Radar target detection method based on comparative learning under multi-terrain condition

Abstract

The invention discloses a radar target detection method based on comparative learning under multiple topographic conditions, which comprises the following steps of: constructing a data set and dividing a training set and a test set after preprocessing, constructing an enhanced sample set after data enhancement, and constructing positive and negative sample pairs; the method is used for subsequent model training, radar target detection performance testing and network training, and based on distribution of training clutter samples in a feature space, a decision threshold is obtained by sorting training scores and taking quantiles. And making a final decision according to the relationship between the calculation score of the to-be-detected sample through the multi-topographic feature extraction detection network and the decision threshold. The deep network is trained by using clutter data in different terrains as different categories, feature distribution of clutters in different terrains can be learned, target detection in different terrains can be adapted, and the actual application performance of radar target detection in a multi-terrain scene is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target detection, and in particular to a radar target detection method based on contrast learning under multi-terrain conditions. Background Art

[0002] High-resolution radar refers to a radar system that can obtain high-resolution distribution information of target scattering points along the distance direction. The echo formed by high-resolution radar is called high-resolution range profile (HRRP), which is the vector sum of the complex sub-echoes of the target scattering points obtained by broadband radar signals projected on the radar ray. It provides detailed distribution characteristics of the target scattering points along the distance direction, and can provide information on the difference in scattering characteristics between clutter and target. In addition, target HRRP can reduce the clutter energy in each distance unit and improve the target signal-to-clutter ratio. Therefore, HRRP is often used for target detection.

[0003] In recent years, HRRP target detection technology based on machine learning has been greatly developed. Some manual feature extraction detection algorithms based on anomaly detection ideas are used for target detection. This type of method first extracts manual features of clutter training data, then constructs a single classifier (such as convex hull learning, decision tree, support vector machine, etc.) to learn the characteristic distribution of clutter, and establishes a decision hypersphere in the feature space to include the clutter features. Finally, according to whether the features of the sample to be detected are within the hypersphere, it is judged whether the sample to be detected is a target or clutter, thereby achieving target detection.

[0004] However, in multi-terrain scenarios, due to the significant differences in the distribution of clutter features under different terrains, their distribution in the feature space is far apart. The decision hypersphere containing all clutter feature distributions constructed by a single classifier will be very large. When the features of the target sample are in the feature space of the non-clutter distribution area inside it, it will be misjudged as clutter, resulting in a decrease in detection performance. Summary of the invention

[0005] Due to the differences in clutter feature distributions under different terrains, the existing algorithm based on manual feature extraction and single classifier has poor applicability in multi-terrain scenarios and is difficult to form an accurate decision boundary, resulting in a significant reduction in the probability of radar target detection. In view of the above defects, the present invention provides a radar target detection method based on contrastive learning under multi-terrain conditions. By using the idea of ​​anomaly detection, clutter is regarded as a normal sample, and the target is regarded as an abnormal sample. The clutter distribution under various terrains is clustered based on the terrain type, and multiple clutter feature distribution clusters are formed in a feature space. Multiple decision thresholds are determined according to the size of the distribution clusters; then, the relationship between the sample to be detected and each distribution cluster is determined according to the threshold, and then the sample to be detected is determined to be a target or clutter, so as to achieve robust detection of radar targets under multi-terrain conditions, including the following steps:

[0006] S1. Obtain the measured polarization high-resolution radar echo data of radar targets under multiple terrains. The acquired data includes clutter samples and samples to be tested. Complete the construction of the data set, perform preprocessing such as data slicing and segmentation, amplitude normalization, and form a data set by arranging and adding labels. Then divide the data set into training set and test set to facilitate subsequent model training and radar target detection performance testing. Specifically:

[0007] Firstly, polarization high-resolution radar is used to collect measured data of radar targets under multiple terrains and perform data preprocessing.

[0008] The radar echo data of measured targets and clutter under multiple terrains are obtained, and the radar echo data are preprocessed by segmentation, amplitude normalization and other operations based on prior knowledge such as target position and physical size.

[0009] S1-1. Construction of the data set and division of the training set and test set:

[0010] S1-1-1. Divide the radar echo data into a clutter sample set and a sample set to be tested:

[0011] In the acquired radar echo data, several frames of clutter echo samples constitute a clutter data sample set X clutter ={x1, x2, ..., x M}, where X clutter The total number of frames in is denoted as M, X clutter The number of terrains contained in the data is denoted as T, X clutter Each frame signal x in m is a 4×N matrix, where m represents the frame number of the polarization high-resolution radar echo signal, m=1, 2, ..., M, N represents the number of range units, and the first dimension of the matrix represents the echo data of the four polarization channels HH, HV, VH, and VV. m (i)=[S HH , S HV , S VH , S VV ] T , i represents the distance unit number, i=1,2,...,N,x m (i) represents x m The element on the i-th distance unit, [·] T represents transposition, H represents horizontal polarization, V represents vertical polarization, S HH Indicates the echo data of horizontal transmission and horizontal reception, S HV Indicates the echo data transmitted horizontally and received vertically, S VH Indicates the echo data transmitted vertically and received horizontally, S VVRepresents the echo data of vertical transmission and vertical reception; for each clutter radar echo data sample x under the terrain m , each with a label y m ,y m ∈{1, 2, …, T} corresponds to the terrain where the corresponding echo is collected, and T is the total number of terrains where data is collected;

[0012] The target echo data collected under multiple terrains constitute the target data sample set X target ={x1, x2, ..., x O}, X target The total number of frames in is denoted as O, X target Each frame signal x n The matrix dimension size and the physical meaning of each dimension are respectively clutter The signal in each frame is the same.

[0013] S1-1-2. Divide the data set into training set and test set:

[0014] Since the target data sample set X target Do not participate in the training process, so the target data samples are all used as test samples, and the clutter data sample set X clutter The training set and the test set are randomly divided in a 1:1 ratio to form the training clutter sample set X clutter_train And the test clutter sample set X clutter_test .

[0015] S1-2. Data enhancement: construct an enhanced sample set and construct positive and negative sample pairs.

[0016] Using contrastive learning method, the training clutter samples are enhanced by two data enhancement methods, cyclic shift and flipping, to form an enhanced sample set; the enhanced samples of the same training clutter sample are defined as positive sample pairs, and the enhanced samples of different training clutter samples are defined as negative sample pairs.

[0017] S1-2-1. Perform data enhancement on the training clutter samples to generate an enhanced sample set.

[0018] The training clutter sample set X clutter_train Divide into different batches, the batch size is recorded as P, consider a random batch For a sample x in this batch k , applying two independent data enhancement methods, we get two enhanced samples and Forming an augmented batch

[0019] S1-2-2. Construct positive and negative sample pairs to facilitate subsequent training.

[0020] For the same sample x k Enhanced sample and Enhanced sample labels and With the original label y k Stay consistent. and Constitutes a positive sample pair, enhancing the sample or The enhanced samples obtained with other samples in the batch constitute negative sample pairs.

[0021] S2. Perform network training:

[0022] S2-1. Construct a multi-terrain feature extraction and detection network.

[0023] S2-1-1. Construction of encoder network:

[0024] The encoder network uses the classic ResNet50 network for feature extraction, and the network input sample is a clutter sample x of size 4×N m , after passing through the encoder, an N1×1 embedding representation is formed for feature extraction.

[0025] S2-1-2. Construction of projection head network:

[0026] The projection head network consists of a fully connected layer. After the N1×1 embedding representation is input into the projection head network, the network outputs an N2×1 feature vector, which is used to map the extracted features into the feature space.

[0027] The multi-terrain feature extraction and detection network is denoted as φ, then and The corresponding vector representation and

[0028] S2-2. Construct a loss function based on contrastive learning.

[0029] The loss function introduces and adjusts the temperature hyperparameter, which reduces the repulsion of clutter samples under the same terrain in the feature space and increases the repulsion of clutter samples under different terrains, thereby further improving the clustering effect of clutter sample features under the same terrain.

[0030] S2-2-1. Add the category label of the enhanced sample as a condition to the calculation of the training loss function, assign different temperature hyperparameters to samples of different categories, and for the sample feature z of the i-th sample and the j-th sample i and z j , the loss function l1(i, j) is as follows:

[0031]

[0032] Where sim(z i , z j ) represents z i and z j The cosine similarity of is the indicator function. If the conditions in the brackets are met, it takes 1, and if not, it takes 0. N is the total number of original samples. and To enhance the sample labels, τ is an adjustable temperature hyperparameter.

[0033] In order to give different enhanced samples of the same category a larger temperature hyperparameter to reduce the repulsion between them, the new loss function Defined as:

[0034]

[0035] S2-2-2. Adjust the temperature hyperparameter to improve the clustering effect:

[0036] Although the repulsive force between samples of the same type is reduced, the above method is still similar to SimCLR in essence. It is still learning the unique features of a single sample. After a long period of training, it is still difficult to form concentrated clusters. In order to further improve the clustering effect, the temperature hyperparameters between samples of the same type are adjusted so that different enhanced samples of the same type are attractive to each other.

[0037] For a sample x in a batch i An enhanced sample of The rest of the augmented samples from the same category in the same batch can be represented as:

[0038]

[0039] The symbol \ represents an integer quotient operator. Based on the original loss function, this embodiment introduces an additional loss term:

[0040]

[0041] The l2 loss function makes the enhanced samples and Random positive sample augmentation batch Positive samples in Mutual attraction, Represents a uniform distribution.

[0042] S2-2-3. Integrate the loss function based on contrastive learning:

[0043] The overall network loss is as follows:

[0044]

[0045] Where λ represents the weight hyperparameter of l2 loss;

[0046] In order to ensure the effectiveness of l2 loss, α should satisfy the following formula:

[0047]

[0048] Where C represents the total number of training categories and N represents the training batch size.

[0049] S2-3. Input the enhanced sample set of the training clutter samples into the multi-terrain feature extraction detection network, and train the network based on the contrast loss function to construct clutter distribution clusters of different terrains. Input the enhanced sample into the network for feature extraction and feature mapping to obtain the final representation, and calculate the contrast loss based on the enhanced sample set. Train the network based on the loss function, and use the back propagation algorithm to update the weight parameters.

[0050] S3. Threshold determination and final decision.

[0051] Based on the distribution of training clutter samples in the feature space, the decision threshold is obtained by sorting the training score and taking the quantile. The final decision is made based on the relationship between the calculated score of the sample to be tested through the encoder and projection head network and the decision threshold.

[0052] S3-1. Detection of threshold determination:

[0053] The present invention adopts negative Mahalanobis distance -dM as the scoring function s, so for sample x, its final score s under the i-th type of terrain is i It can be expressed by the following formula:

[0054]

[0055] Where z is the feature vector obtained after sample x passes through the network, μ i and∑ i The mean and covariance matrices calculated for all training samples in the i-th class, for i, i∈I={1,2,…,C main}.

[0056] S(x) contains the score of each category of sample x. The present invention selects the category with the highest score as the final prediction label.

[0057] i*=argmax n∈I S n (x),

[0058]

[0059] In order to achieve target detection, the present invention needs to use the Mahalanobis distance to obtain the detection threshold, and the threshold of the i-th category containing samples is as follows:

[0060] First, all enhanced sample sets in the i-th class are sorted in ascending order of their scores to form S i :S i =[S i1 , S i2 , …, S iw-1 ]. Secondly, the value of the index n is set according to the adjustable parameter p and the number of feature samples w: n = round(w·(1-p)), where round(·) represents the rounding operator. Finally, by i Selecting the nth element in , we can get δ i :δ i =S in In the above, p is the rejection rate of the training sample, and its value ranges from 0 to 1.

[0061] In order to test the false alarm probability of 10 -4 For the detection probability under this condition, the present invention sets p=0.9999.

[0062] S3-2. Final judgment of the sample to be tested.

[0063] For the test samples, the following decision rules are followed:

[0064]

[0065] Among them, S i* (x) represents the highest score of the test sample, δ i* is the threshold corresponding to the highest scoring class.

[0066] If the score of the test sample is greater than or equal to the threshold, it is classified as a clutter sample; conversely, if the score of the sample is less than the threshold, it is classified as a target sample.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] (1) The present invention proposes a loss function based on contrastive learning, in which a temperature hyperparameter is introduced and adjusted. The improved loss function takes into account the terrain of the clutter samples, transforms unsupervised learning into supervised learning, and assigns different temperature hyperparameters to the cosine similarity between clutter enhanced samples of the same terrain and different terrains. By aggregating sample features based on temperature hyperparameters and cosine similarity, compact clustering in the multi-terrain clutter feature space is finally achieved, and distribution clusters of clutter of different terrains are constructed.

[0069] (2) The present invention proposes a threshold construction method based on Mahalanobis distance, which generates a score based on the Mahalanobis distance between the clutter training sample features and the distribution cluster center of the corresponding terrain clutter, and calculates the decision threshold of the corresponding terrain according to the score distribution and the set false alarm probability. According to the relationship between the score of the sample features to be tested relative to the center of each clutter distribution cluster and each decision threshold, it is judged whether the sample to be detected is a target or clutter.

[0070] (3) The present invention uses clutter data under different terrains as different categories to train the deep network, so as to learn the characteristic distribution of clutter under different terrains, thereby being able to adapt to target detection under different terrains and greatly improving the actual application performance of radar target detection in multi-terrain scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Flowchart of radar target detection based on contrastive learning under multi-terrain conditions

[0072] Figure 2 Block diagram of radar target detection model based on contrastive learning under multi-terrain conditions;

[0073] Figure 3 : is the target detection performance curve of the vehicle 1 in the Gobi terrain in the embodiment.

[0074] Figure 4 : is the target detection performance curve of the vehicle 2 in the Gobi terrain in the embodiment.

[0075] Figure 5 : is the target detection performance curve of the vehicle 1 on the highway terrain in the embodiment.

[0076] Figure 6 : is the target detection performance curve of the vehicle 2 on the highway terrain in the embodiment. DETAILED DESCRIPTION

[0077] In order to facilitate the understanding of the present invention, the device of the present invention will be described more fully below with reference to the relevant drawings. Embodiments of the device are given in the drawings. However, the device can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0078] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "set" should be understood in a broad sense, for example, they can be fixedly connected or set, or detachably connected or set, or integrally connected or set. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0079] Example

[0080] like Figure 1-2 As shown, this embodiment provides a radar target detection method based on contrast learning under multi-terrain conditions, comprising the following steps:

[0081] S1. Obtain the measured polarization high-resolution radar echo data of radar targets under multiple terrains. The acquired data includes clutter samples and samples to be tested. Complete the construction of the data set, perform preprocessing such as data slicing and segmentation, amplitude normalization, and form a data set by arranging and adding labels. Then divide the data set into training set and test set to facilitate subsequent model training and radar target detection performance testing. Specifically:

[0082] Firstly, polarization high-resolution radar is used to collect measured data of radar targets under multiple terrains and perform data preprocessing.

[0083] The radar echo data of measured targets and clutter under multiple terrains are obtained, and the radar echo data are preprocessed by segmentation, amplitude normalization and other operations based on prior knowledge such as target position and physical size.

[0084] S1-1. Construction of the data set and division of the training set and test set:

[0085] S1-1-1. Divide the radar echo data into a clutter sample set and a sample set to be tested:

[0086] In the acquired radar echo data, several frames of clutter echo samples constitute a clutter data sample set X clutter ={x1, x2, ..., x M}, where X clutter The total number of frames in is denoted as M, X clutter The number of terrains contained in the data is denoted as T, X clutter Each frame signal x in mis a 4×N matrix, where m represents the frame number of the polarization high-resolution radar echo signal, m=1, 2, ..., M, N represents the number of range units, and the first dimension of the matrix represents the echo data of the four polarization channels HH, HV, VH, and VV. m (i)=[S HH , S HV , S VH , S VV ] T , i represents the distance unit number, i = 1, 2, ..., N, x m (i) represents x m The element on the i-th distance unit, [·] T represents transposition, H represents horizontal polarization, V represents vertical polarization, S HH Indicates the echo data of horizontal transmission and horizontal reception, S HV Indicates the echo data transmitted horizontally and received vertically, S VH Indicates the echo data transmitted vertically and received horizontally, S VV Represents the echo data of vertical transmission and vertical reception; for each clutter radar echo data sample x under the terrain m , each with a label y m ,y m ∈{1, 2, …, T} corresponds to the terrain where the corresponding echo is collected, and T is the total number of terrains where data is collected;

[0087] The target echo data collected under multiple terrains constitute the target data sample set X target ={x1, x2, ..., x O}, X target The total number of frames in is denoted as O, X target Each frame signal x n The matrix dimension size and the physical meaning of each dimension are respectively clutter The signal in each frame is the same.

[0088] S1-1-2. Divide the data set into training set and test set:

[0089] Since the target data sample set X target It does not participate in the training process, so the target samples to be tested are all used as test samples, and the clutter data sample set X clutter The training set and the test set are randomly divided in a 1:1 ratio to form the training clutter sample set X clutter_train And the test clutter sample set X clutter_test .

[0090] S1-2. Data enhancement: construct an enhanced sample set and construct positive and negative sample pairs.

[0091] like Figure 2 As shown, the contrastive learning method is used to enhance the training clutter samples through two data enhancement methods, namely, cyclic shift and flipping methods, to form an enhanced sample set; the enhanced samples of the same training clutter sample are defined as positive sample pairs, and the enhanced samples of different training clutter samples are defined as negative sample pairs.

[0092] S1-2-1. Perform data enhancement on the training clutter samples to generate an enhanced sample set.

[0093] The training clutter sample set X clutter_train Divide into different batches, the batch size is recorded as P, consider a random batch For a sample x in this batch k , applying two independent data enhancement methods, we get two enhanced samples and Forming an augmented batch

[0094] S1-2-2. Construct positive and negative sample pairs to facilitate subsequent training.

[0095] For the same sample x k Enhanced sample and Enhanced sample labels and With the original label y k Stay consistent. and Constitutes a positive sample pair, enhancing the sample or The enhanced samples obtained with other samples in the batch constitute negative sample pairs.

[0096] S2. Perform network training:

[0097] S2-1. Construct a multi-terrain feature extraction and detection network.

[0098] S2-1-1. Construction of encoder network:

[0099] The encoder network uses the classic ResNet50 network for feature extraction, and the network input sample is a clutter sample x of size 4×N m , after passing through the encoder, an N1×1 embedding representation is formed for feature extraction.

[0100] S2-1-2. Construction of projection head network:

[0101] The projection head network consists of a fully connected layer. After the N1×1 embedding representation is input into the projection head network, the network outputs an N2×1 feature vector, which is used to map the extracted features into the feature space.

[0102] The multi-terrain feature extraction and detection network is denoted as φ, then and The corresponding vector representation and

[0103] S2-2. Construct a loss function based on contrastive learning.

[0104] The loss function introduces and adjusts the temperature hyperparameter, which reduces the repulsion of clutter samples under the same terrain in the feature space and increases the repulsion of clutter samples under different terrains, thereby further improving the clustering effect of clutter sample features under the same terrain.

[0105] S2-2-1. Add the category label of the enhanced sample as a condition to the calculation of the training loss function, assign different temperature hyperparameters to samples of different categories, and for the sample feature z of the i-th sample and the j-th sample i and z j , the loss function l1(ij) is as follows:

[0106]

[0107] Where sim(z i , z j ) represents z i and z j The cosine similarity of is the indicator function. If the conditions in the brackets are met, it takes 1, and if not, it takes 0. N is the total number of original samples. and To enhance the sample labels, τ is an adjustable temperature hyperparameter.

[0108] In order to give different enhanced samples of the same category a larger temperature hyperparameter to reduce the repulsion between them, the new loss function Defined as:

[0109]

[0110] Where n is the total number of samples of the original clutter training samples.

[0111] S2-2-2. Adjust the temperature hyperparameter to improve the clustering effect:

[0112] Although the repulsive force between samples of the same type is reduced, the above method is still similar to SimCLR in essence. It is still learning the unique features of a single sample. After a long period of training, it is still difficult to form concentrated clusters. In order to further improve the clustering effect, the temperature hyperparameters between samples of the same type are adjusted so that different enhanced samples of the same type are attractive to each other.

[0113] For a sample x in a batch i An enhanced sample of The rest of the augmented samples from the same category in the same batch can be represented as:

[0114]

[0115] The symbol \ represents an integer quotient operator. Based on the original loss function, this embodiment introduces an additional loss term:

[0116]

[0117] The l2 loss function makes the enhanced samples and Random positive sample augmentation batch Positive samples in Mutual attraction, Represents a uniform distribution.

[0118] S2-2-3. Integrate the loss function based on contrastive learning:

[0119] The overall network loss is as follows:

[0120]

[0121] Where λ represents the weight hyperparameter of l2 loss;

[0122] In order to ensure the effectiveness of l2 loss, α should satisfy the following formula:

[0123]

[0124] Where C represents the total number of training categories and N represents the training batch size.

[0125] S2-3. Input the enhanced sample set of the training clutter samples into the multi-terrain feature extraction detection network, and train the network based on the contrast loss function to construct clutter distribution clusters of different terrains. Input the enhanced sample into the network for feature extraction and feature mapping to obtain the final representation, and calculate the contrast loss based on the enhanced sample set. Train the network based on the loss function, and use the back propagation algorithm to update the weight parameters.

[0126] S3. Threshold determination and final decision.

[0127] Based on the distribution of training clutter samples in the feature space, the decision threshold is obtained by sorting the training score and taking the quantile. The final decision is made based on the relationship between the calculated score of the sample to be tested through the encoder and projection head network and the decision threshold.

[0128] S3-1. Detection of threshold determination:

[0129] The present invention adopts negative Mahalanobis distance -dM as the scoring function s, so for sample x, its final score s under the i-th type of terrain is i It can be expressed by the following formula:

[0130]

[0131] Where z is the feature vector obtained after sample x passes through the network, μ i and∑ i The mean and covariance matrices calculated for all training samples in the i-th class, for i, i∈I={1,2,…,C main}.

[0132] S(x) contains the score of each category of sample x. The present invention selects the category with the highest score as the final prediction label.

[0133] i*=argmax n∈I S n (x),

[0134]

[0135] In order to achieve target detection, the present invention needs to use the Mahalanobis distance to obtain the detection threshold, and the threshold of the i-th category containing samples is as follows:

[0136] First, all enhanced sample sets in the i-th class are sorted in ascending order of their scores to form S i :S i =[S i1 , S i2 , …, S iw-1 ]. Secondly, according to the adjustable parameter and the number of feature samples w, the value of the index n is set: n = round(w·(1-p)), where round(·) represents the rounding operator. Finally, by i Selecting the nth element in , we can get δ i :δ i =S in In the above, p is the rejection rate of the training sample, and its value ranges from 0 to 1.

[0137] In order to test the false alarm probability of 10-4 For the detection probability under this condition, the present invention sets p=0.9999.

[0138] S3-2. Final judgment of the sample to be tested.

[0139] For the test samples, the following decision rules are followed:

[0140]

[0141] Among them, S i* (x) represents the highest score of the test sample, δ i* is the threshold corresponding to the highest scoring class.

[0142] If the score of the test sample is greater than or equal to the threshold, it is classified as a clutter sample; conversely, if the score of the sample is less than the threshold, it is classified as a target sample.

[0143] At the same time, this embodiment uses the following data for further explanation:

[0144] The clutter data of two types of terrain, Gobi and highway, were used as training data, and two types of vehicle targets were used as test data to test its detection performance under different terrains. The dataset was collected using a broadband radar operating in the w-band frequency range. In addition, for the two types of vehicle targets, high-resolution range profile (HRRP) data of 8 different postures were collected under two different terrain conditions. Each posture corresponds to a different angle of rotation of the vehicle around its own axis, and the data were collected at equal intervals from 0 to 360 degrees;

[0145] Network parameter settings:

[0146] ResNet50 is used as the backbone network. Two data enhancement methods, cyclic shift and flipping, are used to obtain enhanced samples for each sample. This embodiment sets the batch size to 1024, α to 0.008, τ to 0.07, and λ to 1. At the same time, this embodiment uses the SGD optimizer, sets the learning rate to 0.075, the decay to 1e-6, the momentum to 0.9, and uses the cosine annealing scheduler to adjust the learning rate;

[0147] Preprocessing of experimental data:

[0148] In order to better verify the detection performance under different signal-to-noise ratio conditions, this embodiment preprocesses the collected vehicle target data. First, this embodiment uses target segmentation to obtain the target segment data required for the experiment. Then, the target HRRP segment peak power is used as the signal power, and the clutter HRRP segment average power is used as the clutter signal power. The signal-to-noise ratio definition method can be expressed as:

[0149]

[0150] Where N is the length of the range unit of the clutter segment, represents the target peak power, Indicates the power of the clutter segment. Finally, this embodiment generates target data with a signal-to-noise ratio of -10-20dB for testing the detection performance under different signal-to-noise ratio conditions. The specific experimental training and test data division is shown in the table.

[0151] Table 1 Experimental data description

[0152]

[0153] The final experiment obtained the target detection probability curves of vehicle 1 and vehicle 2 in the Gobi and highway terrains under -10-20dB. The results are as follows Figure 3-Figure 6 shown.

[0154] The comparison methods used in the experiment are the classic binary accumulation detection algorithm (M / N detection algorithm) and the generalized likelihood ratio detection algorithm (GLRT detection algorithm), etc. These detection algorithms have good effects in HRRP target detection.

[0155] From the detection results, the target detection curves of vehicle 1 and vehicle 2 in Gobi and highway terrains are better than those of the comparison method, which proves the effectiveness of the algorithm. In addition, when the detection probability reaches 90%, the signal-to-clutter ratio (SCR) required by the method proposed in the present invention is reduced by 1 to 4 dB compared with the GLRT detection algorithm.

[0156] It should be noted that the structure described in the present invention can be implemented in a variety of different forms and is not limited to the described embodiments. Any equivalent transformations made by ordinary technicians in this field using the contents of the present invention description and drawings, or directly or indirectly applied to other related technical fields, such as the loading and unloading of other items, are included in the protection scope of the present invention.

Claims

1. A radar target detection method based on contrastive learning under multi-terrain conditions, characterized by: The following steps are involved: S1. Obtain clutter samples and samples to be tested of measured polarization high-resolution radars of radar targets under multiple terrains, construct data sets and divide training sets and test sets after preprocessing, and then construct enhanced sample sets after data enhancement, and construct positive and negative sample pairs for subsequent model training and radar target detection performance testing; S2. Network training: S21, completing the construction of a multi-terrain feature extraction detection network by establishing an encoder network and a projection head network, which is used to map the extracted features into a feature space; S2-2, adding the category labels of the clutter samples in the training set as conditions to the calculation of the training loss function, assigning different temperature hyperparameters to samples of different categories, and then adjusting the temperature hyperparameters to improve the clustering effect, and integrating to obtain a loss function based on contrastive learning; S2-3, inputting the enhanced sample set of the training clutter samples into the network, training the network based on the contrast loss function, and constructing clutter distribution clusters of different terrains; S3. Based on the distribution of the training clutter samples in the feature space, the Mahalanobis distance is used to obtain the detection threshold, and the final decision is made according to the relationship between the calculated score of the sample to be tested through the encoder and the projection head network and the decision threshold.

2. The radar target detection method based on contrastive learning under multi-terrain conditions as claimed in claim 1, characterized in that: The preprocessing in step S1 includes but is not limited to segmentation and normalization operations.

3. The radar target detection method based on contrastive learning under multi-terrain conditions as claimed in claim 1, characterized in that: In step S1, the samples to be tested obtained under multiple terrains in several frames constitute a set of target data samples to be tested, and the clutter echo samples in several frames constitute a clutter data sample set. Since the target data sample set to be tested does not participate in the training process, the target data samples are all used as test samples, and the clutter data sample set is randomly divided into a training set and a test set in a ratio of 1:1 to form a training clutter sample set and a test clutter sample set.

4. The radar target detection method based on contrastive learning under multi-terrain conditions as claimed in claim 1, characterized in that: In step S1, the contrastive learning method is used to enhance the training clutter samples by two data enhancement methods, namely, cyclic shift and flipping, to form an enhanced sample set; the enhanced samples of the same clutter sample are defined as positive sample pairs, and the enhanced samples of different clutter samples are defined as negative sample pairs.

5. The radar target detection method based on contrastive learning under multi-terrain conditions as claimed in claim 1, characterized in that: In step S2-1, the encoder network uses the classic ResNet50 network for feature extraction, and the projection head network is composed of a fully connected layer for mapping the extracted features into the feature space; In step S2-2, the loss function introduces and adjusts the temperature hyperparameter so that the repulsive force of clutter samples under the same terrain in the feature space is reduced, and the repulsive force of clutter samples under different terrains is increased, so as to improve the clustering effect of clutter sample features under the same terrain; In step S2-3, the enhanced samples are input to the network for feature extraction and feature mapping to obtain the final representation, and the contrast loss is calculated based on the enhanced sample set. At the same time, the network is trained based on the loss function, and the network training adopts the back propagation algorithm to update the weight parameters.

6. The radar target detection method based on contrastive learning under multi-terrain conditions as claimed in claim 1, characterized in that: The step S3 includes using the Mahalanobis distance to obtain the detection threshold, and defining the final decision of the sample to be tested according to the following decision rule: Among them, S i* (x) represents the highest score of the test sample, δ i* is the threshold corresponding to the highest scoring category; If the score of the sample to be tested is greater than or equal to the threshold, it is classified as a clutter sample; on the contrary, if the score of the sample is less than the threshold, it is classified as a target sample.

Images