A vehicle-mounted millimeter-wave radar target classification method, device, system and storage medium

Abstract

The present invention discloses a vehicle-mounted millimeter-wave radar target classification method, device, system, and storage medium. This method belongs to the field of radar target classification technology. The method comprises: performing vector mean cancellation on pre-acquired radar target echo signals to filter out stationary clutter; performing fast Fourier transforms (FFTs) on the radar target echo signals that have been filtered out of stationary clutter in both the velocity and range dimensions to obtain a radar target range-Doppler image; and inputting the radar target range-Doppler image into a pre-built lightweight target classification model for feature extraction and target classification to obtain a target classification result. The lightweight target classification model is constructed based on a pre-acquired Ghost module and a MobileNetV3 network model. This method can eliminate interference from stationary clutter on target signals, reducing network complexity while ensuring classification accuracy.

Description

Technical Field

[0001] The present invention relates to a vehicle-mounted millimeter-wave radar target classification method, device, system and storage medium, and belongs to the technical field of radar target classification. Background Art

[0002] Traffic safety issues have always accompanied the development of the automotive industry. Advanced Driver Assistance Systems (ADAS) are a key means of reducing traffic accidents. Millimeter-wave radar, with its 24 / 7 operation, is a core sensor for ADAS, and target classification technology is a key component of ADAS. In recent years, convolutional neural networks (CNNs) have been widely used, and their powerful image feature extraction capabilities provide insights for radar target classification.

[0003] Different targets, such as pedestrians, bicycles, and cars, have different motion patterns and Doppler characteristics. Radar target classification can be achieved by converting target echo signals into radar RD images and using neural networks to extract and classify RD image features. For example, the paper "Single-Frame Vulnerable Road Users Classification with a 77GHz FMCW Radar Sensor and a Convolutional Neural Network" (2018 19th International Radar Symposium, 2018, pp. 1-10, doi:10.23919 / IRS.2018.8448126.) proposes a radar signal processing chain for target classification. This replaces the traditional radar target classification algorithm with a CNN, enabling classification of single-frame RD images. However, this approach suffers from signal preprocessing limitations, failing to eliminate the effects of stationary objects on the target's Doppler or micro-Doppler characteristics. For slow targets, stationary clutter can interfere with them and even obscure their micro-Doppler characteristics, resulting in low classification accuracy. The paper "Human Detection Based on Time-Varying Signature on Range-Doppler Diagram Using Deep Neural Networks" (IEEE Geoscience and Remote Sensing Letters, vol. 18, no. 3, pp. 426-430, March 2021, doi: 10.1109 / LGRS.2020.2980320.) uses the AlexNet+DRNN network structure to classify multi-frame RD images, which can achieve high classification accuracy. However, storing the model parameter files requires a large amount of memory space, and running the model requires a large amount of memory resources, which is not suitable for in-vehicle mobile devices.

[0004] Therefore, how to eliminate the interference of stationary clutter on the target signal and design a lightweight network suitable for radar RD image classification while ensuring the classification accuracy and reducing the network complexity is a key problem that must be solved. Summary of the Invention

[0005] The purpose of the present invention is to provide a vehicle-mounted millimeter-wave radar target classification method, device, system and storage medium, which can eliminate the interference of stationary clutter on the target signal, while ensuring the classification accuracy and reducing the network complexity.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a vehicle-mounted millimeter-wave radar target classification method, comprising:

[0008] Perform vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter;

[0009] Perform fast Fourier transform of the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter to obtain the radar target range-Doppler image;

[0010] Inputting the radar target range-Doppler image into a pre-built lightweight target classification model for feature extraction and target classification to obtain a target classification result;

[0011] The lightweight target classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model.

[0012] In combination with the first aspect, further performing vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter includes:

[0013] Representing the pre-acquired radar target echo signal as a matrix;

[0014] averaging the values ​​of each row of the matrix to obtain a column vector containing the average values ​​of each row of the matrix;

[0015] Eliminate the column vector containing the average value of each row of the matrix from each column of the matrix, complete the vector mean cancellation of the radar target echo signal, and filter out the stationary clutter.

[0016] In combination with the first aspect, further, building a lightweight target classification model based on the pre-acquired Ghost module and MobileNetV3 network model includes:

[0017] Construct the G-Bottleneck module based on the pre-acquired Ghost module;

[0018] Delete and modify the pre-acquired MobileNetV3 network model to obtain the MobileNet11 network model;

[0019] The Bottleneck module in the MobileNet11 network model is replaced by the G-Bottleneck module, and the Conv2d convolution operation with a convolution kernel size of 1×1 in the MobileNet11 network model is replaced by the Ghost module to obtain a lightweight target classification model.

[0020] In combination with the first aspect, further, constructing a G-Bottleneck module based on the pre-acquired Ghost module includes:

[0021] Use the pre-acquired Ghost module to perform dimension-upgrading operations on the input data;

[0022] Connecting a depthwise separable convolution after the Ghost module to perform a depthwise separable convolution operation on the data after the dimensionality increase operation;

[0023] Connecting a Ghost module after the depthwise separable convolution to perform a dimensionality reduction operation on the data subjected to the depthwise separable convolution operation to obtain output data;

[0024] The G-Bottleneck module is composed of the connected Ghost module, depth-wise separable convolution, and Ghost module.

[0025] In combination with the first aspect, further, after using the G-Bottleneck module to replace the Bottleneck module in the MobileNet11 network model, the input data of the G-Bottleneck module is the image features output by its previous module, and the output data of the G-Bottleneck module is the image features entering its next module.

[0026] In combination with the first aspect, further, the convolution kernel of the depth-separable convolution is 3×3, and the step size of the depth-separable convolution operation is 1 or 2. If the step size of the depth-separable convolution operation is 1, the G-Bottleneck module has a residual connection; if the step size of the depth-separable convolution operation is 2, the G-Bottleneck module has no residual connection.

[0027] In combination with the first aspect, further deleting and modifying the pre-acquired MobileNetV3 network model to obtain the MobileNet11 network model includes:

[0028] The pre-acquired MobileNetV3 network model is deleted and modified, retaining its Conv2d convolution operation and pooling operation, as well as the 7 Bottleneck modules between the Conv2d convolution operation and pooling operation, and reducing the expansion multiples of the last 5 Bottleneck modules from 6 to 3. The LeakyReLU activation function is used to obtain the MobileNet11 network model.

[0029] In a second aspect, the present invention provides a vehicle-mounted millimeter-wave radar target classification device, comprising:

[0030] Data preprocessing module: used to perform vector mean cancellation on the pre-acquired radar target echo signal and filter out stationary clutter;

[0031] Image acquisition module: used to perform fast Fourier transform in the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter, and obtain the radar target range-Doppler image;

[0032] Target classification module: used to input the radar target range-Doppler image into a pre-built lightweight target classification model for feature extraction, and perform target classification to obtain target classification results;

[0033] The lightweight target classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model.

[0034] In a third aspect, the present invention provides a system including a processor and a storage medium;

[0035] The storage medium is used to store instructions;

[0036] The processor is configured to operate according to the instructions to execute the steps of the method according to any one of the first aspects.

[0037] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The vehicle-mounted millimeter-wave radar target classification method proposed in this paper eliminates the influence of stationary clutter on the target's Doppler or micro-Doppler characteristics by performing vector mean cancellation on radar target echo signals. This lightweight target classification model, constructed using the Ghost module and the MobileNetV3 network model, has fewer parameters and floating-point operations than existing models, saving model storage space and the memory required for running the model. This reduces network complexity while maintaining classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a vehicle-mounted millimeter-wave radar target classification method provided by an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of linear frequency modulation continuous wave signal modulation principle provided by an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the principle of the vector mean cancellation algorithm provided by an embodiment of the present invention;

[0043] Figure 4 Schematic diagram of the structure of the Ghost module provided by an embodiment of the present invention;

[0044] Figure 5 : is a schematic diagram of the structure of the G-Bottleneck module provided by an embodiment of the present invention, wherein: Figure 5 (a) is the structure of the G-Bottleneck module when the depthwise separable convolution operation step size is 1. Figure 5 (b) The structure of the G-Bottleneck module when the stride of the depthwise separable convolution operation is 2;

[0045] Figure 6 is a schematic diagram of a radar target range-Doppler (RD) image provided by an embodiment of the present invention, wherein: Figure 6 (a) is the radar target range-Doppler image without vector mean cancellation of the target echo signal. Figure 6 (b) is the radar target range-Doppler image after vector mean cancellation of the target echo signal;

[0046] Figure 7 2 is a schematic diagram of a test accuracy curve of a lightweight target classification model provided by an embodiment of the present invention;

[0047] Figure 8 Schematic diagram of the loss curve of the lightweight target classification model provided by an embodiment of the present invention;

[0048] Figure 9 2 is a schematic diagram of the structure of the comparison network DCNN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present application will be further described in detail below in conjunction with specific implementation methods.

[0050] The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application. The embodiments of the present application and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0051] Example 1:

[0052] Figure 1 This is a flow chart of a vehicle-mounted millimeter-wave radar target classification method provided by the first embodiment of the present invention. This flow chart only shows the logical sequence of the method of this embodiment. In other possible embodiments of the present invention, different methods may be used without conflict. Figure 1 The steps shown or described are accomplished in the order shown.

[0053] The vehicle-mounted millimeter-wave radar target classification method provided in this embodiment can be applied to a terminal and can be executed by a vehicle-mounted millimeter-wave radar target classification device, which can be implemented by software and / or hardware. The device can be integrated into a terminal, for example: any tablet computer or computer device with communication capabilities.

[0054] See also Figure 1 The method of this embodiment specifically includes the following steps:

[0055] Step 1: Perform vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter;

[0056] In this embodiment, first, according to Figure 2 The linear frequency modulated continuous wave (FCMW) signal modulation principle diagram shown in the figure is used to construct the mathematical model of the transmitted signal:

[0057]

[0058] Where x(t) represents the signal waveform, k∈[0,N-1], N is the number of Frequency Modulated Continuous Wave (FCMW) signals in one measurement period, and T is the repetition period of the FCMW signal, including the transmission period and the idle period. The transmission signal S T (t) is in the slow time domain t s = a periodically repeating function on kT. represents the phase of the transmitted signal, expressed as:

[0059]

[0060] Among them, f0, α and are the starting frequency, slope adjustment rate and initial phase of the FCMW signal respectively.

[0061] Define fast time t f :

[0062] t f =tt s =t-kT (3)

[0063] Among them, the transmitted signal S T (t) After being reflected by the target, it is received by the radar receiver and mixed and low-pass filtered. The resulting baseband signal can be expressed as:

[0064]

[0065] Where A represents the amplitude change of the signal after propagation and reflection, and τ(t) refers to the round-trip delay of the signal from the transmitting antenna to the target reflection and reception at the receiving antenna. It is expressed as:

[0066]

[0067] Where R is the distance between the target and the radar, v is the radial velocity of the target, and c is the speed of light. Combining equations (4) and (5), we can get the received signal S B Phase of (t):

[0068]

[0069] Due to c 2 >>2α(R+vt) 2 , so the third term of formula (6) can be ignored. Combining formula (6) and formula (3) we can get:

[0070]

[0071] because Therefore, the third term of equation (7) can be ignored, and the received signal can be expressed as:

[0072]

[0073] In the first exponential factor of formula (8), represents the constant frequency component, is the fast-time frequency component corresponding to the distance, is the slow-time frequency component corresponding to the relative radial velocity, and the second exponential factor corresponds to the range offset and Doppler shift effects, which are usually ignored in current automotive radar systems.

[0074] Then, we set up an experimental platform for data collection. We used the Texas Instruments (TI) AWR1843 Boost radar board and the DCA1000EVM real-time data acquisition board to collect target echo data. We connected the two via a 60-pin high-density cable and transmitted control data via a USB data cable. We configured the millimeter-wave radar sensor using TI's mmWave Studio platform, and transmitted the raw radar data collected by the DCA1000EVM to a PC via Ethernet. Table 1 lists the radar's main configuration parameters.

[0075] Table 1 Radar parameter configuration table

[0076]

[0077] Then, data processing is performed to perform vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter.

[0078] Performing vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter includes the following steps:

[0079] Step ①: Represent the pre-acquired radar target echo signal as a matrix;

[0080] Step 2: Calculate the average value of each row of the matrix and obtain the column vector containing the average value of each row of the matrix;

[0081] Step 3: Eliminate the column vector containing the average value of each row of the matrix from each column of the matrix to complete the vector mean cancellation of the radar target echo signal and filter out the stationary clutter.

[0082] The radar received signal AD sampling data is represented as a matrix Where M is the number of sampling points in the fast time dimension, that is, the number of sampling points for each FMCW signal, and N is the number of sampling points in the slow time dimension, that is, the number of FMCW signals in each observation period. The average value of all FMCW signals is:

[0083]

[0084] in, The reference signal is obtained by averaging all FMCW signals. The mean cancellation process can be expressed as:

[0085] P=RQ(10)

[0086] in, is the signal obtained after vector mean cancellation.

[0087] Figure 3 The schematic diagram of the principle of the vector mean cancellation algorithm provided in this embodiment is as follows: Figure 3 Figure 2 shows a schematic diagram of target echo vector summation when N = 3. m1, m2, and m3 are the vectors for a moving target. Because target motion causes the phases of each FM signal in the echo signal to differ, their phasor summation results in cancellation. n1, n2, and n3 are the vectors for a stationary target. For a stationary target echo, the phases of each FM signal are identical, resulting in a large average value. Therefore, the moving target echo signal component in the reference signal q is small, while the stationary target echo signal component is large and identical to the stationary component in each received signal. Eliminating the reference signal from each FMCW signal in the received signal R eliminates the stationary target echo signal component while retaining the moving target echo signal component.

[0088] Step 2: Perform fast Fourier transform of the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter to obtain the radar target range-Doppler image;

[0089] The radar target range-Doppler image (RD image) obtained after performing fast Fourier transform (FFT) in the distance dimension and velocity dimension is as follows: Figure 6 As shown, Figure 6 (a) is the radar target range-Doppler image without vector mean cancellation of the target echo signal, Figure 6 (a) It can be seen that when the radar echo signal is not processed by vector mean cancellation, when the target speed is low or the target speed spread is large, the echo signal of the stationary target will interfere with the Doppler characteristics of the target. Figure 6 (b) is the radar target range-Doppler image after vector mean cancellation of the target echo signal. Figure 6 (a) and Figure 6 (b) It can be seen that the target echo signal processing process adopted in this embodiment can effectively suppress the interference of stationary clutter on the target Doppler characteristics, while retaining the Doppler characteristics and micro-Doppler characteristics used for vehicle-mounted radar target classification.

[0090] Step 3: Input the radar target range-Doppler image into the pre-built lightweight target classification model for feature extraction and target classification to obtain the target classification result.

[0091] In this embodiment, the lightweight object classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model.

[0092] Building a lightweight target classification model based on the pre-acquired Ghost module and MobileNetV3 network model includes the following steps:

[0093] Step I: Construct the G-Bottleneck module based on the pre-acquired Ghost module;

[0094] The structure of the Ghost module is as follows Figure 4 As shown, Figure 4 In, Φ 11 , Φ 21 ,…,Φ m1 represents the depthwise separable convolution operation, Φ i2 To preserve the identity mapping of the inherent feature map, we first use ordinary convolution to generate m inherent feature maps of the input image, then perform a depthwise separable convolution operation on each inherent feature map to generate a phantom feature map, and finally concatenate it with the identity mapping that preserves the inherent feature map to obtain the output feature map.

[0095] The structure of the G-Bottleneck module constructed in this embodiment is as follows Figure 5 As shown in the figure, the input image features are first increased in dimension by the Ghost module to generate multiple channels, and then a depthwise separable convolution operation is performed on the increased-dimensional data. The convolution kernel size of the depthwise separable convolution is usually 3×3. Finally, the Ghost module is used to reduce the dimension of the data to obtain the output features of the G-Bottleneck module. Figure 5 As shown in (a), if the stride of the depthwise separable convolution operation is 1, the G-Bottleneck module has a residual connection; Figure 5 As shown in (b), if the stride of the depthwise separable convolution operation is 2, the G-Bottleneck module has no residual connection.

[0096] Based on the pre-acquired Ghost module, building the G-Bottleneck module includes the following steps:

[0097] Step i: Use the pre-acquired Ghost module to perform dimension-upgrading operation on the input data;

[0098] Step II: Connect the depthwise separable convolution after the Ghost module to perform the depthwise separable convolution operation on the data after the dimensionality increase operation;

[0099] Step iii: Connect the Ghost module after the depthwise separable convolution to perform dimensionality reduction on the data after the depthwise separable convolution operation to obtain the output data;

[0100] Step IV: The G-Bottleneck module is constructed by connecting the Ghost module, depthwise separable convolution, and Ghost module.

[0101] In this embodiment, after the Bottleneck module in the MobileNet11 network model is replaced by the G-Bottleneck module, the input data of the G-Bottleneck module is the image features output by its previous module, and the output data of the G-Bottleneck module is the image features entering its next module.

[0102] In this embodiment, the convolution kernel of the depth-wise separable convolution is 3×3, and the step size of the depth-wise separable convolution operation is 1 or 2. If the step size of the depth-wise separable convolution operation is 1, the G-Bottleneck module has a residual connection. If the step size of the depth-wise separable convolution operation is 2, the G-Bottleneck module has no residual connection.

[0103] Step II: Delete and modify the pre-acquired MobileNetV3 network model to obtain the MobileNet11 network model;

[0104] The pre-acquired MobileNetV3 network model is pruned and modified to obtain the MobileNet11 network model. This includes pruning the pre-acquired MobileNetV3 network model, retaining the Conv2d convolution and pooling operations, as well as the seven Bottleneck modules between them. The expansion factor of the last five Bottleneck modules is reduced from 6 to 3, and the LeakyReLU activation function is used to obtain the MobileNet11 network model. The structure of the MobileNet11 network model is shown in Table 2.

[0105] Table 2 Structure of MobileNet11 network model

[0106]

[0107] Step III: Replace the Bottleneck module in the MobileNet11 network model with the G-Bottleneck module and replace the Conv2d convolution operation with a kernel size of 1×1 in the MobileNet11 network model with the Ghost module to obtain a lightweight object classification model. The structure of the lightweight object classification model is shown in Table 3.

[0108] Table 3 Structure of lightweight target classification model

[0109]

[0110]

[0111] In order to verify the performance of the lightweight target classification model constructed in this embodiment, the DCNN network and the MobileNet11 network are selected as comparison networks. Figure 9 As shown in the figure, it is a structural diagram of the comparison network DCNN. The experimental hardware platform is CPU: Intel i5-12400F; memory: 16GB; GPU: NVIDIA GeForce GTX1660Spuer 6GB. Experimental comparisons were carried out in four aspects: classification accuracy, spatial complexity, computational complexity and network inference time for radar target range-Doppler images (RD images). The comparison results are shown in Table 4. Since the network depth of DCNN is smaller than that of MobileNet11 and lightweight target classification models, the number of times the CPU accesses the memory is reduced when the computer runs DCNN, and the time consumption is reduced, so the inference time of DCNN is shorter than that of MobileNet11 and lightweight target classification models. However, the spatial complexity of DCNN is large, that is, the number of parameters is large, and a large storage space is required; the computational complexity is high, that is, the number of floating-point operations is large, and a large amount of memory is required when the network is running, and the classification accuracy is low. As Figure 7 As shown in FIG, a schematic diagram of the test accuracy curve of the lightweight target classification model provided in this embodiment is shown in FIG. Figure 8 , which is a schematic diagram of the loss curve of the lightweight target classification model provided in this embodiment.

[0112] Table 4 Network model comparison table

[0113]

[0114]

[0115] As can be seen from Table 4, the classification accuracy of the lightweight target classification model provided by this embodiment is higher than that of the comparison networks DCNN and MobileNet11, and its spatial complexity is reduced by 95.51% and 41.59% compared with DCNN and MobileNet11, respectively, which can save a large amount of memory space required to store the model; its computational complexity is reduced by 72.33% and 42.3% compared with DCNN and MobileNet11, respectively, reducing the running memory occupied by the network during operation.

[0116] This embodiment provides a vehicle-mounted millimeter-wave radar target classification method that eliminates the influence of stationary clutter on the target's Doppler or micro-Doppler characteristics by performing vector mean cancellation on radar target echo signals. This lightweight target classification model, constructed using the Ghost module and the MobileNetV3 network model, has fewer parameters and floating-point operations than existing models, saving model storage space and running memory, while maintaining classification accuracy while reducing network complexity.

[0117] Example 2:

[0118] This embodiment provides a vehicle-mounted millimeter-wave radar target classification device, including:

[0119] Data preprocessing module: used to perform vector mean cancellation on the pre-acquired radar target echo signal and filter out stationary clutter;

[0120] Image acquisition module: used to perform fast Fourier transform in the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter, and obtain the radar target range-Doppler image;

[0121] Target classification module: used to input radar target range-Doppler images into a pre-built lightweight target classification model for feature extraction, target classification, and target classification results;

[0122] Among them, the lightweight target classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model.

[0123] The vehicle-mounted millimeter-wave radar target classification device provided in an embodiment of the present invention can execute the vehicle-mounted millimeter-wave radar target classification method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0124] Example 3:

[0125] This embodiment provides a system, including a processor and a storage medium;

[0126] The storage medium is used to store instructions;

[0127] The processor is configured to operate according to the instructions to execute the steps of the method in the first embodiment.

[0128] Example 4:

[0129] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method in the first embodiment are implemented.

[0130] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0131] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0134] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A vehicle-mounted millimeter-wave radar target classification method, characterized in that: include: Perform vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter; Perform fast Fourier transform of the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter to obtain the radar target range-Doppler image; Inputting the radar target range-Doppler image into a pre-built lightweight target classification model for feature extraction and target classification to obtain a target classification result; The lightweight target classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model; Building a lightweight target classification model based on the pre-acquired Ghost module and MobileNetV3 network model includes: Construct the G-Bottleneck module based on the pre-acquired Ghost module; Delete and modify the pre-acquired MobileNetV3 network model to obtain the MobileNet11 network model; The Bottleneck module in the MobileNet11 network model is replaced by the G-Bottleneck module, and the Conv2d convolution operation with a convolution kernel size of 1×1 in the MobileNet11 network model is replaced by the Ghost module to obtain a lightweight target classification model.

2. The vehicle-mounted millimeter-wave radar target classification method according to claim 1, characterized in that: Perform vector mean cancellation on the pre-acquired radar target echo signal to filter out stationary clutter, including: Representing the pre-acquired radar target echo signal as a matrix; averaging the values ​​of each row of the matrix to obtain a column vector containing the average values ​​of each row of the matrix; Eliminate the column vector containing the average value of each row of the matrix from each column of the matrix, complete the vector mean cancellation of the radar target echo signal, and filter out the stationary clutter.

3. The vehicle-mounted millimeter-wave radar target classification method according to claim 1, characterized in that: Based on the pre-obtained Ghost module, building the G-Bottleneck module includes: Use the pre-acquired Ghost module to perform dimension-upgrading operations on the input data; Connecting a depthwise separable convolution after the Ghost module to perform a depthwise separable convolution operation on the data after the dimensionality increase operation; Connecting a Ghost module after the depthwise separable convolution to perform a dimensionality reduction operation on the data subjected to the depthwise separable convolution operation to obtain output data; The G-Bottleneck module is composed of the connected Ghost module, depth-wise separable convolution, and Ghost module.

4. The vehicle-mounted millimeter-wave radar target classification method according to claim 3, characterized in that: After the Bottleneck module in the MobileNet11 network model is replaced by the G-Bottleneck module, the input data of the G-Bottleneck module is the image features output by its previous module, and the output data of the G-Bottleneck module is the image features entering its next module.

5. The vehicle-mounted millimeter-wave radar target classification method according to claim 3, characterized in that: The convolution kernel of the depth-wise separable convolution is 3×3, and the step size of the depth-wise separable convolution operation is 1 or 2. If the step size of the depth-wise separable convolution operation is 1, the G-Bottleneck module has a residual connection. If the step size of the depth-wise separable convolution operation is 2, the G-Bottleneck module has no residual connection.

6. The vehicle-mounted millimeter-wave radar target classification method according to claim 1, characterized in that: The pre-acquired MobileNetV3 network model is deleted and modified to obtain the MobileNet11 network model, including: The pre-acquired MobileNetV3 network model is deleted and modified, retaining its Conv2d convolution operation and pooling operation, as well as the 7 Bottleneck modules between the Conv2d convolution operation and pooling operation, and reducing the expansion multiples of the last 5 Bottleneck modules from 6 to 3. The LeakyReLU activation function is used to obtain the MobileNet11 network model.

7. A vehicle-mounted millimeter-wave radar target classification device, characterized in that: include: Data preprocessing module: used to perform vector mean cancellation on the pre-acquired radar target echo signal and filter out stationary clutter; Image acquisition module: used to perform fast Fourier transform in the velocity and range dimensions on the radar target echo signal after filtering out the stationary clutter, and obtain the radar target range-Doppler image; Target classification module: used to input the radar target range-Doppler image into a pre-built lightweight target classification model for feature extraction, and perform target classification to obtain target classification results; The lightweight target classification model is constructed based on the pre-acquired Ghost module and MobileNetV3 network model; Building a lightweight target classification model based on the pre-acquired Ghost module and MobileNetV3 network model includes: Construct the G-Bottleneck module based on the pre-acquired Ghost module; Delete and modify the pre-acquired MobileNetV3 network model to obtain the MobileNet11 network model; The Bottleneck module in the MobileNet11 network model is replaced by the G-Bottleneck module, and the Conv2d convolution operation with a convolution kernel size of 1×1 in the MobileNet11 network model is replaced by the Ghost module to obtain a lightweight target classification model.

8. A system, characterized in that: including processors and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images