Multi-target detection method based on radar one-dimensional range profile fusion of CFAR and isolation forest

Abstract

The present invention relates to a radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion. The method comprises the following steps: obtaining radar raw data and performing preprocessing to obtain radar one-dimensional range profile data; detecting targets in the radar one-dimensional range profile data using CA-CFAR detection adaptive threshold technology and normalizing energy values; then calculating target scores using an isolation forest method; performing weighted fusion on CA-CFAR detection and isolation forest results to obtain a fusion result, defining a new detection threshold to confirm the target, setting a new detection threshold #imgabs0#, and determining whether the target exists. The method combines CFAR detection with the isolation forest method to effectively suppress false alarms caused by radar clutter interference, reduce the risks of false detection and missed detection, and improve target detection accuracy. The method is applicable to various clutter environments, has high computational speed and low computational complexity, and meets the requirements for radar multi-target detection in complex scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal multi-target detection, and in particular to a radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion. Background Art

[0002] Target detection is a key task in radar systems. Conventional constant false alarm rate (CFAR) detection is a commonly used radar target detection method with extensive applications in radar signal processing. Based on statistical theory, CFAR assumes that clutter follows a specific probability distribution (such as Gaussian or Weibull). It adaptively adjusts the detection threshold by estimating clutter parameters to maximize the probability of target detection while maintaining a constant false alarm rate.

[0003] However, with the increasing complexity of radar application scenarios, traditional CFAR detection algorithms have gradually exposed some problems and shortcomings in practical applications: they are more sensitive to the assumptions about clutter distribution. When the actual clutter distribution does not match the assumptions, the detection performance will significantly decline. Moreover, in complex scenarios such as sidelobes, azimuth ambiguity, phantoms, strong coherent speckle noise, and multi-target environments, the traditional CFAR detection algorithm has obvious performance bottlenecks in detecting weak targets and is prone to false alarms, which limits the effectiveness of the CFAR detection algorithm in practical applications. In addition, current radar technology is developing towards high resolution, multi-target detection, and adaptability to complex environments. In some emerging radar application fields, such as autonomous driving, drone monitoring, and security monitoring, higher requirements are placed on the accuracy and reliability of radar target detection. Traditional CFAR algorithms and single machine learning algorithms are difficult to meet the needs of these applications for weak target detection.

[0004] With the continuous development of machine learning technology, some machine learning-based anomaly detection algorithms have also been applied to the field of radar target detection. Among them, the isolation forest algorithm, as an unsupervised learning anomaly detection algorithm, constructs multiple isolation trees by randomly splitting the data space and uses path length to measure whether a data point is an outlier. Due to its advantages such as low computational complexity and applicability to high-dimensional data, it has received widespread attention. However, the isolation forest algorithm has insufficient performance when applied to radar target detection and cannot adapt to the characteristics of radar signals, so the algorithm often needs to be improved and optimized. Summary of the Invention

[0005] In view of this, the present invention provides a radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion, which effectively solves the problem that radar target detection in complex scenarios is prone to false alarms in the existing technology, and improves the target detection accuracy and adaptability to the environment.

[0006] To achieve the above object, the present invention provides a radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion, comprising the following steps:

[0007] S1, obtain radar raw data and preprocess the radar one-dimensional range image data;

[0008] S2, performing CFAR detection on the radar one-dimensional range image data to obtain a range unit , the energy value of the distance unit obtained by CFAR detection is normalized to obtain ;

[0009] S3. Use the isolation forest method to calculate the target score corresponding to each range unit index in the radar one-dimensional range image data. ;

[0010] S4. Normalize the CFAR detection result to the energy value of the distance unit The target score corresponding to each distance unit index obtained by the isolation forest Perform weighted processing to obtain a new fusion result , the expression is:

[0011] ;

[0012] in, represents the confidence weight factor of CFAR detection, represents the confidence weight factor of the isolation forest;

[0013] Set a new detection threshold , filter out targets below the detection threshold, when the new fusion result Exceeding the new detection threshold , it is confirmed that the target exists; when the new fusion result The new detection threshold has not been exceeded , then it is determined that there is no target, the expression is:

[0014] ;

[0015] in, Represents a decision function.

[0016] Preferably, the millimeter-wave radar reflection signal received by the receiving antenna is amplified by a low-noise amplifier (LNA) and then mixed with the millimeter-wave radar transmission signal to obtain a complex baseband signal, which is then low-pass filtered and sampled by an analog-to-digital converter (ADC) to obtain discrete radar raw data.

[0017] Preferably, the preprocessing includes performing one-dimensional Fourier transform processing on the radar raw data to obtain one-dimensional range profile data of the radar.

[0018] Preferably, the CFAR detection adopts unit average constant false alarm rate detection CA-CFAR, the CA-CFAR detector includes a detection unit D and 2h reference units, and the sampling values of the reference units are averaged as the background power level estimation , the expression is:

[0019] ;

[0020] The detection threshold of the detection unit D The expression is:

[0021] ;

[0022] in, Indicates the threshold coefficient.

[0023] Preferably, the one-dimensional range image data of the radar is input into the detection unit D of the CA-CFAR detector, and the obtained range unit And the corresponding energy value , normalize the energy value of the distance unit to obtain the normalized energy value , the expression is:

[0024] ;

[0025] in, Indicates the number of detected targets, Indicates the The distance unit corresponding to the detected target, Indicates the The energy value corresponding to the detection target, Indicates the The energy value of the detected target after normalization, express The maximum energy value, express Minimum energy value;

[0026] Compared with the adaptive threshold obtained by the CA-CFAR detector, if the input signal exceeds the adaptive threshold, it is judged that the target exists; if the input signal does not exceed the adaptive threshold, it is judged that the target does not exist.

[0027] Preferably, the expression for calculating the target score corresponding to the distance unit index in the isolation forest is:

[0028] ;

[0029] in, represents the average path length of the isolation forest, represents the expected path length of the isolation forest, the target score The value range is ; Indicates the The target score of the distance unit corresponding to the detected target, the target score The closer it is to 1, the The more likely a detected target is a target point, the higher the target score The closer it is to 0, the The more likely a detected target is to be the target point.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present invention combines CFAR detection and isolation forest detection for radar target detection, giving full play to the advantages of CFAR detection in false alarm probability control and isolation forest's sensitivity to abnormal data distribution, improving the accuracy and stability of radar target detection, reducing the risk of false detection or missed detection by a single algorithm, and effectively solving the problems of poor weak target detection performance and easy false alarm in complex scenarios of traditional CFAR detection methods. It maintains good detection performance in different clutter environments and multi-target scenarios, effectively copes with various complex situations such as sidelobes, azimuth ambiguity, phantoms, strong coherent speckle noise, etc., improves the target detection capability of the radar system in complex environments, and expands the applicability of the radar system in various application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] FIG1 is a flowchart of the fusion of CFAR and isolation forest in the present invention;

[0033] Figure 2 Schematic diagram of the CA-CFAR detector of the present invention;

[0034] Figure 3 This is a flow chart of the isolation forest method of the present invention. DETAILED DESCRIPTION

[0035] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0036] The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion includes the following steps:

[0037] S1. Configure the millimeter-wave radar waveform. The millimeter-wave radar reflected signal received by the receiving antenna is amplified by a low-noise amplifier (LNA) and then mixed with the millimeter-wave radar transmit signal to obtain a complex baseband signal. The signal is then low-pass filtered and sampled by an analog-to-digital converter (ADC) to obtain discrete radar raw data. The radar raw data is processed by a one-dimensional Fourier transform to obtain the radar's one-dimensional range profile data, which can be used to represent the change in the intensity value (power) of the radar echo signal with power.

[0038] S2, using unit average constant false alarm rate detection CA-CFAR to perform CFAR detection on the radar one-dimensional range image data to obtain the range unit , the energy value of the distance unit obtained by CFAR detection is normalized to obtain ;

[0039] Traditional CFAR detection is to perform CFAR detection on the one-dimensional range profile data of the amplitude matrix obtained after radar pre-processing accumulation. The radar signal is detected by adaptive threshold technology. In this embodiment, the unit average constant false alarm rate detection CA-CFAR is used.

[0040] like Figure 2 As shown, the CA-CFAR detector consists of a detection unit D and 2h reference units. The reference units are located on both sides of the detection unit, with h units in front and behind. The main function of the protection unit P is to prevent the target energy from leaking to the reference unit and affecting the detection effect in the case of a single target. The CA-CFAR detector obtains an adaptive threshold through a sliding window, and the sampling values of the reference unit are averaged as the background power level estimate. , the expression is:

[0041] ;

[0042] Detection threshold of detection unit D The expression is:

[0043] ;

[0044] in, represents the threshold coefficient;

[0045] The radar's one-dimensional range image data is input into the detection unit D of the CA-CFAR detector, and the obtained range unit And the corresponding energy value , normalize the energy value of the distance unit to obtain the normalized energy value , the expression is:

[0046] ;

[0047] in, Indicates the number of detected targets, Indicates the The distance unit corresponding to the detected target, Indicates the The energy value corresponding to the detection target, Indicates the The energy value of the detected target after normalization, express The maximum energy value, express Minimum energy value;

[0048] Compared with the adaptive threshold obtained by the CA-CFAR detector, if the input signal exceeds the adaptive threshold, it is judged that the target exists; if the input signal does not exceed the adaptive threshold, it is judged that the target does not exist.

[0049] S3. Use the isolation forest method to calculate the target score corresponding to each range unit index in the radar one-dimensional range image data. ;

[0050] Isolation forest is a boost tree model for unsupervised learning. For each subtree, a random value within the range of eigenvalues in the dataset is randomly selected. Based on the principle that the eigenvalue is greater than or less than the random value, all data is divided into two parts according to the random value of the feature, gradually separating all data. Because abnormal data usually accounts for a small proportion, it can be distinguished from normal data using fewer attributes. Therefore, average path length can be used as a target measurement indicator. Isolation forest mainly includes the following steps:

[0051] Initialization parameters, including the number of isolated trees to construct, the maximum tree depth of each isolated tree, the size of the sample used for each isolated tree, and the seed that controls the random process;

[0052] The greater the number of isolated trees, the higher the stability and accuracy of the model, but the computational cost will also increase;

[0053] The deeper the isolation tree is, the better it can capture the details of the data, but it may lead to overfitting;

[0054] Using smaller sample sizes for isolation trees can improve model efficiency but may reduce accuracy;

[0055] Controlling the seed of the random process ensures the reproducibility of the results;

[0056] Construct an isolation tree model. The isolation forest reduces the deviation caused by the randomness of a single tree by integrating multiple trees, improving the stability and accuracy of the model, including:

[0057] Randomly select a feature at each split node;

[0058] Randomly select a split value within the range of the selected feature;

[0059] Separate the dataset into the left or right subtree according to the split value until each data point is isolated or the maximum tree depth is reached;

[0060] By constructing multiple isolated trees, an isolation forest is formed;

[0061] Calculate the path length, calculate the path length of each data point in the isolation tree, and find the average value. The shorter the path length, the easier it is for the data point to be isolated, and therefore the more likely it is to be the target point. The specific steps include:

[0062] From the training data Random selection Sample points Subset of ;

[0063] For the current node (data subset ) and randomly selected feature dimensions , cutting point The expression is:

[0064] ;

[0065] in, Indicates that the current node data is in dimension All the value sets on , Representation Dimension Down The minimum value of the data, Representation Dimension Down The maximum value of the data, express The values randomly sampled from the uniform distribution;

[0066] Select cutting point Generate a hyperplane, divide the current node data space into 2 subspaces, and divide the currently selected dimension smaller than the cutting point The point is placed on the left branch of the current node, and the points greater than or equal to the cutting point are placed on the left branch of the current node. The point is placed on the right branch of the current node;

[0067] Recursively repeat steps (2) and (3) until all leaf nodes have only one sample point or the isolated tree has reached the specified height;

[0068] Repeat steps (1)-(4) until an isolated tree is generated;

[0069] For each distance unit, traverse each isolated tree and record the path length from the root node to the leaf node In the isolation forest, the construction process of each isolated tree can be regarded as gradually separating data points by random splitting in multidimensional space, and the path length Refers to the distance from the root node to the unit the number of cuts required to be isolated;

[0070] The average path length of all isolated tree species is spherically averaged to obtain the average path length of the data point ;

[0071] Normalize the path length and calculate the expected path length using the harmonic number , the expression is:

[0072] ;

[0073] ;

[0074] in, Indicates the total number of distance units. In this embodiment , Indicates the harmonic numbers;

[0075] Calculate target score based on average path length and expected path length Calculate the target score for each data point , the expression is:

[0076] ;

[0077] in, represents the average path length of the isolation forest, represents the expected path length of the isolation forest, the target score The value range is ;

[0078] Indicates the The target score of the distance unit corresponding to the detected target, the target score The closer it is to 1, the smaller the path length is, and the data points are easily isolated. The more detected objects are, the more likely they are to be target points; target score The closer it is to 0, the longer the path length will be. The more likely a detected target is to be the target point.

[0079] S4. Normalize the CFAR detection result to the energy value of the distance unit The target score corresponding to each distance unit index obtained by the isolation forest Perform weighted processing to obtain a new fusion result , the expression is:

[0080] ;

[0081] in, represents the confidence weight factor of CFAR detection, represents the confidence weight factor of the isolation forest;

[0082] Set a new detection threshold , filter out targets below the detection threshold, when the new fusion result Exceeding the new detection threshold , it is confirmed that the target exists; when the new fusion result The new detection threshold has not been exceeded , then it is determined that there is no target, the expression is:

[0083] ;

[0084] in, Represents a decision function, the input is the fusion result and the detection threshold, and the output is a binary decision (the presence or absence of the target).

[0085] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion, characterized by: The following steps are involved: S1, obtain radar raw data and preprocess the radar one-dimensional range image data; S2, performing CFAR detection on the radar one-dimensional range image data to obtain a range unit , the energy value of the distance unit obtained by CFAR detection is normalized to obtain ; S3. Use the isolation forest method to calculate the target score corresponding to each range unit index in the radar one-dimensional range image data. ; S4. Normalize the CFAR detection result to the energy value of the distance unit The target score corresponding to each distance unit index obtained by the isolation forest Perform weighted processing to obtain a new fusion result , the expression is: ； in, represents the confidence weight factor of CFAR detection, represents the confidence weight factor of the isolation forest; Set a new detection threshold , filter out targets below the detection threshold, when the new fusion result Exceeding the new detection threshold , it is confirmed that the target exists; when the new fusion result The new detection threshold has not been exceeded , then it is determined that there is no target, the expression is: ； in, Represents a decision function.

2. The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion according to claim 1 is characterized in that: The millimeter-wave radar reflected signal received by the receiving antenna is amplified by a low-noise amplifier (LNA) and then mixed with the millimeter-wave radar transmit signal to obtain a complex baseband signal. The signal is then low-pass filtered and sampled by an analog-to-digital converter (ADC) to obtain discrete radar raw data.

3. The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion according to claim 2 is characterized in that: The preprocessing includes performing one-dimensional Fourier transform processing on the radar raw data to obtain one-dimensional range image data of the radar.

4. The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion according to claim 1 is characterized in that: The CFAR detection adopts unit average constant false alarm rate detection CA-CFAR. The CA-CFAR detector includes a detection unit D and 2h reference units. The sampling values of the reference units are averaged as the background power level estimation. , the expression is: ； The detection threshold of the detection unit D The expression is: ; in, Indicates the threshold coefficient.

5. The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion according to claim 4 is characterized in that: The one-dimensional range image data of the radar is input into the detection unit D of the CA-CFAR detector, and the obtained range unit And the corresponding energy value , normalize the energy value of the distance unit to obtain the normalized energy value , the expression is: ； in, Indicates the number of detected targets, Indicates the The distance unit corresponding to the detected target, Indicates the The energy value corresponding to the detection target, Indicates the The energy value of the detected target after normalization, express The maximum energy value, express Minimum energy value; Compared with the adaptive threshold obtained by the CA-CFAR detector, if the input signal exceeds the adaptive threshold, it is judged that the target exists; if the input signal does not exceed the adaptive threshold, it is judged that the target does not exist.

6. The radar one-dimensional range profile multi-target detection method based on CFAR and isolation forest fusion according to claim 1 is characterized in that: The expression for calculating the target score corresponding to the distance unit index of the isolation forest is: ； in, represents the average path length of the isolation forest, represents the expected path length of the isolation forest, the target score The value range is ; Indicates the The object fraction of the distance unit corresponding to each detected object.

Images