A multi-channel target speech extraction method based on initial direction anchoring and continuous identity maintenance

Abstract

The application discloses a multi-channel target speech extraction method based on initial direction anchoring and continuous identity maintenance, and specifically comprises the following steps: performing time-frequency transformation on a mixed speech signal; extracting a spatial clue to generate a candidate direction set and calculating a spatial candidate confidence; selecting a candidate direction satisfying a stability condition in an initialization time window according to a target initial direction and obtaining a local enhancement segment; establishing a history template by using the local enhancement segment; in a subsequent processing process, outputting a current target enhanced speech and a target existence probability based on a current multi-channel observation feature, a target direction reference and the history template; building a continuity control quantity; realizing state migration between states according to the continuity control quantity and a state transition rule, and updating the target direction reference and the history template when a gating condition is satisfied. The application can improve the continuity, stability and robustness of target speaker extraction under the conditions of target short-time pause, slow movement, near-angle competition and unstable candidate direction.

Description

Technical Field

[0001] This invention belongs to the field of speech signal processing technology, and in particular relates to a multi-channel target speech extraction method based on initial direction anchoring and continuous identity preservation. Background Technology

[0002] In multi-person dynamic speech scenarios, multi-channel array systems can typically extract speech components related to a specific spatial direction from mixed speech through beamforming, delay estimation, spatial spectrum scanning, or deep learning enhancement models. However, many existing methods are actually better at solving the problem of which direction is currently strongest or how to enhance a given direction. For the problem of how to maintain the same target identity initially locked when the target pauses briefly, moves slowly, or when competing speakers continue to speak, there is a lack of stable and engineering-feasible closed-loop solutions.

[0003] Common technical approaches in the prior art mainly include: first, directly driving beamformers or directional conditional speech enhancement models based on the current directional peak, spatial response peak, or direction of maximum energy; second, using a single directional conditional network to output enhanced speech under a given target direction; and third, introducing registered speech as an identity template in speaker recognition or speaker verification scenarios to assist in target preservation. All of these technical approaches have limitations in real-world multi-person dynamic environments.

[0004] On the one hand, when the target speaker pauses briefly, speaks intermittently, or experiences significant changes in speech rate, relying solely on the instantaneous directional peak or instantaneous movement direction can easily cause the system to switch the enhanced target to the current strongest competing speaker, resulting in target drift. On the other hand, when the target speaker moves slowly, if the system lacks explicit maintenance of the target's continuous identity and relies solely on a fixed initial direction, it can easily cause directional mismatch, leading to a decrease in output speech quality.

[0005] Furthermore, if the target template is continuously updated using the current enhancement results, it is highly susceptible to contamination when there are strong competing speakers, unstable spatial candidates, or high output leakage. Once the template is contaminated, subsequent enhancements and target determination will deviate further, leading to error accumulation. Therefore, existing technologies generally suffer from the following problems: first, over-reliance on instantaneous directional evidence; second, lack of explicit modeling of the continuous identity of the target; third, lack of gating protection mechanisms against template contamination; and fourth, lack of unified control logic to distinguish between target pauses, target movements, and target loss.

[0006] Therefore, there is an urgent need to propose a new multi-channel target speaker extraction method. Under the premise that only the initial direction of the target is known, the method can establish a continuous identity state of the target, construct an initialization template, suppress interference from competing directions, and perform smooth direction updates when necessary. This will enable the continuous extraction of the same target speaker as much as possible, even when the target pauses briefly, competing speakers continue to speak, and direction candidates fluctuate. Summary of the Invention

[0007] Purpose of the invention: To address the problems of over-reliance on instantaneous direction peaks, difficulty in maintaining continuous target identity, easy switching when the target pauses, and easy template contamination in existing technologies, this invention proposes a multi-channel target speech extraction method based on initial direction anchoring and continuous identity preservation.

[0008] Technical solution: This invention provides a multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation, specifically as follows:

[0009] Preprocessing, framing, windowing, and short-time Fourier transform are performed on N synchronously acquired multi-channel mixed speech signals;

[0010] In the initial time window Below, based on the initial direction of the target speech Building a neighborhood Set a fixed step size, in Candidate directions are extracted from within;

[0011] GCC-PHAT spatial analysis is used to explicitly model the spatial candidate confidence and calculate the spatial candidate confidence of candidate direction d. ;exist The candidate direction that meets the conditions is selected as the initial direction. That is, the measurement direction; candidate directions with speech activity outside the neighborhood of the measurement direction are used as competing candidate directions;

[0012] Calculate the scores of the competing candidate directions, and select the competing candidate direction with the highest score as the strongest competing candidate direction. Based on the probability of voice activity Score against the strongest competitor Enhance the speech in the initial direction; obtain locally enhanced speech segments. ;Will Divided into Individual fragments Calculate the reliability of each sub-segment Therefore, the normalized weights of each sub-segment are calculated. Calculate the initial historical template ;

[0013] Using GCC-PHAT spatial analysis, the measurement direction within the current time window t is obtained. The history template of the previous time window Using the historical template of the current time window, predict the prediction direction at the current time window t. The speech in the current time window is enhanced, and the probability of the target speech existing in the current time window is calculated. ;

[0014] The enhanced speech is input into the speaker representation model to extract the current template. ,calculate and similarity between Level of voice leakage and target direction continuity score ;Will , , as well as As a positive consistency term, As a negative risk term, it generates continuous control quantities. ;

[0015] Based on the state of the previous time window and the state feature vector of the current time window, determine the state of the current time window and whether to update. And historical templates.

[0016] Furthermore, the normalized peak intensity is calculated for any candidate direction. :

[0017] ;

[0018] in, For the candidate direction set, To prevent constants with a denominator of zero; Spatial response for candidate directions:

[0019] ;

[0020] Where i and j represent the i-th and j-th microphones, Indicates a collection of microphones. This represents the weight of the channel pair corresponding to the i-th and j-th microphones. Let represent the time delay of the i-th and j-th microphones in the candidate direction. The weighted cross-correlation function is:

[0021] ;

[0022] in, This represents the cross-power spectrum between the i-th and j-th microphones. Where f represents frequency, This represents the signal after performing a short-time Fourier transform on the signal transmitted from the i-th microphone; * indicates conjugate;

[0023] Calculate the peak sharpness for any candidate direction. :

[0024] ;

[0025] in, The neighborhood of the candidate direction;

[0026] Calculate its cross-frame stability for any candidate direction. ;

[0027] (From the moment of crossing frames to the current frame)

[0028] in, The length of the cross-frame stability window. This represents the minimum resolvable angle of the microphone array. express The direction corresponding to the main peak at any given moment; This refers to the time corresponding to the first frame of the speech signal within the cross-frame stability window. This indicates the time corresponding to the last frame in the current time window. If the current time window is... In this case, cross-frame stability is not calculated;

[0029] Calculate the consistency of speech activity for any candidate direction:

[0030] ;

[0031] in, The length of the voice activity consistency window. Represents an exponential function. express The voice activity detection result corresponding to the candidate direction d at time moment; The time corresponding to the first frame of the speech signal within the speech activity consistency window, if the current time window is In this case, the consistency of speech activity is not calculated;

[0032] Calculate the competition penalty term for any candidate direction. :

[0033] ;

[0034] in, Show set The set of candidate directions other than candidate direction d.

[0035] Calculate the spatial candidate confidence of candidate direction d :

[0036]

[0037] in, All are non-negative weights. For the Sigmoid function, The normalized peak intensity for the candidate direction. The peak sharpness of the candidate direction.

[0038] Furthermore, the following conditions are set to select the initialization direction. :

[0039] (1) ;

[0040] (2) ; The candidate credibility threshold;

[0041] (3) There is a continuous range of time sampling points within the current time window, and the directional deviation between adjacent time points is less than or equal to a preset directional deviation threshold within this continuous range of time sampling points. And each sampling point can detect Or with The deviation between them is less than the threshold. Candidate directions;

[0042] (4) Less than or equal to , This is the threshold for voice activity consistency.

[0043] Furthermore, The corresponding steering vector is Calculate the normalized delay summation beamforming weight vector:

[0044] ;

[0045] in, To prevent constants with a denominator of 0; f represents frequency;

[0046] Calculation direction directional beamforming frequency domain:

[0047] ;

[0048] right Enhance the direction of the speech to obtain :

[0049] ;

[0050] in, For soft masking:

[0051] ;

[0052] in, Let be the probability of speech activity, that is, the probability of speech occurring. The difference in scores:

[0053] ;

[0054] in, This represents the strongest competing candidate direction within the current time window t. The direction with the highest score in the neighborhood of the measurement direction under the current time window t; For the scoring function:

[0055] ;

[0056] in, Indicates the direction within the neighborhood of the measurement direction. , Non-negative weights for In the time window Consistency in direction:

[0057] ;

[0058] in, For time window The predicted direction, when hour, for ;

[0059] ;

[0060] in, Indicates the threshold for determining voice activity. For time window The mean logarithmic energy is:

[0061] ;

[0062] in, This indicates that the nth channel is within the time window. Time-frequency representation at frequency f Represents a set of frequencies.

[0063] Furthermore, the reliability of each sub-segment is calculated. :

[0064] ;

[0065] in, Non-negative weights Let be the average spatial candidate confidence of the k-th sub-segment. For the first The average speech activity probability of each sub-segment No. Average leakage level of each sub-segment:

[0066] ;

[0067] in, For the k-th sub-segment in the time window The leakage level below, For the k-th sub-segment in the time window Output energy in the neighborhood of the measurement direction below. For the k-th sub-segment in the time window The maximum energy in the set of competing candidate directions;

[0068] Calculate the normalized weights:

[0069] ;

[0070] in, This is a weighted sensitivity coefficient used to control the degree of differentiation between the reliability scores of different sub-segments;

[0071] Calculate the initial history template:

[0072] ;

[0073] in, , For speaker representation model.

[0074] Furthermore, a Kalman filter is used to predict the direction in the current time window t based on the prediction direction of the previous time window. That is, the target candidate direction, will , The speech and state feature vectors at the current time window t are input into the time-frequency mask prediction network to obtain the enhanced speech at the current time. The probability of the target speech existing at the current moment. .

[0075] Furthermore, calculate the similarity between the current template and the current historical templates:

[0076] ;

[0077] Calculate the continuity score of the target candidate direction :

[0078] ;

[0079] in, ,in The measurement direction within the current time window. The speed of directional change is preset by humans. ;

[0080] The positive consistency term is:

[0081] ;

[0082] in, , , as well as All are non-negative exponential weight parameters;

[0083] Let the negative risk term be:

[0084] ;

[0085] in ;

[0086] The difference between the score of the strongest competitive candidate direction and the score of the measured direction in the current time window;

[0087] Calculate the continuous control quantity:

[0088] .

[0089] Furthermore, the states include locked state, tracking state, holding state, recapture state, and lost state;

[0090] If continuous control quantity Greater than or equal to the preset locking threshold The probability of the target speech exists Greater than or equal to the preset existence threshold The similarity between the current template and the historical templates at the current moment. Greater than or equal to the preset template threshold Target direction continuity score Greater than or equal to the preset direction threshold When this happens, the locked state is used as a candidate state for the current time window;

[0091] When the target speech exists, probability Greater than or equal to the threshold , Greater than or equal to the acceptable threshold ,when Spatial candidate credibility is greater than or equal to the threshold The absolute value of the difference between the measured direction and the predicted direction within the current time window is greater than or equal to the stationary threshold. And the speed of change of direction Less than or equal to the smoothing threshold When this happens, the tracking status is used as a candidate status for the current time window;

[0092] When continuous control quantity Less than the holding threshold or the probability of the target speech existing Less than the existence threshold And the duration of maintaining the state. Less than or equal to keep window When this happens, the current state will be retained as a candidate state for the current time window;

[0093] When the system is in a hold state, and the hold duration is... Larger than the hold window Simultaneously, if the spatial candidate reappears in the neighborhood of the measurement direction or the neighborhood of the prediction direction, the confidence level is greater than or equal to the threshold. And the target speech has a probability Greater than or equal to the threshold When considering candidate directions, the recapture state is used as the candidate state for the current time window;

[0094] When the system is in recapture state, and the recapture duration is... Exceeding the recapture window If no candidate direction that satisfies the locked state is detected, the lost state is taken as the candidate state of the current time window.

[0095] Calculate the state transition score for each candidate state. The highest scorer is selected as the current state.

[0096] ;

[0097] in, Represents the set of candidate states. , and These are weights and biases, respectively. For state feature vectors: ,in For cooling count.

[0098] Furthermore, when the current time window is in tracking mode, a smooth update is performed on the measured direction:

[0099] ;

[0100] in, This indicates the updated predicted direction, which is also the updated target direction. Indicates the measurement direction within the current time window. Indicates the adaptive smooth update coefficients:

[0101] ;

[0102] in, This represents the Sigmoid function. , where is a non-negative weight;

[0103] Perform gated updates on historical templates:

[0104] ;

[0105] in, This is the updated history template for the current time window. Indicates the update strength of the historical template at the current moment:

[0106] ;

[0107] in, , , Non-negative weights Update the gated variables for the template:

[0108] ;

[0109] in, , as well as All indicate that they are related to , , The corresponding threshold; It is an exponential function.

[0110] Beneficial effects:

[0111] 1. This invention uses local enhancement segment initialization and historical template reliability weighting to make the initial template independent of a single instantaneous frame, reducing the risk of mis-locking caused by accidental peaks, noise, or competing speakers during the initialization stage.

[0112] 2. This invention does not directly regard the current strongest direction as the target, but establishes a continuous identity entry for the target by anchoring the initial direction neighborhood, thus distinguishing the two technical issues of direction enhancement and identity preservation at the methodological level.

[0113] 3. This invention helps reduce the risk of template contamination by introducing competition suppression terms, leakage penalty terms, and template gating update mechanisms.

[0114] 4. This invention, through continuous control variables and state transition scoring mechanisms, helps to distinguish between situations such as short-term target pauses, smooth directional changes, and target loss, thereby reducing false target switching.

[0115] 5. This invention can be implemented with only the initial direction of the target known and no registered voice, and has low engineering difficulty. It is suitable for scenarios such as conference sound pickup, human-computer interaction, array terminals and voice acquisition equipment. Attached Figure Description

[0116] Figure 1 This is the overall flowchart of the present invention.

[0117] Figure 2 This is the state transition diagram of the present invention. Detailed Implementation

[0118] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0119] In this embodiment, the system input includes Array speech signals acquired synchronously and the initial direction of the target speaker The system output includes target-enhanced speech. Historical templates Continuous control quantity and the current status .

[0120] The first in the array The time-domain signal of each channel is denoted as Its short-time Fourier transform is expressed as All channels at time ,frequency The observations below are written as column vectors

[0121] ;

[0122] Target continuous identity state is defined as ,in This represents auxiliary control variables such as cooling count, hold duration, and recapture duration.

[0123] like Figure 1 The specific process of the present invention is as follows:

[0124] Step S1: Multi-channel preprocessing and time-frequency representation

[0125] First, DC removal, amplitude normalization, bad sector detection and compensation, and optional lightweight noise pre-suppression are performed on each channel signal. Then, the frame length is used. Frame shift The signal from each channel is framed, windowed, and then subjected to a short-time Fourier transform to obtain:

[0126] ;

[0127] Among them, frame length 256 points, 512 points, or 1024 points can be selected, frame shift Half the frame length can be used; the window function can be a Hamming window, a Hanning window, or an equivalent form.

[0128] Step S2: Spatial candidate generation and explicit modeling of spatial candidate credibility

[0129] For time windows In a preferred embodiment of the present invention, spatial cue extraction is fixedly performed using the Generalized Cross-Correlation Phase Transform (GCC-PHAT) algorithm. For the microphone array... The first channel and the first One channel, in ,frequency The cross power spectrum is defined as follows: , This represents the signal after performing a short-time Fourier transform on the signal transmitted from the i-th microphone; * indicates conjugate; the corresponding PHAT weighted cross-correlation function. Defined as:

[0130] ;

[0131] in, To prevent positive numbers with a denominator of zero, for any candidate direction... The theoretical arrival delay is calculated based on the array geometry. The formula is as follows:

[0132] ;

[0133] in, This represents the coordinates of the i-th and j-th microphones. Let d represent the unit direction vector of the candidate direction d, and c represent the speed of sound.

[0134] The GCC-PHAT responses of each channel at the theoretical time delay are then aggregated to obtain the spatial responses of the candidate directions:

[0135] ;

[0136] in, This represents the set of microphone channel pairs participating in the aggregation. This represents the channel pair weights. Preferably, channel pairs with appropriate element spacing can be selected for calculation to reduce the impact of long-distance channels on time delay jitter under reverberation conditions. Based on the aggregated spatial response... Generate a set of candidate directions .

[0137] For any candidate direction The corresponding spatial response is defined as To avoid using a single peak value as the sole criterion for judgment, this invention explicitly models the credibility of spatial candidates.

[0138] The normalized peak intensity of the candidate direction is defined as:

[0139] ;

[0140] The peak sharpness of the candidate direction is defined as:

[0141] ;

[0142] in, For direction The local neighborhood.

[0143] The cross-frame stability of the candidate direction is defined as follows (where the cross-frame refers to the case where the measurement directions of two adjacent time windows are inconsistent):

[0144] ;

[0145] in, The length of the cross-frame stability window. This represents the minimum resolvable angle of the microphone array. express The direction corresponding to the main peak at any given moment; This refers to the time corresponding to the first frame of the speech signal within the cross-frame stability window. This indicates the time corresponding to the last frame in the current time window. If the current time window is... In this case, cross-frame stability is not calculated.

[0146] The consistency between candidate directions and speech activity is defined as:

[0147] ;

[0148] in, The length of the voice activity consistency window. Represents an exponential function. express The voice activity detection result corresponding to the candidate direction d at time moment; The time corresponding to the first frame of the speech signal within the speech activity consistency window, if the current time window is In this case, the consistency of speech activity is not calculated;

[0149] To suppress the influence of competing speakers at close angles, a competition penalty term is defined:

[0150] ;

[0151] Therefore, the spatial candidate credibility is defined as:

[0152] ;

[0153] in, Non-negative weights This is the Sigmoid function. Through this modeling approach, spatial candidate confidence simultaneously reflects peak intensity, peak sharpness, cross-frame stability, speech activity consistency, and the level of contention interference.

[0154] Step S3: Initial orientation neighborhood anchoring and local enhancement fragment generation

[0155] Based on the initial direction of the target (Given) Construct the initial directional neighborhood . To use a fixed step size, the initial time window is denoted as . Within this time window, the system only... Internally search for candidate directions and select the candidate direction that meets the following conditions as the initial direction. :

[0156] (1) ;

[0157] (2) ; The candidate credibility threshold;

[0158] (3) Within a continuous period of time, the directional deviation between adjacent time intervals is less than or equal to a preset directional deviation threshold. And all of them can be detected Or with The deviation between them is less than Candidate directions;

[0159] (4) Less than or equal to , This is the threshold for voice activity consistency.

[0160] After obtaining the initial direction, local enhancement is performed based on that direction. Let the candidate direction be... The corresponding steering vector is Then the normalized delay summation beamforming weight vector can be written as:

[0161] ;

[0162] Calculation direction directional beamforming frequency domain:

[0163] ;

[0164] This embodiment uses a soft mask instead of a binary mask to avoid speech breaks and template distortion caused by hard decisions during the initialization phase. In a simplified implementation, the soft mask can also be thresholded to obtain a binary mask, but the preferred embodiment of this invention uses a soft mask. For example, the soft mask is defined as:

[0165] ;

[0166] in, For the probability of voice activity, The score difference between the strongest competing candidate direction and the target candidate direction (candidate directions with speech activity outside the neighborhood of the measurement direction are considered as competing candidate directions, and the competing candidate direction with the highest score is considered as the strongest competing candidate direction);

[0167] ;

[0168] in, This represents the strongest competing candidate direction within the current time window t. The direction with the highest score in the neighborhood of the measurement direction under the current time window t; For the scoring function:

[0169] ;

[0170] in, Indicates the direction within the neighborhood of the measurement direction. , Non-negative weights for In the time window Consistency in direction:

[0171] ;

[0172] in, This indicates a manually preset directional deviation scale parameter;

[0173] ;

[0174] in, Indicates the threshold for determining voice activity. For time window The average logarithmic energy at time:

[0175] ;

[0176] in, This indicates that the nth channel is within the time window. Time-frequency representation at frequency f Represents a set of frequencies.

[0177] right Enhance the direction of the speech to obtain :

[0178] ;

[0179] Through the above design, the local enhancement segment is not simply a cut-out of the original audio, but a target-related enhancement segment obtained jointly based on the initial directional neighborhood, spatial candidate confidence, and local beam output.

[0180] Step S4: Reliability-weighted initialization of historical templates

[0181] Local enhancement fragments Divided into categories based on fixed time length Individual fragments Preferably, each sub-segment is 0.5 seconds long, and 50% temporal overlap is allowed between adjacent sub-segments to improve robustness under conditions of short pauses, local distortion, and residual noise. Each sub-segment is then input into a pre-trained speaker representation model. (This embodiment uses the x-vector model) to obtain the corresponding sub-fragment embedding vector:

[0182] ;

[0183] To reduce the impact of residual noise, competing speech, or local distortion in a single sub-segment on the initial historical template, this invention does not directly perform an equal-weighted average of the embedding vectors of each sub-segment, but instead performs a weighted fusion based on the reliability score of each sub-segment. The reliability score of each sub-segment is denoted as: Preferably, it is determined jointly based on the average spatial candidate confidence, average speech activity intensity, and average leakage level corresponding to the sub-segment. For example, it can be defined as:

[0184] ;

[0185] in, Non-negative weights Let be the average spatial candidate confidence of the k-th sub-segment. For the first The average speech activity probability of each sub-segment No. Average leakage level of each sub-segment:

[0186] ;

[0187] in, For the k-th sub-segment in the time window The leakage level below, For the time window within the k-th sub-segment Output energy in the neighborhood of the target direction (the direction obtained from GCC-PHAT spatial analysis). For the time window within the k-th sub-segment The maximum energy in the set of competing candidate directions;

[0188] To ensure that more reliable sub-fragments have a larger weighting in the initial historical template construction, the reliability score is mapped to a normalized weight using Softmax normalization:

[0189] ;

[0190] in, This is a weighted sensitivity coefficient used to control the degree of differentiation between the reliability scores of different sub-segments; The larger the value, the more concentrated the weight is on the high-reliability sub-segments. As a result, the embedding vectors corresponding to the higher-reliability sub-segments contribute more to the initial historical template, while the contributions of the lower-reliability sub-segments are suppressed.

[0191] The initial history template is defined as:

[0192] ;

[0193] in, To prevent positive numbers with a denominator of zero, the initial historical template obtained by this invention is not derived from a single instantaneous segment, but rather from the reliability-weighted fusion of multiple target-related sub-segments. This reduces the risk of template deviation caused by accidental peaks, local noise, or interference from competing speakers during the initialization phase. In subsequent steps, the system directly uses the initial historical template. This serves as an initial reference for the target's continuous identity within the session, eliminating the need to continue using normalized weights separately. .

[0194] Step S5: Output of target speaker extraction model and target existence probability

[0195] After the initialization phase, a Kalman filter is used to predict the target candidate direction in the current time window t based on the prediction direction of the previous time window. ;Will , Pre-state coding In a preferred embodiment of the present invention, the target speaker extraction model employs a time-frequency mask prediction network.

[0196] The time-frequency mask prediction network adopts an encoder-bottleneck-decoder structure. The encoder receives the multi-channel time-frequency features of the current processing block, which include at least the logarithmic amplitude spectrum of the target reference channel, multi-channel complex spectrum features, and spatial features derived from GCC-PHAT. The bottleneck layer is used to fuse target orientation condition features, historical template condition features, and state coding features. The decoder outputs the target time-frequency mask. .

[0197] Preferably, First, it is mapped to a direction condition vector through direction encoding, then it becomes a historical template. After linear transformation, it is mapped to a speaker condition vector, and then encoded in the pre-state. The vector is embedded and mapped into a state condition vector. The orientation condition vector, speaker condition vector, and state condition vector are concatenated or modulated channel by channel at the bottleneck layer so that the network can simultaneously obtain spatial constraints, identity constraints, and temporal control constraints.

[0198] Output target time-frequency mask After that, the current target enhanced speech is... Obtain, among which Indicates the reference channel spectrum;

[0199] At the same time, by the auxiliary output head Output the probability of the current target existing. During training, the target vocalization label is used as the supervision signal, and binary cross-entropy loss is employed for optimization.

[0200] ;

[0201] The target existence probability describes whether the target speaker is still in an effective vocal state at the current moment. This variable plays an important role in distinguishing between a short pause in the target and the loss of the target or the dominance of a competing speaker.

[0202] Step S6: Calculation of continuous control quantities under competition suppression and leakage penalty

[0203] In this invention, the continuity control quantity is not obtained using a simple linear weighting method, but is preferably constructed using a nonlinear fusion method. The reason for this is that template consistency, directional continuity, target existence probability, and the spatial candidate credibility corresponding to the current target direction represent positive evidence that the current output still belongs to the initially locked target from four aspects: identity consistency, directional movement rationality, target vocal effectiveness, and spatial observation reliability. The leakage penalty term and the competition dominance term represent the risk of non-target speech being mixed into the current output and the risk of competing candidate directions replacing the target direction, respectively. If only a linear weighting method is used, when one positive evidence is significantly insufficient, other higher-scoring terms may still raise the overall score, leading the system to misjudge the target's continued existence when there is a target pause, high output leakage, or a competing speaker dominance. Therefore, this invention preferably uses a nonlinear fusion method that constructs positive evidence as a multiplicative consistency term and negative risks as an exponential suppression term. Through this method, the continuity control quantity remains high only when multiple positive conditions are simultaneously and well satisfied; while when leakage increases or the competing direction becomes dominant, the continuity control quantity will be rapidly suppressed. This improves the sensitivity of continuous control variables to template contamination, erroneous switching, and target loss, making them more suitable for serving subsequent state transitions and template gating update mechanisms.

[0204] First, the speaker representation model is pre-trained using the current target augmented speech input to obtain the current template:

[0205]

[0206] The similarity between the current template and historical templates is defined as:

[0207] ;

[0208] Directional deviation is defined as ,in The measurement direction using GCC-PHAT spatial analysis within the current time window is defined as follows: the rate of change of direction is defined as... The directional continuity score is defined as:

[0209] ;

[0210] in, The directional deviation scale parameter is preset by humans. The speed of directional change is a pre-defined parameter.

[0211] Let the output energy in the neighborhood of the target direction be The maximum output energy in the set of competing directions is The leakage rate is then defined as:

[0212] ;

[0213] To calculate speech activity information, let the time-frequency representation of the nth channel at processing block t and frequency f be as follows: The average logarithmic energy of processing block t is defined as:

[0214]

[0215] The probability of voice activity is defined as:

[0216]

[0217] in, Represents the set of frequencies involved in the calculation. To prevent taking abnormally positive numbers from logarithms, This indicates the threshold for determining voice activity. The larger the value, the higher the probability that there is valid speech activity in the current processing block.

[0218] To measure the dominance of competing candidates relative to the target candidate, let the candidate direction be... The directional consistency in processing block t is as follows:

[0219] ;

[0220] when hour, for .

[0221] in, for Directions within the neighborhood (including) ), This represents the orientation deviation scale parameter. Candidate orientation. The overall score is defined as follows:

[0222]

[0223] in, Indicate candidate direction spatial candidate credibility, , The weights are non-negative.

[0224] Let the set of candidate directions in the reference neighborhood of the target direction be . The target direction references competing candidate directions outside the neighborhood. Then the current target candidate direction and the strongest competing candidate direction They are respectively:

[0225]

[0226]

[0227] The score difference between the strongest competing candidate and the target candidate is defined as:

[0228]

[0229] when A score >0 indicates that the overall score of the strongest competing candidate direction is higher than that of the target candidate direction, posing a risk of mistaken target switching; when When ≤0, it means that the target candidate direction is not weaker than the strongest competing candidate direction.

[0230] The continuous control quantity is defined using a nonlinear fusion method. Let the positive consistency term be:

[0231]

[0232] Where α, β, γ, and δ are all non-negative exponential weight parameters, used to adjust the contribution strength of template similarity, directional continuity, target existence probability, and spatial candidate credibility in the positive consistency term, respectively.

[0233] Let the negative risk term be:

[0234]

[0235] in .

[0236] The continuous control quantity is defined as follows:

[0237]

[0238] Among them, positive consistency terms A multiplicative structure is used to represent joint constraints where multiple conditions must be met simultaneously; negative risk items. An exponential decay structure is employed to reflect the rapid suppression effect of continuous control inputs when risk increases. Through this design, the continuous control input... The rate is higher when the target remains stable, and decreases significantly when the target pauses briefly, when a competing speaker dominates, or when output leakage is high, thus providing a unified and stable control basis for state determination and template gating updates.

[0239] Step S7: State transition scoring and state determination mechanism

[0240] like Figure 2As shown, this invention sets at least five states: locked state, tracking state, holding state, recapture state, and lost state. Specifically: Locked state: Indicates that the current output has high consistency with the historical template, the probability of the target's existence is high, and the target direction is stable; the system considers the current target to be in a stable and continuous holding phase. Tracking state: Indicates that the current target is still identified as the initially locked target, but the target direction has changed smoothly relative to the previous moment; the system needs to dynamically track the target direction reference. Holding state: Indicates that the current target's existence probability or continuous control quantity has decreased; the system cannot temporarily confirm that the target is still effectively emitting sound, but it is not enough to determine that the target is lost, so the target identity file is maintained and not released. Recapture state: Indicates that the system has transitioned from the holding state to the active recovery phase, searching for candidate directions that meet the conditions again in the target's original neighborhood or predicted neighborhood. Lost state: Indicates that a target candidate direction that meets the conditions cannot be obtained again within a preset time; the system determines that the current target has been lost.

[0241] To avoid triggering state transitions based solely on a single threshold, this invention preferably employs a hybrid mechanism of rule-based pre-screening and state transition scoring and ranking. That is, candidate states are first screened based on the basic conditions for each state to be valid; then, state transition scores are calculated for the selected candidate states; and the state with the highest score is chosen as the current state. The constructed state feature vector is as follows:

[0242] ;

[0243] in, This represents the duration of the current state. For any candidate state s, its state transition score is defined as:

[0244]

[0245] in, and These are the weights and the biases, respectively.

[0246] In a preferred embodiment, the basic conditions for each state to enter the candidate set are defined as follows:

[0247] (1) Lock-up state candidate condition: when the continuous control quantity Not lower than the lock threshold Target Existence Probability Not lower than the existence threshold Template similarity Not lower than the template threshold and directional continuity Not lower than the direction threshold When the target is locked, it enters the candidate set. This state indicates that the target is in the stable holding phase.

[0248] (2) Tracking state candidate condition: when the target exists Not lower than the threshold Template similarity Not lower than the acceptable threshold (In this embodiment, the value is 0.50), and the spatial candidate confidence of the current candidate direction is not lower than the threshold. (In this embodiment, we take 0.6), the deviation between the actual target direction and the candidate target direction at the current moment. Greater than the rest threshold And the speed of change of direction Not higher than the smoothing threshold When the tracking state is entered into the candidate set, it indicates that the target is still the original target, but its direction has shifted smoothly.

[0249] (3) Hold state candidate condition: when the continuous control quantity Below the holding threshold (In this embodiment, the value is 0.45) or the probability of the target's existence. Below the existence threshold And currently maintains a duration Not exceeded the hold window When this happens, the target is retained in the candidate set. This state indicates that the target may be briefly paused or the current observation is unreliable, but the target identity file is still preserved.

[0250] (4) Candidate conditions for recapture state: when the system is already in the hold state and the hold duration is [not specified]. Exceeding the hold window Meanwhile, the confidence level of spatial candidates reappearing in the original neighborhood or predicted neighborhood of the target is not lower than the threshold. (In this embodiment, the value is 0.55) and the probability of the target's existence is not lower than the threshold. When the candidate direction is (in this embodiment, 0.50), the reacquisition state is added to the candidate set. This state indicates that the system is attempting to resume continuous tracking of the target.

[0251] (5) Candidate condition for lost state: When the system is in the recapture state and the recapture duration is Exceeding the recapture window If no candidate direction satisfying the lock state is detected, the lost state is added to the candidate set. This state indicates that the system temporarily releases the effective output for continuous target tracking, but can still re-enter the re-acquisition process later.

[0252] Calculate the state transition score for each of the pre-screened candidate states. The highest scorer is selected as the current state.

[0253]

[0254] in, This represents the set of candidate states that have been pre-filtered by rules.

[0255] It should be noted that the locked state and tracking state in this invention do not simply correspond to being stationary and moving, but rather to continuous holding with stable direction and continuous holding with smooth direction changes, respectively. The holding state does not indicate that the system has determined the target to be lost, but rather that when the target is briefly paused or the observation is unreliable, the system temporarily freezes template updates and retains the target's identity file. The reacquisition state is used to actively recover the target after the holding timeout. The lost state indicates that the target could not be recovered within a preset window. Thus, this invention can clearly distinguish between different situations such as target pause, target movement, dominant competing speaker, and true target loss.

[0256] Through the above mechanism, the system can maintain lock when the target is stable, enter tracking when the target moves smoothly, enter hold when the target pauses briefly, perform recapture when the candidate direction reappears, and determine loss when the target has not recovered for a long time, thereby avoiding state jitter and false switching caused by a single threshold judgment.

[0257] Explanation of state determination threshold and counting variables:

[0258] The continuity control threshold for the locked state: when the continuity control amount is not lower than this threshold, it indicates that the current output has a high degree of confidence in maintaining the target continuity.

[0259] The target has a probability threshold. When the probability of the target's existence is not lower than this threshold, it indicates that the target is still in an effective vocalization state.

[0260] Template consistency threshold: When the similarity between the current representation and the historical template is not lower than this threshold, it indicates that the current output still has a high degree of consistency with the initially locked target.

[0261] : Directional continuity threshold, used to determine whether the change in the current candidate direction conforms to the characteristics of smooth target movement;

[0262] The directional deviation threshold for determining effective target movement is used to distinguish between a stationary target and a target that has made effective movement.

[0263] : Upper limit of the direction change rate, used to suppress false tracking caused by sudden changes in direction;

[0264] The maximum duration window for the hold state is incremented by processing block after the system enters the hold state and cleared to zero when exiting the hold state.

[0265] The maximum duration window for recapture state. When the recapture duration exceeds this window and the target is still not recovered, the system enters the lost state.

[0266] Cooling counter is used to suppress frequent switching within a preset short window after a state change. The counter starts when the system enters a new state and is reset or restarted when the count is full.

[0267] Step S8: Smooth update of target direction reference and gating update of historical template

[0268] When the current state is tracking, a smooth update is performed on the target direction reference, and the target direction can be updated as follows:

[0269]

[0270] in, Indicates the target direction after the current time window is updated. Indicates the target direction of the time window (if the previous time window was not in tracking mode, then...). The direction is determined by Kalman filtering. If it is in tracking mode, then... (Indicates the updated target direction) This represents the candidate directions selected by the GCC-PHAT spatial analysis module within the target neighborhood at the current moment. The adaptive smoothing update coefficients represent the direction reference. Preferably, the adaptive smoothing update coefficients are defined as:

[0271]

[0272] in, This represents the Sigmoid function. , where is a non-negative weight. Through the above design, when the continuous control quantity is high, the probability of the target's existence is high, and the change in the candidate direction is small, the system tends to update the target direction reference. When the continuous control quantity is low, the probability of the target's existence decreases, or the candidate direction changes significantly, the update magnitude of the target direction reference is reduced accordingly, thereby avoiding mistracking caused by interference from competing speakers or sudden changes in the candidate direction.

[0273] After completing the target direction reference update, when the continuous control quantity meets the update conditions and the leakage rate is below the set threshold, a gated update is performed on the historical template. Let the template update gating variable be:

[0274] ;

[0275] in, Indicates an indicator function, Indicates the continuity control threshold. This indicates that the target has a probability threshold. This represents the leakage threshold. The gating variable is used to control whether the current output result is allowed to update the historical template at the current moment.

[0276] Define the adaptive template update rate as:

[0277] ;

[0278] in, , , Non-negative weights This indicates the update strength of the historical template at the current moment. When the continuity control is high, the consistency between the current representation and the historical template is high, and leakage is low, A larger value indicates that the current output has a high reliability value for updating historical templates; conversely, a smaller value indicates a lower value. A value that is small or even zero indicates that the current result is not suitable for template updating.

[0279] The final formula for the recursive update of the historical template is:

[0280] ;

[0281] in, This indicates the history template updated in the current window. This represents the historical template of the previous time window. This represents the current representation vector extracted from the speech augmented by the current target. Represents the L2 norm, To prevent positive numbers with a denominator of zero, the above update method does not directly accumulate the original speech segments. Instead, it recursively merges the current representation vector with the historical template from the previous time step according to an adaptive update rate, so that the historical template always remains a stable representation vector that reflects the continuous identity of the target in the current session.

[0282] When the template update gating condition is not met, or when the current state is in a hold or recapture state, historical template updates are prohibited, and the following is maintained:

[0283] ;

[0284] The reason for this design is that if the historical template is forcibly updated when the target pauses briefly, a competing speaker temporarily gains an advantage, or the current output leakage is high, non-target speech information may be written into the template, leading to template contamination and subsequent error accumulation. Therefore, this invention uses a joint mechanism of smooth direction reference update and historical template gating update to enable the system to continuously track the target direction when the target moves smoothly, and to suppress template drift when the output is unreliable, thereby improving the stability and robustness of continuous extraction of the same target speaker.

[0285] Training and Inference Examples

[0286] During the training phase, multi-channel training samples are constructed, including scenarios such as continuous target speech, short target pauses, competing speakers continuing to speak during pauses, slow target movement, multiple speakers speaking at close angles, reverberation, and background noise. Each sample includes at least multi-channel mixed speech, clean target speech, target initial direction, target presence label, and target directional trajectory. Preferably, the duration of a single training sample is 4 to 8 seconds, and it contains at least one short target pause segment and one active competing speaker segment to enhance the model's ability to learn scenarios with no mis-cutting of pauses and suppression of competition.

[0287] During data construction, the initial target orientation can be directly provided by the geometric relationship of the simulated array, or it can be provided by the external pointing module, user initial specification, or independent positioning module in the actual acquisition system. Target presence labels are preferably generated by processing blocks: when the proportion of effective speech frames with clean speech of the target within the corresponding processing block is not less than a preset ratio. If the target exists, the label is set to 1; otherwise, it is set to 0. The target's orientation trajectory is preferably given directly from the known trajectory in the simulation scenario, or calculated from the calibration pose and array geometry relationship in the actual data.

[0288] The primary loss of the target speaker extraction model is preferably a weighted combination of SI-SDR loss and STFT loss, while the auxiliary loss is the binary cross-entropy loss corresponding to the target presence probability. For example, the total loss can be written as:

[0289]

[0290] in The weights are non-negative. Through joint training, the network simultaneously learns the ability to reconstruct target speech and the ability to determine the presence of targets.

[0291] Preferably, during training, an initialization link generation history template consistent with inference is used. This involves first obtaining the initial direction based on the GCC-PHAT response within the target's initial direction neighborhood, then generating local enhancement segments through local beamforming and initialization soft masks, and finally obtaining the initial history template through reliability-weighted aggregation. This reduces performance bias caused by training-inference inconsistency.

[0292] During the inference phase, the system only needs to input multi-channel mixed speech and the initial direction of the target; all other variables are obtained through online estimation. When the target experiences a short pause, it is preferable to output silence or low-energy retained output rather than retaining the complete speech waveform of the previous frame; simultaneously, the current target identity file and historical templates are not immediately released to avoid accidental target switching caused by natural pauses. When the state transitions to hold or recapture state, historical template updates are prohibited.

[0293] XI. Example of Optimal Parameter Settings

[0294] Under a set of preferred parameters, the frame length Use 256 points, and employ a Hamming window in the window function; initialize the window. Take 1 second, initial neighborhood half-width Continuous and stable number of blocks processed Continuity control threshold of locked state Set to 0.75, the probability threshold for the existence of the target. Set to 0.5, template threshold Set the value to 0.65 to determine the directional deviation threshold for effective target movement. The maximum angle is set to 5°, the maximum speed of direction change is set to 4°, and the maximum duration window for maintaining the state is set. Take 3 seconds, maximum duration window of recapture state Set the template update control threshold to 1 second. Set the threshold to 0.7. The value is 0.3; the above parameters are not limitations of the present invention and can be adjusted equivalently for different array sizes, sampling rates and application scenarios.

[0295] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation, characterized in that, Specifically: Preprocessing, framing, windowing, and short-time Fourier transform are performed on N synchronously acquired multi-channel mixed speech signals; In the initial time window Below, based on the initial direction of the target speech Building a neighborhood Set a fixed step size, in Candidate directions are extracted from within; GCC-PHAT spatial analysis is used to explicitly model the spatial candidate confidence and calculate the spatial candidate confidence of candidate direction d. ;exist The candidate direction that meets the conditions is selected as the initial direction. That is, the measurement direction; candidate directions with speech activity outside the neighborhood of the measurement direction are used as competing candidate directions; Calculate the scores of the competing candidate directions, and select the competing candidate direction with the highest score as the strongest competing candidate direction. Based on the probability of voice activity Score against the strongest competitor Enhance the voice in the initialization direction; Locally enhanced speech segments ;Will Divided into Individual fragments Calculate the reliability of each sub-segment Therefore, the normalized weights of each sub-segment are calculated. Calculate the initial historical template ; Using GCC-PHAT spatial analysis, the measurement direction within the current time window t is obtained. The history template of the previous time window Using the historical template of the current time window, predict the prediction direction at the current time window t. The speech in the current time window is enhanced, and the probability of the target speech existing in the current time window is calculated. ; The enhanced speech is input into the speaker representation model to extract the current template. ,calculate and similarity between Voice leakage level and target direction continuity score ;Will , , as well as As a positive consistency term, As a negative risk term, it generates continuous control quantities. ; Based on the state of the previous time window and the state feature vector of the current time window, determine the state of the current time window and whether to update. And historical templates.

2. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 1, characterized in that: Calculate the normalized peak intensity for any candidate direction. : ； in, For the candidate direction set, To prevent constants with a denominator of zero; Spatial response for candidate directions: ； Where i and j represent the i-th and j-th microphones, Indicates a set of microphones. This represents the weight of the channel pair corresponding to the i-th and j-th microphones. Let represent the time delay of the i-th and j-th microphones in the candidate direction. The weighted cross-correlation function is: ； in, This represents the cross-power spectrum between the i-th and j-th microphones. Where f represents frequency, This represents the signal after performing a short-time Fourier transform on the signal transmitted from the i-th microphone; * indicates conjugate; Calculate the peak sharpness for any candidate direction. : ； in, The neighborhood of the candidate direction; Calculate its cross-frame stability for any candidate direction. ; (From the moment of crossing frames to the current frame) in, The length of the cross-frame stability window. This represents the minimum resolvable angle of the microphone array. express The direction corresponding to the main peak at any given moment; This refers to the time corresponding to the first frame of the speech signal within the cross-frame stability window. This indicates the time corresponding to the last frame in the current time window. If the current time window is... In this case, cross-frame stability is not calculated; Calculate the consistency of speech activity for any candidate direction: ； in, The length of the voice activity consistency window. Represents an exponential function. express The voice activity detection result corresponding to the candidate direction d at time moment; The time corresponding to the first frame of the speech signal within the speech activity consistency window, if the current time window is In this case, the consistency of speech activity is not calculated; Calculate the competition penalty term for any candidate direction. : ； in, Show set The set of candidate directions other than candidate direction d. Calculate the spatial candidate confidence of candidate direction d : in, All are non-negative weights. For the Sigmoid function, The normalized peak intensity for the candidate direction. The peak sharpness of the candidate direction.

3. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 2, characterized in that, Select the initialization direction by setting the following conditions. : (1) ； (2) ; The candidate credibility threshold; (3) There is a continuous range of time sampling points within the current time window, and the directional deviation between adjacent time points is less than or equal to a preset directional deviation threshold. And each sampling point can detect Or with The deviation between them is less than the threshold. Candidate directions; (4) Less than or equal to , This is the threshold for voice activity consistency.

4. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 2, characterized in that, The corresponding steering vector is Calculate the normalized delay summation beamforming weight vector: ； in, To prevent constants with a denominator of 0; f represents frequency; Calculation direction directional beamforming frequency domain: ； right Enhance the direction of the speech to obtain : ； in, For soft masking: ； in, Let be the probability of speech activity, that is, the probability of speech occurring. The difference in scores: ； in, This represents the strongest competing candidate direction within the current time window t. The direction with the highest score in the neighborhood of the measured direction under the current time window t; For the scoring function: ； in, Indicates the direction within the neighborhood of the measurement direction. , Non-negative weights for In the time window Consistency in direction: ； in, For time window The predicted direction, when hour, for ; ； in, Indicates the threshold for determining voice activity. For time window The mean logarithmic energy is: ； in, This indicates that the nth channel is within the time window. Time-frequency representation at frequency f Represents a set of frequencies.

5. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 1, characterized in that, Calculate the reliability of each sub-segment : ； in, Non-negative weights Let be the average spatial candidate confidence of the k-th sub-segment. For the first The average speech activity probability of each sub-segment No. Average leakage level of each sub-segment: ； in, For the k-th sub-segment in the time window The leakage level below, For the k-th sub-segment in the time window Output energy in the neighborhood of the measurement direction below. For the k-th sub-segment in the time window The maximum energy in the set of competing candidate directions; Calculate the normalized weights: ； in, This is a weighted sensitivity coefficient used to control the degree of differentiation between the reliability scores of different sub-segments; Calculate the initial history template: ； in, , For speaker representation model.

6. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 1, characterized in that, Using a Kalman filter, the prediction direction for the current time window t is predicted based on the prediction direction of the previous time window. That is, the target candidate direction, will , The speech and state feature vectors at the current time window t are input into the time-frequency mask prediction network to obtain the enhanced speech at the current time. The probability of the target speech existing at the current moment. .

7. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 6, characterized in that, Calculate the similarity between the current template and current historical templates: ； Calculate the continuity score of the target candidate direction : ； in, ,in The measurement direction within the current time window. The speed of directional change is preset by humans. ; The positive consistency term is: ； in, , , as well as All are non-negative exponential weight parameters; Let the negative risk term be: ； in ; The difference between the score of the strongest competitive candidate direction and the score of the measured direction in the current time window; Calculate the continuous control quantity: 。 8. The multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 7, characterized in that, The states include locked state, tracking state, holding state, recapture state, and lost state; If continuous control quantity Greater than or equal to the preset locking threshold The probability of the target speech exists Greater than or equal to the preset existence threshold The similarity between the current template and the historical templates at the current moment. Greater than or equal to the preset template threshold Target direction continuity score Greater than or equal to the preset direction threshold When this happens, the locked state is used as a candidate state for the current time window; When the target speech exists, probability Greater than or equal to the threshold , Greater than or equal to the acceptable threshold ,when Spatial candidate credibility is greater than or equal to the threshold The absolute value of the difference between the measured direction and the predicted direction within the current time window is greater than or equal to the stationary threshold. And the speed of change of direction Less than or equal to the smoothing threshold When this happens, the tracking status is used as a candidate status for the current time window; When continuous control quantity Less than the holding threshold or the probability of the target speech exists Less than the existence threshold And the duration of maintaining the state. Less than or equal to keep window When this happens, the current state will be retained as a candidate state for the current time window; When the system is in a hold state, and the hold duration is... Larger than the hold window Simultaneously, if the spatial candidate reappears in the neighborhood of the measurement direction or the neighborhood of the prediction direction, the confidence level is greater than or equal to the threshold. And the target speech has a probability Greater than or equal to the threshold When considering candidate directions, the recapture state is used as the candidate state for the current time window; When the system is in recapture state, and the recapture duration is... Exceeding the recapture window If no candidate direction that satisfies the locked state is detected, the lost state is taken as the candidate state of the current time window. Calculate the state transition score for each candidate state. The highest scorer is selected as the current state. ； in, Represents the set of candidate states. , and These are weights and biases, respectively. For state feature vectors: ,in For cooling count.

9. A multi-channel target speech extraction method based on initial orientation anchoring and continuous identity preservation according to claim 8, characterized in that, When the current time window is in tracking mode, perform a smooth update on the measured direction: ； in, This indicates the updated predicted direction, which is also the updated target direction. Indicates the measurement direction within the current time window. Indicates the adaptive smooth update coefficients: ； in, This represents the Sigmoid function. , where is a non-negative weight; Perform gated updates on historical templates: ； in, This is the updated history template for the current time window. Indicates the update strength of the historical template at the current moment: ； in, , , Non-negative weights Update the gated variables for the template: ； in, , as well as All indicate that they are related to , , The corresponding threshold; It is an exponential function.

Images