Similarity measurement detection method and device based on radar target feature vectors

Abstract

The invention provides a similarity measurement detection method and device based on a radar target feature vector, and belongs to the technical field of radar intelligent detection, and the method comprises the steps: inputting a radar target detection data set into an intelligent detection model, and obtaining a two-dimensional feature map and a one-dimensional feature vector; performing class center point calculation on the one-dimensional feature vector based on a target class to obtain a plurality of prototype vectors corresponding to different target classes, and storing the prototype vectors in a storage space; performing similarity measurement calculation on the two-dimensional feature map and the prototype vector of the storage space to obtain a key point thermodynamic diagram corresponding to each target category, and extracting a key point coordinate to obtain a prediction result of each target category; and calculating a loss function value between a prediction result and an actual result, and iteratively updating the model parameters until the detection accuracy of the training data set under the model parameters meets a set condition or the model reaches an iterative training round number, thereby obtaining a trained intelligent detection model. According to the invention, the identification degree of the target signal under the complex background can be improved.

Description

Technical Field

[0001] This invention belongs to the field of radar intelligent detection technology, and more specifically, it relates to a similarity measurement and detection method and device based on radar target feature vectors. Background Technology

[0002] With the increasing application of radar technology in fields such as national defense, meteorological observation, and autonomous driving, its working environment is becoming increasingly complex. Traditional detection and classification methods based on statistical theory and fixed thresholds experience a sharp decline in performance when dealing with highly non-stationary, nonlinear, and complex electromagnetic scenarios with insufficient prior knowledge. In recent years, artificial intelligence technologies, represented by deep learning, have effectively improved the performance of radar identification and detection systems in complex electromagnetic scenarios due to their powerful feature extraction and complex pattern recognition capabilities.

[0003] Current radar identification and detection systems utilize intelligent radar identification models with functions such as feature extraction, classification, and regression. After accurate and efficient feature extraction, the ability to quickly and effectively classify and decide on targets with limited data samples is crucial for achieving high-precision and high-accuracy radar detection. The powerful backbone network of the feature extraction module in current intelligent identification models can extract high-dimensional and rich feature information from training data. However, the classifier and regressor need to extract a small subset of features directly related to the category judgment from a massive amount of feature information. A large amount of irrelevant feature information interferes with the classification decision, leading to a decline in the detection performance of the intelligent model. Therefore, there is an urgent need for a detection method to address the insufficient ability of existing methods to finely distinguish features between radar target signals and environmental signals such as active / passive interference and clutter. Summary of the Invention

[0004] The purpose of this invention is to provide a similarity measurement and detection method and device based on radar target feature vectors, which can solve the problem that intelligent target detection systems have insufficient ability to finely identify the features between radar target signals and environmental signals such as active / passive interference and clutter in complex backgrounds.

[0005] A first aspect of this invention provides a similarity measurement and detection method based on radar target feature vectors, comprising:

[0006] The radar target detection dataset is input into the intelligent detection model to obtain a two-dimensional feature map and a one-dimensional feature vector. The radar target detection dataset is the training dataset, which includes data from multiple different target categories.

[0007] Based on the target category, the class center point is calculated on the one-dimensional feature vector to obtain the prototype vector of target feature information corresponding to multiple different target categories, and then stored in the storage space;

[0008] The similarity measurement calculation is performed between the two-dimensional feature map and the prototype vector of the target feature information corresponding to multiple different target categories in the storage space to obtain the key point heat map corresponding to each target category. The key point coordinates on the key point heat map are extracted to obtain the prediction result of each target category.

[0009] Calculate the loss function value between the predicted result and the actual result for each target category, and iteratively update the model parameters of the intelligent detection model based on the loss function value until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

[0010] In one embodiment, the process of constructing the intelligent detection model includes:

[0011] Two independent convolutional feature extraction network models are constructed as intelligent detection models;

[0012] The input to both convolutional feature extraction network models is a radar target detection dataset, which contains two-dimensional radar time-frequency domain signal processing images. The outputs of the two convolutional feature extraction network models are a two-dimensional feature map and a one-dimensional feature vector, respectively.

[0013] In one embodiment, the convolutional feature extraction network model includes a model structure, model initialization parameters, and training hyperparameters; both independent convolutional feature extraction network models are feature pyramid structures.

[0014] In one embodiment, class center points are calculated on a one-dimensional feature vector based on the target category to obtain prototype vectors of target feature information corresponding to multiple different target categories, including:

[0015] Determine the one-dimensional feature vector corresponding to each target category based on the target category;

[0016] One-dimensional feature vectors corresponding to each target category are extracted using a preset extraction algorithm to obtain prototype vectors of target feature information corresponding to multiple different target categories; each target category corresponds to one prototype vector.

[0017] In one embodiment, storing prototype vectors of target feature information corresponding to multiple different target categories into a storage space includes:

[0018] The number of storage channels is determined based on the number of target categories;

[0019] Convert the prototype vectors of target feature information corresponding to each target category into serialized data binary file format;

[0020] The prototype vectors of each binary file format are stored according to the number of storage channels.

[0021] In one embodiment, a similarity measurement is performed between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to obtain a keypoint heatmap corresponding to each target category, including:

[0022] For each target category:

[0023] The distance between the prototype vector of the target feature information corresponding to the target category and the feature vector of each pixel in the two-dimensional feature map is calculated using a similarity algorithm, and then processed based on the activation function to obtain the key point heatmap.

[0024] In one embodiment, each keypoint heatmap corresponds to a target category, and the value of the keypoint coordinates on the keypoint heatmap is the target confidence level of the target category corresponding to that coordinate;

[0025] Extract the coordinates of key points from the key point heatmap to obtain the prediction results for each target category, including:

[0026] For each keypoint heatmap, the keypoint coordinates on the keypoint heatmap are extracted based on the width and height prediction algorithm to obtain the target confidence, target category, target center point coordinates, and target width and height values.

[0027] In one embodiment, iteratively updating the model parameters of the intelligent detection model based on the loss function value includes:

[0028] The model parameters of the intelligent detection model are updated iteratively in reverse using a gradient descent-based method based on the loss function value.

[0029] In one embodiment, the similarity measurement and detection method based on radar target feature vectors further includes:

[0030] The performance status of the trained intelligent detection model is described based on the test dataset. The performance status description metrics include recall and precision.

[0031] A second aspect of the present invention provides a similarity measurement and detection device based on radar target feature vectors, comprising:

[0032] The feature extraction module is used to input the radar target detection dataset into the intelligent detection model to obtain two-dimensional feature maps and one-dimensional feature vectors. The radar target detection dataset is the training dataset, which includes data from multiple different target categories.

[0033] The vector calculation module is used to calculate the class center point of a one-dimensional feature vector based on the target category, obtain the prototype vector of target feature information corresponding to multiple different target categories, and store it in the storage space;

[0034] The result prediction module is used to perform similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space, to obtain the key point heatmap corresponding to each target category, extract the key point coordinates on the key point heatmap, and obtain the prediction result for each target category.

[0035] The model training module is used to calculate the loss function value between the predicted results and the actual results for each target category. Based on the loss function value, the model parameters of the intelligent detection model are iteratively updated until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

[0036] In one embodiment, the feature extraction module is specifically used for:

[0037] Two independent convolutional feature extraction network models are constructed as intelligent detection models;

[0038] The input to both convolutional feature extraction network models is a radar target detection dataset, which contains two-dimensional radar time-frequency domain signal processing images. The outputs of the two convolutional feature extraction network models are a two-dimensional feature map and a one-dimensional feature vector, respectively.

[0039] In one embodiment, the convolutional feature extraction network model includes a model structure, model initialization parameters, and training hyperparameters; both independent convolutional feature extraction network models are feature pyramid structures.

[0040] In one embodiment, the vector calculation module is specifically used for:

[0041] Determine the one-dimensional feature vector corresponding to each target category based on the target category;

[0042] One-dimensional feature vectors corresponding to each target category are extracted using a preset extraction algorithm to obtain prototype vectors of target feature information corresponding to multiple different target categories; each target category corresponds to one prototype vector.

[0043] In one embodiment, the vector calculation module is specifically used for:

[0044] The number of storage channels is determined based on the number of target categories;

[0045] Convert the prototype vectors of target feature information corresponding to each target category into serialized data binary file format;

[0046] The prototype vectors of each binary file format are stored according to the number of storage channels.

[0047] In one embodiment, the result prediction module is specifically used for:

[0048] For each target category:

[0049] The distance between the prototype vector of the target feature information corresponding to the target category and the feature vector of each pixel in the two-dimensional feature map is calculated using a similarity algorithm, and then processed based on the activation function to obtain the key point heatmap.

[0050] In one embodiment, each keypoint heatmap corresponds to a target category, and the value of the keypoint coordinates on the keypoint heatmap is the target confidence level of the target category corresponding to that coordinate;

[0051] The result prediction module is specifically used for:

[0052] For each keypoint heatmap, the keypoint coordinates on the keypoint heatmap are extracted based on the width and height prediction algorithm to obtain the target confidence, target category, target center point coordinates, and target width and height values.

[0053] In one embodiment, the model training module is specifically used for:

[0054] The model parameters of the intelligent detection model are updated iteratively in reverse using a gradient descent-based method based on the loss function value.

[0055] In one embodiment, the similarity measurement and detection device based on radar target feature vectors further includes a performance detection module, which is used to describe the performance status of the trained intelligent detection model based on a test dataset. The performance status description metrics include recall and precision.

[0056] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described similarity measurement detection method based on radar target feature vectors.

[0057] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described similarity measurement and detection method based on radar target feature vectors.

[0058] The beneficial effects of the similarity measurement and detection method and apparatus based on radar target feature vectors provided in this invention are as follows:

[0059] This invention inputs a radar target detection dataset into an intelligent detection model's feature extraction network to obtain a two-dimensional feature map and a one-dimensional feature vector. Based on the target category, class center points are calculated on the one-dimensional feature vector to obtain prototype vectors of target feature information corresponding to multiple different target categories, which are then stored in the storage space. By introducing category prototype vectors with saliency features during the data classification process, this invention significantly increases the amount of input information to the intelligent model's classification module. This effectively improves the precision and recall of radar target detection under various interference backgrounds, supporting the realization of intelligent classification, recognition, and localization detection tasks for radar targets in complex scenarios with high performance requirements. This invention also performs similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to obtain key point heatmaps for each target category. Key point coordinates are extracted from the key point heatmaps to obtain prediction results for each target category. The loss function value between the predicted and actual results for each target category is calculated. Based on the loss function value, the model parameters of the intelligent detection model feature extraction network model and the intelligent detection model are iteratively updated until the detection accuracy of the training dataset under the network model parameters meets the set conditions or the model reaches the preset number of iterations. This yields the trained intelligent detection model feature extraction network model and the intelligent detection model. This invention designs, optimizes, and updates the intelligent detection model based on training data, test data, and validation data, which can greatly improve its ability to identify radar target signals in complex backgrounds. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating a similarity measurement and detection method based on radar target feature vectors according to an embodiment of the present invention.

[0062] Figure 2 This is a structural block diagram of a similarity measurement and detection device based on radar target feature vectors provided in an embodiment of the present invention.

[0063] Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0064] Existing radar intelligent detection signal processing and recognition models include three modules: a feature extraction module, a classifier, and a regressor. The model first extracts features from the input image using the feature extraction module, then feeds these features into the classifier and regressor for category classification and localization, completing the detection. The feature extraction module utilizes target characteristics to extract features from the radar signal that help determine the target category and type in the test set. The feature extraction methods differ for different target ranges and signal types. The classifier and regressor are steps that classify unknown targets based on the similarity between features in the training data and features in the test data, and output prediction results.

[0065] In complex electromagnetic scenarios, current intelligent detection models still suffer from low efficiency in extracting effective information from high-dimensional features. While the powerful backbone network of the feature extraction module in current intelligent models can extract high-dimensional and rich feature information from training data, classifiers and regressors need to extract a small subset of features directly related to class judgment from massive amounts of feature information. A large amount of irrelevant feature information interferes with classification decisions, leading to a decline in the detection performance of intelligent models. Current classic classification and regressor optimization designs still lack supervised utilization of structured prior knowledge, failing to leverage knowledge with key class-specific information to guide the network parameter learning process, resulting in insufficient information utilization efficiency of intelligent models.

[0066] Based on the above, this invention provides a similarity measurement and detection method and apparatus based on radar target feature vectors, which can effectively solve the above problems and improve the model's ability to identify radar target signals in complex backgrounds.

[0067] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0068] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0069] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a similarity measurement and detection method based on radar target feature vectors according to an embodiment of the present invention. The method may include steps S102 to S108.

[0070] S102: Input the radar target detection dataset into the intelligent detection model to obtain a two-dimensional feature map and a one-dimensional feature vector. The radar target detection dataset is the training dataset, which includes data from multiple different target categories.

[0071] In this embodiment, the collected radar data is first divided into training dataset, validation dataset and test dataset according to the proportion. The radar data is a two-dimensional radar signal time-frequency image after time-frequency analysis signal processing. All radar data has the target category number id, the center point coordinates (x,y) of the target signal on the data, the width w of the target signal on the data, and the height label information h.

[0072] In this paper, the radar target detection dataset specifically refers to the training dataset, which can be divided in a 6:2:2 ratio. The training dataset is used to calculate the loss function value and support the model parameter update process; the validation dataset is used to determine whether to stop the model parameter update process; and the test dataset is used to describe the performance status of the model. In this embodiment, the radar target detection dataset includes data from multiple different target categories, such as data numbered 1, 2, ..., 38.

[0073] In one embodiment, the construction process of the intelligent detection model feature extraction network model includes:

[0074] Two independent convolutional feature extraction network models are constructed as intelligent detection models. The input of both convolutional feature extraction network models is a radar target detection dataset, in which the data is a two-dimensional radar time-frequency domain signal processing image. The outputs of the two convolutional feature extraction network models are a two-dimensional feature map and a one-dimensional feature vector, respectively.

[0075] In one embodiment, the convolutional feature extraction network model includes a model structure, model initialization parameters, and training hyperparameters; both independent convolutional feature extraction network models are feature pyramid structures.

[0076] In this embodiment, the one-dimensional feature vector output by the convolutional feature extraction network model corresponds one-to-one with the input radar target detection data, and is the representation vector of the data in the one-dimensional feature space; the output two-dimensional feature map corresponds one-to-one with the input radar target detection data, and its size is the same as that of the radar target detection data, and is the representation of the data in the two-dimensional feature space.

[0077] S104: Calculate the class center point of the one-dimensional feature vector based on the target category to obtain the prototype vector of target feature information corresponding to multiple different target categories, and store it in the storage space.

[0078] In this embodiment, the prototype vector is, in principle, the central position of the distribution of data of the same target category in the feature space; the prototype vector is calculated using the intelligent extraction method of radar time-frequency domain data feature prototype vector based on the one-dimensional feature vector corresponding to data with the same target category ID. Each target category has one prototype vector. If the radar target detection dataset has N... C If there are categories, then the prototype vector P is In one embodiment, class center points are calculated on a one-dimensional feature vector based on the target category to obtain prototype vectors of target feature information corresponding to multiple different target categories, including:

[0079] Determine the one-dimensional feature vector corresponding to each target category based on the target category;

[0080] One-dimensional feature vectors corresponding to each target category are extracted using a preset extraction algorithm to obtain prototype vectors of target feature information corresponding to multiple different target categories; each target category corresponds to one prototype vector.

[0081] In this embodiment, the preset extraction algorithm is the intelligent extraction method of radar time-frequency domain data feature prototype vector.

[0082] One specific implementation of S104 can be:

[0083] Assume the training dataset is N S y is the number of samples in the training dataset. i For the training dataset, sample x i If the category label is given, then the prototype vector c of the k-th target category is... k The calculation formula is:

[0084]

[0085] Among them, f φ (·) represents the convolutional feature extraction network model, S k For the label to satisfy y i =k,k∈[1,N c The sample set N Sk This represents the number of samples in the training dataset with category ID k.

[0086] The above formula can be used to obtain the prototype vector corresponding to each target category, thereby enabling the extraction and summarization of category-specific features.

[0087] In one embodiment, storing prototype vectors of target feature information corresponding to multiple different target categories into a storage space includes:

[0088] The number of storage channels is determined based on the number of target categories;

[0089] Convert the prototype vectors of target feature information corresponding to each target category into serialized data binary file format;

[0090] The prototype vectors of each binary file format are stored according to the number of storage channels.

[0091] In this embodiment, the number of target categories is equal to the number of storage channels. The storage format of the processed prototype vector P is (C,1,L), where C represents the number of channels and L represents the vector length. In this embodiment, the prototype vectors corresponding to each target category are stored by channel, and the processed prototype vector P is stored in the serialized data binary file format specific to the PyTorch framework.

[0092] S106: Perform similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to obtain the key point heatmaps corresponding to each target category. Extract the key point coordinates on the key point heatmaps to obtain the prediction results for each target category.

[0093] In one embodiment, a similarity measurement is performed between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to obtain a keypoint heatmap corresponding to each target category, including:

[0094] For each target category:

[0095] S110. Calculate the distance between the prototype vector of the target feature information corresponding to the target category and the feature vector of each pixel in the two-dimensional feature map using a similarity algorithm, and then process it based on the activation function to obtain the key point heatmap.

[0096] In this embodiment, the similarity algorithm can be the cosine similarity algorithm. The number of keypoint heatmap channels is consistent with the number of target categories, that is, each target category corresponds to one keypoint heatmap, and the size of the keypoint heatmap on each channel is consistent with the size of the radar target detection data; the value at each coordinate of the keypoint heatmap is the confidence level of the presence of a target with the corresponding category number at that coordinate, which can be used as the prediction result for each target category.

[0097] Specifically, one possible implementation of S110 is as follows:

[0098] S111, Determine the two-dimensional feature map F;

[0099] S112. Read in the prototype vector P, and extract the prototype vector set sequentially according to the category number id in the first dimension.

[0100] S113. Use a similarity algorithm for calculation, such as using cosine similarity to calculate the prototype vector c for each category. k The feature vector F of each pixel in the two-dimensional feature map sample x,y The distances between them are processed by the softmax activation function to output the heatmap images of each channel of the keypoint heatmap. The formula for calculating cosine similarity is:

[0101]

[0102] Where v is a one-dimensional vector data, and ||||2 represents the vector L2 norm.

[0103] Key point heatmap heatmap images of each channel The thermal value at position (x, y) on channel k is:

[0104]

[0105] Considering that the gain generated by the cosine similarity distance is relatively small, the cosine similarity distance is multiplied by a factor α to improve the similarity guidance effect. The default value of α is 20. Heatmap of key points The thermal value at position (x,y) on channel k.

[0106] S114, the calculated result Arranged by channel, the number of channels is N. C Key point heatmap

[0107] Following steps S111-S114, a heatmap of key points reflecting the confidence level of the current category target at each location in the input data can be obtained. By introducing category feature supervision, the radar intelligent detection model is guided to further strengthen its attention to the key features of the category during the detection and classification process, thereby enhancing the information utilization rate of the intelligent model and improving the target classification and positioning capabilities.

[0108] In one embodiment, each keypoint heatmap corresponds to a target category, and the value of the keypoint coordinates on the keypoint heatmap is the target confidence level of the target category corresponding to that coordinate;

[0109] Extract the coordinates of key points from the key point heatmap to obtain the prediction results for each target category, including:

[0110] S120. For each key point heatmap, extract the key point coordinates on the key point heatmap based on the width and height prediction algorithm to obtain the target confidence, target category, target center point coordinates and target width and height values.

[0111] Specifically, one possible implementation of S120 is as follows:

[0112] S121, Input key point heatmap Extract the heatmap image of each channel separately. The peak point is used as the predicted center point location for target category k.

[0113] S122. Continue to refine the keypoint heatmap using a width and height prediction algorithm based on the keypoint heatmap. The processing begins with analyzing the heatmap images of each channel. Perform a 3×3 max pooling layer operation, and set a threshold for width and height prediction based on the state of the radar target detection dataset. Retain points near the center key point that are greater than or equal to the threshold. Finally, extract the key point distribution and output the width and height prediction results for target of category k.

[0114] S123. Combine the outputs of S121 and S122 to output the target location prediction result. The target location prediction result includes the target confidence level, target category, target center point coordinates, and target width and height values.

[0115] S108: Calculate the loss function value between the predicted result and the actual result for each target category, and iteratively update the model parameters of the intelligent detection model based on the loss function value until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

[0116] In one embodiment, iteratively updating the model parameters of the intelligent detection model based on the loss function value includes:

[0117] The model parameters of the intelligent detection model are updated iteratively in reverse using a gradient descent-based method based on the loss function value.

[0118] Specifically, one possible implementation of S108 is as follows:

[0119] S131. Calculate the loss function value between the predicted result and the actual result for each target category. The calculation formula is as follows:

[0120] L(p,p * ,t,t * ) = L cls (p,p * )+l×L reg (t,t * ,p * );

[0121] L cls (p,p * )=-[p×log p* +(1-p)log(1-p * )];

[0122] L reg (t,t * ,p * ) = p * ×Smooth L1(tt * );

[0123]

[0124] Where p is the confidence level of the model output that the current predicted target is of this type; p * The true label of the current target, and satisfying p * ∈{0,1}, set to 1 if the current target belongs to this category, otherwise set to 0; t={t x ,t y ,t w ,t h} represents the center point coordinates and width / height values ​​of the currently predicted target output by the model, t * ={t * x ,t * y ,t * w ,t * h} represents the true center point coordinates and width / height values ​​of the current target.

[0125] S132. Based on the loss function value, the network model parameters are updated in reverse using the gradient descent method based on adaptive moment estimation. The update rules of the adaptive moment estimation algorithm are as follows:

[0126]

[0127] Where, θ t+1 Let θ represent the parameters to be optimized at time t+1. t This represents the parameter to be optimized at time t. Let represent the first moment and the second moment, respectively; η represent the learning rate; and ε is a positive constant used to avoid cases where the denominator is zero.

[0128] S133. After each iteration of parameter update, input validation data. When the detection accuracy of the validation data under the model parameters does not increase after the set number of iterations, or the detection accuracy reaches the preset value, or the set number of iterations is reached, stop the parameter iteration update training process. The set number of iterations is determined by manual experience or by analyzing the results of repeated experiments.

[0129] In this embodiment, the detection accuracy meets a set condition, such as a detection accuracy greater than 90%. The model reaches a preset number of training iterations, such as 50 iterations. Model training ends only when the detection accuracy of the training dataset under the model parameters of the network model meets the set condition or the model reaches the preset number of training iterations.

[0130] In one embodiment, the similarity measurement and detection method based on radar target feature vectors further includes:

[0131] S109. Describe the performance status of the trained intelligent detection model based on the test dataset. The performance status description metrics include recall and precision.

[0132] In this embodiment, the trained model also needs to undergo model performance status analysis.

[0133] Specifically, one possible implementation of S109 is as follows:

[0134] S141. Input the test dataset and output the target position prediction result of the model after the parameter iteration update training. Use the test data to describe the model performance status.

[0135] S142. Draw the detection boxes of the model's predicted target location and the target's true label. Calculate the Intersection over Union (IoU) between the predicted target detection boxes and the true label detection boxes. The formula for calculating the IoU is:

[0136]

[0137] Where S1 and S2 represent the areas of the intersection and difference regions of the predicted detection box and the real label detection box, respectively.

[0138] S143. Use recall and precision metrics to describe the performance status of the model.

[0139] The formulas for recall (P) and precision (R) are as follows:

[0140]

[0141] Wherein, TP represents the number of correct predictions, that is, the number of prediction detection boxes with an IoU greater than or equal to a specified threshold; FP represents incorrect localizations, that is, the number of prediction detection boxes with an IoU less than a specified threshold; FN represents the number of undetected targets; and TN represents the correct prediction of negative samples, with the IoU threshold set to 0.5.

[0142] In summary, this embodiment of the invention divides the radar target detection dataset into training, testing, and validation datasets proportionally; it constructs two independent convolutional feature extraction networks to obtain two-dimensional feature maps and one-dimensional feature vectors of the data; it calculates class centers for the one-dimensional feature vectors of targets according to their categories to obtain prototype vectors of the targets; it stores the prototype vectors corresponding to each category of targets; it performs similarity measurement calculations on the two-dimensional feature maps of the data and the stored prototype vectors of different target categories to obtain key point heatmaps for each category; it extracts the key point coordinates from the key point heatmaps to obtain the position prediction results for each category of targets; it calculates the loss function value between the position prediction results of each category of targets and the true labels of the data, and uses a reverse iterative update method to update the network model parameter values ​​until the set conditions are met; it inputs the updated convolutional feature extraction network with test data, and uses the stored prototype vectors of different target categories to perform similarity measurement calculations to finally obtain the target detection results, and calculates the precision and recall of the trained model. This invention addresses the problem of insufficient fine-grained feature identification capability of intelligent target detection models in complex backgrounds when distinguishing between radar target signals and environmental signals such as active / passive interference and clutter. By adding metric guidance and introducing category prototype vectors with saliency features in the data classification process, the amount of input information to the intelligent model's classification module is significantly increased. This effectively improves the precision and recall of radar target detection under various interference backgrounds, supporting the realization of intelligent classification, recognition, and localization detection tasks for radar targets with high performance requirements in complex scenarios.

[0143] Corresponding to the similarity measurement and detection method based on radar target feature vectors in the above embodiment, Figure 2 This is a structural block diagram of a radar target feature vector-based similarity measurement and detection device according to an embodiment of the present invention. For ease of explanation, only the parts relevant to the embodiment of the present invention are shown. References Figure 2 The similarity measurement and detection device 20 based on radar target feature vectors includes: a feature extraction module 21, a vector calculation module 22, a result prediction module 23, and a model training module 24.

[0144] Feature extraction module 21 is used to input the radar target detection dataset into the intelligent detection model to obtain two-dimensional feature maps and one-dimensional feature vectors. The radar target detection dataset is a training dataset, which includes data of multiple different target categories.

[0145] The vector calculation module 22 is used to calculate the class center point of the one-dimensional feature vector based on the target category, obtain the prototype vector of the target feature information corresponding to multiple different target categories, and store it in the storage space;

[0146] The result prediction module 23 is used to perform similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space, to obtain the key point heatmaps corresponding to each target category, extract the key point coordinates on the key point heatmaps, and obtain the prediction results for each target category.

[0147] The model training module 24 is used to calculate the loss function value between the predicted results and the actual results of each target category, and iteratively update the model parameters of the intelligent detection model based on the loss function value until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

[0148] In one embodiment, the feature extraction module 21 is specifically used for:

[0149] Two independent convolutional feature extraction network models are constructed as intelligent detection models;

[0150] The input to both convolutional feature extraction network models is a radar target detection dataset, which contains two-dimensional radar time-frequency domain signal processing images. The outputs of the two convolutional feature extraction network models are a two-dimensional feature map and a one-dimensional feature vector, respectively.

[0151] In one embodiment, the convolutional feature extraction network model includes a model structure, model initialization parameters, and training hyperparameters; both independent convolutional feature extraction network models are feature pyramid structures.

[0152] In one embodiment, the vector calculation module 22 is specifically used for:

[0153] Determine the one-dimensional feature vector corresponding to each target category based on the target category;

[0154] One-dimensional feature vectors corresponding to each target category are extracted using a preset extraction algorithm to obtain prototype vectors of target feature information corresponding to multiple different target categories; each target category corresponds to one prototype vector.

[0155] In one embodiment, the vector calculation module 22 is specifically used for:

[0156] The number of storage channels is determined based on the number of target categories;

[0157] Convert the prototype vectors of target feature information corresponding to each target category into serialized data binary file format;

[0158] The prototype vectors of each binary file format are stored according to the number of storage channels.

[0159] In one embodiment, the result prediction module 23 is specifically used for:

[0160] For each target category:

[0161] The distance between the prototype vector of the target feature information corresponding to the target category and the feature vector of each pixel in the two-dimensional feature map is calculated using a similarity algorithm, and then processed based on the activation function to obtain the key point heatmap.

[0162] In one embodiment, each keypoint heatmap corresponds to a target category, and the value of the keypoint coordinates on the keypoint heatmap is the target confidence level of the target category corresponding to that coordinate;

[0163] The result prediction module 23 is specifically used for:

[0164] For each keypoint heatmap, the keypoint coordinates on the keypoint heatmap are extracted based on the width and height prediction algorithm to obtain the target confidence, target category, target center point coordinates, and target width and height values.

[0165] In one embodiment, the model training module 24 is specifically used for:

[0166] The model parameters of the intelligent detection model are updated iteratively in reverse using a gradient descent-based method based on the loss function value.

[0167] In one embodiment, the similarity measurement and detection device 20 based on radar target feature vectors further includes a performance detection module, which is used to describe the performance status of the trained intelligent detection model based on a test dataset. The performance status description metrics include recall and precision.

[0168] In summary, this embodiment innovatively incorporates a category prototype knowledge guidance mechanism with saliency features into the general radar intelligent target detection network model. It uses a similarity measurement method to slide the prototype vector on the target's two-dimensional feature map to measure similarity and outputs a key point heatmap. This can guide the radar intelligent detection model to further strengthen its focus on key category features during the detection and classification process, enhance the information utilization rate of the intelligent model, and improve the target classification and localization capabilities.

[0169] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of the present invention. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the feature extraction module 21, vector calculation module 22, result prediction module 23, and model training module 24 are shown.

[0170] It should be understood that, in this embodiment of the invention, the processor 301 may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0171] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0172] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory.

[0173] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of the present invention can execute the implementation method described in the similarity measurement and detection method based on radar target feature vectors provided in the embodiments of the present invention, or they can execute the implementation method of the electronic device described in the embodiments of the present invention, which will not be repeated here.

[0174] In another embodiment of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described in the above embodiments. Alternatively, the computer program can instruct related hardware to implement these processes. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0175] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0176] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0178] In the several embodiments provided by this invention, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or modules, or it may be an electrical, mechanical, or other form of connection.

[0179] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the objectives of the embodiments of the present invention, depending on actual needs.

[0180] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0181] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A similarity measurement and detection method based on radar target feature vectors, characterized in that, include: The radar target detection dataset is input into the intelligent detection model to obtain a two-dimensional feature map and a one-dimensional feature vector. The radar target detection dataset is a training dataset, which includes data of multiple different target categories. Based on the target category, the class center point of the one-dimensional feature vector is calculated to obtain the prototype vector of target feature information corresponding to multiple different target categories, and stored in the storage space; The two-dimensional feature map is compared with the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to calculate similarity, so as to obtain the key point heat map corresponding to each target category. The key point coordinates on the key point heat map are extracted to obtain the prediction results of each target category. Calculate the loss function value between the predicted result and the actual result for each target category, and iteratively update the model parameters of the intelligent detection model based on the loss function value until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

2. The method as described in claim 1, characterized in that, The construction process of the intelligent detection model includes: Two independent convolutional feature extraction network models are constructed as intelligent detection models; The inputs to both convolutional feature extraction network models are radar target detection datasets, which contain two-dimensional radar time-frequency domain signal processing images; the outputs of the two convolutional feature extraction network models are two-dimensional feature maps and one-dimensional feature vectors, respectively.

3. The method as described in claim 2, characterized in that, The convolutional feature extraction network model includes a model structure, model initialization parameters, and training hyperparameters; both of the two independent convolutional feature extraction network models are feature pyramid structures.

4. The method as described in claim 1, characterized in that, The calculation of class centers based on the target category of the one-dimensional feature vector yields prototype vectors of target feature information corresponding to multiple different target categories, including: Determine the one-dimensional feature vector corresponding to each target category based on the target category; One-dimensional feature vectors corresponding to each target category are extracted using a preset extraction algorithm to obtain prototype vectors of target feature information corresponding to multiple different target categories; each target category corresponds to one prototype vector.

5. The method as described in claim 1, characterized in that, The step of storing prototype vectors of target feature information corresponding to multiple different target categories into storage space includes: The number of storage channels is determined based on the number of target categories; Convert the prototype vectors of target feature information corresponding to each target category into serialized data binary file format; The prototype vectors of each binary file format are stored according to the number of storage channels.

6. The method as described in claim 1, characterized in that, The step of performing similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space to obtain key point heatmaps corresponding to each target category includes: For each target category: The distance between the prototype vector of the target feature information corresponding to the target category and the feature vector of each pixel in the two-dimensional feature map is calculated using a similarity algorithm, and then processed based on an activation function to obtain a key point heatmap.

7. The method as described in claim 1, characterized in that, Each key point heatmap corresponds to a target category, and the value of the key point coordinates on the key point heatmap is the target confidence level of the target category corresponding to that coordinate; The step of extracting the coordinates of key points on the key point heatmap to obtain prediction results for each target category includes: For each key point heatmap, the key point coordinates on the key point heatmap are extracted based on the width and height prediction algorithm to obtain the target confidence, target category, target center point coordinates and target width and height values.

8. The method as described in claim 1, characterized in that, The iterative update of the model parameters of the intelligent detection model based on the loss function value includes: Based on the loss function value, the model parameters of the intelligent detection model are updated iteratively using a gradient descent-based method.

9. The method as described in claim 1, characterized in that, Also includes: The performance status of the trained intelligent detection model is described based on the test dataset, and the performance status description metrics include recall and precision.

10. A similarity measurement and detection device based on radar target feature vectors, characterized in that, include: The feature extraction module is used to input the radar target detection dataset into the intelligent detection model to obtain a two-dimensional feature map and a one-dimensional feature vector. The radar target detection dataset is a training dataset, which includes data of multiple different target categories. The vector calculation module is used to calculate the class center point of the one-dimensional feature vector based on the target category, obtain the prototype vector of target feature information corresponding to multiple different target categories, and store it in the storage space; The result prediction module is used to perform similarity measurement calculations between the two-dimensional feature map and the prototype vectors of target feature information corresponding to multiple different target categories in the storage space, to obtain the key point heatmaps corresponding to each target category, extract the key point coordinates on the key point heatmaps, and obtain the prediction results for each target category. The model training module is used to calculate the loss function value between the predicted results and the actual results of each target category, and iteratively update the model parameters of the intelligent detection model based on the loss function value until the detection accuracy of the training dataset under the model parameters of the network model meets the set conditions or the model reaches the preset number of iteration training rounds, thus obtaining the trained intelligent detection model.

Images