A bird information statistics method based on the fusion of bird detection radar data and line sampling method

Abstract

The present invention discloses a method for counting bird information by fusing bird detecting radar data with a sample line method, which solves the problem that when the bird detecting radar data and the manual survey data are fused in the existing method, the bird population density cannot be counted, and the radar data and the manual survey data are not fully fused. The method comprises taking the bird population density as a priori probability, fusing an enhanced data set with the bird population density, and a target recognition model extracting features of the data fusion set based on micro-Doppler features, and outputting the statistical results of the bird information in the bird flock; in the present invention, the target recognition model realizes an accurate evaluation of the statistical results of the bird information by fusing the bird detecting radar data with the sample line method, and taking the bird population density as a priori probability to fuse the enhanced data set with the bird population density can effectively reduce the uncertainty in the statistical process, thereby improving the accuracy of the bird information statistics, and can also use the enhanced data set to improve the stability and reliability of the target recognition model.

Description

Technical Field

[0001] The invention belongs to the technical field of target detection, and in particular relates to a bird information statistics method integrating bird detection radar data with a sample line method. Background Art

[0002] Airport bird strike refers to the abbreviation of collision accidents between aircraft and birds, which is a traditional safety hazard that has long existed in the civil aviation industry. With the continuous development of civil aviation and the continuous improvement of the ecological environment, the pressure on airport bird strike prevention is increasing. Bird activities are highly sudden, especially during the migration season, when a large number of birds will gather and cross the airport area in a short period of time. This suddenness makes it difficult to make accurate predictions and preparations in advance for bird strike prevention. At the same time, because the flight paths and behavior patterns of birds are complex and changeable, they are easily affected by multiple factors such as weather, food sources, and breeding habits, which increases the difficulty of bird strike prevention. Bird information statistics provide detailed data support for airport managers. By statistically analyzing the patterns of bird activities, it is possible to predict the activities of birds in and around the airport, so as to take bird repellent measures in advance and reduce the probability of bird strikes.

[0003] Chinese patent CN116430379A discloses a bird information statistics method that integrates bird detection radar data and the line sampling method, the method comprising determining a bird survey area; dividing the bird survey area into grids and sub-areas; determining bird activity hotspots in each sub-area; delineating a bird survey line and conducting manual surveys to obtain manual survey data; integrating the bird detection radar data with the manual survey data to obtain the total number of birds and the number of different bird species in the bird survey area. However, when the existing method integrates the bird detection radar data with the manual survey data, the bird population density cannot be counted, and the integration of the radar data and the manual survey data is insufficient, which increases the complexity and difficulty of data analysis. To address the above problems, we propose a bird information statistics method that integrates the bird detection radar data with the line sampling method. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and provide a bird information statistics method that integrates bird detection radar data and the sample line method, thereby solving the problems that when the prior method integrates bird detection radar data with manual survey data, the bird population density cannot be counted, and the fusion of radar data and manual survey data is insufficient, which increases the complexity and difficulty of data analysis.

[0005] The present invention is implemented as follows: a bird information statistics method integrating bird detection radar data and sample line method, the bird information statistics method integrating bird detection radar data and sample line method comprises:

[0006] 16 sampling lines were set up within a radius of 10±0.1km in the statistical area. The total number of single bird populations and the number of species in the statistical area were recorded using infrared cameras combined with the variable distance sampling line method. The bird population density was calculated based on the total number of single bird populations and the number of species in the statistical area. The total number of single bird populations, the number of species, and the bird population density in the statistical area were uploaded to the information database.

[0007] Obtain bird detection radar data, enhance the bird detection radar data to obtain an enhanced data set, obtain bird population density information from the information database, use the bird population density as a priori probability, fuse the enhanced data set with the bird population density to obtain a data fusion set, and upload the data fusion set to the information database;

[0008] Pre-build a target recognition model that combines micro-Doppler features and the YOLO model, iteratively train the target recognition model, and output a converged target recognition model;

[0009] The data fusion set is loaded and taken as input. The target recognition model extracts features of the data fusion set based on micro-Doppler features and outputs statistical results of bird information in the bird flock. The bird information statistics include the number, type and location information of flying birds.

[0010] Preferably, the method of recording the total number of bird populations and the number of species in a statistical area based on an infrared camera combined with a variable distance transect method includes:

[0011] Based on the vegetation type and landform in the statistical area, the sampling lines are randomly arranged in layers. The width of a single sampling line is 50-200m, and the length of a single sampling line is 0.5-5.5km.

[0012] Camera points were set up at equal intervals of 200 m in a single sampling line, and infrared cameras were set up at the camera points. The sampling line survey time was 7:30-11:30 am and 3:30-5:30 pm. The infrared camera was set to take 3 photos and 30 seconds of video each time it was triggered;

[0013] Integrate at least one set of infrared photos and videos, and upload the infrared photos, videos, camera coordinates, and acquisition time to the information database in real time;

[0014] The non-metric multidimensional scaling method is used to count the total number of single bird populations and the number of species in the area, and the total number of single bird populations and the number of species in the area are output.

[0015] Preferably, the method for calculating the bird population density based on the total number of bird single populations and the number of species in the statistical area specifically includes:

[0016] The total number of bird populations and the number of species in the loading area are calculated, and the initial population density is calculated based on the Fourier intercept method;

[0017] The initial population density is calculated using the following formula:

[0018] (1)

[0019] (2)

[0020] in, represents the initial population density, Represents the length of a single sample line, is the number of splines, is the probability density function of finding a bird individual when the vertical distance is 0, is the width of a single sample line, represents the total number of a single population, It represents the vertical distance between the individual bird and the transect;

[0021] Obtain the initial population density, correct the initial population density based on the density correction function, and output the bird population density;

[0022] The bird population density is calculated by the following formula:

[0023] (3)

[0024] in, is the bird population density, is the correction coefficient of the density correction function, is the number of splines.

[0025] Preferably, the method for enhancing the processing of bird detection radar data comprises:

[0026] Obtain bird detection radar data and identify the three-dimensional coordinates of radar points based on the bird detection radar data , based on the coordinate transformation matrix, the three-dimensional coordinates of the radar point are converted into two-dimensional coordinates ;

[0027] Get the 2D coordinates of the radar point , transform the radar point into a square representation with 80 pixels, and convert the radar point signal into a three-channel color value represented by the square based on the peak conversion criterion;

[0028] Load the square representation containing the three-channel color values, randomly add noise points and brightness to the square representation of the radar point signal, and obtain the enhanced data set.

[0029] Preferably, the coordinate transformation matrix is ​​expressed as:

[0030] (4)

[0031] (5)

[0032] in, is the tilt error of the radar point in three-dimensional coordinates, They are the pitch vector, attitude angle vector and flapping frequency vector of the target object pointed out by the radar respectively;

[0033] Among them, when the radar point signal is converted into three-channel color values ​​represented by a square based on the peak conversion criterion, the three-channel color values ​​are ;

[0034] The three-channel color value is calculated by the following formula:

[0035] (6)

[0036] (7)

[0037] (8)

[0038] in, Indicates the speed value of the radar point target. is the peak value of the intermediate frequency signal of the target object, is the radar point distance value.

[0039] Preferably, the method of fusing the enhanced data set with the bird population density comprises:

[0040] Obtain bird population density, identify camera coordinates, acquisition time, camera focal length, and optical axis vector associated with bird population density;

[0041] The infrared camera parameters are annotated based on OcamCalib, including the optical axis vector, pixel level and height-to-width ratio in the infrared camera parameters;

[0042] Extract the annotated optical axis vector, pixel level and height-to-width ratio, transform the camera coordinate system using Euclidean transformation, rotate and scale the camera coordinate system, and combine the camera coordinate system external parameter rotation matrix , scaling matrix Optical axis vector, pixel level and height aspect ratio, and radar point 2D coordinates Fusion, realizing the fusion of radar point two-dimensional coordinates and camera coordinates, and outputting the fusion result of radar point two-dimensional coordinates and camera coordinates;

[0043] The fusion results of the two-dimensional coordinates of the radar point and the camera coordinates, the bird population density, and the three-channel color values ​​represented by the square are obtained. The time series of the radar and the camera are matched using the timestamp alignment method to obtain the data fusion set.

[0044] Preferably, the method for iteratively training the target recognition model comprises:

[0045] Traverse the data fusion set in the information database, disrupt the order of the data fusion set, and divide the data fusion set into a training set and a validation set, with the ratio of the training set to the validation set being 3:1;

[0046] Initialize the hyperparameters of the initial baseline model of the target recognition model, define the activation function, loss function, learning rate, and training rounds of the initial baseline model;

[0047] Obtain a training set, perform time-frequency analysis on the training set to obtain a target time-frequency diagram, detect the target time-frequency diagram based on micro-Doppler features, and obtain a flapping frequency curve of the training set;

[0048] Perform Fourier transform processing on the flapping frequency curve, and perform threshold judgment, retain the weights corresponding to the high-frequency components in the Fourier transform processing results, prune the weights corresponding to the low-frequency components in the Fourier transform processing results, iteratively train the initial baseline model after weight pruning until convergence, and output the initial baseline model;

[0049] Get the test set, use the test set as input, execute the initial baseline model, output the test results, and determine whether the test results meet the preset accuracy threshold. If they meet the preset accuracy threshold, output the converged target recognition model.

[0050] Preferably, the target recognition model uses the YOLO7 model as the initial baseline model, the initial baseline model consists of a backbone, a neck and a head, the backbone of the initial baseline model is frozen, and the backbone is replaced by the AlexNet network architecture, the AlexNet network architecture includes six convolutional layers, three groups of fully connected layers, and a pooling layer, the six groups of convolutional layers are the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer, and the fully connected layers are the first fully connected layer, the second fully connected layer, and the third fully connected layer. The first convolutional layer, the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer are connected to the pooling layer respectively, the first fully connected layer is connected to the second fully connected layer, and the second fully connected layer is connected to the third fully connected layer. The neck is improved to obtain an improved neck. A bird flock scene recognition network architecture is introduced into the neck. The improved neck is used to focus on the local features of the bird flock, and to perform statistical evaluation on the number and species of flying birds in the scene. The K-Means++ anchor frame is introduced into the head, and the K-Means++ anchor frame is used to mark the position and species of flying birds and output the statistical results of bird information.

[0051] Preferably, the method for extracting features of the target recognition model from the data fusion set based on micro-Doppler features includes:

[0052] Load the data fusion set, extract the bird micro-Doppler characteristic signal in the data fusion set based on the encoder, perform Fourier transform on the bird micro-Doppler characteristic signal, and obtain the Fourier transform result;

[0053] Analyze the micro-Doppler motion distribution of birds based on the Choi-Williams algorithm and output the motion feature distribution results;

[0054] Perform threshold judgment on the motion feature distribution results to determine whether the target object exists. If the target object exists, perform convolution fusion, pooling and full connection processing on the motion feature distribution results based on the AlexNet network architecture, focus on the local features of the bird flock, and output the statistical evaluation results of the number and species of birds in the scene;

[0055] Obtain the statistical evaluation results of the number and species of birds in the scene, annotate the location and species of the birds based on the K-Means++ anchor frame, and output the statistical results of bird information;

[0056] When Fourier transform is performed on the bird micro-Doppler characteristic signal, the transformation function is defined as:

[0057] (9)

[0058] in, is the Fourier transform output representation, It is the time-series micro-Doppler characteristic signal of birds. is the short-time Fourier transform frequency, is a window function;

[0059] The motion feature distribution result is expressed as:

[0060] (10)

[0061] in, is the motion feature distribution result, is the time delay parameter, Represents the scaling factor.

[0062] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0063] In the embodiment of the present invention, the target recognition model realizes the accurate evaluation of the statistical results of bird information by fusing bird detection radar data with the sample line method. At the same time, the enhanced processing of bird detection radar data can adaptively enhance the robustness of the target recognition model, ensure the recognition effectiveness of the target recognition model algorithm, and use the bird population density as the prior probability to fuse the enhanced data set with the bird population density. This can effectively reduce the uncertainty in the statistical process, thereby improving the accuracy of bird information statistics, and can also use the enhanced data set to improve the stability and reliability of the target recognition model. This overcomes the problem that when the existing method of bird detection radar data is fused with manual survey data, the bird population density cannot be counted, and the radar data and manual survey data are not fully fused, which increases the complexity and difficulty of data analysis.

[0064] In an embodiment of the present invention, when calculating the bird population density, the initial population density is first calculated based on the Fourier intercept method, and then the initial population density is corrected based on the density correction function, and finally an accurate bird population density is obtained, which can eliminate or reduce errors and deviations in the data, thereby improving the accuracy of bird population density estimation, and at the same time, it can also promote the integration of bird population density and bird detection radar data, so as to have a more comprehensive understanding of the distribution and dynamic changes of bird populations.

[0065] In an embodiment of the present invention, when enhancing the bird detecting radar data, converting the three-dimensional coordinates of the radar point into two-dimensional coordinates based on the coordinate transformation matrix helps to better and more deeply integrate the bird detecting radar data with the bird population density, and also enables the bird detecting radar data to be recognized and analyzed by the target recognition model, while ensuring the important features in the bird detecting radar data. At the same time, based on the peak conversion criterion, the radar point signal is converted into a three-channel color value represented by a square, thereby enhancing the adaptability of the target recognition model to radar data of different resolutions and improving the recognition efficiency and accuracy of the target recognition model.

[0066] In an embodiment of the present invention, a target recognition model and a training method thereof are provided. The target recognition model uses the YOLO7 model as the initial baseline model, introduces the AlexNet network architecture and the bird flock scene recognition network architecture, combines the bird population density as the prior probability, and finally uses the K-Means++ anchor frame to annotate the small moving target birds at different distances from the airport, which is suitable for the position positioning and normalized monitoring of birds at the airport. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 The present invention provides a schematic diagram of the implementation process of a bird information statistics method that integrates bird detection radar data with a sample line method.

[0068] Figure 2 The present invention shows a schematic diagram of the implementation process of a method for recording the total number of bird populations and the number of species in a statistical area based on an infrared camera combined with a variable distance sample line method.

[0069] Figure 3 The present invention shows a schematic diagram of the implementation process of a method for calculating bird population density based on the total number of single bird populations and the number of species in a statistical area.

[0070] Figure 4 The present invention shows a schematic diagram of the implementation process of the bird detection radar data enhancement processing method.

[0071] Figure 5 A schematic diagram of the implementation process of the method for fusing the enhanced dataset with the bird population density is shown.

[0072] Figure 6 A schematic diagram of the implementation process of an iterative training method for a target recognition model is shown.

[0073] Figure 7 It is a structural schematic diagram of the target recognition model provided by the present invention.

[0074] Figure 8 The figure shows a schematic diagram of the implementation process of the feature extraction method of the target recognition model based on the micro-Doppler feature pair data fusion set. DETAILED DESCRIPTION

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0076] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0077] When the existing bird-detecting radar data is integrated with the manual survey data, the bird population density cannot be counted, and the fusion of radar data and manual survey data is not sufficient, which increases the complexity and difficulty of data analysis. To address the above problems, we proposed a bird information statistics method that integrates bird-detecting radar data with the sample line method. When the method is implemented, the total number of single bird populations and the number of species in the statistical area are first recorded based on an infrared camera combined with a variable distance sample line method. The bird population density is calculated based on the total number of single bird populations and the number of species in the statistical area. The bird radar data is enhanced to obtain an enhanced data set, and the bird population density information is obtained from the information database. The bird population density is used as a priori probability to fuse the enhanced data set with the bird population density. At the same time, a target recognition model combining micro-Doppler features and the YOLO model is pre-constructed. Finally, the data fusion set is used as input, and the target recognition model extracts features of the data fusion set based on the micro-Doppler features, and outputs the statistical results of the bird information in the bird flock. In the embodiment of the present invention, the target recognition model realizes the accurate evaluation of the statistical results of bird information by fusing bird detection radar data with the sample line method. At the same time, the enhanced processing of bird detection radar data can adaptively enhance the robustness of the target recognition model, ensure the recognition effectiveness of the target recognition model algorithm, and use the bird population density as the prior probability to fuse the enhanced data set with the bird population density. This can effectively reduce the uncertainty in the statistical process, thereby improving the accuracy of bird information statistics, and can also use the enhanced data set to improve the stability and reliability of the target recognition model. This overcomes the problem that when the existing method of bird detection radar data is fused with manual survey data, the bird population density cannot be counted, and the radar data and manual survey data are not fully fused, which increases the complexity and difficulty of data analysis.

[0078] The embodiment of the present invention provides a bird information statistics method integrating bird detection radar data with a sample line method. Figure 1 The schematic diagram of the implementation process of the bird information statistics method integrating bird detection radar data and the sample line method is shown. The bird information statistics method integrating bird detection radar data and the sample line method specifically includes:

[0079] Step S10, 16 sampling lines are arranged within a radius of 10±0.1km in the statistical area, and the total number of single bird populations and the number of species in the statistical area are recorded based on an infrared camera combined with a variable distance sampling line method, and the bird population density is calculated based on the total number of single bird populations and the number of species in the statistical area, and the total number of single bird populations, the number of species, and the bird population density in the statistical area are uploaded to the information database;

[0080] It should be noted that the information database is used to provide Web publishing services, data interface interaction services, and Windows component services. The basic architecture of the information database is the SQL Server 2008R2 database. The information database performs basic operations such as adding, deleting, modifying, and checking data, and converts the results of bird information statistics into JSON (JavaScript object notation) data format for visual presentation.

[0081] Step S20, obtaining bird detection radar data, enhancing the bird detection radar data to obtain an enhanced data set, obtaining bird population density information from the information database, taking the bird population density as a priori probability, fusing the enhanced data set with the bird population density to obtain a data fusion set, and uploading the data fusion set to the information database;

[0082] Step S30, pre-building a target recognition model combining micro-Doppler features and a YOLO model, iteratively training the target recognition model, and outputting a converged target recognition model;

[0083] Step S40, loading the data fusion set, taking the data fusion set as input, the target recognition model extracts features of the data fusion set based on micro-Doppler features, and outputs statistical results of bird information in the bird flock, wherein the bird information statistics include the number, species and location information of flying birds.

[0084] In the embodiment of the present invention, the target recognition model realizes the accurate evaluation of the statistical results of bird information by fusing bird detection radar data with the sample line method. At the same time, the enhanced processing of bird detection radar data can adaptively enhance the robustness of the target recognition model, ensure the recognition effectiveness of the target recognition model algorithm, and use the bird population density as the prior probability to fuse the enhanced data set with the bird population density. This can effectively reduce the uncertainty in the statistical process, thereby improving the accuracy of bird information statistics, and can also use the enhanced data set to improve the stability and reliability of the target recognition model. This overcomes the problem that when the existing method of bird detection radar data is fused with manual survey data, the bird population density cannot be counted, and the radar data and manual survey data are not fully fused, which increases the complexity and difficulty of data analysis.

[0085] The embodiment of the present invention provides a method for recording the total number of bird populations and the number of species in a statistical area based on an infrared camera combined with a variable distance sample line method. Figure 2 The present invention shows a schematic diagram of the implementation process of a method for recording the total number of bird populations and the number of species in a statistical area based on an infrared camera combined with a variable distance sample line method. The method for recording the total number of bird populations and the number of species in a statistical area based on an infrared camera combined with a variable distance sample line method specifically includes:

[0086] Step S101, randomly arranging sampling lines in layers based on the vegetation types and landforms in the statistical area, with the width of a single sampling line being 50-200m and the length of a single sampling line being 0.5-5.5km;

[0087] It should be noted that the statistical area may be an airport or a power transmission line area, the landforms include but are not limited to woodlands, cultivated land, hills, and residential areas, and the vegetation types include but are not limited to forests, grasslands, and wetlands. GIS software is used to divide the study area according to the determined stratification standards and generate a stratification map. In this embodiment, the collected sample lines are divided into the first sample line to the sixteenth sample line.

[0088] Step S102: camera points are arranged at equal intervals of 200 m in a single sampling line, and infrared cameras are set up at the camera points. The sampling line survey time is 7:30-11:30 am and 3:30-5:30 pm. The infrared camera is set to continuously take 3 photos and 30 seconds of video each time it is triggered;

[0089] In this embodiment, in addition to infrared cameras, binoculars and recording equipment can also be set up at the camera points to ensure the observation effect. The installation height of the infrared camera is 20-200cm, and the camera points are distributed in different vegetation types such as shrub forests, evergreen broad-leaved forests, evergreen coniferous and broad-leaved mixed forests, and evergreen deciduous broad-leaved mixed forests.

[0090] Step S103, integrating at least one set of infrared photos and videos, and uploading the infrared photos, videos, camera coordinates, and acquisition time to an information database in real time;

[0091] It should be noted that the infrared camera may be a Ltl Acorn-6210MC infrared camera, and the size of a single set of photos of the infrared camera is 5-10M, and the video specification is 1080p. The infrared photos also record the placement time, altitude, temperature, humidity, longitude, latitude, slope direction, slope, and the types of trees, shrubs and herbs, forest coverage and density of each infrared camera.

[0092] Step S104, using non-metric multidimensional scaling method to count the total number of single bird populations and the number of species in the area, and output the total number of single bird populations and the number of species in the area.

[0093] In this embodiment, when the total number of single bird populations and the number of species in the region are counted using non-metric multidimensional scaling, a suitable similarity coefficient is selected to measure the similarity between different birds according to the characteristics of the data. Commonly used similarity coefficients include Euclidean distance, Manhattan distance, cosine similarity, etc. In this embodiment, the Manhattan distance similarity coefficient is selected to measure the similarity between different birds, and then the distance matrix is ​​calculated. The selected similarity coefficient is used to calculate the distance matrix between all birds. For each identified single population, the number of birds and the number of species contained therein are counted. This can be achieved by simple counting or SPSS statistical analysis, when SPSS26.0 software is used for analysis. If a continuous variable is normally distributed, the mean ± standard deviation is used to represent it.

[0094] It should be noted that the bird species in the area include but are not limited to starlings, thrushes, magpies, brown-backed shrikes, black-collared starlings, kestrels, redfinches, blackbirds, black drongos, yellow-clawed falcons, silk-tailed starlings, sparrows, gray starlings, silver-faced long-tailed tits, and gray-backed thrushes.

[0095] The embodiment of the present invention provides a method for calculating the bird population density based on the total number of bird single populations and the number of species in a statistical area. Figure 3 The present invention shows a schematic diagram of the implementation process of a method for calculating the bird population density based on the total number of single bird populations and the number of species in a statistical area. The method for calculating the bird population density based on the total number of single bird populations and the number of species in a statistical area specifically includes:

[0096] Step S201, loading the total number of bird populations and the number of species in the area, and calculating the initial population density based on the Fourier intercept method;

[0097] The initial population density is calculated using the following formula:

[0098] (1)

[0099] (2)

[0100] in, represents the initial population density, Represents the length of a single sample line, is the number of splines. In this embodiment, the number of splines is the number of sample lines, which is 16. is the probability density function of finding a bird individual when the vertical distance is 0, is the width of a single sample line, represents the total number of a single population, It represents the vertical distance between the individual bird and the transect;

[0101] Step S202, obtaining the initial population density, correcting the initial population density based on a density correction function, and outputting the bird population density;

[0102] The bird population density is calculated by the following formula:

[0103] (3)

[0104] in, is the bird population density, is the correction coefficient of the density correction function. In this embodiment, the correction coefficient is 1.05-1.1. is the number of splines.

[0105] In an embodiment of the present invention, when calculating the bird population density, the initial population density is first calculated based on the Fourier intercept method, and then the initial population density is corrected based on the density correction function, and finally an accurate bird population density is obtained, which can eliminate or reduce errors and deviations in the data, thereby improving the accuracy of bird population density estimation, and at the same time, it can also promote the integration of bird population density and bird detection radar data, so as to have a more comprehensive understanding of the distribution and dynamic changes of bird populations.

[0106] The embodiment of the present invention provides a method for enhancing processing of bird detection radar data. Figure 4 The following is a schematic diagram of the implementation process of the bird detection radar data enhancement processing method, wherein the bird detection radar data enhancement processing method specifically includes:

[0107] Step S301, obtaining bird detection radar data, and identifying the three-dimensional coordinates of the radar point based on the bird detection radar data , based on the coordinate transformation matrix, the three-dimensional coordinates of the radar point are converted into two-dimensional coordinates ;

[0108] It should be noted that bird-detecting radar data includes but is not limited to target location data, flight speed data, altitude trajectory data, micro-motion feature data, and risk assessment data. The radar system can provide information on the altitude of birds at different time points, thereby depicting their flight trajectories. These data are helpful in analyzing the flight habits and migration routes of birds. The bird-detecting radar can be a K12T bird-detecting radar, a K12 low-altitude surveillance radar, or an L8000 detection and early warning radar.

[0109] In this embodiment, the coordinate transformation matrix is ​​expressed as:

[0110] (4)

[0111] (5)

[0112] in, is the tilt error of the radar point in three-dimensional coordinates, They are the pitch vector, attitude angle vector and flapping frequency vector of the target object pointed out by the radar respectively.

[0113] Step S302, obtaining the two-dimensional coordinates of the radar point , transform the radar point into a square representation with 80 pixels, and convert the radar point signal into a three-channel color value represented by the square based on the peak conversion criterion;

[0114] Among them, when the radar point signal is converted into three-channel color values ​​represented by a square based on the peak conversion criterion, the three-channel color values ​​are ;

[0115] The three-channel color value is calculated by the following formula:

[0116] (6)

[0117] (7)

[0118] (8)

[0119] in, Indicates the speed value of the radar point target. is the peak value of the intermediate frequency signal of the target object, is the radar point distance value.

[0120] Step S303 , loading a square representation containing three-channel color values, and randomly adding noise points and brightness to the square representation of the radar point signal to obtain an enhanced data set.

[0121] In an embodiment of the present invention, when enhancing the bird detecting radar data, converting the three-dimensional coordinates of the radar point into two-dimensional coordinates based on the coordinate transformation matrix helps to better and more deeply integrate the bird detecting radar data with the bird population density, and also enables the bird detecting radar data to be recognized and analyzed by the target recognition model, while ensuring the important features in the bird detecting radar data. At the same time, based on the peak conversion criterion, the radar point signal is converted into a three-channel color value represented by a square, thereby enhancing the adaptability of the target recognition model to radar data of different resolutions and improving the recognition efficiency and accuracy of the target recognition model.

[0122] The embodiment of the present invention provides a method for fusing an enhanced data set with bird population density. Figure 5 A schematic diagram of the implementation process of a method for fusing an enhanced data set with bird population density is shown. The method for fusing an enhanced data set with bird population density specifically includes:

[0123] Step S401, obtaining bird population density, identifying camera coordinates, acquisition time, camera focal length, and optical axis vector associated with the bird population density;

[0124] It should be noted that when setting up an infrared camera, the specific location coordinates of the camera need to be recorded. This can usually be done with a GPS device. Ensuring the accuracy of the coordinates is critical for subsequent data analysis. The focal length of the camera refers to the distance from the optical center of the lens to the imaging plane, which directly affects the viewing angle and magnification of the image. When photographing birds, choosing the right focal length can help better observe and record the details of the birds. The optical axis vector includes the camera optical axis offset. Combining the optical axis vector data of multiple cameras can achieve three-dimensional reconstruction of bird populations, thereby more accurately estimating the density and distribution of bird populations.

[0125] Step S402, annotating the infrared camera parameters based on OcamCalib, annotating the optical axis vector, pixel level and height-to-width ratio in the infrared camera parameters;

[0126] It should be noted that OcamCalib, or OCam model, is a community abundance model based on observational data. It is mainly used in ecology and wildlife research to estimate the population density of birds or other organisms in a specific area.

[0127] Step S403: extract the annotated optical axis vector, pixel level and height-to-width ratio, transform the camera coordinate system using Euclidean transformation, rotate and scale the camera coordinate system, and combine the camera coordinate system external parameter rotation matrix , scaling matrix Optical axis vector, pixel level and height aspect ratio, and radar point 2D coordinates Fusion, realizing the fusion of radar point two-dimensional coordinates and camera coordinates, and outputting the fusion result of radar point two-dimensional coordinates and camera coordinates;

[0128] Step S404, obtaining the fusion result of the two-dimensional coordinates of the radar point and the camera coordinates, the bird population density, and the three-channel color values ​​represented by the square, and using the timestamp alignment method to perform timing matching between the radar and the camera to obtain a data fusion set.

[0129] In this embodiment, the timestamp refers to an identifier that records the time when an event occurs, usually in seconds or milliseconds. In radar and camera data acquisition, each data packet is accompanied by a timestamp to mark the specific moment when the data was generated. Aligning timestamps is to convert the timestamps of the radar and camera to the same time base. This usually involves converting both timestamps to UTC (Coordinated Universal Time) or a common reference time point. If the sampling frequencies of the radar and camera are different, the timestamps need to be interpolated or adjusted to ensure that they have corresponding data points at the same time interval.

[0130] The embodiment of the present invention provides a method for iterative training of a target recognition model. Figure 6 A schematic diagram of the implementation process of the iterative training method for the target recognition model is shown. The method for iterative training of the target recognition model specifically includes:

[0131] Step S501, traverse the data fusion set in the information database, disrupt the order of the data fusion set, and divide the data fusion set into a training set and a verification set, and the ratio of the training set to the verification set is 3:1;

[0132] Step S502, initializing the hyperparameters of the initial baseline model of the target recognition model, defining the activation function, loss function, learning rate and training rounds of the initial baseline model;

[0133] In this embodiment, the activation function of the initial baseline model is the Softmax function, the loss function is the Hinge loss function, the learning rate is 0.001, and the training rounds are 200-240 times.

[0134] Step S503, obtaining a training set, performing time-frequency analysis on the training set to obtain a target time-frequency diagram, detecting the target time-frequency diagram based on micro-Doppler features, and obtaining a flapping frequency curve of the training set;

[0135] Step S504, performing Fourier transform processing on the flapping frequency curve, and performing threshold judgment, retaining the weights corresponding to the high-frequency components in the Fourier transform processing result, pruning the weights corresponding to the low-frequency components in the Fourier transform processing result, iteratively training the initial baseline model after weight pruning until convergence, and outputting the initial baseline model;

[0136] Step S505, obtaining a test set, taking the test set as input, executing the initial baseline model, and outputting the test results;

[0137] Step S506, determining whether the test result meets a preset accuracy threshold;

[0138] Step S507: if the preset accuracy threshold is met, output a converged target recognition model.

[0139] If it does not meet the preset accuracy threshold, return to step S503 to continue model iterative training.

[0140] In this embodiment, the preset accuracy threshold is set to 0.9-0.92. After iterative training of the target recognition model, the accuracy of the model is improved to 92.8%, and the recall rate is improved to 93.2%, which is a significant improvement compared to the traditional convolutional network model and YOLO model, thereby improving the accuracy of bird information statistics and recognition analysis. At the same time, the method of pruning weights corresponding to low-frequency components in the Fourier transform processing results can significantly reduce the running load of the model and improve the statistical efficiency of the model without affecting the accuracy and precision during model training.

[0141] In an embodiment of the present invention, a target recognition model and a training method thereof are provided. The target recognition model uses the YOLO7 model as the initial baseline model, introduces the AlexNet network architecture and the bird flock scene recognition network architecture, combines the bird population density as the prior probability, and finally uses the K-Means++ anchor frame to annotate the small moving target birds at different distances from the airport, which is suitable for the position positioning and normalized monitoring of birds at the airport.

[0142] like Figure 7 As shown, a structural diagram of the target recognition model is shown. The target recognition model uses the YOLO7 model as the initial baseline model. The initial baseline model consists of a backbone, a neck, and a head. The backbone of the initial baseline model is frozen, and the backbone is replaced by the AlexNet network architecture. The AlexNet network architecture includes six convolutional layers, three groups of fully connected layers, and a pooling layer. The six groups of convolutional layers are the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer. The fully connected layers are the first fully connected layer, the second fully connected layer, and the The first convolutional layer, the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer are connected to the pooling layer respectively, the first fully connected layer is connected to the second fully connected layer, and the second fully connected layer is connected to the third fully connected layer. The neck is improved to obtain an improved neck. A bird flock scene recognition network architecture is introduced into the neck. The improved neck is used to focus on the local features of the bird flock, and statistically evaluate the number and species of flying birds in the scene. The K-Means++ anchor frame is introduced in the head, and the K-Means++ anchor frame is used to mark the position and type of flying birds and output the statistical results of bird information.

[0143] It should be noted that the K-Means++ anchor frame uses the K-Means++ algorithm with a dynamic K value to cluster the positions and types of flying birds, avoiding the influence of the annotation frame size and making the annotation frame more suitable for the small target of the bird flock, thereby improving the accuracy of bird information statistics.

[0144] The embodiment of the present invention provides a method for extracting features of a data fusion set based on micro-Doppler features by a target recognition model. Figure 8The following is a schematic diagram of the implementation process of a method for extracting features of a data fusion set based on a micro-Doppler feature pair of a target recognition model. The method for extracting features of a data fusion set based on a micro-Doppler feature pair of a target recognition model specifically includes:

[0145] Step S601, loading a data fusion set, extracting bird micro-Doppler characteristic signals in the data fusion set based on an encoder, performing Fourier transform on the bird micro-Doppler characteristic signals, and obtaining Fourier transform results;

[0146] It should be noted that the encoder is arranged in front of the AlexNet network architecture, and the encoder is connected to the AlexNet network architecture. The encoder is used to extract the micro-Doppler characteristic signal of birds in the data fusion set, and the encoder is a self-attention encoder.

[0147] Step S602, analyzing the micro-Doppler motion distribution of birds based on the Choi-Williams algorithm, and outputting the motion feature distribution result;

[0148] It should be noted that Choi-Williams distribution (CWD) is a type of bilinear time-frequency distribution. It uses an exponential kernel to reduce the cross-term components in the distribution, thereby improving the accuracy and clarity of the analysis. Compared with other time-frequency distribution methods, CWD has better performance when processing signals with large Doppler shifts and can more accurately reflect the motion characteristics of the target. Using CWD to perform time-frequency analysis on radar echo signals, detailed distribution maps of bird micro-Doppler motion can be obtained. These distribution maps can clearly show the frequency, amplitude, and time-varying patterns of bird wing flapping.

[0149] Step S603, threshold judgment is performed on the motion feature distribution results to determine whether the target object exists. If the target object exists, convolution fusion, pooling and full connection processing are performed on the motion feature distribution results based on the AlexNet network architecture, focusing on the local features of the bird flock, and outputting the statistical evaluation results of the number and species of birds in the scene;

[0150] Step S604, obtain the statistical evaluation results of the number and species of birds in the bird flock in the scene, annotate the position and species of the birds based on the K-Means++ anchor frame, and output the statistical results of the bird information.

[0151] In this embodiment, when Fourier transform is performed on the bird micro-Doppler characteristic signal, the transformation function is defined as:

[0152] (9)

[0153] in, is the Fourier transform output representation, It is the time-series micro-Doppler characteristic signal of birds. is the short-time Fourier transform frequency, is a window function;

[0154] The motion feature distribution result is expressed as:

[0155] (10)

[0156] in, is the motion feature distribution result, is the time delay parameter, represents a scaling factor. In this embodiment, the scaling factor may be 0.02-0.4.

[0157] In summary, the present invention provides a bird information statistics method that integrates bird detection radar data and the sample line method. In the embodiment of the present invention, the target recognition model realizes the accurate evaluation of the bird information statistics results by integrating the bird detection radar data and the sample line method. At the same time, the enhanced processing of the bird detection radar data can adaptively enhance the robustness of the target recognition model, ensure the recognition effectiveness of the target recognition model algorithm, and use the bird population density as the prior probability to fuse the enhanced data set with the bird population density. This can effectively reduce the uncertainty in the statistical process, thereby improving the accuracy of bird information statistics, and can also use the enhanced data set to improve the stability and reliability of the target recognition model. This overcomes the problem that when the bird detection radar data and the manual survey data are integrated in the existing method, the bird population density cannot be counted, and the radar data and the manual survey data are not fully integrated, which increases the complexity and difficulty of data analysis.

[0158] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0159] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A bird information statistics method integrating bird detection radar data with the line sample method, characterized in that: The bird information statistics method integrating bird detection radar data with the transect method comprises: 16 sampling lines were set up within a radius of 10±0.1km in the statistical area. The total number of single bird populations and the number of species in the statistical area were recorded using infrared cameras combined with the variable distance sampling line method. The bird population density was calculated based on the total number of single bird populations and the number of species in the statistical area. The total number of single bird populations, the number of species, and the bird population density in the statistical area were uploaded to the information database. Obtain bird detection radar data, enhance the bird detection radar data to obtain an enhanced data set, obtain bird population density information from the information database, use the bird population density as a priori probability, fuse the enhanced data set with the bird population density to obtain a data fusion set, and upload the data fusion set to the information database; Among them, when enhancing the bird detection radar data, the three-dimensional coordinates of the radar points of the bird detection radar data are transformed into Convert to 2D coordinates , transformed to 80-pixel square representation and converted to three-channel color values ​​based on the peak conversion criterion, and then randomly added noise and brightness adjustment to generate an enhanced data set; The coordinate transformation matrix is ​​expressed as: （4） （5） in, is the tilt error of the radar point in three-dimensional coordinates, They are the pitch vector, attitude angle vector and flapping frequency vector of the target object pointed out by the radar respectively; Pre-build a target recognition model that combines micro-Doppler features and the YOLO model, iteratively train the target recognition model, and output a converged target recognition model; Load the data fusion set, take the data fusion set as input, and extract the features of the data fusion set based on the micro-Doppler feature of the target recognition model, and output the statistical results of the bird information in the bird flock, wherein the bird information statistics include the number, type and location information of the flying birds; The method for fusing the enhanced dataset with the bird population density comprises: Obtain bird population density, identify camera coordinates, acquisition time, camera focal length, and optical axis vector associated with bird population density; The infrared camera parameters are annotated based on OcamCalib, including the optical axis vector, pixel level and height-to-width ratio in the infrared camera parameters; Extract the annotated optical axis vector, pixel level and height-to-width ratio, transform the camera coordinate system using Euclidean transformation, rotate and scale the camera coordinate system, and combine the camera coordinate system external parameter rotation matrix , scaling matrix Optical axis vector, pixel level and height aspect ratio, and radar point 2D coordinates Fusion, realizing the fusion of radar point two-dimensional coordinates and camera coordinates, and outputting the fusion result of radar point two-dimensional coordinates and camera coordinates; The fusion results of the two-dimensional coordinates of the radar point and the camera coordinates, the bird population density, and the three-channel color values ​​represented by the square are obtained. The time series of the radar and the camera are matched using the timestamp alignment method to obtain the data fusion set.

2. The bird information statistics method of bird detection radar data and line sample method as claimed in claim 1, characterized in that: The total number of bird species and the number of bird species in the statistical area based on infrared cameras combined with variable distance transect method include: Based on the vegetation type and landform in the statistical area, the sampling lines are randomly arranged in layers. The width of a single sampling line is 50-200m, and the length of a single sampling line is 0.5-5.5km. Camera points were set up at equal intervals of 200 m in a single sampling line, and infrared cameras were set up at the camera points. The sampling line survey time was 7:30-11:30 am and 3:30-5:30 pm. The infrared camera was set to take 3 photos and 30 seconds of video each time it was triggered; Integrate at least one set of infrared photos and videos, and upload the infrared photos, videos, camera coordinates, and acquisition time to the information database in real time; The non-metric multidimensional scaling method is used to count the total number of single bird populations and the number of species in the area, and the total number of single bird populations and the number of species in the area are output.

3. The bird information statistics method of bird detection radar data and line sample method as claimed in claim 2 is characterized by: The method for calculating the bird population density based on the total number of single bird populations and the number of species in the statistical area specifically includes: The total number of bird populations and the number of species in the loading area are calculated, and the initial population density is calculated based on the Fourier intercept method; The initial population density is calculated using the following formula: （1） （2） in, represents the initial population density, Represents the length of a single sample line, is the number of splines, is the probability density function of finding a bird individual when the vertical distance is 0, is the width of a single sample line, represents the total number of a single population, It represents the vertical distance between the individual bird and the transect; Obtain the initial population density, correct the initial population density based on the density correction function, and output the bird population density; The bird population density is calculated by the following formula: （3） in, is the bird population density, is the correction coefficient of the density correction function, is the number of splines.

4. The bird information statistics method of combining bird detection radar data with the line sample method as claimed in claim 1, characterized in that: The method for enhancing the processing of bird detection radar data comprises: Obtain bird detection radar data and identify the three-dimensional coordinates of radar points based on the bird detection radar data , based on the coordinate transformation matrix, the three-dimensional coordinates of the radar point are converted into two-dimensional coordinates ; Get the 2D coordinates of the radar point , transform the radar point into a square representation with 80 pixels, and convert the radar point signal into a three-channel color value represented by the square based on the peak conversion criterion; Load the square representation containing the three-channel color values, randomly add noise points and brightness to the square representation of the radar point signal, and obtain the enhanced data set.

5. The bird information statistics method of bird detection radar data and line sample method as claimed in claim 4, characterized in that: When the radar point signal is converted into three-channel color values ​​represented by a square based on the peak conversion criterion, the three-channel color values ​​are ; The three-channel color value is calculated by the following formula: （6） （7） （8） in, Indicates the speed value of the radar point target. is the peak value of the intermediate frequency signal of the target object, is the radar point distance value.

6. The bird information statistics method of combining bird detection radar data with the line sample method according to any one of claims 1 to 5, characterized in that: The method for iteratively training the target recognition model includes: Traverse the data fusion set in the information database, disrupt the order of the data fusion set, and divide the data fusion set into a training set and a validation set, with the ratio of the training set to the validation set being 3:1; Initialize the hyperparameters of the initial baseline model of the target recognition model, define the activation function, loss function, learning rate, and training rounds of the initial baseline model; Obtain a training set, perform time-frequency analysis on the training set to obtain a target time-frequency diagram, detect the target time-frequency diagram based on micro-Doppler features, and obtain a flapping frequency curve of the training set; Perform Fourier transform processing on the flapping frequency curve, and perform threshold judgment, retain the weights corresponding to the high-frequency components in the Fourier transform processing results, prune the weights corresponding to the low-frequency components in the Fourier transform processing results, iteratively train the initial baseline model after weight pruning until convergence, and output the initial baseline model; Get the test set, use the test set as input, execute the initial baseline model, output the test results, and determine whether the test results meet the preset accuracy threshold. If they meet the preset accuracy threshold, output the converged target recognition model.

7. The bird information statistics method of combining bird detection radar data with the line sample method as claimed in claim 6, characterized in that: The target recognition model uses the YOLO7 model as the initial baseline model, and the initial baseline model consists of a backbone, a neck, and a head. The backbone of the initial baseline model is frozen, and the backbone is replaced by the AlexNet network architecture. The AlexNet network architecture includes six convolutional layers, three groups of fully connected layers, and a pooling layer. The six groups of convolutional layers are the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer. The fully connected layers are the first fully connected layer, the second fully connected layer, and the third fully connected layer. The first convolutional layer, the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer are connected to the pooling layer respectively, the first fully connected layer is connected to the second fully connected layer, and the second fully connected layer is connected to the third fully connected layer. The neck is improved to obtain an improved neck. A bird flock scene recognition network architecture is introduced into the neck. The improved neck is used to focus on the local features of the bird flock, and to perform statistical evaluation on the number and species of flying birds in the scene. The K-Means++ anchor frame is introduced in the head, and the K-Means++ anchor frame is used to mark the position and species of flying birds and output the statistical results of bird information.

8. The bird information statistics method of combining bird detection radar data with the line sample method as claimed in claim 7, characterized in that: The target recognition model extracts features of the data fusion set based on micro-Doppler features, including: Load the data fusion set, extract the bird micro-Doppler characteristic signal in the data fusion set based on the encoder, perform Fourier transform on the bird micro-Doppler characteristic signal, and obtain the Fourier transform result; Analyze the micro-Doppler motion distribution of birds based on the Choi-Williams algorithm and output the motion feature distribution results; Perform threshold judgment on the motion feature distribution results to determine whether the target object exists. If the target object exists, perform convolution fusion, pooling and full connection processing on the motion feature distribution results based on the AlexNet network architecture, focus on the local features of the bird flock, and output the statistical evaluation results of the number and species of birds in the scene; Obtain the statistical evaluation results of the number and species of birds in the scene, annotate the location and species of the birds based on the K-Means++ anchor frame, and output the statistical results of the bird information.

9. The bird information statistics method of combining bird detection radar data with the line sample method as claimed in claim 8, characterized in that: When Fourier transform is performed on the bird micro-Doppler characteristic signal, the transform function is defined as: （9） in, is the Fourier transform output representation, It is the time-series micro-Doppler characteristic signal of birds. is the short-time Fourier transform frequency, is a window function; The motion feature distribution result is expressed as: （10） in, is the motion feature distribution result, is the time delay parameter, Represents the scaling factor.

Images