A dynamic single sound source DOA tracking method and system

Abstract

The application discloses a dynamic single sound source DOA tracking method and system, and relates to the technical field of signal processing, and comprises the following steps: stacking a full-frequency spatial spectrum diagram and a plurality of sub-band spatial spectrum diagrams of each frame along a channel dimension to obtain a multi-channel spatial spectrum diagram, stacking the multi-channel spatial spectrum diagrams of continuous multiple frames along a time dimension to obtain a space-time input tensor; extracting maximum value position coordinates of the full-frequency spatial spectrum diagrams of the continuous multiple frames, and constructing a maximum value coordinate tensor based on the maximum value position coordinates; the position coordinates comprise an azimuth angle coordinate and a pitch angle coordinate; inputting the space-time input tensor and the maximum value coordinate tensor into a pre-trained tracking network to obtain a dynamic single sound source DOA tracking result on continuous time frames. The application can explicitly retain spatial structure differences under different frequency bands, and avoid weakening of effective information caused by premature aggregation of a single full-frequency spatial spectrum diagram in a frequency dimension.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a dynamic single-source DOA tracking method and system. Background Technology

[0002] Direction of Arrival (DOA) estimation and tracking is a crucial research area in microphone array signal processing, with wide applications in service robots, voice interaction, conferencing systems, and indoor acoustic sensing. For moving sound sources in dynamic scenes, the direction changes continuously over time. The system needs not only to provide the sound source direction result at a single frame but also to output a stable, continuous, and temporally consistent direction trajectory across consecutive time frames. Therefore, achieving robust dynamic DOA tracking in complex indoor environments has become one of the key challenges in array signal processing.

[0003] Most existing dynamic sound source tracking methods are based on frame-by-frame localization results or spatial spectrum analysis, combined with filtering or temporal modeling methods to achieve continuous tracking. Traditional methods have a relatively mature research foundation in state-space modeling and recursive estimation. However, in complex indoor acoustic environments, DOA observations are easily affected by noise, reverberation, multipath propagation, and intermittent sound generation, resulting in problems such as observation drift, spurious peak enhancement, and trajectory discontinuity. In recent years, deep learning methods have begun to be used for dynamic DOA estimation and continuous tracking. By modeling spatial spectrum sequences, they have improved the stability of direction estimation in dynamic scenes to some extent.

[0004] In existing deep learning-based dynamic tracking methods, the spatial spectrum of Steered Response Power using Phase Transform (SRP-PHAT) is a commonly used input representation. This type of method can explicitly give the spatial response intensity on the candidate orientation grid and provide a basic input for subsequent network modeling. However, when inputting a single full-frequency SRP-PHAT spatial spectrum, the spatial structure differences existing in different frequency bands are difficult to preserve, and the effective information related to the true sound source orientation in some frequency bands is easily weakened during the front-end aggregation process. At the same time, the existing tracking networks have limited temporal modeling capabilities on continuous time frames. Therefore, when there are strong spurious peaks or local peak drift in the spatial spectrum, existing methods are prone to problems such as discontinuous orientation estimation, large trajectory fluctuations, and insufficient temporal consistency. Summary of the Invention

[0005] In view of the defects of the existing technology, the present invention provides a dynamic single-source DOA tracking method and system, which solves the existing problems.

[0006] The present invention adopts the following technical solution: In a first aspect, the present invention provides a dynamic single-source DOA tracking method, comprising the following steps: Acquire multi-channel signals from multiple consecutive frames of the dynamic single sound source to be tracked; The SRP-PHAT spatial spectrum of each frame of multi-channel signal is obtained in the full frequency band and multiple sub-frequency bands to obtain the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum. The full frequency spatial spectrum and multiple sub-frequency band spatial spectrum of each frame are stacked along the channel dimension to obtain the multi-channel spatial spectrum. The multi-channel spatial spectrum of multiple consecutive frames is stacked along the time dimension to obtain the spatiotemporal input tensor. The maximum position coordinates are extracted from the full-frequency spatial spectrum of multiple consecutive frames, and a maximum coordinate tensor is constructed based on the maximum position coordinates; the position coordinates include azimuth coordinates and elevation coordinates; The spatiotemporal input tensor and the maximum value coordinate tensor are input into a pre-trained tracking network, which includes a bi-branch coding module, a bi-branch spatial feature extraction module, and a mapping module. The bi-branch coding module extracts features from the spatiotemporal input tensor and the maximum value coordinate tensor respectively and then fuses them to obtain a fusion result. The bi-branch spatial feature extraction module extracts spatial features from the fusion result in the azimuth and elevation directions respectively and then concatenates them to obtain a concatenated result. The mapping module maps the concatenated result to obtain the dynamic single-source DOA tracking result on continuous time frames.

[0007] Preferably, acquiring the SRP-PHAT spatial spectrum of each frame of multi-channel signal in multiple sub-frequency bands specifically includes the following steps: Within the half-spectrum of the short-time Fourier transform, the frequency axis is uniformly divided into multiple sub-bands; For each sub-band, the SRP-PHAT spatial spectrum of each sub-band is calculated by using the frequency components that fall within the range of that sub-band.

[0008] Preferably, the dual-branch coding module includes a spatial spectrum coding branch, a coordinate coding branch, and a fusion module. The spatial spectrum coding branch extracts features from the spatiotemporal input tensor, the coordinate coding branch extracts features from the maximum coordinate tensor, and the fusion module fuses the two feature extraction results.

[0009] Preferably, the spatial spectrum encoding branch includes a first convolutional module and a second convolutional module, with the number of channels of the first and second convolutional modules set to 5-24-24 respectively, and the kernel size of the convolutional module being 5×3×3; the coordinate encoding branch includes a third convolutional module and a fourth convolutional module, with the number of channels of the third and fourth convolutional modules set to 2-8-8 respectively, and the kernel size of the convolutional module being 5×1×1; the fusion module uses a 1×1×1 3D convolution to fuse the two feature extraction results, resulting in 32 channels after fusion.

[0010] Preferably, the dual-branch spatial feature extraction module includes a pitch angle branch, an azimuth angle branch, and a stitching module. The pitch angle branch extracts spatial features from the fused result in the pitch angle direction, the azimuth angle branch extracts spatial features from the fused result in the azimuth angle direction, and the stitching module stitches the two spatial feature extraction results together.

[0011] Preferably, the pitch angle branch includes a fifth convolution module, the azimuth angle branch includes a sixth convolution module, the fifth and sixth convolution modules have a total of 32 convolution channels, the convolution kernel size is 5×3×3, the fifth convolution module uses 1×1×2 max pooling, and the sixth convolution module uses 1×2×1 max pooling.

[0012] Preferably, the mapping module includes a seventh convolution module, an eighth convolution module, and a ninth convolution module, wherein the kernel length of the seventh convolution module, the eighth convolution module, and the ninth convolution module is 5, and the dilation coefficients are 1, 2, and 4, respectively.

[0013] Preferably, the pre-training of the tracking network specifically includes the following steps: The microphone array positions are randomly generated in a room of random size. The starting and ending points of the sound source trajectory are randomly selected in the feasible area of ​​the room and sinusoidal perturbation is superimposed to generate a continuous motion trajectory. Sound propagation simulation is performed by image source method to obtain multi-channel observation signals as training samples. The training samples are divided into frames, and the spatiotemporal input tensor and maximum coordinate tensor of each frame are obtained. The input tensors of multiple consecutive frames are used as network input, and the real DOA at the corresponding time is converted into a direction vector in a three-dimensional Cartesian coordinate system as supervision label. The mean squared error (MSE) is used as the loss function, and the tracking network is trained in a supervised learning manner to optimize the network parameters.

[0014] Preferably, a multi-channel signal of multiple consecutive frames from the dynamic single sound source to be tracked is acquired through a microphone array.

[0015] Secondly, the present invention provides a dynamic single-source DOA tracking system, comprising: The acquisition module is used to acquire multi-channel signals of multiple consecutive frames from the dynamic single sound source to be tracked. The stacking module is used to acquire the SRP-PHAT spatial spectrum of each frame of multi-channel signal in the full frequency band and multiple sub-frequency bands, respectively, to obtain the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum; the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum of each frame are stacked along the channel dimension to obtain the multi-channel spatial spectrum; the multi-channel spatial spectrum of multiple consecutive frames are stacked along the time dimension to obtain the spatiotemporal input tensor. The module is used to extract the position coordinates of the maximum value from the full-frequency spatial spectrum of multiple consecutive frames, and to construct the maximum value coordinate tensor based on the position coordinates of the maximum value; the position coordinates include azimuth coordinates and elevation coordinates; The tracking module is used to input the spatiotemporal input tensor and the maximum value coordinate tensor into a pre-trained tracking network. The tracking network includes a bi-branch encoding module, a bi-branch spatial feature extraction module, and a mapping module. The bi-branch encoding module extracts features from the spatiotemporal input tensor and the maximum value coordinate tensor respectively and then fuses them to obtain a fusion result. The bi-branch spatial feature extraction module extracts spatial features from the fusion result in the azimuth and elevation directions respectively and then concatenates them to obtain a concatenated result. The mapping module maps the concatenated result to obtain the dynamic single-source DOA tracking result on continuous time frames.

[0016] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention constructs a Spectral-SRP multi-channel input consisting of a full-frequency spatial spectrum and multiple sub-band spatial spectra. This explicitly preserves the spatial structure differences across different frequency bands, avoiding the weakening of effective information caused by premature aggregation of a single full-frequency spatial spectrum in the frequency dimension. This enhances the representational and directional discrimination capabilities of the input spatial spectrum in complex indoor environments. Simultaneously, a dual-branch coding module encodes and fuses the spatial spectrum input and peak coordinate input separately, while a dual-branch spatial feature extraction module extracts spatial features in the azimuth and elevation directions respectively. A mapping module performs causal temporal modeling, enabling the network to simultaneously utilize multi-band spatial structure differences, peak position priors, and temporal information across continuous time frames. This suppresses spurious peaks, main peak diffusion, and peak drift caused by noise, reverberation, and multipath propagation, reduces jitter and jumps in frame-by-frame DOA estimation, and improves the stability, trajectory continuity, and temporal consistency of dynamic single-source DOA tracking. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 For an SNR of 30dB, Examples of SRP-PHAT power graphs at different resolutions under the given conditions; red dots represent the true DOA of the sound source, and black dots represent the locations of the maximum values ​​in the power graph; in, Figure 1 (a): 4x8 resolution Figure 1(b): 8 x 16 resolution Figure 1 (c): 16x32 resolution Figure 1 (d): 32x64 resolution; Figure 2 For an SNR of 5dB, Examples of SRP-PHAT power plots at different resolutions under different conditions; in, Figure 2 (a): 4x8 resolution Figure 2 (b): 8 x 16 resolution Figure 2 (c): 16x32 resolution Figure 2 (d): 32x64 resolution; Figure 3 This is a flowchart of the preprocessing process for Cross3D and SpecResCross3D of the present invention; in, Figure 3 (a): Cross3D preprocessing flowchart, Figure 3 (b): Preprocessing flowchart of SpecResCross3D; Figure 4 For an SNR of 30dB, Example of SRP-PHAT power diagrams for the same scene and the same time frame under 32x64 resolution with full bandwidth and different sub-band numbers; in, Figure 4 (a): Full-band Figure 4 (b): Subband 1, Figure 4 (c): Subband 2, Figure 4 (d): Subband 3, Figure 4 (e): Subband 4; Figure 5 For an SNR of 5dB, Example of SRP-PHAT power diagrams for the same scene and the same time frame under 32x64 resolution with full bandwidth and different sub-band numbers; in, Figure 5 (a): Full-band Figure 5 (b): Subband 1, Figure 5 (c): Subband 2, Figure 5 (d): Subband 3, Figure 5 (e): Subband 4; Figure 6 The ResCross3D network structure diagram of the present invention corresponds to a 32x64 Spectral-SRP power diagram; Figure 7This is a flowchart of the dynamic sound source simulation and Spectral-SRP feature construction process of the present invention; Figure 8 This invention provides a comparison of the RMSAE simulation datasets of SpecResCross3D and SRP-PHAT at different reverberation times under different SNR conditions at a fixed 32x64 spatial spectral resolution. in, Figure 8 (a): SNR=30dB, Figure 8 (b): SNR=5dB; Figure 9 This invention provides a comparison of the RMSAE simulation datasets of SpecResCross3D and SRP-PHAT under different SNR conditions at a fixed 32x64 spatial spectral resolution. in, Figure 9 (a): , Figure 9 (b): ; Figure 10 Example diagrams of trajectory estimation for real DOA, SRP-PHAT peak estimates and SpecResCross3D estimates under different simulation scenarios of the present invention; in, Figure 10 of (a): SNR=30dB Figure 10 (b): SNR=5dB Figure 10 (c): SNR=30dB; Figure 11 This invention provides a comparison of the tracking RMSAE of SpecResCross3D and various baseline methods under different SNR conditions at a fixed 32x64 spatial spectral resolution and different reverberation time conditions. in, Figure 11 of (a): , Figure 11 (b): ; Figure 12 This paper compares the DOA estimation results of SpecResCross3D with other methods in the second recording of Task 3 in the LOCATA dataset at 32x64 resolution. Figure 13 This is a flowchart of a dynamic single-source DOA tracking method according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To address the problems in the background technology, this invention proposes a dynamic single-source DOA tracking method, specifically a dynamic single-source DOA tracking method based on a combination of multi-sub-band SRP-PHAT spatial spectrum input and a residual bi-branch causal three-dimensional convolutional network, termed SpecResCross3D. Here, Spec represents the Spectral-SRP spatial spectrum input, Res represents the residual improvement network, and Cross3D represents the naming convention inherited from the Cross3D tracking framework. This method aims to enhance the representation capability of the input spatial spectrum and improve the spatiotemporal joint modeling capability across consecutive time frames, thereby improving the stability, robustness, and temporal consistency of dynamic DOA tracking in complex indoor environments, making the output directional trajectory more continuous and smooth. This scheme is suitable for single-source DOA tracking in complex indoor dynamic scenes, especially exhibiting better stability, trajectory continuity, and temporal consistency under noise, reverberation, and multipath conditions.

[0021] The core of this invention lies in constructing a multi-sub-band SRP-PHAT spatial spectrum input and combining it with a residual bi-branch causal three-dimensional convolutional network to achieve continuous and stable tracking of the DOA of a single sound source in complex indoor dynamic scenes. (Refer to...) Figure 13 Specifically, it includes the following steps:

[0022] S1: Constructing the Network Input. The multi-channel signals received by the microphone array are converted into SRP-PHAT spatial spectra, and the corresponding spatial spectra are calculated in different sub-bands. The full-frequency spatial spectra and the multiple sub-band spatial spectra are stacked along the channel dimension to form a multi-channel Spectral-SRP spatial spectrum input. Simultaneously, the coordinates of the maximum value location are extracted from the full-frequency spatial spectrum and constructed as peak coordinate auxiliary inputs. These, along with the multi-channel Spectral-SRP spatial spectrum inputs, serve as network inputs to provide explicit directional prior information.

[0023] To illustrate how the SRP-PHAT spatial spectrum forms the input for subsequent networks, its basic construction process is first given.

[0024] Assume the direction of the sound source is determined by the azimuth angle. With pitch angle Indicated, denoted as For array-received signals, the direction of the sound source can be obtained by SRP maximization, that is, searching in the candidate direction space for the direction that maximizes the array output power as the DOA estimation result:

[0025] (1); in, This is an estimate of the direction of the sound source. As the sound source, Indicate candidate direction The turning response power is calculated. To enhance robustness under noise and reverberation conditions, this invention further employs a phase-transform-weighted SRP-PHAT representation. The core idea is to apply PHAT weighting to the cross-power spectrum in the frequency domain to suppress the interference of amplitude information on the estimation of correlated peaks, making the spatial spectrum construction more reliant on the cross-channel phase difference—a spatial clue closely related to propagation delay. Its frequency domain form can be written as:

[0026] (2); in, This represents the number of microphone sensors in the array. for Fourier transform, For the first n The time-domain signal received by each sensor Angular frequency, For the first The frequency response of the channel filter for To the The propagation delay corresponding to each sensor. Further, equation (2) can be rewritten using GCC functions as:

[0027] (3); in, ,and For the first The and the first GCC between sensor signals: (4); in, For the first The spectrum of signals received by each sensor The complex conjugate, For weighted functions, For the first Frequency response of each channel filter The complex conjugate of the phase difference. When PHAT weighting is used, the spatial spectrum construction can focus more on the time delay information corresponding to the phase difference, thereby obtaining the SRP-PHAT directional response representation used for subsequent input construction.

[0028] After obtaining the SRP-PHAT directional response, candidate directions can be further sampled on a discrete angle grid, thus representing the continuous direction search results as a two-dimensional spatial spectrum. Let the candidate directions be determined by both azimuth and elevation angles, denoted as […]. ,in Indicates azimuth. This represents the pitch angle. Therefore, at discrete angular resolution, each frame can be constructed as a... The two-dimensional power diagram, in which and These represent the number of discrete sampling points in the azimuth and elevation directions, respectively. Each element in the figure corresponds to the SRP-PHAT response value in a specific direction. (Continuous) The spatial spectrograms of frames can be further stacked along the time dimension, thus providing an input basis for subsequent unified modeling of the spatial structure and temporal evolution of the network. Combined with... Figure 1 and Figure 2 As can be seen, as the resolution is increased from 4×8 to 32×64, the power map significantly improves the characterization of the main peak and its surrounding spatial structure. While further increasing the resolution can still bring some refinement, the benefits are relatively limited, and it also increases the input size and subsequent computational overhead. Therefore, this implementation adopts 32×64 as the default spatial spectrum resolution to achieve a reasonable trade-off between spatial structure representation capability and computational complexity.

[0029] Under complex acoustic conditions, SRP-PHAT power maps may exhibit main peak diffusion and multi-peak structures, limiting the stability of DOA point estimation based on a single full-frequency power map. Although the full-frequency SRP-PHAT power map essentially aggregates the frequency dimension, the spatial spectrum morphology and reliability of different frequency bands often differ under noise, reverberation, and array geometric constraints. Premature fusion of frequency band information at the front end may mask frequency band structural features useful for DOA determination. Therefore, based on the SRP-PHAT spatial spectrum construction, this implementation proposes a multi-sub-band SRP-PHAT spatial spectrum input feature construction method. Unlike using only a single full-frequency SRP-PHAT power map as input, this implementation constructs the input as a multi-channel stacked representation of the full-frequency power map and multiple sub-band power maps, while employing a maximum value coordinate tensor input method, maintaining the network backbone structure unchanged. The full-frequency band has one element, and the sub-frequency band has four elements.

[0030] like Figure 3 As shown, let the first... The SRP-PHAT full-band power diagram of the frame is as follows To construct the sub-band power map, the frequency axis is discretized within the half-spectrum of the short-time Fourier transform. Let the sampling rate be... The STFT length is The number of bins corresponding to the half-spectral frequencies of the real signal is The half-spectrum is uniformly divided into... part, The bin index represents the frequency range within the half-spectrum of the short-time Fourier transform, and the first bin index is defined. The frequency bin set corresponding to each sub-band is:

[0031] (5); Within each sub-band, the SRP-PHAT power map is calculated using only the frequency components of that band, resulting in the sub-band power map. , The spectral value is a real number. Therefore, the Spectral-SRP spatial spectral representation constructed in this invention is defined as a stack along the channel dimension:

[0032] (6); in, This indicates a channel stacking operation. It will continuously... Frames Stacking them along the time dimension yields the spatiotemporal tensor representation of the network input.

[0033] The above definition gives the form of the Spectral-SRP input representation. For ease of implementation, this invention sequentially calculates the full-band power map and the power maps of each sub-band in each time frame, and stacks the channels according to equation (6); then the continuous... The multi-channel power map of the frame is organized along the time dimension to form a spatiotemporal input tensor. The specific process for constructing Spectral-SRP features is as follows:

[0034] S11: Speech framing and activity detection: The multi-channel input signal is framed and combined with Voice Activity Detection (VAD) to remove silent frames or low-energy frames, so as to reduce unstable estimations during silent segment inference.

[0035] S12: Full-frequency SRP-PHAT power map calculation: Based on the SRP-PHAT calculation method described above, the full-frequency power map is calculated on the angle grid for each frame. This power plot depicts the phase-consistent accumulation results along the candidate directions, with its peak position ideally corresponding to the true DOA.

[0036] S13: Subband SRP-PHAT power diagram calculation: Divide the frequency axis into... Each sub-band is used, and the power map is calculated separately for each sub-band. Compared to the full-band power map, the sub-band power map preserves the band-dependent spatial spectrum structure, allowing the effects of noise, reverberation, and array geometric constraints at different frequency bands to be presented in an explicit manner.

[0037] S14: Channel stacking to form a multi-channel spatial spectrum input: [This is followed by a seemingly unrelated sentence fragment: "will..."] and Stacked along the channel dimension And further construct a continuous The spatiotemporal input tensor consists of frames. It is important to emphasize that this invention does not perform front-end weighted summation or fusion of power maps from different frequency bands, but rather directly entrusts the frequency band differences to subsequent network adaptive learning.

[0038] S15: Max Tensors: Since the peak position of the power map plays an important indicative role in DOA determination, and the Arg Max operation has highly nonlinear and non-differentiable characteristics, it is difficult for the network to stably approximate it during end-to-end training. Therefore, this invention refers to the explicit peak encoding approach in existing research and performs the Arg Max operation on the full-frequency power map to extract the peak coordinates. It is then constructed as two additional input channels and fed into the network to provide explicit directional cues for subsequent time series estimation.

[0039] S16: Input network: The Spectral-SRP multi-channel power map tensor and the maximum value coordinate tensor can be uniformly represented as a four-dimensional tensor. .in, Indicates the length of the input sequence. and These represent the number of discrete grids in the depression and azimuth directions, respectively. The number of input channels is given. Finally, this four-dimensional tensor is used as the input to the backbone network ResCross3D for subsequent time-series estimation and tracking of DOA.

[0040] like Figure 4 and Figure 5 As shown, to intuitively demonstrate the necessity of multi-sub-band spatial spectrum, this invention selects the same scene and the same time frame, and at a fixed angular resolution of 32x64, respectively provides the full-frequency SRP-PHAT power map and the sub-band power map under the example setting of B=4. Because it is the same scene and the same time frame, the positions of the true values ​​(red dots) in the figures are the same. It should be noted that... Figure 4 and Figure 5The actual values ​​(red dots) in the graph do not actually exist; they are marked in the graph for visualization purposes. It can be observed that even under favorable conditions, the peak shapes and spurious peak distributions of the power maps for each sub-band still differ; under unfavorable conditions, the differences between bands are more significant, with some sub-bands potentially retaining a clearer main peak structure, while others show stronger spurious peaks or main peak diffusion. For example... Figure 5 As shown in Figure (b), although it is sub-band 1, the estimated maximum value is closer to the true DOA value, and the information carried is more accurate than that of the full-band. Therefore, Spectral-SRP explicitly preserves the frequency band-related spatial spectrum differences by stacking channels of the "full-band + multi-sub-band" spatial spectrum, providing the network with richer discrimination clues and helping to improve the stability and robustness of DOA estimation under complex conditions such as noise and reverberation.

[0041] Step 2: Construct the ResCross3D tracking network and complete continuous DOA output. The multi-channel Spectral-SRP spatial spectrum input and peak coordinate auxiliary input are fed into the residual bi-branch causal 3D convolutional network ResCross3D for joint modeling. Through front-end bi-branch encoding, bi-branch spatial feature extraction, and causal temporal modeling, continuous regression output of single sound source DOA is achieved.

[0042] This implementation proposes a ResCross3D network for joint modeling of multi-channel spatial spectrum inputs across consecutive time frames. The network consists of three parts: a front-end dual-branch coding module, a dual-branch spatial feature extraction module, and a temporal modeling and orientation regression module (mapping module). The network structure is as follows: Figure 6 As shown. Unlike the method of extracting the input spatial spectrum sequence by shallow convolution along a single path, ResCross3D reorganizes the network backbone structure around the Spectral-SRP multi-channel input, enabling the network to simultaneously utilize multi-band spatial spectrum differences, peak position priors, and temporal continuity in historical frames to achieve continuous and stable tracking of the DOA of a single sound source in complex indoor dynamic scenes.

[0043] In the specific implementation, the multi-channel Spectral-SRP spatial spectrum input and peak coordinate auxiliary input obtained in the first part are jointly fed into the residual dual-branch causal 3D convolutional network ResCross3D. Considering that the two types of inputs are semantically different, the Spectral-SRP input mainly carries the local spatiotemporal structure information in the multi-band spatial spectrum, while the peak coordinate auxiliary input mainly provides explicit directional priors. Therefore, in the front-end encoding stage, they are first fed into two independent branches for feature extraction to avoid premature mixing of information from different sources in the shallow layer stage. The spatial spectrum encoding branch uses two residual 3D causal convolutional modules with channels set to 5-24-24 and kernel size of 5×3×3, used to extract local spatial patterns on the angular grid and their structural features that change over time. The coordinate encoding branch also uses two residual 3D causal convolutional modules with channels set to 2-8-8 and kernel size of 5×1×1, used to perform lightweight encoding of the peak coordinate prior and retain its continuous change information in the time dimension. After completing the dual-branch front-end encoding, the two features are concatenated in the channel dimension and mapped to a unified channel number of 32 through a 1×1×1 3D convolution to form the shared representation required for the subsequent spatial feature extraction stage.

[0044] In the spatial feature extraction stage, ResCross3D further feeds the fused shared representation into the pitch and azimuth branches respectively. This design does not simply repeat two network paths, but rather addresses the differences in spatial distribution between pitch and azimuth in DOA tracking, preserving the positional information in their corresponding angular dimensions. Both branches employ residual 3D causal convolution modules for feature extraction, with a fixed number of 32 convolutional channels and a kernel size of 5×3×3. Simultaneously, to avoid weakening spatial positional information directly related to angle discrimination during downsampling, the two branches use different pooling methods: the pitch branch uses 1×1×2 max pooling, and the azimuth branch uses 1×2×1 max pooling, thereby enhancing the spatial representation capabilities in the pitch and azimuth directions respectively. Through this dual-branch spatial modeling approach, the network can reduce redundant spatial dimensions while preserving orientation-sensitive structures, improving the representation efficiency of dynamic changing patterns in subsequent temporal modeling stages.

[0045] In the temporal modeling stage, the dual-branch outputs are concatenated and recombined in the channel dimension to form a temporal representation. Then, 1D causal convolution and two residual 1D causal convolution modules are used to model the dynamic changes across continuous time frames. The convolution kernel length is 5, and the dilation coefficients are set to 1, 2, and 4, respectively. The purpose of using causal convolution is to ensure that the prediction at the current moment depends only on the information of the current and historical frames, without using information from future frames, thus meeting the real-time constraints in online dynamic DOA tracking tasks. Through this temporal header, the network can further integrate the spatial spectrum change trend across continuous time frames, suppressing frame-by-frame estimation jitter and jumps caused by noise, reverberation, multipath interference, and spurious peak interference. Finally, the temporal features are mapped to a three-dimensional direction vector output, and the single-source DOA regression results across continuous time frames are further obtained. The network models the Spectral-SRP spatial spectrum of the input continuous time frames and outputs the three-dimensional direction vector representation results corresponding to each time frame. x , y , z This direction vector is further mapped to DOA angle parameters through spherical coordinate transformation, where Defined as atan2( y , x ), Defined as arccos( z / r ), r Given the vector magnitude, it enables DOA tracking output across continuous time frames in dynamic scenes.

[0046] Through the above structure, ResCross3D can simultaneously utilize multi-band spatial spectrum differences, peak position priors, and temporal continuity in historical frames to obtain more stable, continuous, and temporally consistent single-source DOA tracking results in complex indoor dynamic scenes.

[0047] S3: Training and Inference Implementation.

[0048] In this implementation, due to the difficulty in obtaining dynamic single-source data with precise manual annotations collected by a microphone array, this method uses simulation to construct training samples. The sound source signals are selected from speech segments in the LibriSpeech corpus, with a sampling rate of 16kHz. To reduce the interference of background components in silent speech segments on the model learning process, the WebRTC speech activity detection method is used to remove speech signals in silent intervals. Room dimensions are randomly sampled within the range of 3m×3m×2.5m to 10m×8m×6m. The microphone array positions are randomly generated inside the room, requiring that its minimum distance from the wall be no less than 10% of the corresponding room size, and that it be located in the lower half of the room in the vertical direction. Simultaneously, the signal-to-noise ratio (SNR) and reverberation time (T) are... 60Random sampling was performed within the ranges of 5–30 dB and 0.2–1.3 s to ensure that the training samples had good coverage of different room sizes, noise levels, and reverberation intensities.

[0049] To construct a continuous motion sound source trajectory suitable for dynamic tracking tasks, the starting and ending points of the trajectory were randomly selected within the feasible area of ​​the room. A small sinusoidal perturbation was superimposed on the straight path to generate a continuously changing dynamic sound source trajectory. Sound propagation simulation was implemented using the image source method in the gpuRIR toolkit, with the array being a 12-channel microphone array from the NAO robot's head. After completing the propagation simulation, the overlap-add method was used to synthesize multi-channel observation signals, and isotropic Gaussian white noise was superimposed to achieve the preset SNR. Subsequently, the multi-channel signals underwent short-time framing processing, with a window function length of 4096 sampling points and a frame shift of 1024 sampling points. Corresponding Spectral-SRP input features were constructed for each analysis frame. For time frames determined to be silent by VAD, the corresponding inputs were set to zero to reduce the risk of mistracking by the model to non-target directional interference.

[0050] During the training phase, the true DOA at each time step is converted from spherical coordinates to a unit direction vector in a three-dimensional Cartesian coordinate system, and the three components of this unit direction vector are used as network supervision labels. The deviation between the network output and the true label is constrained by the mean squared error (MSE). Supervised learning is used to optimize the parameters of ResCross3D during training, with Adam as the optimizer. The entire training process consists of 80 epochs and employs a phased training strategy: initially, a higher SNR condition is used for initial parameter learning; as training progresses to later stages, the SNR sampling range is expanded, the batch size is increased, and the learning rate is adjusted to improve training stability and final convergence performance. In this implementation, the input spatial spectrum resolution is 32×64, and the input channels consist of one full-frequency spatial spectrum, four sub-band spatial spectra, and two peak coordinate auxiliary channels.

[0051] During the inference phase, the Spectral-SRP multi-channel spatial spectrum input and peak coordinate auxiliary input from consecutive time frames are fed into the trained ResCross3D network. The network outputs a sequence of three-dimensional direction vectors corresponding to the time frames, which is further converted into azimuth and pitch angle sequences, thereby achieving continuous tracking of a single sound source DOA in complex indoor dynamic scenes. Figure 7 The process of dynamic sound source simulation and Spectral-SRP feature construction is presented, which can be used to illustrate the overall implementation process of training sample generation, input construction and subsequent network processing in this embodiment.

[0052] This invention provides a dynamic single-source DOA tracking method, SpecResCross3D, which offers significant advantages compared to existing technologies. SpecResCross3D constructs a Spectral-SRP multi-channel input consisting of a full-frequency spatial spectrum and multiple sub-band spatial spectra. This explicitly preserves the spatial structure differences across different frequency bands, avoiding the weakening of effective information caused by premature aggregation of a single full-frequency spatial spectrum in the frequency dimension. This enhances the characterization and direction discrimination capabilities of the input spatial spectrum in complex indoor environments. Especially under noise, reverberation, and multipath conditions, the main peak structure and pseudo-peak distribution in different frequency bands often exhibit significant differences. SpecResCross3D can more fully utilize these differences, providing a more stable input foundation for dynamic tracking. Based on this, SpecResCross3D employs the ResCross3D tracking network, using front-end dual-branch coding, dual-branch spatial feature extraction, and causal temporal modeling to jointly model multi-channel spatial spectrum information and dynamic changes across continuous time frames. This network can not only utilize the local structure and peak prior information in the spatial spectrum, but also combine the temporal continuity in historical frames to suppress pseudo-peaks, main peak diffusion and peak drift caused by noise and reverberation in complex indoor environments, thereby reducing jitter and jumps in frame-by-frame DOA output and improving the stability, trajectory continuity and temporal consistency of dynamic DOA tracking.

[0053] Example In addition to evaluating the proposed SpecResCross3D method, this experiment also introduces several representative comparative models to systematically compare different methods from the perspectives of input representation, temporal modeling capability, and network complexity. Since the models differ in input form, parameter size, computational complexity, inference latency, and causality, the main comparative models and their key configurations involved in this invention are summarized in Table 1 for ease of subsequent experimental result analysis. Computational complexity is measured by the number of floating-point operations, the number of parameters is represented by the total number of model parameters, and inference latency reflects the time cost of the model during the actual inference stage. Table 1 provides a clearer picture of the fundamental differences in modeling capability and implementation cost among the different methods, and also provides the necessary structural background for subsequent performance comparisons.

[0054] Table 1. Comparative models and their main configurations used in the experiments of this invention. To verify the dynamic DOA tracking capability of the proposed SpecResCross3D under complex acoustic conditions on simulation datasets, this invention trained the model at various spatial spectral resolutions and covered different reverberation times. Performance was evaluated under test conditions of SNR. Several representative samples were selected. Reverberation time (RUT) serves as a reverberation intensity indicator and indirectly reflects the difficulty of temporal expansion caused by multipath reflections. Combined with SNR, the results are analyzed in a comprehensive manner, reflecting the impact of complex indoor acoustic conditions on tracking performance. Based on this, subsequent experiments will focus on the performance under varying noise and reverberation conditions, analyzing the performance advantages of the proposed SpecResCross3D method compared to the traditional SRP-PHAT, to examine its direction estimation and dynamic tracking capabilities in complex acoustic environments. It should be noted that this invention primarily verifies the overall effectiveness of the final method; the specific contributions of the Spectral-SRP input design and the reconstructed tracking network ResCross3D will be further analyzed through subsequent ablation experiments. According to the subsequent ablation experiment results, unless otherwise specified, all experiments related to SpecResCross3D in this invention use the default input configuration of 1 full-band and 4 sub-bands. Since traditional SRP-PHAT is essentially a frame-by-frame localization method and does not have explicit temporal tracking capabilities, while SpecResCross3D achieves continuous orientation estimation through spatial spectral sequence modeling, RMSAE statistics do not include silent frames in this comparison with traditional methods to ensure consistency in evaluation criteria between different methods.

[0055] like Figure 8 and Figure 9 As shown, at a fixed spatial spectral resolution of 32×64, SpecResCross3D outperforms the traditional SRP-PHAT method under different reverberation times and SNR conditions. While the tracking error increases with reverberation enhancement or SNR reduction, the RMSAE increase of SpecResCross3D is significantly smaller and remains at a low level overall, indicating that the proposed method has better stability and robustness under complex acoustic conditions. These results demonstrate that by introducing multi-subband Spectral-SRP input and combining it with subsequent spatiotemporal joint modeling, the impact of noise, reverberation, and multipath propagation on dynamic DOA tracking can be more effectively reduced, thus maintaining good tracking accuracy even under adverse acoustic environments.

[0056] Based on the verification of the overall error trend, the performance of the method is further analyzed from the results of typical trajectory estimation. For example... Figure 10As shown, under simulated dynamic trajectory conditions, the azimuth and pitch angle estimation results of SpecResCross3D are generally closer to the actual DOA trajectory, and the output is smoother and more continuous. Compared with the SRP-PHAT peak estimation results, SpecResCross3D can still better suppress local fluctuations and short-term jumps under high reverberation and low SNR conditions, reducing the instability in direction estimation caused by spurious peaks, main peak diffusion, and peak drift. This demonstrates that the proposed method not only reduces the overall error but also maintains better trajectory coherence and temporal consistency across continuous time frames, which is particularly important for dynamic single-source DOA tracking.

[0057] Furthermore, to illustrate the comprehensive advantages of the proposed method compared to other baseline methods, it was compared with methods such as Cross3D, GCCs, Max, Spectrograms, and SELDnet under the same simulation conditions. Figure 11 As shown, at a fixed spatial spectral resolution of 32×64, SpecResCross3D maintains the lowest or lowest RMSAE under different reverberation times and SNR conditions, and its overall performance is superior to the comparative methods mentioned above. This result demonstrates that the advantages of the proposed method are not only reflected in its improvement over the traditional SRP-PHAT, but also in its better overall performance compared to other deep learning tracking methods. Figures 8 to 11 The results demonstrate that SpecResCross3D, through the combination of multi-subband spatial spectrum input and ResCross3D network structure, can achieve more stable, continuous, and accurate single-source DOA tracking in complex indoor dynamic scenes.

[0058] To verify the generalization ability of SpecResCross3D in real indoor environments, after training on simulation data, further testing was conducted on the LOCATA challenge dataset. The experiment used the LOCATA development dataset, focusing on evaluating Task 3 and Task 5. Task 3 corresponds to a moving speaker scenario under static array conditions, while Task 5 corresponds to a moving speaker scenario under moving array conditions. Since the training phase used simulation data generated under static array conditions, Task 5 places higher demands on the model's generalization ability and temporal stability. Regarding the array configuration used in this implementation, the LOCATA dataset corresponds to the robothead microphone array. This array is a pseudo-spherical microphone array integrated into the head of the NAO humanoid robot prototype, containing 12 microphone channels. Using Task 3 and Task 5 corresponding to this array for testing allows for a good assessment of the applicability of the proposed method to the problem of localizing and tracking real dynamic sound sources under robot auditory acquisition conditions.

[0059] Table 2. RMSAE results of ablation experiments with Spectral-SRP power map input configuration. Table 2 presents the ablation experiment results for different Spectral-SRP input configurations on the LOCATA dataset. Overall, the model performance is improved by introducing multi-subband spatial spectrum input, but this improvement does not increase continuously with the number of subbands. Instead, a suitable balance is struck between frequency band division granularity, input stability, and information redundancy. When the number of subbands is 4 and the full-band channel is retained, Task 3, Task 5, and the overall average result all achieve the best performance, indicating that this configuration can better take into account global directional information and spatial structure differences in different frequency bands, thus providing a more stable and effective input representation for subsequent dynamic tracking. Further comparison shows that after removing the full-band channel, the errors of Task 5 and the overall average increase significantly, indicating that the full-band spatial spectrum is not a redundant input but an important reference for stable localization. This global directional reference is even more significant in scenarios where real recording conditions are more complex and both the sound source and the array may move. On the other hand, as the number of sub-bands continued to increase, the results did not improve further; instead, they decreased to some extent. This indicates that excessively fine band division can introduce redundant perturbations, and the newly added input information failed to be converted into effective gain. Instead, it may affect the model's ability to discriminate the main spatial structures. The results in Table 2 show that the input configuration of 1 full-band and 4 sub-bands in this implementation achieves better overall results.

[0060] Table 3. Comparison of RMSAE results in ablation experiments with different input and network combinations on the LOCATA dataset. Section 3 presents ablation experiments on the LOCATA dataset with different input and network combinations. The results show that SpecResCross3D outperforms SpecCross3D, SRPResCross3D, and Cross3D in Task 3, Task 5, and the overall average RMSAE, indicating that the proposed method has the best overall performance under real-world recording conditions and can simultaneously balance direction estimation accuracy, trajectory continuity, and temporal consistency. Further comparison shows that, with the network structure unchanged, using Spectral-SRP input yields better results than the original SRP-PHAT input, demonstrating that the multi-subband spatial spectrum design at the input enhances the expressive power of direction-related features, enabling the model to explicitly utilize the spatial structure differences across different frequency bands, thereby mitigating the information loss caused by premature aggregation of a single full-frequency spatial spectrum. While maintaining a consistent input format, the introduction of the ResCross3D backbone further enhances tracking performance, demonstrating that the proposed network structure can more effectively utilize spatial and temporal information across consecutive time frames to suppress short-term fluctuations caused by spurious peaks, main peak diffusion, and peak drift. The results in Table 3 show that both Spectral-SRP and ResCross3D contribute to performance improvement, with the combination yielding the best results. This indicates that the advantage of the proposed method stems from the combined effect of input representation enhancement and network structure design, rather than from localized improvements achieved by a single module.

[0061] Table 4. Comparison of RMSAE results of different methods on the LOCATA dataset under different tasks and recording conditions. R1, R2, and R3 represent three recordings, Recording1, Recording2, and Recording3, respectively; Ave represents the average value; the statistical results in the table include silent frames.

[0062] Table 4 presents the overall comparison results of different methods on the LOCATA dataset. The results show that SpecResCross3D achieves the best performance in Task 3, Task 5, and overall average RMSAE. Specifically, the average RMSAE for Task 3 is 6.69°, for Task 5 it is 9.18°, and the overall average RMSAE is 7.93°, all superior to Cross3D's 9.49°, 11.19°, and 10.34°, respectively. It also significantly outperforms the compared methods such as 1DCNN, 2DCNN, and SELDnet. These results demonstrate that the proposed method not only preserves the spatial structure differences across different frequency bands at the input end through Spectral-SRP, but also enhances the spatiotemporal joint modeling capability across continuous time frames at the network end through ResCross3D, thus achieving better dynamic tracking performance under realistic recording conditions. Especially in the more complex Task 5 scenario, SpecResCross3D maintains the lowest error, indicating that the method still possesses strong stability and robustness under the combined conditions of array motion, sound source motion, and real-world environmental interference. Furthermore, as... Figure 12 As shown, under real-world recording conditions in LOCATA, the DOA estimation sequence from SpecResCross3D more closely approximates the actual direction change process, with smoother and more continuous trajectories and fewer local offsets and short-term jumps. This indicates that the proposed method not only outperforms existing methods in terms of overall error metrics but also demonstrates significant advantages in continuous tracking performance and temporal consistency in real-world scenarios.

[0063] Experimental results show that SpecResCross3D exhibits superior dynamic tracking performance on the real-world public dataset LOCATA. On Task 3 and Task 5, SpecResCross3D achieves average RMSAEs of 6.69° and 9.18°, respectively, with an overall average RMSAE of 7.93°, both superior to Cross3D's 9.49°, 11.19°, and 10.34°, and significantly outperformed comparative methods such as 1D CNN, 2D CNN, and SELDnet. Ablation experiments further demonstrate that both the Spectral-SRP input and the ResCross3D backbone contribute to performance improvement, with the combination of both yielding the best results. Therefore, SpecResCross3D not only improves the accuracy of dynamic single-source DOA tracking but also enhances stability and robustness in complex indoor dynamic scenes, achieving smoother, continuous, and reliable directional trajectories. This provides an effective and feasible technical solution for applications such as service robots, voice interaction, and indoor acoustic perception.

[0064] Based on the same concept, the present invention also provides a dynamic single-source DOA tracking system, including an acquisition module, a stacking module, a construction module and a tracking module.

[0065] The acquisition module is used to acquire multi-channel signals of multiple consecutive frames from the dynamic single sound source to be tracked.

[0066] The stacking module is used to acquire the SRP-PHAT spatial spectrum of each frame of multi-channel signal in the full frequency band and multiple sub-frequency bands, respectively, to obtain the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum; the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum of each frame are stacked along the channel dimension to obtain the multi-channel spatial spectrum; the multi-channel spatial spectrum of multiple consecutive frames is stacked along the time dimension to obtain the spatiotemporal input tensor.

[0067] The construction module is used to extract the maximum position coordinates of the full-frequency spatial spectrum of multiple consecutive frames, and construct the maximum coordinate tensor based on the maximum position coordinates; the position coordinates include azimuth coordinates and elevation coordinates.

[0068] The tracking module is used to input the spatiotemporal input tensor and the maximum coordinate tensor into the pre-trained tracking network. The tracking network includes a bi-branch coding module, a bi-branch spatial feature extraction module, and a mapping module. The bi-branch coding module extracts features from the spatiotemporal input tensor and the maximum coordinate tensor respectively and then fuses them to obtain a fusion result. The bi-branch spatial feature extraction module extracts spatial features from the fusion result in the azimuth and elevation directions respectively and then concatenates them to obtain a concatenated result. The mapping module maps the concatenated result to obtain the dynamic single-source DOA tracking result on continuous time frames.

[0069] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0070] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dynamic single-source DOA tracking method, characterized in that, Includes the following steps: Acquire multi-channel signals from multiple consecutive frames of the dynamic single sound source to be tracked; The SRP-PHAT spatial spectrum of each frame of multi-channel signal is obtained in the full frequency band and multiple sub-frequency bands to obtain the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum. The full frequency spatial spectrum and multiple sub-frequency band spatial spectrum of each frame are stacked along the channel dimension to obtain the multi-channel spatial spectrum. The multi-channel spatial spectrum of multiple consecutive frames is stacked along the time dimension to obtain the spatiotemporal input tensor. The maximum position coordinates are extracted from the full-frequency spatial spectrum of multiple consecutive frames, and a maximum coordinate tensor is constructed based on the maximum position coordinates; the position coordinates include azimuth coordinates and elevation coordinates; The spatiotemporal input tensor and the maximum value coordinate tensor are input into a pre-trained tracking network, which includes a bi-branch coding module, a bi-branch spatial feature extraction module, and a mapping module. The bi-branch coding module extracts features from the spatiotemporal input tensor and the maximum value coordinate tensor respectively and then fuses them to obtain a fusion result. The bi-branch spatial feature extraction module extracts spatial features from the fusion result in the azimuth and elevation directions respectively and then concatenates them to obtain a concatenated result. The mapping module maps the concatenated result to obtain the dynamic single-source DOA tracking result on continuous time frames.

2. The dynamic single-source DOA tracking method as described in claim 1, characterized in that, Obtaining the SRP-PHAT spatial spectrum of each frame of multi-channel signal in multiple sub-frequency bands specifically includes the following steps: Within the half-spectrum of the short-time Fourier transform, the frequency axis is uniformly divided into multiple sub-bands; For each sub-band, the SRP-PHAT spatial spectrum of each sub-band is calculated by using the frequency components that fall within the range of that sub-band.

3. The dynamic single-source DOA tracking method as described in claim 1, characterized in that, The dual-branch coding module includes a spatial spectrum coding branch, a coordinate coding branch, and a fusion module. The spatial spectrum coding branch extracts features from the spatiotemporal input tensor, the coordinate coding branch extracts features from the maximum coordinate tensor, and the fusion module fuses the two feature extraction results.

4. The dynamic single-source DOA tracking method as described in claim 3, characterized in that, The spatial spectral coding branch includes a first convolutional module and a second convolutional module, with the number of channels in the first and second convolutional modules set to 5-24-24 respectively, and the kernel size being 5×3×3; the coordinate coding branch includes a third convolutional module and a fourth convolutional module, with the number of channels in the third and fourth convolutional modules set to 2-8-8 respectively, and the kernel size being 5×1×1; the fusion module uses a 1×1×1 3D convolution to fuse the two feature extraction results, resulting in 32 channels after fusion.

5. The dynamic single-source DOA tracking method as described in claim 4, characterized in that, The dual-branch spatial feature extraction module includes a pitch angle branch, an azimuth angle branch, and a stitching module. The pitch angle branch extracts spatial features from the fused result in the pitch angle direction, the azimuth angle branch extracts spatial features from the fused result in the azimuth angle direction, and the stitching module stitches the two spatial feature extraction results together.

6. The dynamic single-source DOA tracking method as described in claim 5, characterized in that, The pitch angle branch includes a fifth convolution module, and the azimuth angle branch includes a sixth convolution module. The fifth and sixth convolution modules have a total of 32 convolution channels and a kernel size of 5×3×3. The fifth convolution module uses 1×1×2 max pooling, and the sixth convolution module uses 1×2×1 max pooling.

7. The dynamic single-source DOA tracking method as described in claim 1, characterized in that, The mapping module includes a seventh convolution module, an eighth convolution module, and a ninth convolution module. The kernel length of the seventh, eighth, and ninth convolution modules is 5, and the dilation coefficients are 1, 2, and 4, respectively.

8. The dynamic single-source DOA tracking method as described in claim 1, characterized in that, The pre-training of the tracking network specifically includes the following steps: The microphone array positions are randomly generated in a room of random size. The starting and ending points of the sound source trajectory are randomly selected in the feasible area of ​​the room and sinusoidal perturbation is superimposed to generate a continuous motion trajectory. Sound propagation simulation is performed by image source method to obtain multi-channel observation signals as training samples. The training samples are divided into frames, and the spatiotemporal input tensor and maximum coordinate tensor of each frame are obtained. The input tensors of multiple consecutive frames are used as network input, and the real DOA at the corresponding time is converted into a direction vector in a three-dimensional Cartesian coordinate system as supervision label. The mean squared error (MSE) is used as the loss function, and the tracking network is trained in a supervised learning manner to optimize the network parameters.

9. The dynamic single-source DOA tracking method as described in claim 1, characterized in that, Multi-channel signals from multiple consecutive frames of the dynamic single sound source to be tracked are acquired using a microphone array.

10. A dynamic single-source DOA tracking system, characterized in that, include: The acquisition module is used to acquire multi-channel signals of multiple consecutive frames from the dynamic single sound source to be tracked. The stacking module is used to acquire the SRP-PHAT spatial spectrum of each frame of multi-channel signal in the full frequency band and multiple sub-frequency bands, respectively, to obtain the full frequency spatial spectrum and multiple sub-frequency band spatial spectrum; The full-frequency spatial spectrum of each frame and multiple sub-band spatial spectra are stacked along the channel dimension to obtain a multi-channel spatial spectrum. The multi-channel spatial spectra of multiple consecutive frames are stacked along the time dimension to obtain a spatiotemporal input tensor. The module is used to extract the coordinates of the maximum value position from the full-frequency spatial spectrum of multiple consecutive frames, and to construct the maximum value coordinate tensor based on the coordinates of the maximum value position. Position coordinates include azimuth coordinates and elevation coordinates; The tracking module is used to input the spatiotemporal input tensor and the maximum value coordinate tensor into a pre-trained tracking network. The tracking network includes a bi-branch encoding module, a bi-branch spatial feature extraction module, and a mapping module. The bi-branch encoding module extracts features from the spatiotemporal input tensor and the maximum value coordinate tensor respectively and then fuses them to obtain a fusion result. The bi-branch spatial feature extraction module extracts spatial features from the fusion result in the azimuth and elevation directions respectively and then concatenates them to obtain a concatenated result. The mapping module maps the concatenated result to obtain the dynamic single-source DOA tracking result on continuous time frames.

Images