -0.2 m / s 2 , and a z-axis acceleration of 9.8 m / s 2These data are transmitted to the host unit through the I2C or SPI communication protocol inside the device. Then, the system automatically attaches a time stamp to each set of data, assuming the current timestamp is 2023-10-15 14:30:00.123, and through a high-precision clock synchronization mechanism, ensures that the time stamp accuracy reaches milliseconds, forming a data record: timestamp 2023-10-15 14:30:00.123, pitch (15.2, -8.3, 45.7), acceleration (0.5, -0.2, 9.8). Subsequently, the system stores the continuous data stream as a time series and uses a sliding window algorithm to perform preliminary smoothing processing, such as taking the average of 5 consecutive sampling points to reduce noise. The calculation result shows that the average value of the roll angle in a certain window is 15.1 degrees, and the standard deviation is 0.3, indicating that the data stability is high. Finally, these processed data sequences are stored in the local cache of the device and compressed by data compression algorithms (such as Huffman encoding) to reduce storage space occupation, with a compression rate of up to 30%, providing reliable input for subsequent attitude estimation or motion analysis. Through the above process, a complete chain from data acquisition to processing is formed, ensuring data accuracy and real-time performance, and combining with the attitude recognition business to provide accurate head movement feedback to users.

[0029] Step S102, apply a low-pass filter algorithm to the pitch values in the original data sequence for denoising, obtaining filtered pitch measurement values and corresponding time stamp sequences.

[0030] Raw sensor data, especially accelerometer data, is easily disturbed by device vibration and user's small motion. If the pitch data is directly integrated from angular velocity, there will also be cumulative error. This step uses a low-pass filter (such as a Butterworth filter, a first-order RC filter, or a moving average filter) to process the original pitch values. For example, a first-order low-pass filter with a cutoff frequency of 5Hz can be used, and its transfer function can be designed as H(z) = (1-α) / (1 - αz⁻ 1 ), where α is determined according to the sampling frequency and the cutoff frequency. After processing, smoother pitch measurement values are obtained, and the time stamp remains unchanged.

[0031] Specifically, a low-pass filter algorithm is applied to the original tilt angle data sequence collected by the augmented reality smart glasses for denoising processing, generating a filtered tilt angle measurement value and a corresponding time tag sequence. This can be achieved in the following way. Assuming that the gyroscope sensor collects tilt angle data at a frequency of 100 Hz, for example, the timestamps of the first five consecutive sampling points in a certain time period are 2023-10-16 09:00:00.100 to 2023-10-16 09:00:00.140, and the roll angle data are 16.5 degrees, 16.8 degrees, 17.0 degrees, 16.3 degrees, and 16.9 degrees, respectively, containing high-frequency noise. The system first selects a first-order low-pass filter with a cutoff frequency of 5 Hz, and the filter transfer function is H(z) = 0.1 / (1-0.9z⁻ 1 ), where 0.1 is the gain coefficient and 0.9 is the filter coefficient, ensuring that noise signals above 5 Hz are filtered out. For the first data point of 16.5 degrees, the initial filter output is set to its own value, and then the formula y(n) = 0.1x(n) + 0.9y(n-1) is applied to each subsequent data point, where x(n) is the current input and y(n-1) is the previous filter output.

[0032] For example, for the second point of 16.8 degrees, calculate y(2) = 0.1x16.8 + 0.9x16.5 = 16.55 degrees, and then process in turn to obtain the filtered sequence as 16.50 degrees, 16.53 degrees, 16.58 degrees, 16.54 degrees, and 16.56 degrees. The corresponding timestamps remain unchanged, forming a new sequence: 2023-10-16 09:00:00.100 to 2023-10-16 09:00:00.140, with tilt angles of 16.50 degrees, 16.53 degrees, 16.58 degrees, 16.54 degrees, and 16.56 degrees, respectively. To verify the filtering effect, the standard deviation of the original sequence is 0.28 degrees, and the standard deviation after filtering is reduced to 0.03 degrees, indicating that the data fluctuation is significantly reduced. The filtered data is transmitted to the pose analysis module through the internal bus of the embedded system for subsequent head pose tracking, and the data storage uses JSON format to ensure compatibility with the augmented reality scene rendering module. Through this process, the system effectively removes high-frequency noise and retains low-frequency effective signals, providing stable input for pose recognition.

[0033] In step S103, the head motion frequency is calculated based on the filtered tilt angle measurement value and the time tag sequence. If the head motion frequency exceeds the pre-set threshold of 5 times per second, a high-sensitivity mode flag is generated. If the tilt angle change rate is less than 0.5 degrees per second and lasts for 5 seconds, a stability priority mode flag is generated.

[0034] The head motion frequency can be estimated by performing a fast Fourier transform (FFT) on the filtered tilt data to analyze the dominant frequency, or by detecting the number of zero-crossings or peak points in the tilt data per unit time (e.g., 1 second). If the calculated frequency exceeds 5 Hz, it indicates that the user's head motion is relatively intense, and the system is marked as a high sensitivity mode.

[0035] The tilt rate of change is obtained by calculating the difference between the tilt values of two adjacent sampling points and dividing by the time interval. If the absolute value of the tilt rate of change of all sampling points within 5 seconds is less than 0.5 degrees / second, it indicates that the user's head is relatively stationary, and the system is marked as a stability priority mode. These two flags (e.g., Boolean values or enumeration types) are used for subsequent parameter adjustment.

[0036] The tilt measurement values and time marker sequence are obtained by the sensor, and a filtering algorithm is used to obtain smoothed tilt data. According to the smoothed tilt data and the time marker sequence, the head motion frequency is calculated using Fourier transform to obtain a frequency value. If the frequency value exceeds a preset frequency threshold, a high sensitivity mode flag is generated to determine the high sensitivity mode state. According to the smoothed tilt data and the time marker sequence, the tilt rate of change is calculated to obtain a rate value. If the rate value is below a preset rate threshold and the duration meets the conditions, a stability priority mode flag is generated to determine the stability priority mode state. Through a real-time monitoring mechanism, the high sensitivity mode flag and the stability priority mode flag are obtained to determine the current mode state. According to the current mode state, a preset rule is used to adjust the data processing flow to obtain optimized monitoring parameters.

[0037] Specifically, assume that a sequence data containing head tilt measurement values and time markers is received, the tilt measurement values are in degrees, the time markers are in seconds, and the data sampling frequency is 100 Hz, i.e., one data point every 0.01 seconds. The example data is: time sequence [0, 0.01, 0.02, 。。。, 10] seconds, corresponding to the tilt sequence [5.2, 5.3, 5.5, 。。。, 6.0] degrees. First, the tilt data is filtered using a 5-point moving average filtering algorithm, the formula is y[i] = (x[i-2]+ x[i-1] + x[i]+ x[i+1] + x[i+2]) / 5, where x[i] is the original tilt, y[i] is the filtered tilt, and the boundary points are zero-padded.

[0038] For example, for the point at time 0.02 seconds, the tilt angle value is 5.5 degrees, the surrounding points are [5.2, 5.3, 5.5, 5.7, 5.8], and the filtered value is (5.2+5.3+5.5+5.7+5.8) / 5=5.5 degrees, resulting in the filtered sequence [5.2, 5.3, 5.5, 5.7, 5.8, 5.9]. Then the head motion frequency is calculated, and the frequency spectrum of the filtered sequence is analyzed using Fast Fourier Transform (FFT), with the number of sampling points N=1000 and the sampling frequency fs=100Hz, and the frequency resolution is fs / N=0.1Hz. The FFT result shows that the main frequency component is 5.2Hz, which exceeds the preset threshold of 5Hz, so the high sensitivity mode flag is generated, with an output of 1. Next, the tilt angle change rate is calculated using the difference method, with the formula rate[i] = (y[i+1]- y[i]) / 0.01. For example, from time 0.02 seconds to 0.03 seconds, the tilt angle changes from 5.5 degrees to 5.7 degrees, with a change rate of (5.7-5.5) / 0.01=20 degrees / second. Check the change rate of the last 500 points (corresponding to 5 seconds), if the absolute value of all points is less than 0.5 degrees / second, for example, the sequence [0.4, 0.3, 0.2, 0.4] degrees / second, which meets the condition, then the stability priority mode flag is generated, with an output of 1. If the change rate at a certain point is 0.6 degrees / second, the count is terminated, and the flag output is 0. Finally, the system automatically switches modes according to the flag value, with the high sensitivity mode responding to fast motion first, and the stability priority mode optimizing low speed stable scenes, ensuring that the device adapts to different head motion states.

[0039] Step S104, using the high sensitivity mode flag or the stability priority mode flag, combined with the stored past 30 days of user interaction record data, training the adaptive parameter adjustment model, generating personalized response curve parameters and sensitivity adjustment coefficients.

[0040] Please refer to Figure 2 , which is a detailed flowchart of this step.

[0041] First, obtain the past 30 days of user interaction records from the local storage of the AR glasses or the cloud database (S201). These records may include the time the user starts the application, the operation type (such as nodding to confirm, turning to select), the tilt angle data corresponding to the operation, the system response time, the user's satisfaction feedback on the response (if any), etc.

[0042] Next, pre-process and feature engineer these data (S202), extract features related to tilt angle control, such as average head motion speed, commonly used tilt angle range, high / low frequency motion proportion, tilt angle mode in specific interaction scenarios, etc. At the same time, input the mode flag (high sensitivity / stability priority) generated in S103 as the current state.

[0043] Then, depending on whether the current is high sensitivity mode or stability priority mode, a machine learning model is selected or adjusted (S203). For example, if the current is high sensitivity mode, a random forest regression model can be employed, targeting a set of response curve parameters (e.g. steeper slope) and a higher sensitivity adjustment factor that results in a faster response. If the current is stability priority mode, a gradient boosting decision tree model can be employed, targeting a set of response curve parameters (e.g. gentler slope) and a moderate sensitivity adjustment factor that results in a smoother response and stronger anti-interference capability. The training process uses interaction features in historical data and corresponding “ideal” response parameters (which can be back-calculated based on user satisfaction or pre-set rules) as training samples.

[0044] Finally, the model outputs a set of personalized response curve parameters (e.g. coefficients of a first or second order curve) and a sensitivity adjustment factor a_user (S204). These parameters will affect the way sensor data is weighted and processed in subsequent pose solving.

[0045] Specifically, using high sensitivity mode flag and stability priority mode flag, combined with the stored past 30 days of user interaction record data, the adaptive parameter adjustment model is trained to generate personalized response curve parameters and sensitivity adjustment coefficients. The specific implementation method can be realized through the following technical process. First, extract the user interaction record from the database, assuming that the data contains user ID, interaction timestamp, input type (such as text, click), response time (unit: millisecond) and user satisfaction score (1-5 points). Take user A as an example, who has 1000 interactions in 30 days, with an average response time of 200ms and a satisfaction average of 4.2. In the data preprocessing stage, use the Pandas library of Python to clean the data, remove the response time outliers (more than ±2 times the standard deviation of the mean, about 2% of the data), and encode the input type as a numerical value (text = 1, click = 2). Then, construct a feature vector, including interaction frequency (an average of 33 times per day), input type proportion (text 70%, click 30%) and satisfaction distribution. In the model training stage, a random forest regression algorithm (n_estimators=100, max_depth=10) is used, with high sensitivity mode flag (value 1 means on, 0 means off) and stability priority mode flag (same reason) as input features, and the target output as response curve parameters (slope k and intercept b) and sensitivity adjustment coefficient α. Taking user A's data as an example, after dividing the training set (80%) and test set (20%), the model predicts k=0.85, b=50, α=1.2, and the mean square error is 0.03. Analysis shows that in high sensitivity mode, α increases by 10% to speed up the response, while in stability priority mode, k decreases by 5% to smooth the curve. To generate personalized parameters, the system dynamically adjusts α to 1.25 (high-frequency interaction users prefer fast response) based on user A's interaction frequency and satisfaction. Finally, the parameters are stored in the database and called in real time through the API, applied to the front-end response module to ensure logical consistency. If the user interaction mode changes (such as satisfaction falling to 3.8), the system re-trains the model every 7 days to update the parameters, maintaining adaptability. This process is realized through automated scripts without human intervention, and the parameter adjustment is closely related to business goals (such as improving user retention rate).

[0046] Step S105, fuse the light intensity change rate obtained by the ambient light sensor and the angular velocity change rate obtained by the gyroscope. If the light intensity change rate exceeds 50 lux per second or the angular velocity change rate exceeds 10 degrees per second, adjust the personalized response curve parameters and sensitivity adjustment coefficients to generate environment-adaptive corrected control parameters.

[0047] This step aims to adapt the AR glasses to the dynamic changes of the external environment. First, the ambient light sensor monitors the ambient light intensity in real time, and calculates its rate of change dL / dt. The gyroscope outputs the three-axis angular velocity {ωx, ωy, ωz} in real time, and calculates the rate of change of the magnitude of the combined angular velocity d|ω| / dt, or directly uses the angular velocity values of each axis. If dL / dt>50 lux / s (for example, from dark to light, or light flickering), or the absolute value of any axis angular velocity or the combined angular velocity exceeds 10 deg / s (for example, when the user is walking, running or riding a vehicle), it is considered that the environment is highly dynamic.

[0048] At this time, the personalized parameters generated in S104 are adjusted. For example, an environmental adjustment factor β_env can be introduced, so that the final sensitivity coefficient α_final = α_user * β_env. β_env can be a function based on dL / dt and d|ω| / dt, when the environment is highly dynamic, β_env is appropriately increased (if faster response is needed) or decreased (if stronger anti-interference is needed). The response curve parameters may also be fine-tuned according to the environmental dynamics, for example, when the motion is intense, the filtering strength is appropriately increased or the sensitivity to high-frequency acceleration components is reduced. The adjusted parameters are the control parameters after environmental adaptability correction.

[0049] In step S106, according to the control parameters after environmental adaptability correction, the Kalman filter algorithm is used to fuse and calibrate the filtered tilt angle measurement and accelerometer data to generate a predicted tilt angle value. If the deviation between the predicted tilt angle value and the actual measured value exceeds 2 degrees, the error correction process is activated.

[0050] The Kalman filter is an effective multi-sensor data fusion algorithm. Here, the state vector can include the tilt angle, angular velocity, and even the bias of the accelerometer. The measurement vector includes the tilt angle calculated from the gravity component of the accelerometer (more accurate in quasi-static state) and the tilt angle obtained by integrating the gyroscope (good dynamic response but with drift). The control parameters after environmental adaptability correction generated in S105 can be used to adjust the process noise covariance Q and the measurement noise covariance R in the Kalman filter. For example, when the environment is highly dynamic, Q can be increased, indicating that the system state changes quickly, and the degree of trust in the model prediction is reduced; at the same time, if the accelerometer is greatly disturbed by vibration, R corresponding to the measurement of the accelerometer is increased.

[0051] The Kalman filter outputs the fused predicted tilt angle value. This predicted value is compared with a reference "actual" tilt angle measurement (which can be the filtered tilt angle value, or a sensor value that is considered more reliable under certain conditions). If the absolute value of the deviation between the two exceeds 2 degrees, it is considered that the current fusion result may have a large error, and a deeper error correction process needs to be activated.

[0052] The inclination measurement and accelerometer data are obtained from the sensor, and the first filtered data set is obtained after preprocessing. The Kalman filter algorithm is used to fuse and calibrate the first data set to generate a predicted inclination value. By comparing the predicted inclination value with the actual measurement value, the deviation value is calculated. If the deviation value exceeds the preset threshold, the error correction process is activated, the Kalman filter algorithm parameters are adjusted, and the second corrected data set is generated. According to the second data set, the predicted inclination value is regenerated, and the updated inclination prediction accuracy is obtained. Through the error detection mechanism, the matching degree of the updated inclination prediction accuracy and the actual measurement value is analyzed, and the calibrated inclination output value is generated. The calibrated inclination output value is used to update the sensor data processing model, and the optimized data fusion calibration is obtained.

[0053] Specifically, based on the environment-adaptive corrected control parameters, first, it is assumed that the system has corrected the sensor data under different temperature and humidity environments to obtain a set of control parameters, for example, the temperature correction coefficient is 1.05 and the humidity correction coefficient is 0.98. By multiplying the original inclination measurement value 30.5 degrees and the accelerometer data 0.52g by the corresponding coefficients, the corrected inclination value 32.025 degrees and the acceleration value 0.5096g are obtained, which lay the foundation for subsequent fusion. Then, the Kalman filter algorithm is used for data fusion, assuming that the state vector contains inclination and angular velocity, the initial state covariance matrix P is set to [0.1, 0; 0, 0.1], the process noise covariance Q is 0.01, and the measurement noise covariance R is 0.05. Through the state transition equation x(k)=x(k-1)+w(k) and the measurement equation z(k)=x(k)+v(k), the Kalman gain K is calculated as P*H' / (H*P*H'+R), the state estimate value is updated, and finally the predicted inclination value is fused to 31.8 degrees. Analysis shows that the data noise after fusion is reduced by about 30%, improving the accuracy. Then, the predicted inclination value 31.8 degrees is compared with the actual measurement value 30.5 degrees, and the deviation is 1.3 degrees, which does not exceed the threshold of 2 degrees, so the error correction process does not need to be activated, but the system will record this deviation value to the historical database for subsequent trend analysis. If the deviation exceeds the threshold, for example, the predicted value is 33.0 degrees, and the deviation is 2.5 degrees, the system automatically triggers the error correction process, adjusts the parameters of the Kalman filter, such as increasing Q to 0.02, recalculates the state estimate, and analyzes whether the corrected deviation is reduced to within 1.8 degrees to ensure system stability. The whole process is automatically run through the embedded algorithm, and the data is stored in the cloud to support real-time monitoring and parameter optimization, forming a complete logical chain from correction to fusion to error processing.

[0054] Step S107, generate correction factors in the error correction process, apply a multi-dimensional error analysis model to the predicted inclination value for real-time deviation adjustment, if the deviation direction and amplitude of the continuous 5 sampling points are greater than 1 degree, execute the reference reset, and generate the inclination data after deviation adjustment.

[0055] Please refer to Figure 3 , which is a detailed flowchart of this step.

[0056] When the error correction process is activated (S301): First, get the current predicted inclination value (from S106) (S302).

[0057] Then, apply a multi-dimensional error analysis model (S303). This model may consider the current motion state (judged by the accelerometer and gyroscope), temperature changes (if the AR glasses have a temperature sensor, because temperature can affect the performance of the IMU), magnetic field environment (if a magnetometer is used to assist in attitude determination), and other dimensions, to comprehensively evaluate the possible error sources and size of the current predicted inclination value. Based on this analysis, a real-time correction factor or correction amount is generated to fine-tune the predicted inclination value.

[0058] At the same time, the system will continuously monitor the deviation between the adjusted predicted inclination and the reference actual inclination. If it is found that the deviation direction of the continuous 5 sampling points is consistent (for example, the predicted value is always about 1.5 degrees larger than the actual value), and the deviation amplitude is greater than 1 degree (S304), it usually means that there is a systematic cumulative error or the sensor reference has drifted.

[0059] In this case, perform the reference reset operation (S305). Reference reset may mean recalibrating the sensor (for example, prompting the user to hold the head level and still for a few seconds to calibrate the zero offset), or using the current most reliable external reference (such as the result of visual SLAM, if available) to force correction of the current attitude reference.

[0060] After real-time deviation adjustment and possible reference reset, generate the inclination data after deviation adjustment (S306). These data are considered to be the most accurate inclination estimates.

[0061] Specifically, in the error correction process, for the real-time deviation adjustment of the predicted inclination value, first, the inclination data collected is processed by a multi-dimensional error analysis model, assuming that the current sampling point predicted inclination value is 35.2 degrees, and the historical reference value is 34.0 degrees, the deviation is calculated to be 1.2 degrees, combined with the multi-dimensional influence factors of environmental temperature 25 degrees Celsius and humidity 60%, through the preset error model formula: deviation correction value = original deviation x (1 + 0.01 x temperature influence coefficient + 0.005 x humidity influence coefficient), the correction value is calculated to be 1.26 degrees, then the correction factor is generated as 1.037, the predicted value is adjusted to 34.0 + 1.26 = 35.26 degrees, forming the preliminary correction data. Next, the system automatically records the deviation data of the continuous 5 sampling points, assuming that they are 1.3 degrees, 1.4 degrees, 1.5 degrees, 1.2 degrees and 1.6 degrees, analysis finds that all deviation directions are positive and the amplitudes are all greater than 1 degree, triggering the reference reset condition, the system recalculates the reference value to be 34.5 degrees through historical data regression analysis, and updates the model parameters, and the subsequent predicted value is corrected based on this as a reference. Finally, the inclination data adjusted by the deviation is generated, assuming that the new round of sampling predicted value is 35.8 degrees, based on the new reference 34.5 degrees, the deviation is calculated to be 1.3 degrees, and after applying the correction factor, it is adjusted to 35.76 degrees, at the same time, the system associates the adjustment result with the device running state, analyzes whether the inclination change is related to the vibration frequency 10 hertz, if related, further optimizes the error model parameters to ensure the correction accuracy. This process is automatically completed by the system algorithm throughout the process, forming a complete logical chain from data collection, deviation analysis to reference update and final correction, ensuring the high reliability of the inclination data.

[0062] Specifically, when the error correction process is activated (S301), it means that the deviation between the Kalman filter predicted inclination value from step S106 and the reference actual inclination measurement value exceeds the preset threshold of 2 degrees. At this time, the system considers that there may be errors that need to be further processed.

[0063] First, the predicted inclination value θ_predicted_kf output by the Kalman filter is obtained (for example, {Roll_kf, Pitch_kf, Yaw_kf}) (S302).

[0064] Then, a multi-dimensional error analysis model (Multi-Dimensional Error Analysis Model, MDEAM) is started and applied to perform real-time and refined deviation adjustment on the predicted inclination value (S303). The construction and application process of this multi-dimensional error analysis model is as follows: 1. Extraction and quantification of multi-dimensional error source characteristics: MDEAM will take into account multiple potential factors that currently affect the accuracy of the tilt measurement, which are quantified as feature inputs to the model: Motion state features: including accelerometer magnitude stability: compute the mean and variance of the magnitude of the 3-axis accelerometer readings over a short time window (e.g. 0.1s). A large deviation from the gravitational acceleration g (about 9.8 m / s2) or a large variance usually indicates that there is significant external acceleration disturbance, and the accelerometer-derived tilt (which is often used as an observation in Kalman filtering) is less reliable.

[0065] Gyroscope angular velocity features: compute the mean, peak, and rate of change of the 3-axis gyroscope angular velocity over a short time window. High angular velocity or large changes in angular velocity can mean that the gyroscope integration error is more likely to accumulate.

[0066] Temperature change features: if the AR smart glasses are equipped with a temperature sensor, get the current ambient temperature T near the IMU chip. Changes in temperature can cause zero-point drift (Bias Drift) and scale factor error (Scale Factor Error) of the IMU sensors, especially the gyroscope. MDEAM may have pre-calibrated temperature-error compensation curves or parameters stored internally.

[0067] Running time features: record the cumulative running time t_elapsed since the last complete sensor calibration or reference reset. Over a long period of time, small drifts that are not fully compensated for can accumulate into significant errors.

[0068] Current mode flag: introduce the high-sensitivity mode flag or stability-priority mode flag generated in step S103. In different modes, users expect different control characteristics, and error tolerance may also be different, and MDEAM can adjust the aggressiveness of its correction strategy accordingly.

[0069] Historical error statistics features: if the system has learning ability, it can record the typical deviation size and direction that has occurred in the past under similar motion state and environmental characteristics as a reference for correction.

[0070] 2. Error analysis and correction quantity generation: The core of MDEAM can be a lightweight, pre-trained machine learning model (e.g. a small feedforward neural network, a set of decision trees, or a rule-based fuzzy logic system), or a lookup table / function based on physical models and empirical parameters.

[0071] Input: Combine the quantized multi-dimension features into a feature vector `V_features = [std(|a|), mean(|ω|), T, t_elapsed, Mode_flag,...]`.

[0072] Process: MDEAM analyzes the input feature vector `V_features`.

[0073] If based on rules: e.g. Rule 1: "IF std(|a|)>Threshold_acc_var AND Mode_flag== Stability_Priority THEN predict that the accelerometer introduces large random errors, should reduce its weight in the tilt correction, or apply a reverse compensation term related to the direction of acceleration change". Rule 2: "IF (T - T_calib)>Threshold_temp_diff THEN calculate a temperature drift correction amount `Δθ_temp` according to the pre-stored temperature compensation coefficient `f_temp(T)`.

[0074] If based on machine learning models: the model directly outputs a predicted error vector `Δθ_error_pred ={ΔRoll_pred, ΔPitch_pred, ΔYaw_pred}`, or a comprehensive correction weight / factor.

[0075] Output: MDEAM outputs a real-time tilt correction vector `Δθ_correction`. This correction vector aims to compensate for the most likely errors under the combined action of multiple dimensions.

[0076] 3. Implementation of real-time bias adjustment: Apply the tilt correction vector `Δθ_correction` output by MDEAM to the predicted tilt value of Kalman filter: `θ_adjusted = θ_predicted_kf - Δθ_correction` (Here, subtraction is used because `Δθ_correction` represents the "error" predicted by the model, so the error needs to be subtracted to get the adjusted value; or `Δθ_correction` is directly the correction amount that should be applied, then use addition).

[0077] This `θ_adjusted` is the tilt estimate after real-time adjustment by the multi-dimension error analysis model.

[0078] Example to illustrate the working scenario of MDEAM: Scenario 1: User is using AR glasses on a bumpy vehicle.

[0079] MDEAM detects: Accelerometer norm variance `std(|a|)` is significantly increased, and gyro angular velocity `|ω|` is also at a high level.

[0080] MDEAM analysis: At this time, external linear acceleration causes serious interference to the accelerometer measurement of the gravity direction, and the accelerometer-based observation update in the Kalman filter may introduce large errors. At the same time, the gyro integration drift under high dynamics may also accelerate.

[0081] MDEAM output: A `Δθ_correction` that may temporarily increase the confidence in short-term integration of the gyro (reduce the dependence on accelerometer observations in Kalman filtering), and try to compensate for a part of the pseudo-pitch angle change introduced by vibration according to the statistical characteristics of acceleration disturbance (if identified). At the same time, a small dynamic adjustment of the gyro scale factor may be made.

[0082] Scenario 2: After a long time of AR glasses running, a small but persistent head-up or head-down drift occurs in a stationary state.

[0083] MDEAM detects: `std(|a|)` and `|ω|` are both small (indicating stationary), but `t_elapsed` is large, and the temperature `T` may have changed compared to the initial calibration. The continuously monitored `θ_adjusted` has a small fixed deviation from the user's subjective feeling or absolute reference (such as the desktop level).

[0084] MDEAM analysis: This is likely the result of slow drift accumulation of gyro zero offset over time and temperature.

[0085] MDEAM output: A slowly changing `Δθ_correction` related to the current temperature change and the running time to offset this accumulated zero offset drift.

[0086] After real-time bias adjustment by MDEAM, the system continues to compare the adjusted tilt angle θ_adjusted with the reference actual tilt measurement. Next is the logic to determine whether a reference reset is needed: the system continuously monitors the deviation δθ = θ_adjusted - θ_reference between the adjusted predicted tilt angle θ_adjusted and the reference actual tilt. If it is found that the direction of the deviation vector δθ is consistent (e.g., the sign of each component is the same or changes little for M consecutive samples, e.g., M = 5) and the magnitude of the deviation (e.g., |δθ| or the absolute value of its dominant component) is larger than a preset magnitude threshold (e.g., 1 degree) for each sample (S304), it usually indicates that real-time fine-tuning by MDEAM is not enough to completely correct the error, there may be a deeper, systematic, cumulative error, or there is a significant drift in the sensor's own reference (zero point) that cannot be completely covered by the model.

[0087] In this case (S304 is "Yes"), the system will perform a reference reset operation (S305). Reference reset is a more "thorough" correction method than real-time bias adjustment, and its purpose is to rebuild a reliable attitude reference. The specific operation may include: Forced sensor recalibration: If the AR glasses support, the internal gyroscope zero offset calibration program can be triggered (e.g., prompt the user to hold the device horizontally still for a few seconds).

[0088] Use external absolute reference (if available): If the system integrates visual SLAM or other absolute positioning / orientation modules, the attitude information output by these modules, which is considered to be of higher accuracy, can be used to forcibly reset the attitude reference calculated by the current IMU.

[0089] Reset the cumulative state: Clear or significantly adjust the state covariance matrix in the Kalman filter and the time-cumulative related error estimates in MDEAM.

[0090] User-involved calibration: In extreme cases, the user may be prompted to perform manual calibration.

[0091] After real-time bias adjustment (performed by MDEAM) and possible reference resetting, the system finally generates bias-adjusted tilt data θ_final_adjusted (S306). If reference resetting is performed, θ_final_adjusted is the tilt under the new reference after resetting; if no reference resetting is performed, θ_final_adjusted is usually θ_adjusted, or the result of further smoothing or confirmation of θ_adjusted after judging that no reference resetting is needed. These bias-adjusted tilt data have higher accuracy and reliability, and will be used as input of the subsequent step S108, the attitude solving module. If S304 judges “No”, i.e. the continuous bias does not meet the reference resetting condition, θ_adjusted is directly output as the bias-adjusted tilt data (S306).

[0092] Step S108, the bias-adjusted tilt data is used to update the attitude solving matrix, and the head three-dimensional space positioning information is calculated through the quaternion rotation algorithm combined with the gyroscope data, to generate the final attitude angle data sequence.

[0093] Attitude solving usually uses Direction Cosine Matrix (DCM), Euler angle or quaternion. Quaternion is widely used because it avoids gimbal lock problem and has higher calculation efficiency.

[0094] Firstly, the bias-adjusted tilt data (usually in the form of Euler angle) obtained in S107 is converted into quaternion q_tilt. At the same time, the angular velocity {ωx, ωy, ωz} output by the gyroscope at the current time is obtained. Using the attitude quaternion q(t-1) at the last time and the current angular velocity, the attitude change predicted by the gyroscope is obtained by integrating the quaternion differential equation dq / dt = 0.5 * q * [0, ωx, ωy, ωz] (for example, using Runge-Kutta method or simple Euler integration).

[0095] Then, the attitude change predicted by the gyroscope is fused with the attitude information q_tilt obtained from the bias-adjusted tilt data (for example, through a complementary filter or a more complex fusion algorithm such as Mahony or Madgwick filter) to update the current attitude solving matrix (usually updating the attitude quaternion q(t)).

[0096] This updated attitude quaternion q(t) represents the accurate attitude of the head in three-dimensional space. It can be converted back to Euler angle form to form the final attitude angle data sequence for AR application.

[0097] Specifically, assume that the initial inclination data is the original sensor output, which contains bias: pitch is 10.5°, roll is 5.2°, and yaw is 15.8°. First, eliminate the sensor zero bias by bias adjustment. Assume that the zero bias values obtained through calibration experiments are 1.2°, 0.8°, and 2.5° respectively, then the adjusted inclination data is: pitch 10.5-1.2=9.3°, roll 5.2-0.8=4.4°, and yaw 15.8-2.5=13.3°. Then update the attitude solving matrix, and use the adjusted inclination data to construct the direction cosine matrix (DCM), whose formula is C = [[cos(yaw)cos(pitch), cos(yaw)sin(pitch)sin(roll)-sin(yaw)cos(roll), cos(yaw)sin(pitch)cos(roll)+sin(yaw)sin(roll)], [sin(yaw)cos(pitch), sin(yaw)sin(pitch)sin(roll)+cos(yaw)cos(roll), sin(yaw)sin(pitch)cos(roll)-cos(yaw)sin(roll)], [-sin(pitch), cos(pitch)sin(roll), cos(pitch)cos(roll)]], substitute the angles (in radian: 9.3°=0.162 rad, 4.4°=0.077 rad, 13.3°=0.232 rad), and calculate C=[[0.970, -0.223, 0.093], [0.242, 0.873,-0.426], [-0.162, 0.427, 0.899]]. Then, combined with the gyroscope data, assume that the gyroscope output angular velocity is [ωx=0.01 rad / s, ωy=0.02 rad / s, ωz=0.015 rad / s], update the attitude by using the quaternion rotation algorithm. The initial quaternion is set to q=[1, 0, 0, 0], the quaternion differential equation is dq / dt=0.5*q⊗[0, ωx, ωy, ωz], and the Euler integration method (time step Δt=0.01s) is used to update the quaternion, obtaining the new quaternion q_new=[0.999999,0.00005, 0.0001, 0.000075], and normalizing the processing. Based on the quaternion, calculate the rotation matrix, and then convert it to Euler angles to obtain the head three-dimensional space positioning information: new pitch 9.31°, roll 4.41°, and yaw 13.32°. Repeat this process to generate a sequence of attitude angle data, such as {[9.31°, 4.41°, 13.32°], [9.32°, 4.42°, 13.34°],.The analysis shows that the bias adjustment improves the data accuracy, the quaternion algorithm effectively fuses the dynamic information of the gyroscope, ensures the continuity and stability of the attitude solution, and is suitable for real-time positioning of the virtual reality head-mounted display.

[0098] In step S109, according to the stability judgment of the final attitude angle data sequence, if the angle change variance is lower than 0.1 degree square, and the improvement amplitude of the user interaction frequency exceeds 10%, the learning weight of the adaptive parameter adjustment model is adjusted, the optimized tilt angle control parameter is generated, and the closed-loop adjustment process is completed.

[0099] The system continuously monitors the stability of the final attitude angle data sequence generated in S108. For example, in the scenario where the user expects the head to be stationary (which can be combined with the stability priority mode flag in S103), the variance of the attitude angle in a period of time is calculated. If the variance is continuously lower than 0.1 degree square, it indicates that the current attitude control is very stable.

[0100] At the same time, the system also monitors the user's interaction frequency. If it is found that the user's interaction frequency with the AR glasses (such as the number of clicks, confirmations, etc. per unit time) has increased by more than 10% compared to the average level in the past period of time, it may mean that the user is satisfied with the current control experience and the interaction is smoother.

[0101] When both the high stability and the high interaction frequency increase conditions are met, the system considers that the current adaptive parameter adjustment model (trained in S104) performs well. At this time, the learning weight or trust degree can be appropriately increased, so that the model is more inclined to maintain the current parameter characteristics in subsequent fine-tuning, or gives higher weight to these successful experiences in the next retraining. On the contrary, if the stability is poor or the interaction frequency decreases, the trust degree of the current model may need to be reduced to promote a larger adjustment in the next training.

[0102] In this way, the optimized tilt angle control parameter is generated, and the entire control system forms a closed loop that can self-optimize according to long-term performance feedback.

[0103] Specifically, in implementing the judgment of the stability of the attitude angle data sequence and the subsequent closed-loop adjustment process, first, continuous attitude angle data is collected by sensors, assuming that the collected 100 angle data point sequence is [1.2, 1.3, 1.1,...] degrees, and the variance is calculated to evaluate the stability. The variance calculation formula is: variance = Σ(xi - xmean) 2 / n, where x mean is the sequence mean value and n is the number of data points. It is calculated that if the mean value is 1.2 degrees, the variance result is 0.05 degrees square, which is lower than the preset threshold of 0.1 degrees square, and it is determined to be stable. Then, the change of user interaction frequency is analyzed, assuming that the initial frequency is 10 times per minute and the current frequency is 11.5 times per minute, the growth rate is (11.5-10) / 10*100%=15%, which exceeds the threshold of 10%, meeting the adjustment condition. Subsequently, the learning weight adjustment link of the adaptive parameter adjustment model is entered, the gradient descent algorithm is used to optimize the weight, the initial weight is set to 0.5, the learning rate is 0.01, and the loss function (assuming it is mean square error) is calculated by iteration to adjust the weight to 0.52 to reduce the error. Based on this weight, the optimized inclination control parameter is generated, assuming that the original inclination parameter is 5 degrees, and the new parameter is output as 5.2 degrees after adjusting the weight. Finally, the closed-loop adjustment process feeds back the new parameter 5.2 degrees to the actuator in real time through the embedded controller, completes the inclination adjustment, and records the adjusted attitude data to verify the effect. If the variance is still lower than 0.1 degrees square and the frequency growth maintains above 10%, it is confirmed that the optimization is effective, otherwise the weight is iteratively adjusted to 0.53 for further optimization. Through the above data-driven automatic process, each link is closely connected to ensure the stability and responsiveness of the system, and historical data analysis (such as the previous adjustment variance of 0.08 degrees square) is used to assist decision-making, forming a complete technical closed loop.

[0104] Example Two Reference Figure 4Fig. 1 is a schematic diagram of a module structure of an AR smart glasses control system provided by an embodiment of the present application. The system can be integrated on the mainboard of AR smart glasses, or part of the functions can be realized by a connected computing unit (such as a smart phone). The system comprises: a data acquisition module for acquiring original inclination data and accelerometer data from the sensors built-in in the augmented reality smart glasses, and generating an original data sequence containing inclination values, acceleration values and corresponding time markers; a low-pass filtering module for applying a low-pass filtering algorithm to the inclination values in the original data sequence for denoising processing, to obtain filtered inclination measurement values and corresponding time marker sequences; a mode flag generation module for calculating a head movement frequency according to the filtered inclination measurement values and time marker sequences, and generating a high sensitivity mode flag if the head movement frequency exceeds a preset threshold of 5 times per second, or generating a stability priority mode flag if the inclination change rate is lower than 0.5 degrees per second and lasts for 5 seconds; an adaptive parameter adjustment module for using the high sensitivity mode flag or the stability priority mode flag, combining the stored past 30-day user interaction record data, and using a built-in machine learning model (such as random forest, gradient boosting tree) for training and prediction to generate personalized response curve parameters and sensitivity adjustment coefficients; an environmental adaptability correction module for fusing the light intensity change rate obtained by the ambient light sensor and the angular velocity change rate obtained by the gyroscope, and adjusting the personalized response curve parameters and sensitivity adjustment coefficients if the light intensity change rate exceeds 50 lux per second or the angular velocity change rate exceeds 10 degrees per second, to generate environment-adapted control parameters; a data fusion calibration module for fusing and calibrating the filtered inclination measurement values and accelerometer data according to the environment-adapted control parameters using a Kalman filtering algorithm, to generate predicted inclination values, and activating an error correction process if the deviation between the predicted inclination values and the actual measured values exceeds 2 degrees; an error correction module for generating a correction factor in the error correction process, applying a multi-dimensional error analysis model to real-time deviation adjustment of the predicted inclination values, and executing a reference reset if the deviation direction and amplitude of the continuous 5 sampling points are both greater than 1 degree, to generate deviation-adjusted inclination data; a posture solving module for updating a posture solving matrix using the deviation-adjusted inclination data, calculating head three-dimensional space positioning information through a quaternion rotation algorithm in combination with the gyroscope data, and generating a final attitude angle data sequence; a closed-loop adjustment module for judging the stability of the final attitude angle data sequence, adjusting the learning weight of the adaptive parameter adjustment model if the angle change variance is lower than 0.1 degrees square and the improvement amplitude of the user interaction frequency exceeds 10%, generating optimized inclination control parameters, and completing the closed-loop adjustment process.

[0105] The above modules can be realized by software programs on the processors (CPU / GPU / NPU) inside the AR smart glasses, and the data can be stored in the memory or flash memory of the glasses.

[0106] Embodiment Three The present application also provides an AR smart glasses, in addition to containing the conventional display components (such as optical waveguide display or Micro-LED display), optical components, battery, communication module (such as Wi-Fi, Bluetooth) and the like, the core control part of which integrates the AR smart glasses control system described above. The system works with the central processing unit (CPU) or dedicated digital signal processor (DSP), neural processing unit (NPU) of the AR smart glasses. The glasses are also configured with high-precision IMU (including three-axis gyroscope, three-axis accelerometer) and ambient light sensor, and the data of these sensors is directly input to the data acquisition module of the control system.

[0107] Referring to Figure 5 , a simplified AR smart glasses network architecture diagram is shown. The sensor group (IMU, ambient light sensor) transmits data to the processor (C503) through the sensor interface. The processor C503 runs the AR smart glasses control system (including Figure 4 each module shown) described in the present application. The pose data processed by the AR smart glasses control system is provided to the AR application rendering engine, which accurately superimposes virtual content into the user's real field of view according to the pose information, and presents it on the display through the display driver. User interaction data (such as interaction generated by gestures, voice or keys on the glasses) is also recorded by the system for learning of the adaptive parameter adjustment module. The system can also communicate with the cloud service through the network interface for data backup, model update or collaborative computing.

[0108] Referring to Figure 6 , a simulation comparison chart of the tilt error changing with time is shown. The horizontal axis represents time, and the vertical axis represents the tilt error (for example, the deviation from the true value). Curve 601 represents the tilt error of the conventional AR glasses without using the method of the present application, which may exhibit large fluctuations and drift. Curve 602 represents the tilt error of the AR glasses after using the method of the present application, which can be seen to be significantly reduced due to the effects of individualized adaptation, environmental correction and multi-level error compensation mechanism, and exhibits better stability. This directly translates into more accurate and smoother user experience.

[0109] In summary, the application generates a raw data sequence by acquiring the original inclination data and accelerometer data collected by the built-in sensors of the AR smart glasses, applies a low-pass filtering algorithm to the inclination values in the raw data sequence to denoise, calculates the head movement frequency and inclination change rate based on the filtered data to generate a high sensitivity mode flag or a stability priority mode flag, trains an adaptive parameter adjustment model in combination with the mode flag and user interaction records to generate personalized response parameters, adjusts the response parameters by fusing environmental light and gyroscope data to generate environment-adaptive corrected control parameters, uses the Kalman filtering algorithm to fuse the calibrated inclination and acceleration data based on the control parameters to generate predicted inclination values and activate the error correction process, adjusts the predicted inclination values in the error correction process to generate inclination data adjusted by deviation, calculates the head three-dimensional positioning information by using the adjusted inclination data and gyroscope data through the quaternion rotation algorithm to generate a final attitude angle data sequence, and learns the weights of the stability adjustment model based on the final attitude data to complete the closed-loop adjustment. The application can automatically evolve the inclination control according to the user's usage habits and environmental changes, and improve the response sensitivity and attitude recognition accuracy of the AR glasses.

[0110] The technical solution of the application has at least the following beneficial effects: 1. By learning the interaction data of the user in the past period of time, the personalized response curve and sensitivity coefficient are dynamically generated in combination with the current head movement mode (high sensitivity or stability priority), so that the inclination control of the AR glasses is more in line with the usage habits of individual users.

[0111] 2. The control parameters are adjusted in real time by fusing the environmental light and gyroscope angular velocity change information, which improves the adaptability and robustness of the AR glasses under different lighting conditions and motion states.

[0112] 3. The Kalman filtering is used for data fusion and preliminary calibration, and a deep error correction process based on multi-dimensional error analysis and benchmark resetting is designed, which significantly improves the accuracy and reliability of the inclination data.

[0113] 4. The accurate inclination data adjusted by deviation is used in combination with the quaternion rotation algorithm to more accurately calculate the head three-dimensional spatial positioning information, which provides high-quality attitude data for the upper AR application.

[0114] 5. By monitoring the stability of the final attitude angle sequence and the change of user interaction frequency, the learning weight of the adaptive parameter adjustment model is optimized, the continuous learning and evolution of the control strategy are realized, and the system performance is continuously improved over time. The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for controlling AR smart glasses, characterized by, The method comprises: obtaining original inclination data and original accelerometer data collected by sensors built in the AR intelligent glasses, and generating an original data sequence containing inclination values, acceleration values and corresponding first time markers; applying a low-pass filtering algorithm to the inclination values in the original data sequence to perform denoising processing, and obtaining filtered inclination measurement values and a corresponding second time marker sequence; calculating a head motion frequency and an inclination change rate according to the filtered inclination measurement values and the second time marker sequence, generating a high-sensitivity mode flag when the head motion frequency exceeds a preset frequency threshold, and generating a stability priority mode flag when the inclination change rate is lower than a preset rate threshold and lasts for a preset length of time; combining the high-sensitivity mode flag or the stability priority mode flag with stored user interaction record data in the past preset number of days to train an adaptive parameter adjustment model, and generating personalized response curve parameters and sensitivity adjustment coefficients; fusing a light intensity change rate obtained by an ambient light sensor and an angular velocity change rate obtained by a gyroscope, and adjusting the personalized response curve parameters and the sensitivity adjustment coefficients when the light intensity change rate exceeds a preset light intensity threshold or the angular velocity change rate exceeds a preset angular velocity threshold to generate environment-adaptive corrected control parameters; using a Kalman filtering algorithm to fuse and calibrate the filtered inclination measurement values and the accelerometer data according to the environment-adaptive corrected control parameters, generating predicted inclination values, and activating an error correction process if a deviation between the predicted inclination values and actual inclination measurement values exceeds a preset deviation threshold; generating a correction factor in the error correction process, performing a reference reset and real-time deviation adjustment on the predicted inclination values if the deviation direction and amplitude of a continuous predetermined number of sampling points are both greater than a preset error analysis threshold, and generating deviation-adjusted inclination data; updating a pose solving matrix using the deviation-adjusted inclination data, calculating head three-dimensional space positioning information through a quaternion rotation algorithm in combination with gyroscope data, and generating a final attitude angle data sequence; according to the stability of the final attitude angle data sequence, adjusting learning weights of the adaptive parameter adjustment model if an angle change variance is lower than a preset variance threshold and an improvement amplitude of a user interaction frequency exceeds a preset amplitude threshold, generating optimized inclination control parameters, and completing a closed-loop adjustment process.

2. The method of claim 1, wherein, The training of the adaptive parameter adjustment model further comprises: obtaining user interaction data in the past preset number of days from a storage system, analyzing time sequence characteristics and interaction modes in the data, and obtaining a user behavior data set; according to the user behavior data set, extracting identification features of the high-sensitivity mode and the stability priority mode, and determining mode classification labels; if the mode classification label is the high-sensitivity mode, training an adaptive parameter model using a random forest algorithm; and if the mode classification label is the stability priority mode, optimizing the adaptive parameter model using a gradient boosting decision tree algorithm. The sensitivity adjustment coefficient is calculated by the initial response curve parameter and the stable response curve parameter, and a personalized response curve is generated.

3. The method of claim 1, wherein, The light intensity change rate obtained by the fusion ambient light sensor and the angular velocity change rate obtained by the gyroscope further comprise: Raw data of the ambient light sensor and the gyroscope are obtained, noise is filtered through preprocessing, and the light intensity change rate and the angular velocity change rate are obtained; If the light intensity change rate or the angular velocity change rate exceeds a preset threshold, a linear interpolation method is used to calculate the change trend to obtain a change trend vector; According to the change trend vector, the support vector machine algorithm is used to classify the environment state and determine the environment dynamic level; According to the environment dynamic level, the preset response curve parameter table is queried to obtain the corresponding response curve parameter; The corresponding response curve parameter and the current sensitivity adjustment coefficient are fused by using the weighted average method to obtain the updated sensitivity adjustment coefficient.

4. The method of claim 3, wherein, The generation of the correction factor in the error correction process further comprises: Obtain the predicted inclination data, collect the raw inclination value by the sensor, store it as the initial data set, and obtain the predicted inclination data; Using a multi-dimensional error analysis model, a feature vector is extracted from the predicted inclination data, the deviation from the true value is calculated, and a real-time deviation calculation result is obtained; If the real-time deviation calculation result shows that the deviation directions of a continuous predetermined number of sampling points are consistent and the amplitudes exceed a preset threshold, the deviation direction consistency and the deviation amplitude over-limit are triggered, and an abnormal deviation state is obtained. For the abnormal deviation state, a reference reset operation is performed, the reference is updated by the calibration algorithm, and the reset reference data is obtained.

5. The method of claim 1, wherein, The inclination data updated by the bias adjustment further comprises: Obtain the bias adjustment data, eliminate the system error in the inclination data by sensor calibration processing, and obtain the calibrated inclination data; Using the calibrated inclination data, the gyroscope data is fused, and an optimized sensor data fusion result is generated by the Kalman filter algorithm; If the sensor data fusion result meets the preset threshold condition, the spatial attitude update matrix is calculated by the quaternion rotation algorithm, and the attitude solution matrix is obtained.

6. An AR smart glasses control system, characterized by, The system comprises: A data acquisition module is configured to acquire raw inclination data and raw accelerometer data collected by sensors built in the AR smart glasses, and generate a raw data sequence containing inclination values, acceleration values, and corresponding first time markers; A low-pass filter module is configured to apply a low-pass filter algorithm to the inclination values in the raw data sequence to perform denoising processing, and obtain filtered inclination measurement values and corresponding second time marker sequences; A mode flag generation module is configured to calculate a head motion frequency and an inclination change rate based on the filtered inclination measurement values and the second time marker sequences, generate a high sensitivity mode flag when the head motion frequency exceeds a preset frequency threshold, and generate a stability priority mode flag when the inclination change rate is lower than a preset rate threshold and lasts for a preset duration. An adaptive parameter adjustment module is configured to combine the high-sensitivity mode flag or the stability priority mode flag, and the stored user interaction record data in the past preset days, train an adaptive parameter adjustment model, and generate personalized response curve parameters and sensitivity adjustment coefficients. An environmental adaptability correction module is configured to fuse the light intensity change rate obtained by the ambient light sensor and the angular velocity change rate obtained by the gyroscope, adjust the personalized response curve parameters and sensitivity adjustment coefficients when the light intensity change rate exceeds a preset light intensity threshold or the angular velocity change rate exceeds a preset angular velocity threshold, and generate environmental adaptability corrected control parameters. A data fusion calibration module is configured to fuse and calibrate the filtered inclination measurement value and the accelerometer data by using a Kalman filtering algorithm according to the environmental adaptability corrected control parameters, generate a predicted inclination value, and activate an error correction process if the deviation between the predicted inclination value and the actual inclination measurement value exceeds a preset deviation threshold. An error correction module is configured to generate a correction factor in the error correction process, perform a reference reset and real-time deviation adjustment on the predicted inclination value to generate deviation-adjusted inclination data if the deviation direction and amplitude of a continuous predetermined number of sampling points are greater than a preset error analysis threshold. A posture solving module is configured to update a posture solving matrix by using the deviation-adjusted inclination data, calculate head three-dimensional space positioning information by using a quaternion rotation algorithm in combination with the gyroscope data, and generate a final attitude angle data sequence. A closed-loop adjustment module is configured to adjust the learning weight of the adaptive parameter adjustment model to generate optimized inclination control parameters and complete a closed-loop adjustment process if the angle change variance of the final attitude angle data sequence is lower than a preset variance threshold and the improvement amplitude of the user interaction frequency exceeds a preset amplitude threshold.

7. The system of claim 6, wherein, The adaptive parameter adjustment module is further configured to: obtain user interaction data in the past preset days from a storage system, analyze the time sequence characteristics and interaction modes in the data to obtain a user behavior data set; extract identification features of the high-sensitivity mode and the stability priority mode according to the user behavior data set, and determine mode classification labels; train an adaptive parameter model by using a random forest algorithm if the mode classification label is the high-sensitivity mode, or optimize the adaptive parameter model by using a gradient boosting decision tree algorithm if the mode classification label is the stability priority mode; calculate the sensitivity adjustment coefficients by using the preliminary response curve parameters and the stable response curve parameters, and generate personalized response curves.

8. The system of claim 6, wherein, The environmental adaptability correction module is further configured to: obtain original data of the ambient light sensor and the gyroscope, filter out noise by preprocessing, and obtain the light intensity change rate and the angular velocity change rate; calculate the change trend by using a linear interpolation method to obtain a change trend vector if the light intensity change rate or the angular velocity change rate exceeds a preset threshold; classify the environmental state by using a support vector machine algorithm according to the change trend vector, and determine the environmental dynamic level; and By the environment dynamic level, a preset response curve parameter table is inquired to obtain corresponding response curve parameters; The corresponding response curve parameters and the current sensitivity adjustment coefficient are fused by using a weighted average method to obtain an updated sensitivity adjustment coefficient.

9. An AR smart glass, characterized by, The AR smart glasses control system of any one of claims 6 to 8 is electrically connected with the processor, the memory, and the display. 10.The AR smart glasses of claim 9, wherein, Further comprising: At least one inclination sensor, at least one accelerometer sensor, at least one gyroscope sensor, and at least one ambient light sensor are connected with the data acquisition module and the environment adaptive correction module, and are used to provide the original inclination data, the original accelerometer data, the gyroscope data, and the light intensity change data.

Images