Shooting training posture correction feedback method and system based on intelligent analysis

Abstract

The invention provides a shooting training posture correction feedback method and system based on intelligent analysis, and the method comprises the steps: firstly obtaining a real-time posture data set of a target object, the real-time posture data set comprises a plurality of groups of posture collection data which are continuous in time and cover motion tracks and physiological state features, then carrying out the motion feature extraction, and obtaining a motion feature extraction result; generating an attitude correction parameter set containing an action adjustment direction parameter and a physiological coordination optimization parameter based on the dynamic difference between the target and the abnormal action feature set, and generating a real-time feedback instruction sequence according to the attitude correction parameter set; and finally, the real-time feedback instruction sequence is sent to a posture adjusting device to execute training posture correction operation, so that actions and physiological states can be comprehensively considered, and intelligent and precise shooting training posture correction is realized.

Description

Technical Field

[0001] The present application relates to the technical field of shooting training, and in particular to a shooting training posture correction feedback method and system based on intelligent analysis. Background Art

[0002] In shooting training, good shooting posture is crucial for improving accuracy and training effectiveness. Traditional methods for correcting shooting posture rely primarily on on-site coaching and the athlete's own accumulated experience. The coach visually observes the athlete's shooting posture, draws on their experience to determine if there are any issues, and provides corrective suggestions. However, this approach has significant limitations.

[0003] On the one hand, a coach's observation is subjective and limited. Different coaches may use different criteria, making it difficult to comprehensively and accurately analyze an athlete's posture. For example, a coach may not be able to detect subtle deviations in movement or changes in physiological state in a timely and accurate manner. On the other hand, manual guidance cannot provide real-time, continuous monitoring and feedback. Athletes may not be aware of any posture issues during training, resulting in poor training results.

[0004] Furthermore, while existing auxiliary training devices can monitor some of an athlete's movements, most only provide single-dimensional information, such as tracking movement trajectory, while ignoring the impact of the athlete's physiological state on their shooting posture. Furthermore, these devices lack intelligent analysis and automatic correction capabilities, unable to generate personalized correction plans based on the athlete's actual situation and provide real-time feedback to the athlete. Consequently, they fail to meet the requirements of modern shooting training for efficiency, precision, and intelligence. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a shooting training posture correction feedback method based on intelligent analysis, the method comprising: Acquire a real-time posture data set of the target object, wherein the real-time posture data set includes multiple sets of time-continuous posture acquisition data, each set of posture acquisition data includes motion trajectory characteristics and physiological state characteristics; Performing motion feature extraction processing on the real-time posture data set to obtain a target motion feature set and an abnormal motion feature set; generating a posture correction parameter set based on a dynamic difference between the target motion feature set and the abnormal motion feature set, the posture correction parameter set including a motion adjustment direction parameter and a physiological coordination optimization parameter; generating a real-time feedback instruction sequence according to the posture correction parameter set; The real-time feedback instruction sequence is sent to the posture adjustment device to perform the training posture correction operation.

[0006] On the other hand, an embodiment of the present application also provides a shooting training posture correction feedback system based on intelligent analysis, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiment of the present invention obtains a real-time posture data set containing multiple sets of time-continuous and covering motion trajectory features and physiological state features, performs motion feature extraction processing on the real-time posture data set to obtain a target motion feature set and an abnormal motion feature set, and generates a posture correction parameter set containing motion adjustment direction parameters and physiological coordination optimization parameters based on the dynamic difference between the two, thereby achieving comprehensive consideration and optimization of the target object's motion and physiological state. A real-time feedback instruction sequence is generated based on the posture correction parameter set, so that the posture adjustment device can perform mechanical movements corresponding to the corresponding parameters, achieving intelligent and automated posture adjustment. The real-time feedback instruction sequence is sent to the posture adjustment device to perform training posture correction operations, breaking the limitations of traditional shooting training posture correction relying on human experience and subjective judgment, and achieving efficient, accurate, and personalized posture correction based on multi-dimensional data intelligent analysis, significantly improving the scientific nature and effectiveness of shooting training, and enhancing the intelligence level and adaptability of the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the execution flow of the shooting training posture correction feedback method based on intelligent analysis provided in an embodiment of the present application.

[0009] Figure 2 This is a schematic diagram of the hardware architecture of the shooting training posture correction feedback system based on intelligent analysis provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a shooting training posture correction feedback method based on intelligent analysis provided by an embodiment of the present application. The shooting training posture correction feedback method based on intelligent analysis is introduced in detail below.

[0011] Step S110: obtaining a real-time posture data set of the target object, wherein the real-time posture data set includes multiple groups of time-continuous posture acquisition data, and each group of posture acquisition data includes motion trajectory features and physiological state features.

[0012] In this embodiment, to obtain a real-time posture data set for the target object, the first step is to collect a raw posture signal stream, which can be accomplished using a distributed posture sensor array. The distributed posture sensor array is composed of multiple types of sensors, which are respectively arranged at key locations on the target object's body. These sensors can collect multiple signals, including joint displacement signals, muscle tension signals, and respiratory rate signals. Joint displacement signals can reflect the position changes of the target object's joints in space, muscle tension signals reflect the degree of muscle tension, and respiratory rate signals reflect the target object's respiratory state.

[0013] Step S111: collecting the original posture signal stream of the target object through a distributed posture sensor array, where the original posture signal stream includes joint displacement signals, muscle tension signals, and respiratory rate signals.

[0014] When deploying a distributed posture sensor array, angle sensors can be installed at various joints, such as the shoulder, elbow, and knee, to collect joint displacement signals. These angle sensors can measure joint rotation angles in real time, generating joint displacement signals. For example, for the shoulder joint, angle sensors can measure shoulder rotation angles in different directions. The changes in these angles over time constitute the joint displacement signal. To collect muscle tension signals, electromyographic (EMG) sensors are attached to the surfaces of key muscles, such as the biceps and triceps in the arms and the quadriceps in the legs. EMG sensors detect the electrical signals generated by muscle contraction. The strength of these signals is correlated with muscle tension. By processing and analyzing these electrical signals, muscle tension signals can be obtained. To collect respiratory rate signals, respiration sensors can be placed on the chest or abdomen. When the subject breathes, the chest or abdomen rises and falls, which are detected by the respiration sensor and converted into a respiratory rate signal.

[0015] Step S112: performing time stamp alignment processing on the original gesture signal stream to obtain a synchronized gesture signal stream, in which data collected by different signal sources in the synchronized gesture signal stream have a unified time stamp sequence.

[0016] Because different sensor types may collect data at different times, timestamp alignment is necessary. The timestamp of one sensor can be selected as the base timestamp, T0. Assume the timestamp of the joint displacement signal is Ta, the timestamp of the muscle tension signal is Tb, and the timestamp of the respiratory rate signal is Tc. For each sensor's data, calculate the time difference from the base timestamp, T0: ΔTa = Ta - T0, ΔTb = Tb - T0, and ΔTc = Tc - T0. Then, adjust the data based on these time differences so that all data timestamps are aligned to a time sequence based on T0. This results in a synchronized posture signal stream, where data collected from different signal sources has a unified timestamp sequence, facilitating subsequent processing and analysis.

[0017] Step S113: performing multi-channel normalization processing on the joint displacement signals, muscle tension signals and respiratory rate signals in the synchronized posture signal stream, and performing signal fusion processing on the normalized synchronized posture signal stream to generate a fused posture data block, wherein each data unit in the fused posture data block contains a unified dimension of motion trajectory code and physiological state code.

[0018] Multi-channel normalization processing is to eliminate the dimensional differences between different signals. For the joint displacement signal, let its original data be Sa, and it can be normalized to the interval [0, 1] through linear transformation. First, determine the maximum value Smax_a and minimum value Smin_a of the joint displacement signal, and then perform normalization processing according to the formula Sa'=(Sa-Smin_a) / (Smax_a-Smin_a) to obtain the normalized joint displacement signal Sa'. For the muscle tension signal, let its original data be Sb, and similarly determine its maximum value Smax_b and minimum value Smin_b, and perform normalization processing according to the formula Sb'=(Sb-Smin_b) / (Smax_b-Smin_b) to obtain the normalized muscle tension signal Sb'. For the respiratory rate signal, let its original data be Sc, determine the maximum value Smax_c and the minimum value Smin_c, and perform normalization processing using the formula Sc'=(Sc-Smin_c) / (Smax_c-Smin_c) to obtain the normalized respiratory rate signal Sc'.

[0019] The normalized synchronized posture signal stream contains the normalized joint displacement signal Sa', muscle tension signal Sb' and respiratory rate signal Sc'. Next, signal fusion processing is performed, which can be done by weighted splicing. Let the weight of the joint displacement signal be Wa, the weight of the muscle tension signal be Wb, the weight of the respiratory rate signal be Wc, and Wa+Wb+Wc=1. The normalized signals are spliced according to the weights, that is, each data unit D in the fused posture data block is [Wa*Sa', Wb*Sb', Wc*Sc']. Each data unit in the generated fused posture data block contains a unified dimension of motion trajectory code and physiological state code. The motion trajectory code is mainly reflected by the joint displacement signal, and the physiological state code is reflected by the muscle tension signal and the respiratory rate signal.

[0020] Step S114: segmenting and cutting the fused posture data block according to a preset shooting action cycle to obtain the multiple groups of time-continuous posture acquisition data, each group of posture acquisition data corresponding to a complete shooting action cycle.

[0021] The preset shooting action cycle is a time interval T pre-set based on the characteristics and requirements of shooting training. When segmenting the fused posture data block, the segmentation is based on time, starting from the start time of the fused posture data block and performed every shooting action cycle T. Assuming the time range of the fused posture data block is from T_start to T_end, the time interval of the first posture acquisition data is [T_start, T_start+T], the time interval of the second posture acquisition data is [T_start+T, T_start+2T], and so on, until the segmentation reaches time T_end. This results in multiple sets of temporally continuous posture acquisition data, each set of posture acquisition data corresponding to a complete shooting action cycle, and each set of data contains motion trajectory characteristics and physiological state characteristics.

[0022] Step S120: performing motion feature extraction processing on the real-time posture data set to obtain a target motion feature set and an abnormal motion feature set.

[0023] After obtaining the real-time posture data set, it is necessary to extract motion features to distinguish target motion features from abnormal motion features. This process requires the use of a pre-trained motion decomposition model and a series of feature processing operations.

[0024] Step S121: calling a pre-trained motion decomposition model to perform hierarchical feature analysis on the motion trajectory encoding in each set of posture acquisition data to obtain an initial motion feature unit set.

[0025] The pre-trained motion decomposition model, trained on a large amount of historical data, is capable of performing effective hierarchical feature analysis on motion trajectory encodings. The motion trajectory encodings contain information about the target's joint displacements during the shooting cycle, reflecting the trajectory of its motion. The motion trajectory encodings from each set of posture acquisition data are input into the pre-trained motion decomposition model, which then performs a multi-level analysis. First, the model performs preliminary feature extraction on the motion trajectory encodings, identifying basic motion elements such as joint rotation direction and amplitude. The model then further combines and analyzes these basic motion elements to identify higher-level motion features, such as the overall gun-holding posture and aiming trajectory. This hierarchical feature analysis yields an initial set of motion feature units, each of which represents a specific feature within the motion trajectory.

[0026] Step S122: performing feature dimension normalization processing on the gun holding stability feature, aiming trajectory smoothness feature, and firing rhythm consistency feature in the initial action feature unit set, mapping features of different dimensions into a unified metric space, and generating a basic action feature unit set.

[0027] The initial action feature unit set contains features of various dimensions. In order to facilitate subsequent processing and analysis, it is necessary to standardize the feature dimensions of the gun holding stability feature, the aiming trajectory smoothness feature, and the firing rhythm consistency feature. For the gun holding stability feature, let its original feature value be F1, and its feature dimension may be a value related to the amplitude of shaking when holding the gun. For the aiming trajectory smoothness feature, let its original feature value be F2, and its feature dimension may be a value related to the degree of tortuosity of the trajectory during the aiming process. For the firing rhythm consistency feature, let its original feature value be F3, and its feature dimension may be a value related to the stability of the firing time interval.

[0028] To map these features of different dimensions into a unified metric space, a standardization method can be used. First, the mean and standard deviation of each feature are calculated. For the gun-holding stability feature F1, its mean is μ1 and its standard deviation is σ1; for the aiming trajectory smoothness feature F2, its mean is μ2 and its standard deviation is σ2; and for the firing rhythm consistency feature F3, its mean is μ3 and its standard deviation is σ3. The features are then standardized according to the formulas F1'=(F1-μ1) / σ1, F2'=(F2-μ2) / σ2, and F3'=(F3-μ3) / σ3, resulting in the standardized feature values F1', F2', and F3'. These standardized features are combined to generate a set of basic action feature units. Each element in this set of basic action feature units is in a unified metric space, facilitating subsequent pattern matching and classification operations.

[0029] Step S123: performing pattern matching processing on the basic action feature unit set, clustering the action feature units that meet the preset shooting standard into target action clusters, and marking the action feature units that deviate from the preset shooting standard as abnormal action clusters.

[0030] The preset shooting standard is a series of action patterns and feature ranges that are pre-set according to the requirements and specifications of shooting training. When performing pattern matching processing on the basic action feature unit set, each action feature unit in the set is compared with the preset shooting standard. For each action feature unit, if its feature value is within the range of the preset shooting standard, it is considered that the action feature unit meets the preset shooting standard and is clustered into the target action cluster. For example, for the gun holding stability feature, if its standardized feature value F1' is within the preset gun holding stability standard range, the gun holding stability feature of the action feature unit meets the requirements and the action feature unit can be clustered into the target action cluster. If the feature value of the action feature unit deviates from the range of the preset shooting standard, it is marked as an abnormal action cluster. For example, if the standardized feature value F2' of the aiming trajectory smoothness feature exceeds the preset aiming trajectory smoothness standard range, the aiming trajectory smoothness feature of the action feature unit does not meet the requirements and is marked as an abnormal action cluster.

[0031] Step S124: performing feature enhancement processing on the action feature units in the target action cluster to generate the target action feature set, wherein the feature enhancement processing includes noise suppression and feature dimension expansion.

[0032] Feature enhancement is performed on the action feature units in the target action cluster to improve the quality and discernibility of the target action features. First, noise suppression is performed. Since noise may be introduced during data acquisition and processing, this noise can affect the accuracy of the target action features. Filtering can be used to suppress noise in the action feature units. For example, the gun-holding stability feature may contain some random, subtle fluctuations, which can be considered noise. Using a low-pass filter, the gun-holding stability feature is filtered to remove high-frequency noise, making the feature smoother and more stable.

[0033] Feature dimension expansion is then performed, enriching the target action feature information by adding relevant feature dimensions. For example, for the gun-holding stability feature, the hand muscle tension feature can be added as a new feature dimension. The original gun-holding stability feature and the newly added hand muscle tension feature are combined to form a higher-dimensional feature vector. This feature enhancement process is performed on all action feature units in the target action cluster, ultimately generating a target action feature set. Each feature vector in the target action feature set contains the enhanced target action feature.

[0034] Step S125: performing deviation quantization processing on the motion feature units in the abnormal motion cluster to generate the abnormal motion feature set, wherein the deviation quantization processing includes motion offset angle calculation and physiological state fluctuation amplitude statistics.

[0035] Deviation quantification is performed on the action feature units in the abnormal action cluster to accurately describe the degree and characteristics of the abnormal action. First, the action offset angle is calculated. For the action trajectory features, such as the gun holding posture and the aiming trajectory, the offset angle from the preset standard action is calculated. Assuming that the gun holding posture of the preset standard action is vector V0, and the gun holding posture of the abnormal action is vector V1, the action offset angle is determined by calculating the angle θ between the two vectors. For example, the dot product formula of the vector can be used to calculate the cosine value of the angle cosθ=(V0·V1) / (|V0|*|V1|), and then the angle θ is obtained by the inverse trigonometric function. Such action offset angle calculation is performed for the action trajectory features of each abnormal action feature unit.

[0036] Next, we perform statistics on the fluctuation amplitude of physiological states. For physiological state features in abnormal motion feature units, such as muscle tension signals and respiratory rate signals, we calculate their fluctuation amplitudes during periods of abnormal motion. Assuming the muscle tension signal's mean value under normal conditions to be μ_m, its maximum value to be M_m, and its minimum value to be m_m during periods of abnormal motion, we then calculate the fluctuation amplitude of the muscle tension signal, ΔM_m = M_m - m_m. Similarly, for the respiratory rate signal, assuming its mean value under normal conditions to be μ_r, its maximum value to be Mr_r, and its minimum value to be m_r during periods of abnormal motion, we then calculate the fluctuation amplitude of the respiratory rate signal, ΔM_r = Mr_r - m_r. The motion offset angle and the physiological state fluctuation amplitude are combined to form the deviation quantization result for each abnormal motion feature unit. This deviation quantization process is performed on all motion feature units in the abnormal motion cluster, ultimately generating an abnormal motion feature set. Each feature vector in the abnormal motion feature set contains the deviation quantization information for the abnormal motion.

[0037] Step S130: generating a posture correction parameter set based on the dynamic difference between the target motion feature set and the abnormal motion feature set, wherein the posture correction parameter set includes a motion adjustment direction parameter and a physiological coordination optimization parameter.

[0038] In order to generate a set of posture correction parameters, it is necessary to first build a dynamic difference evaluation model, through which the dynamic difference between the target action feature set and the abnormal action feature set is calculated, and then the corresponding posture correction parameters are generated based on the difference.

[0039] Step S131: constructing a dynamic difference evaluation model, wherein the dynamic difference evaluation model includes a motion space difference measurement module and a physiological state coupling analysis module.

[0040] The dynamic difference evaluation model consists of two main modules: the motion space difference measurement module and the physiological state coupling analysis module. The motion space difference measurement module calculates the difference between the target motion feature set and the abnormal motion feature set in the motion space, focusing on differences in gun stability, aiming trajectory, and firing rhythm. The physiological state coupling analysis module analyzes the correlation between multidimensional difference vectors and physiological state features, determining the correlation weight matrix to comprehensively consider the impact of physiological state on motion differences.

[0041] Step S132: Calculating the multi-dimensional difference vectors between the target action feature set and the abnormal action feature set in the gun holding stability dimension, the aiming trajectory dimension, and the firing rhythm dimension through the action space difference measurement module.

[0042] For the gun holding stability dimension, let the gun holding stability feature vector in the target action feature set be Gt, and the gun holding stability feature vector in the abnormal action feature set be Ga. To calculate the difference between the two vectors, the difference vector ΔG = Gt-Ga can be obtained by vector subtraction. Similarly, for the aiming trajectory dimension, let the aiming trajectory feature vector in the target action feature set be Pt, and the aiming trajectory feature vector in the abnormal action feature set be Pa, and the difference vector ΔP = Pt-Pa. For the firing rhythm dimension, let the firing rhythm feature vector in the target action feature set be Rt, and the firing rhythm feature vector in the abnormal action feature set be Ra, and the difference vector ΔR = Rt-Ra. The difference vectors of these three dimensions are combined to form a multidimensional difference vector D = [ΔG, ΔP, ΔR].

[0043] Step S133: determining the correlation weight matrix between the multi-dimensional difference vector and the physiological state feature through the physiological state coupling analysis module.

[0044] Step S1331: performing multi-channel normalization processing on the physiological state characteristics to generate a set of standardized physiological indicators with consistent unit dimensions, wherein the set of standardized physiological indicators includes a normalized value of respiratory rate, a normalized value of muscle tension, and a normalized value of heart rate variability coefficient.

[0045] When determining the association weight matrix between the multidimensional difference vector and physiological state characteristics, the physiological state characteristics must first be subjected to multi-channel normalization. Physiological state characteristics include different types of data, such as respiratory rate, muscle tension, and heart rate variability, which have different dimensions and value ranges. To eliminate dimensional differences and facilitate subsequent calculations and analysis, normalization is required. For respiratory rate, let the original data be R, determine the maximum value Rmax and minimum value Rmin of respiratory rate, and normalize them to the interval [0, 1] using a formula to obtain the normalized respiratory rate value R'. For muscle tension, let the original data be M, determine the maximum value Mmax and minimum value Mmin of muscle tension, and similarly perform normalization to obtain the normalized muscle tension value M'. For heart rate variability, let the original data be H, determine the maximum value Hmax and minimum value Hmin of heart rate variability, and normalize them to obtain the normalized heart rate variability value H'. These three normalized values are combined to generate a standardized physiological indicator set S = [R', M', H'] with consistent unit dimensions.

[0046] Step S1332: extracting the action difference parameters of each dimension in the multi-dimensional difference vector, wherein the action difference parameters include the gun holding stability offset, the aiming trajectory discreteness, and the firing rhythm deviation rate.

[0047] The motion difference parameters of each dimension are extracted from the multidimensional difference vector. The multidimensional difference vector is calculated in the motion space difference measurement module. It contains the difference information between the target motion feature set and the abnormal motion feature set in the gun holding stability dimension, the aiming trajectory dimension, and the firing rhythm dimension. For the gun holding stability dimension, its offset is extracted, that is, the difference between the gun holding stability features in the target motion feature set and the gun holding stability features in the abnormal motion feature set, recorded as the gun holding stability offset ΔGp. For the aiming trajectory dimension, its discreteness is extracted. The discreteness can reflect the irregularity of the aiming trajectory and is recorded as the aiming trajectory discreteness ΔPp. For the firing rhythm dimension, its deviation rate is extracted, that is, the deviation ratio between the firing rhythm features in the target motion feature set and the firing rhythm features in the abnormal motion feature set, recorded as the firing rhythm deviation rate ΔRp.

[0048] Step S1333: Calculate the dynamic correlation coefficient between each of the standardized physiological indicators and each of the motion difference parameters to generate a physiological-motion correlation matrix, wherein the dynamic correlation coefficient is determined by dividing the covariance within the sliding time window by the product of the standard deviations of the two variables.

[0049] The dynamic correlation coefficient between each indicator in the set of standardized physiological indicators and the motion difference parameter is calculated. Taking the normalized respiratory rate value R' and the gun-holding stability offset ΔGp as examples, a sliding time window method is used to calculate the dynamic correlation coefficient between them. Within a sliding time window, the covariance of the normalized respiratory rate value R' and the gun-holding stability offset ΔGp is calculated. The covariance can measure the degree of linear relationship between the two variables. The standard deviation of the normalized respiratory rate value R' and the gun-holding stability offset ΔGp is then calculated. The covariance is divided by the product of the standard deviations of the two variables to obtain the dynamic correlation coefficient rR_G within the sliding time window. This calculation is performed for each combination of the indicator and the motion difference parameter in the set of standardized physiological indicators, ultimately generating a physiological-motion correlation matrix C. The element Cij in the matrix represents the dynamic correlation coefficient between the i-th standardized physiological indicator and the j-th motion difference parameter.

[0050] Step S1334: determining the physiological state influence weight factor according to the module length of each row vector in the physiological-action association matrix, wherein the module length is calculated by the Euclidean norm and compressed to a preset weight interval by a logarithmic function.

[0051] The physiological state influence weight factor is determined based on the modulus of each row vector in the physiological-motion correlation matrix. For the i-th row vector Ci of the physiological-motion correlation matrix C, it contains the dynamic correlation coefficient between the i-th standardized physiological index and each motion difference parameter. To calculate the modulus of the row vector, the Euclidean norm is used. The Euclidean norm is the square root of the sum of the squares of the vector elements. Each element of the row vector can be squared and then these squared values are added to obtain a sum value. This sum value represents a certain "length" measure of the row vector in multidimensional space.

[0052] However, the calculated modulus may have a large value range, which is not conducive to the subsequent weight distribution. Therefore, the modulus needs to be compressed by a logarithmic function. The logarithmic function can compress a large numerical range into a relatively small preset weight interval. The preset weight interval is a range pre-set according to actual needs, for example, it can be set to the [0, 1] interval. After compression by the logarithmic function, a value within the preset weight interval is obtained, which is the physiological state influence weight factor Wi corresponding to the i-th standardized physiological indicator. By performing such processing on each row vector in the physiological-action association matrix C, a set of physiological state influence weight factors {W1, W2, W3} corresponding to all standardized physiological indicators is obtained, which correspond to the physiological state influence weight factors of the normalized value of respiratory frequency, the normalized value of muscle tension and the normalized value of heart rate variability coefficient, respectively.

[0053] Step S1335: Based on the physiological state influence weight factor and the confidence score of the corresponding indicator in the preset expert experience weight table, a priority weighted fusion strategy is adopted to generate the associated weight matrix, wherein the expert experience weight table stores the priority coefficients of physiological indicators to action differences in different shooting stages.

[0054] The pre-set expert experience weighting table, developed by shooting experts based on their extensive experience and research, stores the priority coefficients for each physiological indicator in relation to movement differences during different shooting phases. Each physiological indicator has varying degrees of influence on movement differences at different shooting phases, and the expert experience weighting table captures these differences. For example, during the aiming phase, breathing rate may have a greater impact on the accuracy of the aiming trajectory, so its corresponding priority coefficient in the expert experience weighting table may be relatively high. Meanwhile, during the firing phase, muscle tension may be more critical to the stability of the shot, so its corresponding priority coefficient may be relatively high.

[0055] For each standardized physiological indicator, its corresponding physiological state impact weight factor is combined with the confidence score of the indicator in the current shooting phase from the expert experience weight table. The confidence score is a quantitative representation of the priority coefficient in the expert experience weight table, which reflects the expert's confidence in the physiological indicator's impact on the action difference in the current shooting phase. A priority weighted fusion strategy is adopted, which weights and fuses the physiological state impact weight factor and confidence score based on their importance. Specifically, for the i-th standardized physiological indicator, its physiological state impact weight factor Wi is weighted and calculated with the confidence score Si of the indicator in the current shooting phase from the expert experience weight table to obtain a new weight value Wij, where j represents the corresponding action difference parameter.

[0056] This calculation is performed for each combination of standardized physiological indicators and each action difference parameter, ultimately generating the association weight matrix W. The element Wij in the association weight matrix W represents the association weight of the i-th standardized physiological indicator to the j-th action difference parameter. This matrix takes into account the dynamic correlation of the physiological state characteristics themselves and the expert experience's assessment of their impact at different shooting stages, and can more accurately reflect the association between physiological state characteristics and action differences.

[0057] Step S134: performing weighted fusion on the multi-dimensional difference vectors according to the association weight matrix to generate a comprehensive difference score.

[0058] Perform a weighted fusion of the association weight matrix W and the multidimensional difference vector D. Let Wi be the i-th row vector of the association weight matrix W, and Di be the i-th element of the multidimensional difference vector D. By calculating the weighted sum Sd = Σ(Wi * Di) (where i ranges from 1 to the number of dimensions), we obtain the comprehensive difference score Sd. This comprehensive difference score comprehensively considers both spatial differences in motion and the relationship between physiological state and motion differences, and can more comprehensively reflect the degree of difference between the target motion feature set and the abnormal motion feature set.

[0059] Step S135: matching the motion adjustment direction parameter and the physiological coordination optimization parameter in a preset correction parameter mapping table based on the comprehensive difference score.

[0060] The preset correction parameter mapping table is a pre-established mapping relationship table, which stores the motion adjustment direction parameters and physiological coordination optimization parameters corresponding to different comprehensive difference scores. According to the calculated comprehensive difference score Sd, the corresponding motion adjustment direction parameter A and physiological coordination optimization parameter P are searched in the preset correction parameter mapping table. The motion adjustment direction parameter A is used to indicate the direction and amplitude of the target object's motion adjustment, such as whether the gun holding posture needs to be fine-tuned to the left or right, or whether the aiming angle needs to be adjusted up or down, etc. The physiological coordination optimization parameter P is optimized and adjusted according to the physiological state of the target object, such as adjusting the breathing rhythm, balancing muscle tension, etc. In this way, a posture correction parameter set including the motion adjustment direction parameter and the physiological coordination optimization parameter is obtained.

[0061] Step S140: generating a real-time feedback instruction sequence according to the posture correction parameter set.

[0062] After obtaining the posture correction parameter set, it needs to be converted into a real-time feedback instruction sequence so that the posture adjustment device can make corresponding motion adjustments based on these instructions.

[0063] Step S141: parsing the mechanical motion constraint conditions corresponding to the motion adjustment direction parameters, wherein the mechanical motion constraint conditions include a motion amplitude threshold and a motion direction priority.

[0064] In this embodiment, when analyzing the mechanical motion constraints corresponding to the action adjustment direction parameters, the motion amplitude threshold must first be determined. The motion amplitude threshold is intended to ensure that the posture adjustment device remains within a safe and reasonable range when performing mechanical motion. Different motion adjustment direction parameters have different corresponding motion amplitude thresholds. For example, in a support surface tilt operation, the horizontal tilt amplitude may have a maximum threshold and a minimum threshold, and the vertical tilt amplitude also has a corresponding threshold.

[0065] Assuming that the action adjustment direction parameter indicates that the support surface needs to be tilted in the horizontal direction, it is necessary to determine the maximum allowable angle and the minimum allowable angle of the horizontal tilt. These two angle values are the motion amplitude thresholds for the horizontal tilt operation. These thresholds can be obtained by querying a pre-set parameter table or according to the physical properties of the posture adjustment device. The parameter table stores the motion amplitude threshold information corresponding to different action adjustment directions. According to the current action adjustment direction parameter, the corresponding threshold range is searched in the parameter table. The physical properties of the posture adjustment device will also limit the motion amplitude threshold. For example, the maximum stroke of the hydraulic actuator, the strength of the structure and other factors will determine the upper limit of the motion amplitude.

[0066] In addition to the motion amplitude threshold, it's also necessary to analyze the motion direction priority corresponding to the action adjustment direction parameters. In actual shooting training, posture correction may require adjustments in multiple directions simultaneously. For example, the support surface may need to be tilted both horizontally and vertically, or fine-tuned at multiple angles. In this case, it's necessary to determine which direction of adjustment has a higher priority.

[0067] The priority of movement direction can be determined based on a variety of factors. One way is to determine it based on the urgency of adjustments in each direction in the action adjustment direction parameters. If adjustments in a certain direction are more critical to improving the shooting posture, then the adjustment priority in that direction will be higher. For example, when it is found that the aiming deviation of the target object is mainly caused by incorrect tilt of the support surface in the horizontal direction, then the priority of the horizontal tilt adjustment will be higher than the vertical adjustment. Another way is to determine the priority based on preset rules. The preset rules can be derived from a large amount of experimental data and experience, for example, stipulating that under normal circumstances, adjustments in the horizontal direction take precedence over adjustments in the vertical direction, or setting different priority orders according to different shooting stages.

[0068] Step S142: generating a first control signal sequence according to the mechanical motion constraint condition, wherein the first control signal sequence is used to drive the hydraulic actuator of the posture adjustment device to complete a support surface tilting operation at a specified angle.

[0069] When generating the first control signal sequence based on mechanical motion constraints, the motion amplitude threshold must be considered. For a hydraulic actuator's support surface tilt operation, the first control signal sequence must include information about the tilt amplitude, and this amplitude information must be within the motion amplitude threshold.

[0070] When the motion adjustment direction parameter requires the support surface to tilt, the required tilt amplitude is checked and adjusted based on the motion amplitude threshold. If the required tilt amplitude exceeds the maximum threshold, the first control signal sequence adjusts the tilt amplitude to the maximum threshold; if the required tilt amplitude is less than the minimum threshold, the first control signal sequence adjusts the tilt amplitude to the minimum threshold. For example, if the motion adjustment direction parameter requires the support surface to tilt horizontally by an angle A, but this angle A is greater than the maximum threshold for horizontal tilt, Amax, the first control signal sequence sets the horizontal tilt angle to Amax. This ensures that the hydraulic actuator does not exceed a safe and reasonable range when performing support surface tilt operations.

[0071] In addition to the amplitude constraint, the first control signal sequence also needs to determine the execution order based on the motion direction priority. When there are adjustment requirements in multiple directions, the first control signal sequence will generate control instructions in the order of motion direction priority.

[0072] Assume both horizontal and vertical tilt adjustments are required, with horizontal adjustment taking priority over vertical. The first control signal sequence generates control instructions for the horizontal tilt operation first. Once the horizontal tilt operation is complete, the vertical tilt control instruction is generated. This ensures that the posture adjustment device executes mechanical movements in a reasonable order, improving the efficiency and accuracy of posture correction.

[0073] Step S143: analyzing the biofeedback regulation strategy corresponding to the physiological coordination optimization parameter, wherein the biofeedback regulation strategy includes a respiratory rhythm synchronization mode and a muscle tension balance mode.

[0074] When analyzing the biofeedback regulation strategy corresponding to the physiological coordination optimization parameters, it is necessary to clarify the specific regulation method and goal of the respiratory rhythm synchronization mode. The purpose of the respiratory rhythm synchronization mode is to guide the target subject to adjust his or her respiratory rhythm to match the preset optimal respiratory rhythm.

[0075] The preset optimal breathing rhythm is developed based on the characteristics and requirements of shooting training. Different shooting scenarios and phases may have different optimal breathing rhythms. For example, during the aiming phase, a slower, more stable breathing rhythm may be required to improve aiming stability; during the firing phase, a short breath hold may be required to prevent breathing from interfering with the shooting action. Physiological coordination optimization parameters include information about the optimal breathing rhythm, such as breathing frequency and the ratio of inhalation to exhalation duration.

[0076] When analyzing respiratory rhythm synchronization patterns, this information needs to be translated into specific regulation strategies. For example, if the optimal breathing rhythm requires X breaths per minute and a Y:Z ratio of inhalation to exhalation duration, the biofeedback regulation strategy will generate corresponding control instructions based on these parameters, which are used to drive the tactile feedback module to guide the target subject's breathing in a specific way.

[0077] In addition to the respiratory rhythm synchronization mode, the physiological coordination optimization parameters also correspond to the muscle tension balance mode. This mode aims to apply different degrees of stimulation to the target subject's local muscle groups by adjusting the tactile feedback module of the posture adjustment device to balance muscle tension.

[0078] Physiological coordination optimization parameters include target information regarding muscle tension balance, such as which muscle groups need to relax and which need to be more contracted. When analyzing muscle tension balance patterns, a specific adjustment strategy needs to be determined based on this target information. For example, if the target subject's arm muscles are found to be overly tense while their leg muscles are relatively relaxed, the biofeedback adjustment strategy will generate control instructions, causing the tactile feedback module to apply a smaller pressure stimulus to the arm to promote relaxation of the arm muscles, and a larger pressure stimulus to the leg to strengthen the leg muscles' contraction, thereby achieving balanced muscle tension.

[0079] Step S144: generating a second control signal sequence according to the biofeedback regulation strategy, wherein the second control signal sequence is used to drive the tactile feedback module of the posture adjustment device to perform vibration frequency modulation and pressure gradient distribution operations.

[0080] When generating the second control signal sequence according to the biofeedback regulation strategy, a corresponding breathing guidance portion is generated for the breathing rhythm synchronization mode. This portion of the control signal is used to drive the tactile feedback module to guide the target subject's breathing at a specific vibration frequency and pattern.

[0081] Based on the optimal breathing rhythm information obtained through analysis, such as breathing frequency, the ratio of the duration of inspiration and exhalation, etc., the corresponding vibration frequency and pattern are generated. For example, if the optimal breathing rhythm requires X breaths per minute, the tactile feedback module will vibrate at a frequency of X times per minute, and the duration and interval of each vibration will be adjusted according to the ratio of the duration of inspiration and exhalation. During the inspiration phase, the tactile feedback module may vibrate in a gradually increasing vibration pattern to guide the target object to inhale; during the exhalation phase, it may vibrate in a gradually decreasing vibration pattern to guide the target object to exhale. In this way, the tactile feedback module can effectively guide the target object to adjust its breathing rhythm to match the optimal breathing rhythm.

[0082] For the muscle tension balance mode, a muscle stimulation portion of the second control signal sequence is generated, which is used to drive the tactile feedback module to apply different degrees of pressure stimulation to different muscle groups of the target object.

[0083] Based on the analyzed muscle tension balance target information, the amount of pressure required for each muscle group is determined. For example, if a muscle group needs to relax, the tactile feedback module applies a lower pressure stimulus; if a muscle group needs to contract more strongly, the tactile feedback module applies a higher pressure stimulus. This pressure information is converted into specific control instructions and added to the second control signal sequence. After receiving these control instructions, the tactile feedback module applies the specified amount of pressure to the corresponding muscle group to achieve balanced muscle tension.

[0084] Step S145: Adding timestamp synchronization marks to the first control signal sequence and the second control signal sequence respectively, and calculating the dynamic time offset compensation value based on the mechanical execution delay and the biofeedback response delay of the posture adjustment device, interleaving the first control signal sequence and the second control signal sequence after delay compensation according to the time window to generate the real-time feedback instruction sequence.

[0085] When processing the first and second control signal sequences to generate the real-time feedback instruction sequence, a timestamp synchronization marker is first added to each of them. The timestamp synchronization marker records the time when each control signal was generated, ensuring the accuracy and consistency of the two signal sequences in time.

[0086] When generating the first and second control signal sequences, the generation time of each control signal is recorded simultaneously. A high-precision clock system can be used to obtain accurate time information, which is then added to the header of each control signal as a timestamp synchronization marker. For example, for each control instruction in the first control signal sequence, a timestamp field is added to its data structure, and the time the instruction was generated is written into this field. The same process is repeated for the second control signal sequence. This allows for accurate time alignment and arrangement of the two signal sequences in subsequent processing based on the timestamp synchronization marker.

[0087] To perform dynamic time offset compensation, it's necessary to calculate the mechanical execution delay and biofeedback response delay of the posture adjustment device. Mechanical execution delay refers to the time it takes for the hydraulic actuator to actually complete mechanical movement after receiving a control signal. Biofeedback response delay refers to the time it takes for the tactile feedback module to generate the corresponding biofeedback effect after receiving a control signal.

[0088] Mechanical actuation delay can be calculated through multiple experimental measurements. In this experiment, a control signal is sent to the hydraulic actuator, and the time it takes for the hydraulic actuator to complete the corresponding mechanical movement is recorded. The difference between the two times is the measured mechanical actuation delay. By repeating the experiment multiple times and performing statistical analysis on these measurements, the average mechanical actuation delay is calculated.

[0089] The calculation method for biofeedback response delay is similar. A control signal is sent to the tactile feedback module. The time between the signal transmission and the time when the target subject feels the corresponding biofeedback effect is recorded. The time difference between the two is the measured value of the biofeedback response delay. Similarly, through multiple experiments and statistical analysis, the average biofeedback response delay time is calculated.

[0090] A dynamic time offset compensation value is calculated based on the calculated mechanical execution delay and biofeedback response delay. The dynamic time offset compensation value is used to ensure that the first control signal sequence and the second control signal sequence are executed at the correct time point to avoid uncoordinated movements caused by delays.

[0091] Because the first control signal sequence controls the mechanical motion of the hydraulic actuator, time offset compensation is required based on the mechanical execution delay. If the mechanical execution delay is Tm, each control signal in the first control signal sequence must be sent Tm in advance to ensure that the hydraulic actuator can accurately execute the mechanical motion at the required time.

[0092] The second control signal sequence controls the biofeedback operation of the tactile feedback module, so time offset compensation is required based on the biofeedback response delay. If the biofeedback response delay is Tb, each control signal in the second control signal sequence must be sent Tb in advance to ensure that the target subject experiences the corresponding biofeedback effect at the desired time.

[0093] These time offsets are used as dynamic time offset compensation values and applied in subsequent signal arrangement.

[0094] After obtaining the dynamic time offset compensation value, the first control signal sequence and the second control signal sequence are interleaved and arranged after delay compensation according to a time window. The time window is a pre-set time period within which the two signal sequences are arranged and scheduled.

[0095] First, the timestamps of each control signal in the first and second control signal sequences are adjusted based on the dynamic time offset compensation value. For the first control signal sequence, the mechanical execution delay time Tm is subtracted from the timestamp of each control signal; for the second control signal sequence, the biofeedback response delay time Tb is subtracted from the timestamp of each control signal.

[0096] Then, within a time window, the two signal sequences are interleaved according to the adjusted timestamp order. For example, within a shorter time window, the control signal with the earliest timestamp in the first control signal sequence is first selected and added to the real-time feedback instruction sequence. Next, the control signal with the earliest timestamp in the second control signal sequence that falls within the same time window is selected and added to the real-time feedback instruction sequence. The next control signal in the first control signal sequence is then selected, and this process continues in this alternating manner until the end of the time window. The interleaving process then continues in the next time window, ultimately generating a complete real-time feedback instruction sequence.

[0097] Step S150: sending the real-time feedback instruction sequence to the posture adjustment device to perform a training posture correction operation.

[0098] After the real-time feedback instruction sequence is generated, it needs to be sent to the posture adjustment device so that the posture adjustment device performs the training posture correction operation according to the instruction.

[0099] Step S151: transmitting the real-time feedback instruction sequence to the central control unit of the posture adjustment device.

[0100] For example, the real-time feedback command sequence is transmitted to the central control unit of the attitude adjustment device via a suitable communication method. This communication method can be wired communication, such as Ethernet or serial communication, or wireless communication, such as Bluetooth or Wi-Fi. The central control unit is the core control component of the attitude adjustment device and is responsible for receiving, parsing, and processing the real-time feedback command sequence.

[0101] Step S152: calling the signal analysis module of the central control unit to decompose the real-time feedback instruction sequence into mechanical motion control instructions and biofeedback control instructions.

[0102] After receiving the real-time feedback command sequence, the central control unit invokes its signal parsing module to decompose the command sequence. Since the real-time feedback command sequence is composed of an interleaved sequence of first and second control signal sequences, the signal parsing module needs to decompose it into mechanical motion control commands and biofeedback control commands. Mechanical motion control commands are used to control the movement of the hydraulic actuator, while biofeedback control commands are used to control the operation of the tactile feedback module.

[0103] Step S153: sending the mechanical motion control instruction to the hydraulic actuator through the mechanical drive interface to trigger the support surface tilting operation to change the body center of gravity distribution of the target object.

[0104] The central control unit transmits the decomposed mechanical motion control instructions to the hydraulic actuator via the mechanical drive interface. The mechanical drive interface serves as a bridge between the central control unit and the hydraulic actuator, ensuring accurate transmission of the control instructions. Upon receiving the mechanical motion control instructions, the hydraulic actuator executes the support surface tilting operation according to the instructions. This tilting of the support surface changes the target's center of gravity, thereby guiding the target to adjust their posture to better meet the requirements of shooting training.

[0105] Step S154: sending the biofeedback control instruction to the tactile feedback module through the biofeedback interface to trigger the electrical stimulation pulse and breathing rhythm guidance signal of the local muscle group.

[0106] The central control unit sends biofeedback control commands to the tactile feedback module via the biofeedback interface. The biofeedback interface is responsible for transmitting the control commands to the tactile feedback module, enabling it to execute the corresponding operations. Upon receiving the biofeedback control commands, the tactile feedback module triggers electrical stimulation pulses and respiratory rhythm guidance signals for the local muscle groups. The electrical stimulation pulses adjust the tension of the local muscle groups, and the respiratory rhythm guidance signals guide the target subject to adjust their breathing rhythm, thereby optimizing their physiological state.

[0107] Step S155: Monitor the execution status data of the posture adjustment device in real time. When execution delay or parameter drift is detected, trigger a dynamic calibration operation in an asynchronous calibration thread independent of the real-time feedback thread, generate an updated feedback instruction sequence and cache it until the next execution cycle call.

[0108] In order to ensure that the posture adjustment device can accurately and stably perform the training posture correction operation, it is necessary to monitor its execution status data in real time. The execution status data includes the motion state of the hydraulic actuator, the working state of the tactile feedback module, etc. When an execution delay or parameter drift is detected, for example, the hydraulic actuator fails to complete the support surface tilt operation on time, or the vibration frequency of the tactile feedback module deviates, a dynamic calibration operation is triggered in an asynchronous calibration thread independent of the real-time feedback thread. The dynamic calibration operation will recalculate and adjust the control instructions according to the deviation of the execution status data to generate an updated feedback instruction sequence. The updated feedback instruction sequence is cached until the next execution cycle call to ensure that subsequent training posture correction operations can be more accurate and effective.

[0109] Furthermore, the method may further include: Step S210: Acquire a historical shooting training dataset, where the historical shooting training dataset includes sample posture data labeled with standard action labels and abnormal action labels.

[0110] To train the action decomposition model, a historical shooting training dataset is required. This dataset can be constructed by collecting a large amount of target pose data from shooting training. During this process, each pose data sample is labeled to indicate whether it is a standard or abnormal action. A standard action label indicates that the shooting action corresponding to the pose data meets the preset shooting standards, while an abnormal action label indicates that the shooting action corresponding to the pose data deviates from the preset shooting standards. This labeling process can be performed by professional shooting coaches or trained personnel to ensure accuracy.

[0111] Step S220: constructing an initial action decomposition network, wherein the initial action decomposition network includes a spatiotemporal feature encoder and an action pattern classifier.

[0112] The initial action decomposition network consists of two main components: a spatiotemporal feature encoder and an action pattern classifier. The spatiotemporal feature encoder extracts multi-scale features from the sample pose data, while the action pattern classifier classifies the extracted features to determine whether the action corresponding to the sample pose data is a standard action or an abnormal action.

[0113] Step S221: Design a hierarchical structure of the spatiotemporal feature encoder, wherein the hierarchical structure includes three parallel feature extraction branches, the first branch is used to extract the gun holding stability feature, the second branch is used to extract the aiming trajectory smoothness feature, and the third branch is used to extract the firing rhythm consistency feature.

[0114] The hierarchical structure of the spatiotemporal feature encoder is designed to include three parallel feature extraction branches. Each branch is responsible for extracting features from different aspects. The first branch focuses on extracting gun holding stability features. It can judge the stability of gun holding by analyzing the joint displacement information and muscle tension information in the sample posture data. The second branch mainly extracts the aiming trajectory smoothness feature. By processing the trajectory information of the aiming process in the sample posture data, the smoothness of the aiming trajectory is evaluated. The third branch is used to extract the firing rhythm consistency feature. It analyzes the consistency of the firing rhythm based on the time series information of the firing action in the sample posture data.

[0115] Specifically, the first branch can adopt a convolutional neural network (CNN) architecture. First, the joint displacement and muscle tension information in the sample pose data is taken as input. The input data can be a multidimensional tensor, where different dimensions represent different sensor data and time series. In the CNN, multiple convolutional layers are set up, each containing multiple convolution kernels. The convolution kernels perform sliding convolution operations on the input data to extract local features in the data. For example, the convolution kernels can capture subtle joint movements and changes in muscle tension, which are closely related to the stability of holding a gun.

[0116] After the convolutional layer, a pooling layer, such as a max pooling layer or an average pooling layer, is placed. The pooling layer downsamples the feature maps output by the convolutional layer, reducing the dimensionality of the data while retaining important feature information. Through the pooling layer, higher-level features can be extracted, improving feature robustness.

[0117] Finally, at the end of the first branch, a fully connected layer is set to map the feature vectors output by the pooling layer to obtain feature units related to gun stability. The fully connected layer can further combine and abstract features, making them more representative.

[0118] The second branch also uses a CNN architecture, taking as input the joint displacement information related to the aiming trajectory in the sample pose data. The input data can be a two-dimensional or three-dimensional tensor containing information about the position changes of the target object during the aiming process. Within the CNN, convolutional and pooling layers are configured. The convolutional layer uses convolution kernels to extract local features from the trajectory data, such as the curvature and slope changes of the trajectory. The pooling layer downsamples the feature maps output by the convolutional layer to reduce the data dimensionality.

[0119] To better capture the time series information of the trajectory, a recurrent neural network (RNN) or its variants, such as a long short-term memory network (LSTM), can be added after the CNN. RNNs or LSTMs can process sequential data and remember past information, allowing for more accurate analysis of the smoothness of the aiming trajectory. Following the RNN or LSTM layer, a fully connected layer is placed to map the output feature vector into feature units related to the smoothness of the aiming trajectory.

[0120] The input data for the third branch is the muscle tension and joint displacement information related to the firing action in the sample posture data. An RNN or LSTM structure can be used to process this time series data. RNN or LSTM can handle long-term dependencies in sequence data and is suitable for analyzing the consistency of firing rhythm. In the RNN or LSTM layer, the temporal pattern and rhythm changes of the firing action are captured by continuously updating the hidden state.

[0121] After the RNN or LSTM layer, a fully connected layer is placed to map the output feature vector into feature units related to firing rhythm consistency. The fully connected layer can further process and abstract the features to obtain more representative features.

[0122] Step S222: a feature fusion layer is set at the end of each feature extraction branch, and the feature fusion layer uses an attention mechanism to dynamically weight the output features of the three parallel feature extraction branches.

[0123] A feature fusion layer is placed at the end of each feature extraction branch to effectively fuse the features extracted by the three parallel branches. This feature fusion layer employs an attention mechanism, which dynamically weights the output features based on their importance. For example, in some cases, the stability of the gun may be more important for action classification, and the attention mechanism will assign a higher weight to this feature; in other cases, the smoothness of the aiming trajectory may be more critical, and the attention mechanism will adjust its weight accordingly. This allows for more flexible and integrated utilization of the features extracted by each branch.

[0124] The attention mechanism first calculates the importance score for each feature. For each of the three feature extraction branches, the importance scores are calculated. Importance scores can be calculated using a fully connected layer and an activation function. The fully connected layer maps the feature vectors to a low-dimensional space, and the activation function converts the mapped vectors into importance scores. A commonly used activation function is the softmax function, which converts the scores into a probability distribution such that the sum of all scores is 1.

[0125] Next, the feature vectors are weighted based on the calculated importance scores. Each feature vector is multiplied by its corresponding importance score to obtain a weighted feature vector. Finally, the weighted feature vectors are concatenated or added together to obtain a fused feature vector. In this way, the attention mechanism can dynamically adjust the weights of each feature based on different input data, improving the effectiveness of feature fusion.

[0126] After completing weighted feature fusion through the attention mechanism, the feature fusion layer needs to further process the output fused feature vector. A batch normalization layer can be added after the fused feature vector. The batch normalization layer normalizes the input feature vector so that its mean is 0 and its variance is 1. This accelerates model training and improves model stability.

[0127] After the batch normalization layer, an activation function, such as the ReLU activation function, is set. The ReLU activation function introduces nonlinearity, increasing the model's expressiveness. Through these processes, the feature vector output by the feature fusion layer is more suitable for subsequent classification tasks.

[0128] Step S223: Setting a multi-task learning module in the action pattern classifier, wherein the multi-task learning module performs standard type recognition and abnormal type recognition simultaneously.

[0129] A multi-task learning module is incorporated into the action pattern classifier to simultaneously identify both standard and abnormal patterns. This module can share some feature information, improving model training efficiency and classification accuracy. For example, while recognizing standard actions, it can also identify specific abnormal actions, such as abnormal gun holding posture and aiming deviation.

[0130] The architectural design of the multi-task learning module needs to consider how to share feature information to improve the training efficiency and classification accuracy of the model.

[0131] An architecture that shares the underlying feature extraction layer and separates the upper classification layer can be used. At the bottom layer, a set of convolutional and pooling layers is used to extract common features from the input. This common feature information can be used for both standard and abnormal type recognition tasks. At the middle layer, a feature sharing layer is set up to share the common features extracted from the bottom layer.

[0132] At the top level, two classification layers are set up: one for standard action identification and the other for abnormal action identification. Each classification layer consists of multiple fully connected layers and an output layer. The output layer uses a softmax function to output classification probabilities, representing the probability that a sample belongs to a standard action type or an abnormal action type, respectively.

[0133] Step S224: inserting an adaptive feature dimensionality reduction layer after the feature fusion layer of the spatiotemporal feature encoder, the adaptive feature dimensionality reduction layer dynamically selects the retained feature dimensions according to the weighted feature importance output by the attention mechanism, and uses a nonlinear dimensionality reduction method to compress it to a preset dimension range.

[0134] An adaptive feature dimensionality reduction layer is inserted after the feature fusion layer of the spatiotemporal feature encoder. This layer dynamically selects the feature dimensions to retain based on the weighted feature importance output by the attention mechanism. It removes less important feature dimensions to reduce feature redundancy. Nonlinear dimensionality reduction methods, such as kernel principal component analysis, are then used to compress the retained feature dimensions to a preset dimension range. This reduces model complexity and improves training speed and generalization.

[0135] Step S230: performing multi-scale feature extraction on the sample posture data through the spatiotemporal feature encoder to obtain a set of training action feature units.

[0136] The acquired sample pose data is input into the spatiotemporal feature encoder, which performs multi-scale feature extraction. During this multi-scale feature extraction process, the spatiotemporal feature encoder analyzes the sample pose data from different scales and perspectives. For example, at a small scale, it focuses on local joint movements and changes in muscle tension; at a large scale, it considers the overall posture and rhythm of the entire shooting action. This multi-scale feature extraction yields a set of training action feature units, each of which represents the characteristics of the sample pose data at a specific scale and aspect.

[0137] Step S240: performing label prediction on the training action feature unit set by the action pattern classifier to obtain a predicted action category.

[0138] The training action feature unit set is input into the action pattern classifier. Based on its internal classification rules and parameters, the action pattern classifier predicts labels for the training action feature unit set. Based on the input feature units, the action pattern classifier determines whether the corresponding action is a standard action or an abnormal action and outputs a predicted action category. The predicted action category can be a category label, such as "standard action" or "abnormal action," or a more detailed abnormal action type label.

[0139] Step S250: Calculating the classification error between the predicted action category and the standard action label and abnormal action label.

[0140] To evaluate the accuracy of the action pattern classifier, it is necessary to calculate the classification error between the predicted action category and the standard action label and the abnormal action label. This error can be calculated using methods such as the cross-entropy loss function. The cross-entropy loss function measures the degree of difference between the predicted action category and the true label. The predicted action category and the true standard action label and the abnormal action label are substituted into the cross-entropy loss function to calculate the classification error. The smaller the classification error, the closer the action pattern classifier's prediction is to the true label, and the higher the model's accuracy.

[0141] For example, for a sample, assume its true label is y and its predicted probability distribution is p. The true label y is a one-hot encoded vector in which only one element is 1, indicating the true category to which the sample belongs, and the rest are 0. The predicted probability distribution p is the probability vector output by the action pattern classifier, in which each element represents the probability of the sample belonging to each category.

[0142] The cross-entropy loss function is calculated by taking the negative logarithm of the product of the corresponding elements of the true label vector and the predicted probability distribution vector, and then summing them. Specifically, for each category i, if the i-th element of the true label vector is 1, the negative logarithm of the i-th element of the predicted probability distribution vector is taken; if the i-th element of the true label vector is 0, the negative logarithm of the i-th element is zero. Finally, the results for all categories are summed to obtain the cross-entropy loss for that sample.

[0143] For the entire training set, the cross entropy loss is the average of the cross entropy losses of all samples. By minimizing the cross entropy loss, the predicted probability distribution can be made as close as possible to the true label distribution, thereby improving the classification accuracy of the model.

[0144] After calculating the cross entropy loss for each sample, the classification errors of all samples need to be accumulated and counted. You can set a cumulative variable and add the cross entropy loss of each sample to the cumulative variable after calculating it.

[0145] At the same time, record the total number of samples in the training set. After completing the calculation for all samples, divide the accumulated cross entropy loss by the total number of samples to obtain the average classification error. The average classification error can more accurately reflect the classification performance of the model on the entire training set.

[0146] In addition, you can count the number of misclassified samples and the number of correctly classified samples. By comparing the number of misclassified samples with the total number of samples, you can calculate the classification error rate; by comparing the number of correctly classified samples with the total number of samples, you can calculate the classification accuracy rate. These statistical metrics can help evaluate model performance and training results.

[0147] Step S260: Jointly optimize the spatiotemporal feature encoder and the motion pattern classifier based on the classification error until the classification error is lower than a preset threshold, and output a trained motion decomposition model.

[0148] Based on the calculated classification error, the spatiotemporal feature encoder and motion pattern classifier are jointly optimized. Optimization algorithms such as gradient descent can be used to adjust the parameters of the spatiotemporal feature encoder and motion pattern classifier based on the gradient information of the classification error. With each iteration, the parameters are continuously updated to gradually reduce the classification error. The optimization process continues until the classification error falls below a preset threshold. The threshold is a pre-set error standard. When the classification error falls below this threshold, the model is considered to have achieved satisfactory training results, and the trained motion decomposition model is output. This trained motion decomposition model can be used for subsequent motion feature extraction processing of real-time posture data sets.

[0149] In detail, the core idea of the gradient descent optimization algorithm is to update the parameters of the model along the negative gradient direction of the loss function so that the loss function gradually decreases.

[0150] For the parameters of the spatiotemporal feature encoder and action pattern classifier, we first calculate the gradient of the classification error with respect to these parameters. The gradient represents the rate and direction of change of the loss function in parameter space. Using the chain rule, we can backpropagate the gradient of the classification error to the various parameters of the model.

[0151] In each iteration, the model parameters are updated according to the calculated gradients at a certain learning rate. The learning rate is a hyperparameter that controls the step size of the parameter update. If the learning rate is too large, the model may skip the optimal solution, causing the loss function to fail to converge. If the learning rate is too small, the model training speed will be very slow.

[0152] Specifically, for each parameter, its current value is subtracted from the learning rate multiplied by the gradient of the parameter to obtain the updated parameter value. This process is repeated until the classification error is lower than the preset threshold.

[0153] In each iteration, the parameters of the spatiotemporal feature encoder and the action pattern classifier are updated using the gradient descent optimization algorithm. The update process involves all layers of the model, including convolutional layers and fully connected layers.

[0154] The convolutional layer parameters, such as the kernel weights and biases, are updated based on the calculated gradients. The kernel weights determine the feature extraction capability of the convolution operation, while the biases adjust the output of the convolution layer. By updating the kernel weights and biases, the convolution layer can better extract features from the sample pose data.

[0155] The parameters of the fully connected layer, such as the weight matrix and bias vector, are also updated based on the gradient. The weight matrix of the fully connected layer maps the input feature vector to the output score vector, and the bias vector adjusts the output score. By updating the parameters of the fully connected layer, the fully connected layer can better classify and predict features.

[0156] After each parameter update, the classification error is recalculated. If the classification error is still higher than the preset threshold, the next iteration is continued until the classification error meets the requirements.

[0157] The validation set plays an important role in the model training process. It is used to evaluate the performance of the model during training, adjust the model's hyperparameters, and avoid overfitting.

[0158] After each iteration, the validation set is fed into the trained model, and the classification error and accuracy on the validation set are calculated. If the classification error on the validation set begins to increase while the classification error on the training set continues to decrease, the model may be overfitting. In this case, you can take measures to prevent overfitting, such as adjusting the learning rate or adding regularization.

[0159] At the same time, based on the evaluation results on the validation set, you can adjust the model's hyperparameters. For example, if you find that the model's training speed is too slow, you can increase the learning rate appropriately. If you find that the model's classification accuracy is low, you can adjust the network structure, such as increasing the number of convolutional layers or the number of neurons in the fully connected layer.

[0160] When the classification error is below the preset threshold, the model is considered to have been trained. At this point, the test set is input into the trained model for final evaluation.

[0161] The test set is data that has not been used during model training and tuning. It can more objectively evaluate the model's generalization ability. Metrics such as classification error, accuracy, recall, and F1 score are calculated on the test set. These metrics can comprehensively reflect the model's performance on unseen data.

[0162] If the evaluation results on the test set meet the requirements, the trained motion decomposition model is output. This motion decomposition model can be used for subsequent motion feature extraction and processing of real-time posture data sets to help achieve correction and feedback of shooting training postures.

[0163] In the above embodiments, attention must be paid to data privacy protection and leakage prevention. When collecting real-time posture data of a target object, some privacy-sensitive data, such as muscle tension signals and respiratory rate signals, is involved. To protect the privacy of this data, encryption technology is employed to encrypt the collected data. For example, immediately after the sensor collects the data, it is symmetrically encrypted, using an encryption key to convert the data into ciphertext. Only authorized devices and systems can decrypt the ciphertext using the corresponding decryption key.

[0164] Furthermore, during data transmission, secure communication protocols, such as SSL / TLS, are employed to prevent data theft or tampering during transmission. Regarding data storage, encrypted data is stored on secure servers, with strict access control mechanisms in place to ensure that only authorized personnel can access and process this data. These privacy protection and anti-leakage technologies ensure that the target user's private data is effectively protected.

[0165] Figure 2The hardware structure of the shooting training posture correction feedback system 100 based on intelligent analysis for implementing the shooting training posture correction feedback method based on intelligent analysis provided in an embodiment of the present application is shown as follows: Figure 2 As shown, the shooting training posture correction feedback system 100 based on intelligent analysis may include a processor 110 , a machine-readable storage medium 120 , a bus 130 and a communication unit 140 .

[0166] In a possible design, the shooting training posture correction feedback system 100 based on intelligent analysis can be a single server or a server group. The server group can be centralized or distributed (for example, the shooting training posture correction feedback system 100 based on intelligent analysis can be a distributed system). In some embodiments, the shooting training posture correction feedback system 100 based on intelligent analysis can be local or remote. For example, the shooting training posture correction feedback system 100 based on intelligent analysis can access the information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the shooting training posture correction feedback system 100 based on intelligent analysis can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the shooting training posture correction feedback system 100 based on intelligent analysis can be implemented on the shooting training posture correction feedback system based on intelligent analysis. As an example only, the shooting training posture correction feedback system based on intelligent analysis can include private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-layer cloud, etc. or any convergence thereof.

[0167] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the intelligent analysis-based shooting training posture correction feedback system 100 to execute or perform the exemplary methods described herein.

[0168] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the shooting training posture correction feedback method based on intelligent analysis as described in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0169] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the shooting training posture correction feedback system 100 based on intelligent analysis. The implementation principles and technical effects are similar and will not be repeated here in this embodiment.

[0170] In addition, an embodiment of the present application also provides a readable storage medium, in which computer-executable instructions are set. When the processor runs the computer-executable instructions, the shooting training posture correction feedback method based on intelligent analysis as described above is implemented.

[0171] It should be noted that, in order to simplify the description of the present disclosure and thus facilitate understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present disclosure, multiple features may sometimes be combined into one embodiment, figure, or description thereof. Similarly, it should be noted that, in order to simplify the description of the present disclosure and thus facilitate understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present disclosure, multiple features may sometimes be combined into one embodiment, figure, or description thereof.

Claims

1. A shooting training posture correction feedback method based on intelligent analysis, characterized in that: The method comprises: Acquire a real-time posture data set of the target object, wherein the real-time posture data set includes multiple sets of time-continuous posture acquisition data, each set of posture acquisition data includes motion trajectory characteristics and physiological state characteristics; Performing motion feature extraction processing on the real-time posture data set to obtain a target motion feature set and an abnormal motion feature set; generating a posture correction parameter set based on a dynamic difference between the target motion feature set and the abnormal motion feature set, the posture correction parameter set including a motion adjustment direction parameter and a physiological coordination optimization parameter; generating a real-time feedback instruction sequence according to the posture correction parameter set; The real-time feedback instruction sequence is sent to the posture adjustment device to perform the training posture correction operation.

2. The shooting training posture correction feedback method based on intelligent analysis according to claim 1 is characterized in that: The step of obtaining a real-time posture data set of the target object includes: Collecting an original posture signal stream of the target object through a distributed posture sensor array, wherein the original posture signal stream includes a joint displacement signal, a muscle tension signal, and a respiratory rate signal; Performing timestamp alignment processing on the original gesture signal stream to obtain a synchronized gesture signal stream, wherein data collected by different signal sources in the synchronized gesture signal stream have a unified timestamp sequence; performing multi-channel normalization processing on the joint displacement signals, muscle tension signals, and respiratory rate signals in the synchronized posture signal stream, respectively, and performing signal fusion processing on the normalized synchronized posture signal stream to generate a fused posture data block, wherein each data unit in the fused posture data block contains a unified dimension of motion trajectory code and physiological state code; The fused posture data block is segmented and cut according to a preset shooting action cycle to obtain the multiple groups of time-continuous posture acquisition data, each group of posture acquisition data corresponds to a complete shooting action cycle.

3. The shooting training posture correction feedback method based on intelligent analysis according to claim 1 is characterized in that: The performing motion feature extraction processing on the real-time posture data set to obtain a target motion feature set and an abnormal motion feature set includes: Call the pre-trained action decomposition model to perform hierarchical feature analysis on the action trajectory encoding in each set of posture collection data to obtain the initial action feature unit set; Normalizing the gun holding stability feature, aiming trajectory smoothness feature, and firing rhythm consistency feature in the initial action feature unit set, mapping features of different dimensions into a unified metric space, and generating a basic action feature unit set; Performing pattern matching processing on the basic action feature unit set, clustering the action feature units that meet the preset shooting standard into target action clusters, and marking the action feature units that deviate from the preset shooting standard as abnormal action clusters; Performing feature enhancement processing on the action feature units in the target action cluster to generate the target action feature set, wherein the feature enhancement processing includes noise suppression and feature dimension expansion; Deviation quantization processing is performed on the motion feature units in the abnormal motion cluster to generate the abnormal motion feature set, wherein the deviation quantization processing includes motion offset angle calculation and physiological state fluctuation amplitude statistics.

4. The shooting training posture correction feedback method based on intelligent analysis according to claim 3 is characterized in that: The method further comprises the step of training the action decomposition model: Acquire a historical shooting training dataset, wherein the historical shooting training dataset includes sample posture data labeled with standard action labels and abnormal action labels; Constructing an initial action decomposition network, wherein the initial action decomposition network includes a spatiotemporal feature encoder and an action pattern classifier; Performing multi-scale feature extraction on the sample posture data by the spatiotemporal feature encoder to obtain a set of training action feature units; Performing label prediction on the training action feature unit set by the action pattern classifier to obtain a predicted action category; Calculating the classification error between the predicted action category and the standard action label and the abnormal action label; The spatiotemporal feature encoder and the motion pattern classifier are jointly optimized based on the classification error until the classification error is lower than a preset threshold, and the trained motion decomposition model is output.

5. The shooting training posture correction feedback method based on intelligent analysis according to claim 4 is characterized in that: The constructing of the initial action decomposition network includes: Designing a hierarchical structure of the spatiotemporal feature encoder, the hierarchical structure comprises three parallel feature extraction branches, the first branch is used to extract gun holding stability features, the second branch is used to extract aiming trajectory smoothness features, and the third branch is used to extract firing rhythm consistency features; A feature fusion layer is set at the end of each feature extraction branch, and the feature fusion layer uses an attention mechanism to dynamically weight the output features of the three parallel feature extraction branches; Setting a multi-task learning module in the action pattern classifier, wherein the multi-task learning module simultaneously performs standard type recognition and abnormal type recognition; An adaptive feature dimensionality reduction layer is inserted after the feature fusion layer of the spatiotemporal feature encoder. The adaptive feature dimensionality reduction layer dynamically selects the retained feature dimensions according to the weighted feature importance output by the attention mechanism, and compresses them to a preset dimension range using a nonlinear dimensionality reduction method.

6. The shooting training posture correction feedback method based on intelligent analysis according to claim 1 is characterized in that: The generating of a posture correction parameter set based on the dynamic difference between the target motion feature set and the abnormal motion feature set includes: Constructing a dynamic difference evaluation model, wherein the dynamic difference evaluation model includes a motion space difference measurement module and a physiological state coupling analysis module; Calculating, by the action space difference measurement module, a multidimensional difference vector between the target action feature set and the abnormal action feature set in the dimensions of gun holding stability, aiming trajectory, and firing rhythm; determining, by the physiological state coupling analysis module, a correlation weight matrix between the multidimensional difference vector and the physiological state feature; Performing weighted fusion on the multidimensional difference vectors according to the association weight matrix to generate a comprehensive difference score; The action adjustment direction parameter and the physiological coordination optimization parameter are obtained by matching in a preset correction parameter mapping table based on the comprehensive difference score.

7. The shooting training posture correction feedback method based on intelligent analysis according to claim 6 is characterized in that: The determining, by the physiological state coupling analysis module, a correlation weight matrix between the multidimensional difference vector and the physiological state feature includes: Performing multi-channel normalization processing on the physiological state characteristics to generate a set of standardized physiological indicators with consistent unit dimensions, wherein the set of standardized physiological indicators includes a normalized value of respiratory rate, a normalized value of muscle tension, and a normalized value of heart rate variability coefficient; Extracting action difference parameters of each dimension in the multidimensional difference vector, wherein the action difference parameters include gun holding stability offset, aiming trajectory discreteness, and firing rhythm deviation rate; Calculating a dynamic correlation coefficient between each of the standardized physiological indicators and each of the motion difference parameters to generate a physiological-motion correlation matrix, wherein the dynamic correlation coefficient is determined by dividing the covariance within a sliding time window by the product of the standard deviations of the two variables; Determining a physiological state influence weight factor according to the modulus of each row vector in the physiological-motion association matrix, wherein the modulus is calculated by the Euclidean norm and compressed to a preset weight interval by a logarithmic function; According to the physiological state influence weight factor and the confidence score of the corresponding indicator in the preset expert experience weight table, a priority weighted fusion strategy is adopted to generate the associated weight matrix, wherein the expert experience weight table stores the priority coefficients of physiological indicators to action differences in different shooting stages.

8. The shooting training posture correction feedback method based on intelligent analysis according to claim 1 is characterized in that: Generating a real-time feedback instruction sequence according to the posture correction parameter set includes: Analyzing mechanical motion constraints corresponding to the motion adjustment direction parameters, the mechanical motion constraints including a motion amplitude threshold and a motion direction priority; generating a first control signal sequence according to the mechanical motion constraint condition, wherein the first control signal sequence is used to drive the hydraulic actuator of the posture adjustment device to complete a support surface tilting operation at a specified angle; Analyzing the biofeedback regulation strategy corresponding to the physiological coordination optimization parameter, the biofeedback regulation strategy including a respiratory rhythm synchronization mode and a muscle tension balance mode; generating a second control signal sequence according to the biofeedback regulation strategy, wherein the second control signal sequence is used to drive the tactile feedback module of the posture adjustment device to perform vibration frequency modulation and pressure gradient distribution operations; Timestamp synchronization marks are added to the first control signal sequence and the second control signal sequence respectively, and after calculating the dynamic time offset compensation value based on the mechanical execution delay and biofeedback response delay of the posture adjustment device, the first control signal sequence and the second control signal sequence are interleaved after delay compensation according to the time window to generate the real-time feedback instruction sequence.

9. The shooting training posture correction feedback method based on intelligent analysis according to claim 8, characterized in that: The step of sending the real-time feedback instruction sequence to the posture adjustment device to perform a training posture correction operation includes: transmitting the real-time feedback instruction sequence to the central control unit of the posture adjustment device; calling the signal parsing module of the central control unit to decompose the real-time feedback instruction sequence into mechanical motion control instructions and biofeedback control instructions; sending the mechanical motion control instruction to the hydraulic actuator via a mechanical drive interface to trigger a tilting operation of the support surface to change the body center of gravity distribution of the target object; Sending the biofeedback control instruction to the tactile feedback module via the biofeedback interface to trigger electrical stimulation pulses and respiratory rhythm guidance signals for local muscle groups; The execution status data of the posture adjustment device is monitored in real time. When execution delay or parameter drift is detected, a dynamic calibration operation is triggered in an asynchronous calibration thread independent of the real-time feedback thread, and an updated feedback instruction sequence is generated and cached until the next execution cycle call.

10. A shooting training posture correction feedback system based on intelligent analysis, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the shooting training posture correction feedback method based on intelligent analysis as described in any one of claims 1 to 9.

Images