A self-calibration method for distributed radars

Abstract

The application discloses a self-calibration method for distributed radars in the field of radar detection, and comprises the following steps: S1, creating a self-calibration neural network model, including a feature extraction layer and a full connection layer, to extract corresponding feature data of a distance-angle two-dimensional graph and output a position estimation value according to the feature data; S2, constructing a distance-angle two-dimensional graph based on distributed radar detection data, and constructing a data set, and training the self-calibration network model by using the data set; S3, inputting a distance-angle two-dimensional graph of a radar to be calibrated into the trained self-calibration network model, and outputting a calibration parameter, so as to realize adaptive estimation and calibration of the distributed radars. The self-calibration method can quickly and adaptively estimate and calibrate the position of each radar, and improves the detection precision of the distributed radars.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar, in particular to a self-calibration method for a distributed radar. BACKGROUND

[0002] FMCW (Frequency Modulated Continuous Wave) radars have the advantages of simple structure, small size and easy use, and are widely used in object detection, action recognition, automatic driving and other fields. Due to policy and process restrictions, the transmit power of FMCW radars is often small, which limits the detection range of a single radar. For some targets with weak radar wave reflection ability, the detection effect is not good. A distributed radar system composed of multiple FMCW radars can increase the detection area and improve the target detection performance. The spatial positions of each radar node in the distributed radar system are different, and the data of each radar node cannot be directly utilized. Therefore, it is of great significance to calibrate the data of the distributed radar network.

[0003] Patent No. 202111586447.8 proposes a distributed radar network by connecting a control module to a communication bus in multiple radar transceivers and combining a clock signal to synchronize the data of each radar node. However, this method limits the arrangement and range of the distributed radar system and increases the hardware cost.

[0004] Patent No. 202110832937.5 proposes a distributed radar target positioning method by converting target position estimation into a nonlinear constraint optimization problem to solve the target position parameters. However, this method requires prior acquisition of the position parameters of different radar nodes, which is difficult to measure in some cases. SUMMARY

[0005] The present application provides a self-calibration method for a distributed radar to solve the problem of difficult calibration between distributed radar systems in the prior art, and realizes adaptive estimation and calibration of the position of each radar.

[0006] The present application provides a self-calibration method for a distributed radar, including the following steps:

[0007] S1: create a self-calibration neural network model for a distributed radar, the self-calibration neural network model including a feature extraction layer and a fully connected layer, the feature extraction layer being used to extract corresponding feature data of a range-angle two-dimensional graph and input to the fully connected layer, the fully connected layer being used to output a position estimation value according to the feature data;

[0008] S2: constructing a distance-angle two-dimensional graph based on distributed radar detection data, and constructing a data set, and training the self-calibration neural network model using the data set;

[0009] S3: inputting the distance-angle two-dimensional graph of the radar to be calibrated into the trained self-calibration neural network model, and the self-calibration neural network model outputs the calibration parameters to realize adaptive estimation and calibration of the distributed radar.

[0010] The above embodiment has the beneficial effect that a self-calibration neural network model is pre-established, the distance-angle two-dimensional graph of the radar to be calibrated is input into the trained self-calibration neural network model, the self-calibration neural network model quickly outputs the calibration parameters, and adaptive calibration of the distributed radar is realized. The self-calibration method can quickly and adaptively estimate and calibrate the position of each radar, and improve the detection accuracy of the distributed radar.

[0011] On the basis of the above embodiment, the present application can be further improved as follows:

[0012] In one embodiment of the present application, the feature extraction layer includes a convolution layer, a pooling layer, a batch normalization layer and an activation layer one, the full connection layer includes a flattening layer, a linear layer and an activation layer two, the convolution layer and the pooling layer are used to extract the feature data, the batch normalization layer is used to normalize the feature data, the activation layer one and the activation layer two are used to apply nonlinear features to the feature data, the flattening layer is used to convert the feature data into one-dimensional data, and the linear layer is used to perform linear transformation processing on the one-dimensional data. The distance-angle two-dimensional graph is extracted by the convolution layer and the pooling layer, then normalized by the batch normalization layer, and finally input into the full connection layer after nonlinear processing by the activation layer one. Finally, the flattening layer, the linear layer and the activation layer two are output to estimate the relative position of each radar.

[0013] In one embodiment of the present application, the convolution layer is composed of C convolution kernels, the size of each convolution kernel is N*N, the step size of each convolution kernel is S, the number of 0 padding used at the edge of the input data is P, and the size of the feature map after the convolution layer processing is

[0014] ;

[0015] wherein Input represents the size of the input data, Output represents the size of the output data, and the number of channels of the output data is C. The convolution layer is used to quickly and accurately extract the feature data corresponding to the distance-angle two-dimensional graph of each radar.

[0016] In one embodiment of the present application, the normalization processing formula of the batch normalization layer is as follows:

[0017] ;

[0018] ;

[0019] ;

[0020] wherein, X represents all samples input, X i represents one data in the sample, m represents the number of samples read by the neural network each time, E(X) represents the sample mean, Var(X) represents the sample variance, X input represents the input data of the batch normalization layer, X output represents the output data of the batch normalization layer, and ε is a set value. ε is a very small constant to avoid the denominator being 0. The batch normalization layer normalizes the input data by calculating the mean and variance of the input data to solve the problem of unstable values in the self-calibration network and improve the stability and convergence speed of the self-calibration network.

[0021] In one of the embodiments of the present application, the self-calibration neural network model comprises three feature extraction layers and one full connection layer. The output of the third feature extraction layer is used as the input of the full connection layer, thereby improving the feature extraction accuracy.

[0022] In one of the embodiments of the present application, in the step S2, the self-calibration neural network model is trained using the data set, and the training is specifically as follows:

[0023] The data set is divided into a training set and a validation set. Based on the training set, the loss function of the self-calibration neural network model is calculated, and the formula is as follows:

[0024] ;

[0025] wherein, Loss represents the loss function, output represents the calibrated parameter output by the neural network model after predicting the training set, the calibrated parameter is the distance between radars, label represents the true value of the training set, n represents the number of samples input into the network at one time, ∑ represents summation, and | | represents absolute value;

[0026] Based on the gradient back propagation algorithm and the loss function, the self-calibration network parameters are optimized, and the trained self-calibration neural network model is obtained after being verified by the validation set.

[0027] In one of the embodiments of the present application, in the step S2, the data set is constructed as follows:

[0028] The distances and angles of a plurality of target objects relative to a radar are determined respectively.

[0029] The distance and angle of each target relative to another radar are calculated respectively;

[0030] The distance and angle of each target relative to another radar are calculated respectively;

[0031] The distance and angle of each target relative to another radar are calculated respectively;

[0032] The distance and angle of each target relative to another radar are calculated respectively;

[0033] The distance and angle of each target relative to another radar are calculated respectively;

[0034] The distance and angle of each target relative to another radar are calculated respectively;

[0035] The distance and angle of each target relative to another radar are calculated respectively; BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0037] Figure 1 The flow of the self-calibration method for distributed radars provided by the present application;

[0038] Figure 2 The structure diagram of the self-calibration neural network model provided by the present application;

[0039] Figure 3 The distance-angle used in the example Figure 1 ;

[0040] Figure 4 The distance-angle used in the example Figure 2 ;

[0041] Figure 5 The schematic diagram of the fusion of the distance-angle diagrams of two different radars. DETAILED DESCRIPTION

[0042] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application. The present application provides a data fusion self-calibration method for distributed radar.

[0043] The present application provides a data fusion self-calibration method for distributed radar, and the radar device mainly comprises an antenna module, a radio frequency module and a data processing module. The antenna module is used for transmitting and receiving signals, the radio frequency module is used for realizing signal generation and frequency change, and the data processing module is used for collecting and processing the original signals output by the radio frequency module. The radar device provided by the present application does not need a synchronization and positioning ranging function module.

[0044] The present application provides a self-calibration method for distributed radar, Figure 1 is the overall implementation process of the method proposed in the present application.

[0045] In the implementation process of the present application, the following steps are included:

[0046] S1: create a self-calibration neural network model for distributed radar, the self-calibration neural network model comprises a feature extraction layer and a full connection layer, the feature extraction layer is used to extract corresponding feature data of the distance-angle two-dimensional graph and input to the full connection layer, and the full connection layer is used to output a position estimation value according to the feature data.

[0047] The self-calibration neural network model can be built through a neural network environment such as PyTorch and TensorFlow. The radar device is characterized by comprising functional modules such as an antenna module, a radio frequency module and a data processing module, and does not need signal synchronization and positioning ranging functions.

[0048] The self-calibration neural network model after parameter optimization can realize the prediction of the distance between radars by inputting the distance-angle graph of the radar. Assuming that the distance between radar A and radar B needs to be estimated at present, the parameter settings of radar A and B are the same, the size of the distance-angle graph of radar A and B is X*Y, X represents the number of sampling points of radar in the distance dimension, and Y represents the number of sampling points of radar in the angle dimension, then the data size of radar A and B input into the model should be N*1*X*Y, wherein N is the data size read by the model each time. First, the tensor formed by the distance-angle graph input by radar A and B is spliced together along the second dimension, the input data size after splicing should be N*2*X*Y, and the input data after splicing is used as the original data for subsequent processing of the model.

[0049] The feature extraction layer of the self-calibration neural network model extracts features from the original data to be processed, and the feature extraction layer includes a convolution layer, a pooling layer, a batch normalization layer, and an activation layer.

[0050] The convolution layer is composed of C convolution kernels, each convolution kernel has a size of N*N, the step size of each convolution kernel is S, the number of 0 padding used at the edge of the input data is P, and the size of the feature map after processing by the convolution layer is

[0051] (1);

[0052] In formula (1), Input represents the size of the input data, and Output represents the size of the output data. The number of channels of the output data is C.

[0053] The feature map processed by the convolution layer is again processed by the pooling layer to enhance the features and improve the understanding of the relative position information between the features by the model. The pooling layer is similar to the convolution layer, and each pooling layer is composed of a pooling kernel. Each pooling kernel moves on the feature map and processes the feature map according to a certain rule. In this example, the maximum pooling rule is selected, under which the pooling kernel retains the maximum value within its coverage range and discards other values. The size of the feature map processed by the pooling layer is calculated in the same way as the convolution layer, with the only difference being that the number of channels of the input data and the output data of the pooling layer is the same.

[0054] The feature map output by the pooling layer is input to the batch normalization layer. The batch normalization layer normalizes the input data by calculating the mean and variance of the input data to solve the problem of unstable values in the network and improve the stability and convergence speed of the network. The mean and variance are determined by the following formula

[0055] (2);

[0056] (3);

[0057] In formulas (2) and (3), X represents all samples input to the neural network each time, X i represents a certain data in all samples input to the neural network each time, m represents the number of samples read by the neural network each time, E(X) represents the mean of the samples, Var(X) represents the variance of the samples, and the output value of the feature map processed by the batch normalization layer is determined by the following formula

[0058] (4);

[0059] In formula (4), X input represents the input data of the batch normalization layer, and X outputThe output data of the batch normalization layer is represented as y, and ε is a set value. ε is a very small constant, for example, 0.0001, which is used to avoid the abnormality of the denominator being 0.

[0060] The activation layer applies a nonlinear feature to the feature map after the batch normalization processing, and in this example, a rectified linear unit (ReLU) is used as the activation function. The output of the activation layer is used as the result of the feature extraction layer. In this embodiment, the self-calibration neural network model includes three feature convolution layers and one fully connected layer. The output result of the third layer of the feature extraction layer is used as the input of the fully connected layer.

[0061] The fully connected layer is composed of a flattening layer, a linear layer, and an activation layer two. The flattening layer vectorizes the multidimensional input data into one-dimensional data. The linear layer performs a linear transformation on the input data, which can be represented by the following formula:

[0062] (5);

[0063] In formula (5), y represents the output of the linear layer, x represents the input of the linear layer, A represents a matrix composed of parameters of the linear layer, and T represents transposition, b represents bias, and is a learnable parameter. In this example, the activation layer two selects ReLU as the activation function. Figure 2 The network structure diagram of this example is shown in FIG. 2.

[0064] S2: Construct a distance-angle two-dimensional graph based on the distributed radar detection data, construct a data set, and use the data set to train the self-calibration neural network model.

[0065] The optimization of the network parameters is realized by the back propagation algorithm. The optimization goal of the network parameters is to reduce the difference between the predicted value and the true value of the self-calibration neural network model. In this example, the average absolute error is used to measure the distance between the predicted value and the true value, and the result is used as the loss function. The loss function of the average absolute error is calculated by the following formula:

[0066] (6);

[0067] In formula (6), n represents the number of samples read by the neural network each time, output represents the calibrated parameter output by the neural network model after predicting the training set, the calibrated parameter is the distance between the radars, label represents the true value, Loss represents the loss function, and | | represents the absolute value. In this example, the stochastic gradient descent method is used to optimize the network parameters.

[0068] In the present example, a dataset required for generating a self-calibration neural network model is generated by simulating echo signals of a radar detection target at two different positions, and the dataset is divided into a training set and a validation set. In the process of generating the dataset, first, the distance and angle of the target relative to a radar are determined, and then the distance and angle of the target relative to another radar are calculated. The reflection cross-sectional area of the target relative to different radars can be randomly generated, and the distance and angle of the target relative to different radars are calculated by the following formulas:

[0069] (7);

[0070] (8);

[0071] In formulas (7) and (8), R and represent the distance between the target and different radars, and and represent the angle between the target and different radars, and L represents the distance between the radars. A plurality of targets having different distances and angles relative to a radar are generated, and the distances and angles of these targets relative to another radar at different radar spacings are calculated by the formulas. All targets are superimposed to obtain a range-angle map of each radar. The range-angle map of different radars and the real distance between the radars corresponding thereto jointly constitute a dataset required for training the network.

[0072] The self-calibration neural network model is trained by the training set, and after the loss function of the neural network converges, the self-calibration neural network model is verified in combination with the validation set. After verification, the distance between different radars can be predicted. Figure 3 and Figure 4 is a range-angle map used in the present example. Figure 5 is a result after fusing the range-angle maps from two different radars.

[0073] S3: inputting the range-angle two-dimensional map of the radar to be calibrated into the trained self-calibration neural network model, and the self-calibration neural network model outputs a calibration parameter, realizing adaptive estimation and calibration of the distributed radars.

[0074] The parameter represents the distance between the radars, and in combination with the output of the self-calibration neural network model and formulas (7) and (8), the conversion of the distance and angle of the target relative to different radars can be realized, thereby realizing adaptive estimation and calibration of the distributed radars.

[0075] The following Table 1 is a comparison of the performance of existing papers and the method proposed in this patent. Method 1 is S. Li, J. Guo, R. Xi, C. Duan, Z. Zhai, Y. He, Pedestrian Trajectory based Calibration for Multi-Radar Network, IEEE INFOCOM 2021, May 2021. This method observes the movement trajectory of the test personnel in the overlapping observation area of two radars, simplifies the movement trajectory into a straight line, and calculates the difference in slope and position of the straight line generated by the same trajectory in different radars, to realize the calibration of the radar external parameters. Method 2 is S. Iwata, T. Koda and T. Sakamoto, Multi-radar Data Fusion for Respiratory Measurement of Multiple People, IEEE Sensors Journal, vol.21, no.22, pp.25870-25879, November, 2021. This method extracts the breathing characteristics of the test personnel to realize the positioning of the same target. The same target has different positions in different radars. The affine transformation of the same target between different radars is calculated by combining the Procrustes analysis to achieve the purpose of estimating the radar external parameters. Method 3 is A. Shastri, M. Canil, J. Pegoraro, P. Casari and M. Rossi, mmSCALE: Self-Calibration of mmWave Radar Networks from Human Movement Trajectories, 2022 IEEE Radar Conference, March, 2022. This method considers the time synchronization between different radar nodes, and introduces a time calibration error into the cost function, which realizes the automatic association of the trajectory of the moving target. These methods all need to observe moving targets for online calibration, which greatly limits the application scenarios.

[0076] Table 1 Comparison of performance of the method proposed in this patent and other methods

[0077]

[0078] As can be seen from Table 1, the self-calibration accuracy of the distributed radar system using the method proposed in this patent is significantly improved.

[0079] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. A self-calibration method for distributed radar, characterized in that, Includes the following steps: S1: Create a self-calibrating neural network model for a distributed radar. The self-calibrating neural network model includes three feature extraction layers and one fully connected layer. The feature extraction layers include a convolutional layer, a pooling layer, a batch normalization layer, and an activation layer one. The fully connected layer includes a flattening layer, a linear layer, and an activation layer two. The feature extraction layers are used to extract the corresponding feature data of the distance-angle two-dimensional map and input it into the fully connected layer. The fully connected layer is used to output a position estimate based on the feature data. S2: Construct the range-angle two-dimensional map based on distributed radar detection data, and build a dataset. Use the dataset to train the self-calibrating neural network model. The dataset is constructed as follows: The distance and angle of multiple targets relative to a single radar are determined separately; The distances and angles of the multiple targets relative to another radar were calculated respectively; The range and angle of all the targets are superimposed to obtain the range-angle map of each radar. The range-angle maps of different radars and the actual distances between them together constitute the dataset required to train the network; S3: Input the range-angle two-dimensional map of the radar to be calibrated into the trained self-calibration neural network model. The self-calibration neural network model outputs calibration parameters to realize adaptive estimation and calibration of the distributed radar.

2. The self-calibration method according to claim 1, characterized in that: The convolutional and pooling layers are used to extract the feature data, the batch normalization layer is used to normalize the feature data, the activation layer one and activation layer two are used to apply nonlinear features to the feature data, the flattening layer is used to convert the feature data into one-dimensional data, and the linear layer is used to perform linear transformation on the one-dimensional data.

3. The self-calibration method according to claim 2, characterized in that: The convolutional layer consists of C convolutional kernels, each with a size of N*N. The stride of each kernel movement is S, and the number of zeros padded at the edges of the input data is P. Therefore, the size of the feature map after processing by the convolutional layer is... ； Where Input represents the size of the input data, Output represents the size of the output data, and the number of channels for the output data is C.

4. The self-calibration method according to claim 3, characterized in that: The normalization formula for the batch normalization layer is as follows: ； ； ； Where X represents all input samples, X i Let X represent a data point in the sample, m represent the number of samples read by the neural network each time, E(X) represent the sample mean, Var(X) represent the sample variance, and X... input X represents the input data of the batch normalization layer. output This represents the output data of the batch normalization layer, where ε is a set value.

5. The self-calibration method according to claim 1, characterized in that: In step S2, the self-calibrating neural network model is trained using the dataset, as follows: The dataset is divided into a training set and a validation set. Based on the training set, the loss function of the self-calibrating neural network model is calculated as follows: ； Where Loss represents the loss function, output represents the calibration parameters output by the neural network model after predicting the training set, label represents the true value of the training set, n represents the number of samples input into the network at one time, ∑ represents summation, and || represents taking the absolute value; Then, based on the gradient backpropagation algorithm and the loss function, the parameters of the self-calibrating network are optimized, and the trained self-calibrating neural network model is obtained after verification by the validation set.

6. The self-calibration method according to claim 1, characterized in that: In step S2, the distance and angle of the target relative to different radars are calculated using the following formula: ； ； Where R and θ represent the distances between the target and different radars, θ and θ represent the angles between the target and different radars, and L represents the distances between radars.

Images