Animal identification and localization method and system based on sound events

Abstract

The application discloses an animal recognition and positioning method and system based on sound events, relates to the technical field of bioacoustic monitoring, and comprises collecting multiple original audio data to form a target sound event set. Animal species recognition processing is performed on the set, acoustic features of each audio are extracted and fused, and an animal species determination result is generated. At the same time, a sound source spatial positioning process is performed, and the estimated spatial position of the animal sound source is calculated based on the time difference and intensity difference of sound waves. The species determination result and the estimated position are fused and associated, and an animal activity record containing the animal species and the position is generated. Temporal and spatial continuity analysis is performed on the record, and the record is associated and matched with historical animal activity records, so that the position or species information in the record is corrected or updated. Finally, according to the animal activity record, an animal activity heat map or an animal trajectory map of a monitoring area is generated. The method improves the accuracy and data continuity of animal monitoring in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of bioacoustic monitoring technology, specifically an animal identification and localization method and system based on sound events. Background Technology

[0002] In sound-based wildlife monitoring, existing technologies typically employ separate audio acquisition units and subsequent analysis modules. Species identification relies on audio signals acquired from a single acquisition point, extracting characteristic parameters such as Mel-frequency cepstral coefficients and inputting them into a classification model for identification. Sound source localization is based on multiple deployed acquisition points, estimating the sound source coordinates using the time difference or energy difference of sound reaching different sensors. These steps are usually processed sequentially according to a fixed procedure, outputting independent, timestamped species tags and location data.

[0003] The quality and stability of acoustic features obtained from a single acquisition channel are heavily dependent on the specific acoustic environment at that point. In real-world scenarios with continuous environmental noise, intermittent interference, or multipath propagation of sound waves, single-channel features are prone to distortion or loss of key information, leading to a decline in the recognition model's ability to distinguish similar species, and making it difficult to guarantee the reliability of the overall recognition results under complex conditions. Current technical solutions process each sound event in isolation and instantaneously, lacking correlation between the calculated location information. The system cannot determine whether successive events belong to the continuous activities of the same animal, nor can it utilize historical data to logically verify and smooth out abnormal results that may be caused by signal obstruction or calculation errors.

[0004] This leads to two main problems: the accuracy of species identification is limited by the constraints of single-point signals, resulting in instability in complex acoustic environments; and the output animal activity information is a collection of discrete points, failing to form a coherent and reliable spatiotemporal trajectory, thus limiting its value for in-depth behavioral analysis and long-term ecological research. A method is needed to comprehensively utilize multi-channel audio information to improve feature robustness and to collaboratively optimize and correct the continuity of identification and localization results by establishing spatiotemporal correlations between events. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes an animal identification and localization method based on sound events, including:

[0007] Collect multiple channels of raw audio data to form a set of target sound events;

[0008] Animal species identification processing is performed on the target sound event set to extract and fuse the acoustic features of each audio data in the target sound event set to generate the animal species determination result of the sound event;

[0009] The target sound event set is subjected to sound source spatial localization processing. Based on the sound wave arrival time difference and sound wave intensity difference information between each audio data, the estimated spatial location of the animal sound source in the sound event is calculated.

[0010] The animal species determination result is fused and correlated with the estimated spatial location to generate an animal activity record that includes the animal species and their corresponding location;

[0011] Spatiotemporal continuity analysis is performed on the animal activity records, and they are correlated and matched with historical animal activity records to correct or update the location or species information in the animal activity records.

[0012] Based on the animal activity records, generate a heat map of animal activity or an animal trajectory map of the monitoring area.

[0013] Furthermore, multiple channels of raw audio data are collected to form a target sound event set, including:

[0014] The system receives synchronously acquired raw audio data streams from multiple sound acquisition nodes deployed within the monitoring area, obtains multiple raw audio data streams, performs environmental noise suppression and time-frequency domain enhancement processing on the multiple raw audio data streams, and generates multiple enhanced audio data streams.

[0015] The multi-channel enhanced audio data is subjected to sound event segmentation and feature extraction processing to obtain multiple sound event segments and their corresponding acoustic feature vector sets;

[0016] Cross-channel correlation analysis and event matching are performed on the multiple sound event segments to obtain a set of multiple sound event segments associated with the same animal vocalization event, which constitutes the target sound event set;

[0017] The step of performing environmental noise suppression and time-frequency domain enhancement processing on the multiple original audio data to generate multiple enhanced audio data includes:

[0018] For each raw audio data stream, an adaptive filter is used to remove ambient background noise, resulting in preliminary noise-reduced audio data.

[0019] Perform a short-time Fourier transform on the preliminary noise-reduced frequency data to obtain a time-spectrum graph;

[0020] On the aforementioned time-spectrum graph, a mask-based sound source separation algorithm is applied to further separate the foreground animal sound components from the residual background noise components;

[0021] The energy distribution of the foreground animal sound component in the temporal spectrogram is enhanced to obtain an enhanced temporal spectrogram;

[0022] Perform an inverse short-time Fourier transform on the enhanced time-spectrum to reconstruct the single-channel enhanced audio data.

[0023] Further, the step of performing sound event segmentation and feature extraction processing on the multi-channel enhanced audio data to obtain multiple sound event segments and their corresponding acoustic feature vector sets includes:

[0024] Endpoint detection is performed on each enhanced audio data stream to identify the time interval containing effective sound energy, thus obtaining candidate sound segments;

[0025] For each candidate sound segment, its Mel frequency cepstral coefficients, linear prediction coefficients, and spectral centroid features are calculated to obtain the initial acoustic features;

[0026] The initial acoustic features are input into a pre-trained sound event classification model to determine whether the candidate sound segments belong to animal vocalization events.

[0027] If it belongs to an animal vocalization event, the candidate sound segment is marked as a sound event segment, and the high-dimensional deep acoustic features of the sound event segment are extracted and concatenated with the initial acoustic features to obtain the acoustic feature vector of the sound event segment.

[0028] The sound event segments and their acoustic feature vectors corresponding to all animal vocalization events in each enhanced audio data stream are collected.

[0029] Furthermore, the cross-channel correlation analysis and event matching of the multiple sound event segments yields a set of multiple sound event segments associated with the same animal vocalization event, constituting a target sound event set, including:

[0030] Calculate the acoustic feature similarity and temporal overlap between any two audio event segments from different audio acquisition nodes;

[0031] Based on the acoustic feature similarity and the temporal overlap, calculate the comprehensive association probability that two sound event segments belong to the same animal vocalization event;

[0032] Set a comprehensive correlation probability threshold, and merge two audio event segments whose comprehensive correlation probability exceeds the comprehensive correlation probability threshold to form an initial audio event cluster;

[0033] Each of the initial sound event clusters is expanded by incorporating new sound event segments whose probability of being associated with any segment within the initial sound event cluster exceeds a threshold into the initial sound event cluster.

[0034] Each initial sound event cluster contains a set of sound event segments from different sound acquisition nodes, which are then defined as a target sound event set.

[0035] Further, the step of performing animal species identification processing on the target sound event set, extracting and fusing the acoustic features of each audio data stream in the target sound event set, and generating an animal species determination result for the sounding event includes:

[0036] From the target sound event set, extract acoustic feature vectors of sound event segments from different sound acquisition nodes;

[0037] The acoustic feature vectors from different nodes are input into the feature fusion network, and the weights of the features of each node are calculated through an attention mechanism.

[0038] The acoustic feature vectors from different nodes are weighted and fused according to the weights to obtain the fused global acoustic feature vector.

[0039] The global acoustic feature vector is input into a pre-trained animal species classification model;

[0040] The animal species classification model outputs the probability distribution of animal vocalization events belonging to each known animal species;

[0041] The animal species with the highest probability is used as the animal species determination result.

[0042] Further, the step of performing sound source spatial localization processing on the target sound event set, and calculating the estimated spatial location of the animal sound source in the sounding event based on the sound wave arrival time difference and sound wave intensity difference information between each audio data, includes:

[0043] From the target sound event set, select at least three sound event segments with the best signal quality, and extract the geographical coordinates of their corresponding sound acquisition nodes respectively;

[0044] Calculate the relative time difference of arrival between the at least three sound event segments;

[0045] Based on the speed of sound in the air and the relative time difference of arrival, a system of hyperbolic equations about the location of the sound source is established.

[0046] Simultaneously, the sound wave intensity difference between the at least three sound event segments is calculated, and a set of distance attenuation equations about the sound source location is established in conjunction with the sound propagation attenuation model.

[0047] Solve the hyperbolic equations and the distance attenuation equations simultaneously to obtain the estimated spatial coordinates of the animal sound source.

[0048] Further, the step of fusing and associating the animal species determination result with the estimated spatial location to generate an animal activity record containing the animal species and its corresponding location includes:

[0049] The animal species determination result is bound to the estimated spatial location coordinates to generate an initial activity record, which includes the animal species identifier, location coordinates, and timestamp of the vocalization event.

[0050] The initial activity records are compared with the animal characteristic knowledge base, and the typical activity range and movement speed information of the animal are obtained according to the animal species identifier;

[0051] Using the typical activity range and movement speed information, the location coordinates are validated for reasonableness. If the location coordinates exceed the threshold of the typical activity range, the location coordinates are smoothed and corrected based on historical data.

[0052] The verified or corrected animal species identifiers, location coordinates, and timestamps are integrated to generate the final animal activity record.

[0053] Furthermore, the step of performing spatiotemporal continuity analysis on the animal activity records, associating and matching them with historical animal activity records, and correcting or updating the location or species information in the animal activity records specifically includes:

[0054] Retrieve historical animal activity records from the database that are temporally adjacent to the current animal activity record and geographically close to its location coordinates.

[0055] Calculate the overall matching degree between the current animal activity record and each of the aforementioned historical animal activity records in terms of species consistency, continuity of location movement, and reasonableness of time interval;

[0056] If there is a historical animal activity record with a comprehensive matching degree exceeding a preset threshold, then the current animal activity record and the historical animal activity record are determined to be continuous activities of the same animal.

[0057] Based on the trajectory trends in the historical animal activity records, the position coordinates in the current animal activity records are smoothed using Kalman filtering to correct positioning errors;

[0058] If the animal species determination result of the current animal activity record is inconsistent with the animal species of the associated historical animal activity record, the species re-determination process will be initiated to make a comprehensive determination based on historical data.

[0059] Further, generating an animal activity heatmap or animal trajectory map of the monitoring area based on the animal activity records includes:

[0060] The monitoring area is divided into a uniform grid;

[0061] The number of animal activity records falling into each grid within a preset time period is counted to obtain the grid activity frequency;

[0062] Based on the activity frequency of the grid, a color mapping method is used to render and generate a heat map of animal activity in the monitoring area;

[0063] Alternatively, select the activity records of a specific animal individual, connect their location coordinates in chronological order, and obtain the movement trajectory of the specific animal individual;

[0064] The movement trajectories of multiple individual animals are overlaid on the same base map to generate an animal trajectory map.

[0065] Furthermore, the present invention also includes an animal identification and localization system based on sound events, the system including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the animal identification and localization method based on sound events as described above.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] Animal species identification is achieved by extracting and fusing acoustic features from multiple audio sources. The multi-channel feature fusion process integrates observational information from different spatial locations on the same sound event, forming a more redundant and complementary high-dimensional feature representation. This fusion mechanism mitigates the impact of potential feature distortion or loss of partial frequency band information in a single channel, enhancing the robustness and completeness of the overall feature representation. The classification model makes decisions based on these fused features, providing more comprehensive judgment criteria and reducing the probability of misclassification due to adverse local acoustic environments or transient signal interference. This improves the overall accuracy of species identification and the system's stability under non-ideal recording conditions.

[0068] This method performs spatiotemporal continuity analysis on the generated animal activity records and correlates them with historical records. This process examines each newly identified event within a continuous spatiotemporal context. The system constructs logical constraints for event association based on the animal's possible movement speed, activity range boundaries, and historical behavioral patterns. By comparing the temporal and spatial continuity of the current event with historical trajectories, it can identify anomalous data points that suddenly jump in location or abruptly change in species. Based on this, the system can perform positional smoothing corrections or reassess species information for these anomalous points that do not conform to continuous movement patterns, rather than directly outputting the original instantaneous calculation results.

[0069] Through the aforementioned correlation and correction mechanisms, originally discrete and independent monitoring events are integrated into activity sequences with spatiotemporal logical coherence. This reduces isolated erroneous records caused by transient signal loss, sudden noise, or computational errors, making the output animal activity trajectories more consistent with real animal behavior patterns. The inherent consistency and reliability of the data are enhanced, providing higher-quality and more continuous foundational data for subsequent analysis. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of the animal identification and localization method based on sound events described in this invention.

[0071] Figure 2 A flowchart for cross-channel event matching;

[0072] Figure 3 Comparison chart of animal sound source localization error performance;

[0073] Figure 4 A comparison chart of animal acoustic feature similarity and recognition confidence;

[0074] Figure 5 Heat map of activity frequency in the animal activity monitoring area. Detailed Implementation

[0075] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0076] See Figure 1 Multiple raw audio data streams are collected synchronously from various sound acquisition nodes deployed within the monitoring area to form a target sound event set. Animal species identification processing is performed on the target sound event set. This processing extracts and fuses the acoustic features of each audio data stream in the target sound event set to generate an animal species determination result for the sounding event. Sound source spatial localization processing is performed on the target sound event set. This processing calculates the estimated spatial location of the animal sound source in the sounding event based on the time difference of arrival and intensity difference of the sound waves between each audio data stream. The animal species determination result and the estimated spatial location are fused and correlated to generate an animal activity record containing the animal species and their corresponding locations. Spatiotemporal continuity analysis is performed on the animal activity record, and it is correlated and matched with historical animal activity records to correct or update the location or species information in the animal activity record. Based on the final animal activity record, an animal activity heat map or animal trajectory map of the monitoring area is generated, thereby visually displaying the spatiotemporal distribution pattern of animal activity.

[0077] In one embodiment of the present invention, the specific process of acquiring multiple channels of raw audio data to form a target sound event set involves multiple processing stages. Multiple channels of raw audio data are obtained by receiving synchronously acquired raw audio data streams from multiple sound acquisition nodes deployed within the monitoring area. Environmental noise suppression and time-frequency domain enhancement processing are performed on the multiple channels of raw audio data to generate multiple channels of enhanced audio data. Specifically, for each channel of raw audio data, an adaptive filter is used to filter out environmental background noise to obtain preliminary noise-reduced audio data. A short-time Fourier transform is performed on the preliminary noise-reduced audio data to obtain a time-frequency spectrum. A mask-based sound source separation algorithm is applied to the time-frequency spectrum to further separate the foreground animal sound components and residual background noise components. The energy distribution of the foreground animal sound components in the time-frequency spectrum is enhanced to obtain an enhanced time-frequency spectrum. An inverse short-time Fourier transform is performed on the enhanced time-frequency spectrum to reconstruct a single channel of enhanced audio data. In some embodiments, the coefficients of the adaptive filter can be periodically updated according to the statistical characteristics of the background noise in the monitoring area. It is understood that the window function length and overlap rate of the short-time Fourier transform need to be configured according to the typical duration of the target animal sound.

[0078] Sound event segmentation and feature extraction are performed on multiple enhanced audio data streams to obtain multiple sound event segments and their corresponding acoustic feature vector sets. For each enhanced audio stream, endpoint detection is performed to identify time intervals containing effective sound energy, resulting in candidate sound segments. For each candidate sound segment, its Mel-frequency cepstral coefficients, linear prediction coefficients, and spectral centroid features are calculated to obtain initial acoustic features. These initial acoustic features are input into a pre-trained sound event classification model to determine whether the candidate sound segment belongs to an animal vocalization event. If it does, the candidate sound segment is marked as a sound event segment, and its high-dimensional deep acoustic features are extracted and concatenated with the initial acoustic features to obtain the acoustic feature vector of the sound event segment. All sound event segments and their acoustic feature vectors corresponding to animal vocalization events in each enhanced audio stream are then aggregated. Optionally, endpoint detection can employ a dual-threshold method based on short-time energy and zero-crossing rate. In some embodiments, high-dimensional deep acoustic features can be extracted from the time-spectral graph of the sound event segment using a pre-trained convolutional neural network.

[0079] Cross-channel correlation analysis and event matching are performed on multiple sound event segments to obtain a set of sound event segments associated with the same animal vocalization event, forming a target sound event set. The acoustic feature similarity and temporal overlap between any two sound event segments from different sound acquisition nodes are calculated. Based on the acoustic feature similarity and temporal overlap, the comprehensive association probability that two sound event segments belong to the same animal vocalization event is calculated. A comprehensive association probability threshold is set, and sound event segments with a comprehensive association probability exceeding the threshold are merged to form an initial sound event cluster. The calculation of the comprehensive association probability can be expressed as:

[0080] ;

[0081] in: Indicates the overall correlation probability. This represents the normalized acoustic feature similarity. Indicates the time overlap under normalization. and It is a preset weighting coefficient and satisfies It is understandable that acoustic feature similarity can be calculated based on the cosine distance between the eigenvectors of the Mel-frequency cepstral coefficients of sound event segments. Each initial sound event cluster is expanded by incorporating new sound event segments whose overall association probability with any segment within the initial cluster exceeds a threshold. The set of sound event segments from different sound acquisition nodes contained within each initial sound event cluster is then defined as a target sound event set. Optionally, temporal overlap is calculated based on the ratio of the intersection to the union of the start and end times of two sound event segments.

[0082] In one embodiment of the present invention, see [reference] Figure 2The process involves performing sound event segmentation and feature extraction on multiple enhanced audio data streams to obtain multiple sound event segments and their corresponding acoustic feature vector sets. This process includes endpoint detection for each enhanced audio data stream to identify time intervals containing effective sound energy, resulting in candidate sound segments. For each candidate sound segment, the Mel-frequency cepstral coefficients, linear prediction coefficients, and spectral centroid features are calculated to obtain initial acoustic features. These initial acoustic features are then input into a pre-trained sound event classification model to determine whether the candidate sound segments belong to animal vocalization events. If they do, the candidate sound segments are marked as sound event segments, and the high-dimensional deep acoustic features of the sound event segments are extracted and concatenated with the initial acoustic features to obtain the acoustic feature vector of the sound event segment. All sound event segments and their acoustic feature vectors corresponding to animal vocalization events in each enhanced audio data stream are then aggregated. In some embodiments, endpoint detection can employ a dual-threshold method based on short-time energy and zero-crossing rate, where the energy threshold and zero-crossing rate threshold are dynamically set according to the global statistical characteristics of the enhanced audio data. The pre-trained sound event classification model can be a support vector machine-based classifier, whose training data includes various animal vocalization samples and environmental noise samples. It is understandable that the extraction of high-dimensional deep acoustic features relies on a pre-trained deep neural network model that takes the time-spectrum graph of a sound event segment as input and outputs a feature vector of fixed dimensions.

[0083] Cross-channel correlation analysis and event matching of multiple sound event segments yield a set of sound event segments associated with the same animal vocalization event, forming a target sound event set. This process calculates the acoustic feature similarity and temporal overlap between any two sound event segments from different sound acquisition nodes. Based on acoustic feature similarity and temporal overlap, the comprehensive association probability of two sound event segments belonging to the same animal vocalization event is calculated. A comprehensive association probability threshold is set, and sound event segments with a comprehensive association probability exceeding the threshold are merged to form an initial sound event cluster. The calculation of temporal overlap can be expressed as:

[0084] ;

[0085] in: Represents a segment of a sound event With sound event fragments The degree of time overlap between them Represents a segment of a sound event The time interval Represents a segment of a sound event Time interval, symbol Represents the intersection of intervals, with the symbol... Represents the union of intervals. This represents the interval length. In some embodiments, the acoustic feature similarity is calculated based on the cosine similarity between the eigenvectors of the Mel-frequency cepstral coefficients of the sound event segments. Each initial sound event cluster is expanded by incorporating new sound event segments whose overall association probability with any segment within the initial cluster exceeds a threshold. The set of sound event segments from different sound acquisition nodes contained within each initial sound event cluster is determined as a target sound event set. Optionally, the overall association probability can be calculated by weighted summation of acoustic feature similarity and temporal overlap, where the weighting coefficients are adjusted according to the density of the sound acquisition nodes. It is understood that the merging process of the initial sound event clusters employs an iterative algorithm until no new sound event segments meet the merging criteria.

[0086] In one embodiment of the present invention, animal species identification processing is performed on a target sound event set to extract and fuse acoustic features of audio data from each audio stream in the target sound event set to generate an animal species determination result for the sounding event. This process extracts acoustic feature vectors of sound event segments from different sound acquisition nodes from the target sound event set, inputs the acoustic feature vectors from different nodes into a feature fusion network, calculates the weights of each node's features through an attention mechanism, and performs weighted fusion of the acoustic feature vectors from different nodes according to the weights to obtain a fused global acoustic feature vector. The global acoustic feature vector is input into a pre-trained animal species classification model, which outputs the probability distribution of the animal sounding event belonging to each known animal species. The animal species with the highest probability is taken as the animal species determination result. In some embodiments, the feature fusion network can be a multilayer perceptron structure, whose input is the concatenation of acoustic feature vectors from all nodes, and whose output is the attention weight score corresponding to each node feature. Optionally, the pre-trained animal species classification model can adopt a deep residual network architecture and be trained on a large-scale animal sound dataset. It is understandable that the calculation of attention weights allows the model to pay more attention to the acoustic features extracted from nodes with higher signal quality or better signal-to-noise ratio.

[0087] The process involves spatial localization of the sound source in a target sound event set. Based on the time-of-arrival (TOA) and intensity differences of sound waves between different audio data streams, the estimated spatial location of the animal sound source in the sound event is calculated. This process selects at least three sound event segments with the best signal quality from the target sound event set, extracts the geographical coordinates of their corresponding sound acquisition nodes, calculates the relative TOA between these three sound event segments, and establishes a hyperbolic equation system regarding the sound source location based on the speed of sound in air and the relative TOA. Simultaneously, it calculates the intensity differences between the at least three sound event segments and, combined with a sound propagation attenuation model, establishes a distance attenuation equation system regarding the sound source location. Signal quality can be comprehensively evaluated through the signal-to-noise ratio (SNR) and spectral flatness of the sound event segments. The distance attenuation equation system is constructed based on the propagation attenuation law of sound waves in a homogeneous medium; an example of its form is as follows:

[0088] ;

[0089] in: and They represent the first The and the first The intensity of the sound waves received by each sound acquisition node and Let represent the Euclidean distances from the sound source to these two nodes, respectively. This is the environmental attenuation coefficient of sound propagation. Solving the hyperbolic equations and the distance attenuation equations simultaneously yields the estimated spatial coordinates of the animal sound source. In some embodiments, the equations can be solved numerically using a nonlinear least squares optimization algorithm. Optionally, the speed of sound in the air can be corrected based on the real-time temperature and humidity of the monitoring area. It is understood that simultaneously utilizing the time difference of arrival and the difference in sound intensity helps overcome the limitations of a single method in complex environments and improves the robustness of the localization results.

[0090] See Figure 3 This is a performance comparison chart of animal sound source localization error, showing the error performance of three sound source localization methods at different distances. Figure 3 This study visually demonstrates the performance advantages of the "TDOA + intensity difference" fusion method in complex environments, overcoming the limitations of single methods and maintaining high positioning accuracy even at long distances. The error fluctuations in pure intensity difference positioning mainly stem from environmental attenuation interference during sound wave propagation, while the fusion method effectively suppresses this influence by incorporating time difference information. These results can provide a basis for selecting positioning algorithms in practical animal monitoring systems, prioritizing fusion positioning schemes to ensure monitoring reliability. This figure provides a clear performance baseline for subsequent algorithm iterations; all improved positioning algorithms can be compared to this baseline to quantitatively evaluate their optimization effects.

[0091] In one embodiment of the present invention, the animal species determination result and the estimated spatial location are fused and correlated to generate an animal activity record containing the animal species and its corresponding location. This process binds the animal species determination result and the estimated spatial location coordinates to generate an initial activity record. The initial activity record includes the animal species identifier, location coordinates, and a timestamp of the vocalization event. The initial activity record is compared with an animal characteristic knowledge base. Based on the animal species identifier, the typical activity range and movement speed information of the animal are obtained. The location coordinates are validated for reasonableness using the typical activity range and movement speed information. If the location coordinates exceed the typical activity range threshold, the location coordinates are smoothed and corrected based on historical data. The validated or corrected animal species identifier, location coordinates, and timestamp are integrated to generate the final animal activity record. In some embodiments, the animal characteristic knowledge base can be a structured database that stores ecological parameters of different animal species. Optionally, the smoothing correction based on historical data can use a moving average algorithm to correct the current abnormal location based on the historical location coordinate sequence of the same animal individual over a recent period. It can be understood that the typical activity range threshold is usually defined as a circular geographical area with the center point of the animal's known habitat as the center and the radius of the typical activity range as the distance.

[0092] The process involves performing spatiotemporal continuity analysis on animal activity records and correlating and matching them with historical animal activity records to correct or update the location or species information in the records. This process retrieves historical animal activity records from the database that are temporally adjacent to the current record and geographically close to its location coordinates. The overall matching degree between the current record and each historical record is calculated in terms of species consistency, continuity of location movement, and reasonableness of time intervals. The calculation of the overall matching degree can be expressed as:

[0093] ;

[0094] in: Indicates the current animal activity record and the first The overall matching degree of historical animal activity records, This represents the category consistency index (1 when the category identifiers are the same, 0 otherwise). This indicates the consistency index of movement speed (whether the speed calculated based on the position difference and time difference is within the typical movement speed range). Indicates the time interval between two records. This represents the spatial Euclidean distance between two records. , , and These are preset weights and scaling factors. If there are historical animal activity records with a comprehensive matching degree exceeding a preset threshold, the current animal activity record and the historical animal activity record are determined to be continuous activities of the same animal. Kalman filtering is then applied to the position coordinates in the current animal activity record to smooth the position and correct positioning errors based on the trajectory trends in the historical animal activity records. In some embodiments, the state variables of the Kalman filter may include position coordinates and movement speed. If the animal species determination result of the current animal activity record is inconsistent with the animal species of the associated historical animal activity record, a species re-determination process is initiated to combine historical data for comprehensive determination. Optionally, the species re-determination process can combine the species determination probability distribution of the current record with the species identifier in the historical records for weighted voting. It can be understood that the specific thresholds for "adjacent" and "near" during database retrieval can be set according to the typical activity capabilities of the target animal group. See Table 1, which shows part of the content of the animal characteristic knowledge base.

[0095] Table 1: Animal Characteristics Knowledge Base Table

[0096] ;

[0097] See Figure 4 This is a comparison chart of animal acoustic feature similarity and recognition confidence, showing the similarity performance of five species across three types of acoustic features, and their corresponding recognition confidence. It clarifies the differences in the contribution of different acoustic features to species recognition, validating the necessity of multi-feature fusion. The correlation between high similarity and high confidence of temporal features provides a clear direction for subsequent algorithm optimization. In resource-constrained scenarios, temporal features can be prioritized as the core recognition basis to reduce computational complexity. For species with low spectral similarity, such as species C, the spectral feature extraction algorithm can be optimized to improve recognition performance. Feature performance baselines for different species are established, providing a quantitative comparison standard for subsequent algorithm iterations. The correlation between feature similarity and recognition confidence is revealed, providing data support for the design of multi-feature fusion weights.

[0098] In one embodiment of the present invention, an animal activity heatmap or animal trajectory map of a monitoring area is generated based on animal activity records. This process divides the monitoring area into uniform grids, counts the number of animal activity records falling into each grid within a preset time period to obtain the grid activity frequency, and renders the animal activity heatmap of the monitoring area using a color mapping method based on the grid activity frequency. The grid division is based on the latitude and longitude boundaries of the monitoring area and a set grid resolution, and each grid is defined by the coordinates of its lower left and upper right corners. It can be understood that the preset time period can be the past hour, day, or week, depending on the analysis requirements. The grid activity frequency statistics are a process of spatially querying the location coordinates of all animal activity records, determining the grid to which it belongs, and accumulating the count. The color mapping scheme maps the grid activity frequency values ​​to a predefined color gradient, for example, a gradient from blue to red corresponding to a frequency from low to high. In some embodiments, the calculation of the color mapping can be expressed as:

[0099] ;

[0100] in: Indicates the first Line number The final rendered color of the column grid. This indicates the original activity frequency of the grid. and These represent the maximum and minimum activity frequencies across all grids, respectively. This represents a mapping function from a normalized numerical value to a specific color value. Optionally, the color mapping function can use linear interpolation or a non-linear transformation.

[0101] The process of generating animal trajectory maps involves selecting the activity records of specific animal individuals and connecting their location coordinates in chronological order to obtain the movement trajectory of that specific animal individual. Multiple animal trajectories are then overlaid on the same base map to generate the animal trajectory map. The selection of specific animal individuals is based on unique identifiers in their activity records, which are assigned during the spatiotemporal continuity analysis phase to determine a sequence of records belonging to the same continuous activity. It can be understood that an animal individual's movement trajectory is formed by connecting a series of location coordinate points ordered by timestamps; the connecting lines can be straight lines or smooth curves. When overlaid on the same base map, different animal individuals can be distinguished by using different colors, line styles, or marker symbols for different trajectories. In some embodiments, trajectory map generation may include filtering of trajectory points, such as removing isolated points that significantly deviate from the main path. Alternatively, selecting the activity records of specific animal individuals may require setting a time window, for example, only displaying the trajectories of animal individuals with continuous activity records within the past 24 hours. Optionally, the base map may include the geographical boundaries of the monitoring area, terrain information, and location markers for sound acquisition nodes.

[0102] See Figure 5 This is a heatmap of animal activity frequency within a grid-based monitoring area, serving as the core visualization result of an animal identification and localization system based on sound events. Six white pentagrams mark the locations of sound acquisition nodes, arranged in a ring around the core activity area. This deployment method improves the accuracy of sound source localization through multi-node collaboration. The activity frequency decreases gradually from the center to the edge, consistent with the aggregation patterns of wild animals and validating the spatial coverage effectiveness of the monitoring system. The system first acquires animal vocalization events through multiple audio acquisition nodes, then completes species identification and spatial localization. The monitoring area is divided into latitude and longitude grids, and the animal activity frequency within each grid is statistically analyzed over 24 hours. This data can be used for animal habitat analysis in ecological conservation, providing data support for protection strategies in the core activity area. It can also help evaluate the rationality of the monitoring node layout, providing a basis for optimizing the monitoring network.

[0103] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for animal identification and localization based on sound events, characterized in that, The method includes: Collect multiple streams of raw audio data to form a target sound event set, including: The system receives synchronously acquired raw audio data streams from multiple sound acquisition nodes deployed within the monitoring area, obtains multiple raw audio data streams, performs environmental noise suppression and time-frequency domain enhancement processing on the multiple raw audio data streams, and generates multiple enhanced audio data streams. The multi-channel enhanced audio data is subjected to sound event segmentation and feature extraction processing to obtain multiple sound event segments and their corresponding acoustic feature vector sets; Cross-channel correlation analysis and event matching are performed on the multiple sound event segments to obtain a set of multiple sound event segments associated with the same animal vocalization event, constituting a target sound event set, including: The acoustic feature similarity and temporal overlap between any two sound event segments from different sound acquisition nodes are calculated. The acoustic feature similarity is calculated based on the cosine distance between the eigenvectors of the Mel frequency cepstral coefficients of the sound event segments, and the temporal overlap is calculated based on the ratio of the intersection to the union of the start and end times of the two sound event segments. Based on the acoustic feature similarity and the temporal overlap, calculate the comprehensive association probability that two sound event segments belong to the same animal vocalization event; The calculation of the comprehensive association probability is expressed as follows: ； in: Indicates the overall correlation probability. This represents the normalized acoustic feature similarity. Indicates the time overlap under normalization. and It is a preset weighting coefficient and satisfies ; Set a comprehensive correlation probability threshold, and merge two audio event segments whose comprehensive correlation probability exceeds the comprehensive correlation probability threshold to form an initial audio event cluster; Each of the initial sound event clusters is expanded by incorporating new sound event segments whose probability of being associated with any segment within the initial sound event cluster exceeds a threshold into the initial sound event cluster. Each initial sound event cluster contains a set of sound event segments from different sound acquisition nodes, which is then defined as a target sound event set. Animal species identification processing is performed on the target sound event set to extract and fuse the acoustic features of each audio data in the target sound event set to generate the animal species determination result of the sound event; The target sound event set is subjected to sound source spatial localization processing. Based on the sound wave arrival time difference and sound wave intensity difference information between each audio data, the estimated spatial location of the animal sound source in the sound event is calculated. The animal species determination result is fused and correlated with the estimated spatial location to generate an animal activity record that includes the animal species and their corresponding location; Spatiotemporal continuity analysis is performed on the animal activity records, and they are correlated and matched with historical animal activity records to correct or update the location or species information in the animal activity records. Based on the animal activity records, generate a heat map of animal activity or an animal trajectory map of the monitoring area.

2. The animal identification and localization method based on sound events according to claim 1, characterized in that, The step of performing environmental noise suppression and time-frequency domain enhancement processing on the multiple original audio data to generate multiple enhanced audio data includes: For each raw audio data stream, an adaptive filter is used to remove ambient background noise, resulting in preliminary noise-reduced audio data. Perform a short-time Fourier transform on the preliminary noise-reduced frequency data to obtain a time-spectrum graph; On the aforementioned time-spectrum graph, a mask-based sound source separation algorithm is applied to further separate the foreground animal sound components from the residual background noise components; The energy distribution of the foreground animal sound component in the temporal spectrogram is enhanced to obtain an enhanced temporal spectrogram; Perform an inverse short-time Fourier transform on the enhanced time-spectrum to reconstruct the single-channel enhanced audio data.

3. The animal identification and localization method based on sound events according to claim 2, characterized in that, The process of performing sound event segmentation and feature extraction on the multi-channel enhanced audio data yields multiple sound event segments and their corresponding acoustic feature vector sets, including: Endpoint detection is performed on each enhanced audio data stream to identify the time interval containing effective sound energy, thus obtaining candidate sound segments; For each candidate sound segment, its Mel frequency cepstral coefficients, linear prediction coefficients, and spectral centroid features are calculated to obtain the initial acoustic features; The initial acoustic features are input into a pre-trained sound event classification model to determine whether the candidate sound segments belong to animal vocalization events. If it belongs to an animal vocalization event, the candidate sound segment is marked as a sound event segment, and the high-dimensional deep acoustic features of the sound event segment are extracted and concatenated with the initial acoustic features to obtain the acoustic feature vector of the sound event segment. The sound event segments and their acoustic feature vectors corresponding to all animal vocalization events in each enhanced audio data stream are collected.

4. The animal identification and localization method based on sound events according to claim 1, characterized in that, The process of performing animal species identification processing on the target sound event set, extracting and fusing the acoustic features of each audio data stream in the target sound event set, and generating an animal species determination result for the sounding event includes: From the target sound event set, extract acoustic feature vectors of sound event segments from different sound acquisition nodes; The acoustic feature vectors from different nodes are input into the feature fusion network, and the weights of the features of each node are calculated through an attention mechanism. The acoustic feature vectors from different nodes are weighted and fused according to the weights to obtain the fused global acoustic feature vector. The global acoustic feature vector is input into a pre-trained animal species classification model; The animal species classification model outputs the probability distribution of animal vocalization events belonging to each known animal species; The animal species with the highest probability is used as the animal species determination result.

5. The animal identification and localization method based on sound events according to claim 1, characterized in that, The step of performing sound source spatial localization processing on the target sound event set, based on the sound wave arrival time difference and sound wave intensity difference information between each audio data stream, calculates the estimated spatial location of the animal sound source in the sounding event, including: From the target sound event set, select at least three sound event segments with the best signal quality, and extract the geographical coordinates of their corresponding sound acquisition nodes respectively; Calculate the relative time difference of arrival between the at least three sound event segments; Based on the speed of sound in the air and the relative time difference of arrival, a system of hyperbolic equations about the location of the sound source is established. Simultaneously, the sound wave intensity difference between the at least three sound event segments is calculated, and a set of distance attenuation equations about the sound source location is established in conjunction with the sound propagation attenuation model. Solve the hyperbolic equations and the distance attenuation equations simultaneously to obtain the estimated spatial coordinates of the animal sound source.

6. The animal identification and localization method based on sound events according to claim 1, characterized in that, The step of fusing and associating the animal species determination result with the estimated spatial location to generate an animal activity record containing the animal species and their corresponding locations includes: The animal species determination result is bound to the estimated spatial location coordinates to generate an initial activity record, which includes the animal species identifier, location coordinates, and timestamp of the vocalization event. The initial activity records are compared with the animal characteristic knowledge base, and the typical activity range and movement speed information of the animal are obtained according to the animal species identifier; Using the typical activity range and movement speed information, the location coordinates are validated for reasonableness. If the location coordinates exceed the threshold of the typical activity range, the location coordinates are smoothed and corrected based on historical data. The verified or corrected animal species identifiers, location coordinates, and timestamps are integrated to generate the final animal activity record.

7. The animal identification and localization method based on sound events according to claim 6, characterized in that, The step of performing spatiotemporal continuity analysis on the animal activity records, associating and matching them with historical animal activity records, and correcting or updating the location or species information in the animal activity records specifically includes: Retrieve historical animal activity records from the database that are temporally adjacent to the current animal activity record and geographically close to its location coordinates. Calculate the overall matching degree between the current animal activity record and each of the aforementioned historical animal activity records in terms of species consistency, continuity of location movement, and reasonableness of time interval; If there is a historical animal activity record with a comprehensive matching degree exceeding a preset threshold, then the current animal activity record and the historical animal activity record are determined to be continuous activities of the same animal. Based on the trajectory trends in the historical animal activity records, the position coordinates in the current animal activity records are smoothed using Kalman filtering to correct positioning errors; If the animal species determination result of the current animal activity record is inconsistent with the animal species of the associated historical animal activity record, the species re-determination process will be initiated to make a comprehensive determination based on historical data.

8. The animal identification and localization method based on sound events according to claim 7, characterized in that, The step of generating an animal activity heatmap or animal trajectory map of the monitoring area based on the animal activity records includes: The monitoring area is divided into a uniform grid; The number of animal activity records falling into each grid within a preset time period is counted to obtain the grid activity frequency; Based on the activity frequency of the grid, a color mapping method is used to render and generate a heat map of animal activity in the monitoring area; Alternatively, select the activity records of a specific animal individual, connect their location coordinates in chronological order, and obtain the movement trajectory of the specific animal individual; The movement trajectories of multiple individual animals are overlaid on the same base map to generate an animal trajectory map.

9. An animal identification and localization system based on sound events, characterized in that, The system includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the animal identification and localization method based on sound events as described in any one of claims 1 to 8.

Images