Track establishment system based on multi-source information fusion in complex environment

Abstract

The invention relates to the technical field of track establishment, in particular to a track establishment system based on multi-source information fusion in a complex environment. Comprising a multi-source information acquisition module for acquiring and preprocessing original data of a radar, a photoelectric sensor and an infrared sensor; the feature extraction and analysis module is used for extracting features from the preprocessed data and analyzing and evaluating the features; the multi-source information fusion module is used for fusing radar data by adopting a Kalman filtering algorithm and fusing target recognition results of different sensors by adopting an evidence theory to obtain multi-source fusion features; the true and false trace point identification module is used for constructing and training a true and false trace point identification model and identifying true and false trace points; and the data association and track establishment module is used for carrying out data association by adopting an improved nearest neighbor algorithm, and establishing a target track and carrying out real-time updating and maintenance on the track when an unassociated trace point exceeds a preset threshold value. According to the invention, the accuracy and continuity of the track and the reliability and effectiveness of the system in a complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of track establishment technology, and in particular to a track establishment system based on multi-source information fusion in complex environments. Background Technology

[0002] Current track establishment technologies primarily employ two core methods to identify and continue target trajectories: The genuine / false track identification stage uses the signal-to-clutter ratio and track width as key criteria. By analyzing whether these characteristics match the signal attributes of a real target, it distinguishes genuine target tracks from false tracks such as clutter and noise. The data association stage generally uses nearest-neighbor algorithms. By calculating the spatial distance between the currently acquired track and existing tracks, it matches the track with the smallest distance that meets a preset association threshold with the corresponding track. This enables the updating and continuation of existing tracks, or determines whether to initiate a new track establishment process for unmatched isolated tracks. These two methods work together to support the current track establishment tasks.

[0003] In cluttered environments, such as complex electromagnetic interference and densely packed multi-target scenarios, the aforementioned technical solutions suffer from a series of problems due to limitations in their design. Firstly, the single features relied upon for distinguishing between real and false tracks, such as signal-to-clutter ratio and track width, often closely resemble clutter signal characteristics in cluttered environments. This makes it difficult for the identification algorithm to effectively differentiate between real tracks and clutter, easily misclassifying clutter as real tracks. Secondly, nearest-neighbor-based association algorithms use only spatial distance as the matching criterion. In cluttered environments, many clutter tracks are distributed around real target tracks, making them prone to incorrect association with existing tracks. These two factors combined lead to an increase in false tracks. These false tracks not only consume system computing and storage resources but also mislead subsequent decisions, such as misjudging threat targets in air defense systems, reducing the overall effectiveness of the system. Thirdly, misjudgment and incorrect association of false tracks disrupt the spatiotemporal consistency of tracks, causing interruptions and position jumps in the real target's track, failing to accurately reflect the target's actual trajectory. For example, in missile tracking systems, this lack of track continuity can lead to target loss, directly affecting the reliability of real-time tracking missions. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a track establishment system based on multi-source information fusion in complex environments, so as to solve the problems of insufficient system reliability and effectiveness caused by numerous false tracks and poor track continuity under strong clutter.

[0005] To achieve the above objectives, the present invention provides a trajectory establishment system based on multi-source information fusion in complex environments, comprising:

[0006] The multi-source information acquisition module collects raw data from radar sensors, photoelectric sensors, and infrared sensors, and preprocesses the collected raw data, including a sensor selection and layout unit, a data acquisition unit, and a data preprocessing unit.

[0007] The feature extraction and analysis module extracts features from the preprocessed data and analyzes and evaluates the extracted features, including a data feature extraction unit and a feature analysis and evaluation unit.

[0008] The multi-source information fusion module, based on the features output by the feature extraction and analysis module, uses the Kalman filter algorithm to fuse radar data and uses evidence theory to fuse target recognition results from different sensors to obtain multi-source fusion features, including a radar sensor data fusion unit and a target recognition result fusion unit from different sensors.

[0009] The real and fake dot identification module constructs feature vectors based on multi-source fusion features, uses SVM to construct and train the real and fake dot identification model, and uses the trained model to identify real and fake dots. It includes a feature vector construction unit, a machine learning-based identification model selection unit, a kernel function introduction unit, a model training and parameter adjustment unit, and a dot identification unit.

[0010] The data association and track establishment module uses an improved nearest neighbor algorithm to associate data based on the real track output by the real and fake track identification module. When the number of unassociated tracks exceeds a preset threshold, a target track is established, and the track is updated and maintained in real time. The module includes an improved nearest neighbor algorithm unit, a data association unit, a track establishment and update unit, and a track maintenance unit.

[0011] Preferably, the data feature extraction unit includes radar sensor data feature extraction, photoelectric sensor data feature extraction, and infrared sensor data feature extraction.

[0012] Preferably, the radar sensor data feature extraction specifically includes: signal-to-noise ratio. Dot width and Doppler frequency ,in, For the target signal power, For clutter power, and By solving the equation The solution is obtained. This is the echo signal after pulse compression. Peak value, For Doppler frequency, The radial velocity of the target, The frequency of the signals transmitted by the radar sensor and The speed of light; the photoelectric sensor data feature extraction specifically includes: shape features and texture features, wherein shape features are obtained through geometric moments. Texture features were obtained by extracting gray-level co-occurrence matrices. For a two-dimensional image, the grayscale function is... and These are the coordinates of image pixels. and It is a non-negative integer; the infrared sensor data feature extraction specifically includes: infrared radiation intensity. and radiation distribution characteristics, where radiation distribution characteristics are obtained through thermal moments get, The infrared radiation power received by the infrared sensor from the target. Let be the projected area of ​​the target in the sensor's field of view. For infrared images in pixels The temperature value at that location.

[0013] Preferably, the feature analysis and evaluation unit specifically includes: a statistical analysis subunit, which calculates the mean of each feature using statistical analysis methods. ,variance The analysis focuses on the degree of dispersion of features; the machine learning evaluation subunit uses machine learning algorithms to evaluate the importance of each feature for target recognition and track establishment.

[0014] Preferably, the dot pattern identification unit specifically includes: inputting the newly acquired dot pattern feature vector into the trained model, and using a discriminant function. To distinguish between genuine and fake dots, if If the dot is then considered a real dot; if If the dot is false, then it is determined to be a false dot; where, This represents the total number of training samples; For the first The optimal Lagrange multiplier corresponding to each training sample; , Represents the actual dot pattern. Indicates false dots; Use a Gaussian kernel function to calculate the training samples. With the sample to be classified The inner product in the high-dimensional feature space, i.e. , It is the bandwidth parameter of the Gaussian kernel; This is the optimal bias term.

[0015] Preferably, the improved nearest neighbor algorithm unit specifically includes: calculating the combined distance between point traces and flight paths. Spatial location distance Feature distance , It is the weight of spatial location distance. It is a dot. The three-dimensional coordinates It is a flight path The three-dimensional coordinates of the current predicted location. For characteristic number, The weights of the features and , , They are respectively and No. The eigenvalues ​​of a feature.

[0016] The beneficial effects of this invention are as follows: Compared with the prior art, this invention uses multi-source sensor data fusion and feature extraction, combined with SVM to build a model to accurately identify real and fake points and reduce false tracks; it uses an improved nearest neighbor algorithm to calculate the comprehensive distance associated with points, ensuring track continuity and ultimately improving the reliability and effectiveness of the track system in complex environments. Attached Figure Description

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

[0018] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains.

[0020] like Figure 1 As shown, a trajectory establishment system based on multi-source information fusion in complex environments includes:

[0021] 1. Multi-source information acquisition module

[0022] 1.1 Sensor Selection and Layout Unit

[0023] In complex environments, multi-source information acquisition requires selecting appropriate sensor equipment and deploying it rationally based on the specific characteristics of the application scenario. Taking maritime target monitoring as an example, to comprehensively cover the target area, the characteristics of radar sensors, photoelectric sensors, and infrared sensors must be considered.

[0024] Radar sensors can provide the target's distance. ,speed ,position Information such as antenna gain affects the detection range and accuracy. Transmission power ,wavelength Factors such as [list of factors] affect the maximum detection range of a radar sensor, according to the radar sensor equation. It can be represented as:

[0025]

[0026] in, The radar cross-section of the target is the radar sensor cross-section. This represents the minimum detectable signal power of the radar sensor receiver.

[0027] Photoelectric sensors can acquire image information of targets, which helps in identifying their shape features. To ensure image clarity and coverage, the focal length of the photoelectric sensor must be appropriately selected based on the target's likely location and size. Field of view Parameters such as field of view. With focal length and imaging plane size The relationship is:

[0028]

[0029] Infrared sensors detect the infrared radiation emitted by a target, thus determining its thermal characteristics. Their detection range is affected by the intensity of the target's infrared radiation. Atmospheric transmittance Detector sensitivity Influenced by factors such as [list of factors], the maximum detection range of an infrared sensor [is limited]. It can be represented as:

[0030]

[0031] Based on the above parameters, the number and installation location of radar sensors, photoelectric sensors, and infrared sensors should be reasonably determined to ensure coverage of the target area.

[0032] 1.2 Data Acquisition Unit

[0033] Sensor data is collected at a specific sampling frequency to ensure real-time performance and accuracy. For radar sensors, the sampling frequency... It depends on the target's speed and the pulse repetition frequency of the radar sensor. To determine the sampling frequency, in order to avoid Doppler blurring, the sampling frequency is... Should meet:

[0034]

[0035] in, The maximum radial velocity of the target.

[0036] Photoelectric sensors and infrared sensors can acquire images and infrared radiation data in real time, but in order to reduce the amount of data and processing burden, the acquisition frequency can be adjusted appropriately according to the target's motion state and changes.

[0037] 1.3 Data Preprocessing Unit

[0038] The acquired raw data undergoes preprocessing, including filtering and noise reduction, to improve data quality. For radar sensor data, filters are used to remove noise interference. Commonly used filters include Kalman filters and Wiener filters. Taking the Kalman filter as an example, its state prediction equation is:

[0039]

[0040] in, for The predicted state vector at time t. for The state transition matrix at each time step, for The estimated state vector at time t. for Control the input matrix at all times. for Control the input vector at all times.

[0041] The measurement update equation is:

[0042]

[0043] in, for Update the state vector at each time step. for Moment Kalman gain, for The measurement vector at time, for Time measurement matrix.

[0044] Noise reduction methods can be used to process data from photoelectric sensors and infrared sensors, such as median filtering and Gaussian filtering. Taking median filtering as an example, for each pixel in the image, the median value of its neighboring pixels is taken as the new value of that pixel, thereby removing interference such as salt-and-pepper noise.

[0045] 2. Feature Extraction and Analysis Module

[0046] 2.1 Data Feature Extraction Unit

[0047] (1) Radar sensor data feature extraction

[0048] As an important type of sensor, radar sensors can provide information such as the target's range, velocity, and azimuth. For radar sensor data, we focus on extracting features such as the signal-to-clutter ratio, spot width, and Doppler frequency of the data points.

[0049] Signal-to-Clutter Ratio (SCR): The SCR is the ratio of the radar sensor's echo signal power to the clutter power, reflecting the relative strength of the target signal against a clutter background. Let the target signal power received by the radar sensor be... The clutter power is Then the mixture of information The calculation formula is:

[0050]

[0051] Pulse Width (PW): Pulse width refers to the width of the radar sensor echo signal in the time domain. It is related to the size and shape of the target and the resolution of the radar sensor. The pulse width is determined by measuring the full width at half maximum (FWHM) of the signal after pulse compression processing of the radar sensor echo signal. Let the pulse-compressed echo signal be... Its peak value is Then the width of the dot By solving the equation Two solutions were obtained. and dot width .

[0052] Doppler frequency ( The Doppler frequency is the frequency shift caused by the relative motion between the target and the radar sensor. According to the Doppler effect, the Doppler frequency... radial velocity relative to the target The frequency of signals transmitted by radar sensors and the speed of light The relationship between them is:

[0053]

[0054] (2) Feature extraction of photoelectric sensor data

[0055] Photoelectric sensors can acquire image information of targets, which helps to identify the target's shape features. We mainly extract features such as the target's shape and texture.

[0056] Shape features: Geometric moments are used to describe the shape of the target. Geometric moments are weighted integrals of the image's pixel grayscale values ​​and are invariant to translation, rotation, and scale. For two-dimensional images, let... Let be the grayscale function of the image, and its Order geometric moments Defined as:

[0057]

[0058] in, and These are the coordinates of image pixels. and It is a non-negative integer. By calculating geometric moments of different orders, the shape feature vector of the target can be obtained.

[0059] Texture Features: The gray-level co-occurrence matrix (GLCM) is used to extract the texture features of the target. The GLCM describes the spatial distribution relationship of pixel pairs with different gray levels in an image. Let the image have 100 gray levels. Gray-level co-occurrence matrix Indicates a distance of , direction is Under these conditions, the grayscale value is and The frequency of pixel pairs. Multiple texture features, such as contrast, correlation, energy, and homogeneity, can be extracted from the gray-level co-occurrence matrix.

[0060] (3) Feature extraction from infrared sensor data

[0061] Infrared sensors can detect the infrared radiation of a target to determine its thermal characteristics. We extract features such as the target's infrared radiation intensity and distribution.

[0062] Infrared radiation intensity ( Infrared radiation intensity refers to the infrared radiation power per unit area. Let the infrared radiation power received by the infrared sensor from the target be... The projected area of ​​the target in the sensor's field of view is Then the infrared radiation intensity The calculation formula is:

[0063]

[0064] Radiation distribution characteristics ( The thermal moment is used to describe the infrared radiation distribution of a target. The thermal moment is a weighted integral of the pixel temperature values ​​in an infrared image, similar to a geometric moment. For an infrared image, let... For its pixel The temperature value at that location, First thermal moment Defined as:

[0065]

[0066] By calculating thermal moments of different orders, the radiation distribution characteristic vector of the target can be obtained.

[0067] 2.2 Feature Analysis and Evaluation Unit

[0068] The extracted features are analyzed and evaluated to determine the importance of each feature for target identification and track establishment.

[0069] Statistical analysis: Using statistical analysis methods, calculate the mean of each feature. ,variance Equivalent statistics. Let the eigenvector be... Then the mean and variance The calculation formulas are as follows:

[0070]

[0071] Variance reflects the degree of dispersion of features; the larger the variance, the worse the stability of the features.

[0072] Machine Learning Evaluation: Using machine learning algorithms, such as decision trees and neural networks, the importance of each feature for target recognition and trajectory establishment is evaluated. Taking decision trees as an example, the importance of features is evaluated by calculating the information gain of the features. Let the dataset be... The entropy is ,feature Information gain Defined as:

[0073]

[0074] in, It is a feature The set of possible values, It is a dataset Chinese characteristics Values A subset of the subset. The greater the information gain, the greater the contribution of the feature to the classification.

[0075] 3. Multi-source information fusion module

[0076] 3.1 Radar Sensor Data Fusion Unit

[0077] In complex environments, radar sensor data is subject to noise interference. To improve the accuracy of target position and velocity estimation, a Kalman filter algorithm is employed. Let the target's state vector be... ,in and The target is in Position coordinates at that moment and They are respectively the corresponding Moment velocity component.

[0078] The state transition equation of the target is:

[0079]

[0080] in, yes The state transition matrix at each time step, , The sampling period; yes The control input matrix is ​​always active; it can be set to a zero matrix when there is no control input. yes Control the input vector at all times; yes The noise vector at each time step, and , yes The noise covariance matrix at each time step.

[0081] The measurement equation for the radar sensor is:

[0082]

[0083] in, It is the measurement vector of the radar sensor. It is a measurement matrix. It is a measurement noise vector, and , It is the measurement noise covariance matrix.

[0084] The steps of the Kalman filter algorithm are as follows:

[0085] (1) Prediction steps

[0086] Predicted status:

[0087] Prediction error covariance:

[0088] (2) Update steps:

[0089] Calculate the Kalman gain:

[0090] Update status:

[0091] Update error covariance:

[0092] in for The prediction error covariance at time 10:00 for The prediction error covariance at time 10:00 for The update error covariance at time step, It is an identity matrix. By iterating through the above steps, filtering and prediction of radar sensor data can be achieved, thereby improving the accuracy of target position and velocity estimation.

[0093] 3.2 Fusion Unit for Target Recognition Results from Different Sensors

[0094] The Dempster-Shafer (DS) theory, also known as the evidence theory, can be used to fuse target recognition results from different sensors. Let... The identification framework encompasses all possible target types. The target identification result for each sensor can be represented as a Basic Probability Assignment (BPA) function. , ,in This represents the number of sensors.

[0095] For two BPA functions and Its fusion rule, namely the Dempster combination rule, is: ,in , The conflict coefficient represents the degree of conflict between two pieces of evidence. J, W, and O all represent the identification framework. A subset of.

[0096] When there are multiple BPA functions At this point, fusion can be carried out gradually through pairwise fusion. Assume that the previous fusions have already occurred. A BPA function obtains Then the fusion of the first BPA functions The result is:

[0097] ,in

[0098] The final result This refers to the BPA function, which integrates the target recognition results from all sensors. The final target recognition result can be determined based on this function, for example, by selecting... The largest As the target type.

[0099] In complex environments, based on the features output by the feature extraction and analysis module, Kalman filtering is used to fuse radar sensor data, and evidence theory is used to fuse target recognition results from different sensors. This fully leverages the advantages of each sensor, reduces the limitations of a single sensor, and outputs multi-source fused features from radar, photoelectric, and infrared sensors, providing more accurate information for subsequent identification of real and false points and track establishment.

[0100] 4. Module for distinguishing between genuine and fake dot patterns

[0101] 4.1 Feature Vector Construction Unit

[0102] To distinguish between genuine and fake data points in complex environments, the first step is to construct a feature vector based on the multi-source fusion features output by the multi-source information fusion module. This module fuses data from radar, photoelectric, and infrared sensors, outputting a result containing the fused features from each sensor. Based on this result, features required for data identification are extracted: for radar sensor fusion features, the signal-to-clutter ratio (SCR) of the data points is extracted. Dot width Doppler frequency For photoelectric sensor fusion features, extract the target's shape features. Texture features For infrared sensor fusion features, extract the infrared radiation intensity of the target. Radiation distribution characteristics These features are combined into a feature vector. This feature vector integrates information from multiple sensors, enabling a more comprehensive description of the characteristics of the dot pattern.

[0103] 4.2 Machine Learning-Based Identification Model Selection Unit

[0104] We chose Support Vector Machine (SVM) as the model for distinguishing real and fake dots. SVM is a supervised machine learning algorithm that classifies samples of different categories by finding the optimal hyperplane. For the real and fake dot identification problem, our goal is to find a hyperplane that can distinguish real dots from fake dots.

[0105] Let the training sample set be ,in It is the first The feature vector of each sample Indicates the first The categories of each sample, Represents the actual dot pattern. Indicates false dots; This represents the total number of training samples.

[0106] The goal of SVM is to solve the following optimization problem:

[0107]

[0108] in, It is the normal vector of the hyperplane. It is a bias term. These are slack variables used to handle cases where samples are indivisible. It is a penalty parameter used to balance the complexity of the model and the classification error.

[0109] 4.3 Introduction of Kernel Function Unit

[0110] To handle the case of nonlinear separability, a kernel function is introduced. Commonly used kernel functions include linear kernels, polynomial kernels, and Gaussian kernels. Here, we choose the Gaussian kernel function:

[0111]

[0112] in, It is the bandwidth parameter of the Gaussian kernel.

[0113] Using kernel functions, the original optimization problem can be transformed into a dual problem:

[0114]

[0115] in All are Lagrange multipliers, corresponding to the 1st, 2nd, 3rd, 4th, 5th One sample.

[0116] Solving the above dual problem yields the first... The optimal Lagrange multiplier corresponding to each training sample This allows us to obtain the optimal normal vector of the hyperplane. and optimal bias term .

[0117] 4.4 Model Training and Parameter Tuning Unit

[0118] Known real and fake dot patterns are used as training samples and input into the SVM model for training. During training, the penalty parameter needs to be adjusted. Bandwidth parameters of Gaussian kernel Cross-validation can be used to select the optimal combination of parameters.

[0119] Cross-validation divides the training sample set into... A subset, each time selecting one of them One subset is used as the training set, and the remaining subset is used as the validation set. Through multiple training and validation cycles, the accuracy of the model under different parameter combinations is calculated, and the parameter combination with the highest accuracy is selected as the optimal parameter set.

[0120] 4.5 Dot Identification Unit

[0121] The feature vector of the newly collected points The input is fed into the trained SVM model, and the discrimination is performed according to the following discriminant function:

[0122]

[0123] if If the dot is then considered a real dot; if If the dot is false, then it is determined to be a false dot. .

[0124] This method for identifying true and false tracks, based on multi-source information fusion and support vector machines, makes full use of information from multiple sensors. It can improve the accuracy of the identification model and reduce the generation of false tracks in complex environments, especially in areas with strong clutter.

[0125] 5. Data Association and Track Establishment Module

[0126] 5.1 Improved nearest neighbor algorithm unit

[0127] Based on the traditional nearest neighbor algorithm and the true / false dot identification module outputting the true dot information, we improve it by incorporating target feature information fused from multiple sources. Let there be... a dot trace , already track For dots and flight path We need to calculate the combined distance between them. .

[0128] Traditional spatial location distance It can be calculated using the Euclidean distance formula:

[0129]

[0130] in, It is a dot. Three-dimensional spatial coordinates, It is a flight path Predict the three-dimensional spatial coordinates of the current position.

[0131] Feature distance It is calculated from the target feature information after integrating multi-source information. Assume we have integrated... Features include target shape, velocity, infrared radiation, etc., with each feature having a weight of . ,and For the first Type of feature, dot pattern and flight path The eigenvalues ​​are respectively and Then the feature distance is:

[0132]

[0133] Overall distance It is a weighted sum of spatial distance and feature distance:

[0134]

[0135] in, This is the weight of spatial location distance, which can be adjusted according to the actual application scenario. When there are many clutter points in a strong clutter environment, it can be appropriately reduced. The value of is increased to give more weight to feature distance, thereby improving the accuracy of association.

[0136] 5.2 Data Association Unit

[0137] Based on the calculated comprehensive distance Perform data association. For each point... We calculate it with all tracks The comprehensive distance is used to select the track with the smallest comprehensive distance for matching. Let... ,if ,in If it is a preset association threshold, then the point is considered to be... If the association with the corresponding track is successful, the point will be added to the corresponding track; otherwise, the association will fail.

[0138] During the association process, to avoid situations where a single point is associated with multiple tracks or a single track is associated with multiple points, we employ a greedy algorithm for optimization. The specific steps are as follows:

[0139] (1) Initialize the association status of all points and tracks to unassociated.

[0140] (2) All of them are sorted in ascending order of comprehensive distance. Sort by combining.

[0141] (3) Iterate through the sorted sequences sequentially. Combination, if the dots and flight path They are not related, and Then the dots will be Related to flight path And mark their association status as associated.

[0142] Repeat step (3) until all points and tracks have been traversed or can no longer be associated.

[0143] 5.3 Track Establishment and Update Unit

[0144] When a point association fails, consider whether a new track needs to be established. If continuous... Within a given time period, the number of unrelated points in a certain region exceeds a preset threshold. If these points are identified as potentially originating from a new target, a new trajectory is established. The initial state of the new trajectory can be estimated based on the position and velocity information of these points.

[0145] For existing tracks, update them based on the newly associated point information. Let the track be... exist The state vector at time t is ,in It's location information. It's speed information. Newly associated points. The measurement vector is .

[0146] We use Kalman filtering for track updates. The prediction equation for Kalman filtering is:

[0147]

[0148] in, For the track exist The predicted state vector at time t. For the track exist The predicted state vector at time t. It is the state transition matrix, which can be determined based on the target's motion model.

[0149] The update equation for the Kalman filter is:

[0150]

[0151] in, For the track exist The updated state vector at each time step. It is a measurement matrix. The Kalman gain can be calculated using the following formula:

[0152]

[0153]

[0154]

[0155] in, and They are The prediction error covariance matrix at time t and The update error covariance matrix at time step. It is the process noise covariance matrix. It measures the noise covariance matrix. It is an identity matrix.

[0156] 5.4 Track Maintenance Unit

[0157] Track and manage flight paths; consider deleting a flight path if it has not generated any new associated points for an extended period. Let the flight path... The continuous unrelated time is ,if ,in If the maximum unassociated time is preset, then the track will be deleted.

[0158] When the state of a flight path changes abnormally, such as sudden speed changes or position jumps, timely adjustments and corrections should be made. A state change threshold can be set; when the state change exceeds the threshold, the flight path should be reinitialized or processed using a more complex filtering algorithm.

[0159] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0160] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A track establishment system based on multi-source information fusion in a complex environment, characterized in that, The application relates to a multi-source information fusion method for radar target tracking, which comprises the following steps: a multi-source information acquisition module is used to acquire original data of radar sensors, photoelectric sensors and infrared sensors, and the acquired original data is pretreated, including a sensor selection and layout unit, a data acquisition unit and a data pretreatment unit; a feature extraction and analysis module is used to extract features from the pretreated data and analyze and evaluate the extracted features, including a data feature extraction unit and a feature analysis and evaluation unit; a multi-source information fusion module is used to fuse radar data based on the features output by the feature extraction and analysis module by using a Kalman filtering algorithm, and to fuse target recognition results of different sensors by using evidence theory to obtain multi-source fusion features, including a radar sensor data fusion unit and a different sensor target recognition result fusion unit; a true and false point track discrimination module is used to construct a feature vector based on the multi-source fusion features, to construct and train a true and false point track discrimination model by using SVM, and to discriminate true and false point tracks by using the trained model, including a feature vector construction unit, a machine learning-based discrimination model selection unit, a kernel function introduction unit, a model training and parameter adjustment unit and a point track discrimination unit; a data association and track establishment module is used to associate data based on the true point tracks output by the true and false point track discrimination module by using an improved nearest neighbor algorithm, to establish a target track when unassociated point tracks exceed a preset threshold, and to update and maintain the track in real time, including an improved nearest neighbor algorithm unit, a data association unit, a track establishment and update unit and a track maintenance unit.

2. The system of claim 1, wherein, The data feature extraction unit comprises radar sensor data feature extraction, photoelectric sensor data feature extraction and infrared sensor data feature extraction.

3. The system of claim 2, wherein, The radar sensor data feature extraction specifically includes signal-to-clutter ratio , point track width , and Doppler frequency , wherein is target signal power, is clutter power, and can be solved by solving equation , and is the echo signal after pulse compression, is the peak value, is the Doppler frequency, is the radial velocity of the target, is the frequency of the radar sensor transmitting signal, and is the speed of light; the photoelectric sensor data feature extraction specifically includes shape features and texture features, wherein the shape features are obtained by geometric moments , and the texture features are extracted by using a gray level co-occurrence matrix, is a gray function of a two-dimensional image, and is the coordinate of an image pixel, and are non-negative integers; the infrared sensor data feature extraction specifically includes infrared radiation intensity and radiation distribution features, wherein the radiation distribution features are obtained by thermal moments , is the target infrared radiation power received by the infrared sensor, is the projected area of the target in the field of view of the sensor, is the temperature value of the infrared image at pixel .

4. The system of claim 1, wherein, The feature analysis and evaluation unit specifically includes: a statistical analysis subunit, which calculates the mean value of each feature by using a statistical analysis method , variance , and analyzes the dispersion degree of the feature; and a machine learning evaluation subunit, which evaluates the importance of each feature for target identification and track establishment using a machine learning algorithm.

5. The system of claim 1, wherein, The point trail identification unit specifically comprises: inputting the newly collected point trail feature vector into the trained model, and identifying the function If , it is judged that the point trail is a real point trail; if , it is judged that the point trail is a false point trail; wherein is the total number of training samples; is the optimal Lagrange multiplier corresponding to the th training sample; , represents a real point trail, represents a false point trail; is a Gaussian kernel function, and the inner product of the training sample and the sample to be classified in a high-dimensional feature space is calculated, that is , is a bandwidth parameter of the Gaussian kernel; is an optimal bias term.

6. The system of claim 1, wherein, The improved nearest neighbor algorithm unit specifically includes: calculating the combined distance between point tracks and flight tracks. Spatial location distance Feature distance , It is the weight of spatial location distance. It is a dot. The three-dimensional coordinates It is a flight path The three-dimensional coordinates of the current predicted location. For characteristic number, The weights of the features and , , They are respectively and No. The eigenvalues ​​of a feature.

Images