Target detection method, system and device against intermittent sampling interference and storage medium

Abstract

The application discloses an anti-intermittent sampling interference target detection method, system and device and a storage medium. The method of the application embodiment first converts a radar echo signal into a range-Doppler map, which can significantly compress data and reduce data processing amount. Moreover, more field knowledge is introduced through the range-Doppler map, which can further facilitate the training and prediction analysis of a visual target detection network. Meanwhile, the single detection head structure based on the improved YOLOX-Nano network can greatly reduce the model parameter amount and the calculation amount without affecting the detection accuracy. Moreover, the focal loss function introduced can improve the performance of the anti-intermittent sampling target detection under the condition of positive and negative sample imbalance. Finally, according to the clearer characteristics of the physical meaning of the range-Doppler map relative to the original radar echo signal, the detection result can be efficiently and accurately obtained through the determined target prediction box, and the application can be widely applied to the technical field of radar detection.

Description

Technical Field

[0001] This invention relates to the field of radar detection technology, and in particular to a target detection method, system, device, and storage medium that resists intermittent sampling interference. Background Technology

[0002] Digital radio frequency memory (DRFM) is a common device used to generate deceptive jamming signals. It has two operating modes: full-transfer mode and intermittent sampling-transfer mode. In full-transfer mode, the DRFM intercepts, stores, and forwards the complete radar transmitted signal. In intermittent sampling-transfer mode, the DRFM uses a rectangular window to partially sample the radar's transmitted signal and forwards it multiple times, repeating this process until the end of the radar's transmitted signal is sampled. The deceptive jamming generated in intermittent sampling-transfer mode is usually called interrupted sampling repeater jamming (ISRJ). Compared to jamming in full-transfer mode, ISRJ has advantages such as low power consumption, low latency, fast response, high flexibility, and the ability to generate false target signals that precede the real target, making it one of the most threatening jamming techniques currently available. Intermittent sampling repeater jamming severely limits the normal operation of radar functions; therefore, researching target detection algorithms that resist mainlobe intermittent sampling repeater jamming is of significant importance and application value.

[0003] In the area of ​​intermittent sampling interference resistance based on deep learning, some scholars have proposed an end-to-end target detection method based on CNN (convolutional neural network) that achieves significantly better results than traditional methods. This method first uses a waveform adaptive processing method to convert the transmitted and echo signals into a three-dimensional feature block, and then inputs this feature block into a CNN network. This CNN network detects the real target while suppressing interference. However, because this method outputs a binary classification probability of the presence of a target in each range cell, it is difficult to combine with traditional moving target detection algorithms to calculate the target's velocity information. Furthermore, this method suffers from the problem that the input sample size increases dramatically with the number of range cells to be detected. To address these issues, another scholar recently proposed a method for intermittent sampling target ranging and velocity measurement based on time-frequency maps and CNN. This method first utilizes a computer vision target detection network, YOLO (You Only Look Once), to roughly detect the location of the jammed signal on the time-frequency map of the echo signal. Then, it uses a constructed validation set to obtain the estimation error of the YOLO network and uses this estimation error to crop the echo signal, thereby reducing the input size of subsequent radar target ranging and velocity measurement networks. Next, the method constructs target ranging and target velocity measurement sample sets based on the cropped signals, and uses these two sample sets to train the proposed CNN ranging and CNN velocity measurement models, thus obtaining the final anti-ISRJ target ranging and velocity measurement models. This method achieves significantly better anti-ISRJ target velocity and ranging results than the compared traditional methods on both simulation and experimental datasets. However, this method is cumbersome, processing ranging and velocity separately rather than in an end-to-end manner, which can lead to error accumulation at each step, making it difficult for the system to achieve optimal performance. Furthermore, this method requires constructing multiple datasets and multiple detection networks, which increases the complexity of the method and the consumption of memory and time resources. On the other hand, the inputs of the two aforementioned anti-intermittent sampling target detection models are feature blocks obtained from the original radar signal through simple processing. These feature blocks contain limited prior information, leading to a large number of samples required for network training. Furthermore, although these two methods perform better than traditional methods under extremely low signal-to-noise ratio conditions, their detection accuracy still drops significantly. In addition, these two methods suffer from a lack of intuitiveness, weak interpretability, and unclear physical meaning, which reduces their reliability. Therefore, how to efficiently achieve anti-intermittent sampling interference of radar signals and obtain accurate radar target detection is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a target detection method, system, device and storage medium that resists intermittent sampling interference, which can efficiently realize the resistance to intermittent sampling interference of radar signals and obtain true and accurate radar detection targets.

[0005] On one hand, embodiments of the present invention provide a target detection method resistant to intermittent sampling interference, comprising:

[0006] Acquire radar echo signals;

[0007] Doppler transformation processing is performed on the radar echo signal to obtain the range Doppler map; the Doppler transformation processing includes fast time dimension pulse compression processing and slow time dimension discrete Fourier transform processing.

[0008] A visual target detection network is used to analyze the distance Doppler image to obtain the target feature map; the target feature map includes a preset number of prediction boxes; the visual target detection network is generated by training on labeled samples based on an improved YOLOX-Nano network, and the improved YOLOX-Nano network includes a single detection head structure and a focus loss function;

[0009] Based on the predicted bounding box, the target predicted bounding box is determined by a preset detection threshold; wherein, the preset detection threshold is determined by the false alarm rate index.

[0010] The detection result is determined based on the parameters of the target prediction box; the parameters of the target prediction box include the confidence score of the object to be detected contained in the target prediction box, the class score of the object, and the coordinates of the center point of the target prediction box; the detection result includes the detection score, target distance, and target velocity.

[0011] Optionally, the radar echo signal is subjected to Doppler conversion processing to obtain a range Doppler image, including:

[0012] The radar echo signal is demodulated, and the demodulated radar echo signal is convolved with the matched filter to complete the fast time dimension pulse compression processing of the radar echo signal.

[0013] The radar echo signal after pulse compression is processed by slow-time discrete Fourier transform to obtain the range Doppler spectrum.

[0014] The distance-Doppler spectrum is plotted to obtain the distance-Doppler map.

[0015] Optionally, it also includes:

[0016] A visual object detection network was created based on the improved YOLOX-Nano network.

[0017] Optionally, a visual object detection network is created based on the improved YOLOX-Nano network, including:

[0018] An improved YOLOX-Nano network is set up. Based on training samples, the improved YOLOX-Nano network is trained for prediction using the overall loss function. Based on the training results, the improved YOLOX-Nano network is adjusted to obtain a visual object detection network.

[0019] The improved YOLOX-Nano network consists of a backbone, a connection part, and a detection head part, with the detection head part using a single detection head structure; the overall loss function includes foreground prediction loss, category prediction loss, and bounding box localization loss.

[0020] Optionally, in the step of training the improved YOLOX-Nano network for prediction using the overall loss function, the focus loss function is used to calculate the foreground prediction loss, the cross-entropy loss function is used to calculate the class prediction loss, and the cross-union loss is used to calculate the bounding box localization loss.

[0021] Optionally, the visual target detection network includes a backbone, a connector, and a detection head. The network analyzes the range Doppler image to obtain a target feature map, including:

[0022] By performing multi-level feature extraction on the distance Doppler image through the main part, features at different levels are obtained;

[0023] By fusing features from different levels through the connection component, fused features are obtained;

[0024] The target feature map is obtained by performing regression and classification on the fused features through the detection head.

[0025] Optionally, the detection result is determined based on the parameters of the target prediction bounding box, including:

[0026] The detection score is calculated based on the confidence level of the objects to be detected contained within the target prediction bounding box and the category score of the objects.

[0027] The target distance is calculated based on the x-coordinate of the center point of the target prediction box, combined with the width of the range Doppler image and the radar range detection range.

[0028] The target velocity is calculated based on the ordinate of the center point of the target prediction box, combined with the altitude of the distance Doppler image and the radar velocity range.

[0029] The x-coordinate and y-coordinate of the center point are determined by the center point coordinates, while the radar range and radar speed range are determined based on the radar corresponding to the radar echo signal.

[0030] On the other hand, embodiments of the present invention provide a target detection system resistant to intermittent sampling interference, comprising:

[0031] The first module is used to acquire radar echo signals;

[0032] The second module is used to perform Doppler conversion processing on the radar echo signal to obtain the range Doppler map; the Doppler conversion processing includes fast time dimension pulse compression processing and slow time dimension discrete Fourier transform processing.

[0033] The third module is used to analyze the distance Doppler image using a visual target detection network to obtain the target feature map; wherein, the target feature map includes a preset number of prediction boxes; the visual target detection network is generated by training on labeled samples based on an improved YOLOX-Nano network, and the improved YOLOX-Nano network includes a single detection head structure and a focus loss function;

[0034] The fourth module is used to determine the target prediction box based on the prediction box and a preset detection threshold; wherein the preset detection threshold is determined by the false alarm rate index.

[0035] The fifth module is used to determine the detection results based on the parameters of the target prediction box. The parameters of the target prediction box include the confidence score of the object to be detected contained in the target prediction box, the object's category score, and the coordinates of the center point of the target prediction box. The detection results include the detection score, target distance, and target velocity.

[0036] On the other hand, embodiments of the present invention provide a target detection device resistant to intermittent sampling interference, including a processor and a memory;

[0037] Memory is used to store programs;

[0038] The processor executes the program as described above.

[0039] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0040] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0041] This invention first acquires radar echo signals; then performs Doppler transformation processing on the radar echo signals to obtain a range Doppler map; wherein the Doppler transformation processing includes fast-time dimension pulse compression processing and slow-time dimension discrete Fourier transform processing; the range Doppler map is analyzed using a visual target detection network to obtain a target feature map; wherein the target feature map includes a preset number of prediction boxes; the visual target detection network is generated based on an improved YOLOX-Nano network trained on labeled samples, and the improved YOLOX-Nano network includes a single detection head structure and a focus loss function; based on the prediction boxes, target prediction boxes are determined through a preset detection threshold; wherein the preset detection threshold is determined by the false alarm rate index; based on the parameters of the target prediction boxes, the detection result is determined; wherein the parameters of the target prediction boxes include the confidence score of the object to be detected contained in the target prediction box, the object's category score, and the coordinates of the center point of the target prediction box; the detection result includes the detection score, target distance, and target velocity. This invention firstly converts radar echo signals into range Doppler images, which significantly compresses data and reduces data processing volume. Furthermore, by incorporating more domain knowledge through the range Doppler images, it facilitates the training and predictive analysis of visual target detection networks. Simultaneously, the single-detection-head structure based on the improved YOLOX-Nano network drastically reduces the number of model parameters and computational load without affecting detection accuracy. Moreover, the introduced focus loss function improves the performance of target detection under conditions of imbalanced positive and negative samples, thus enhancing its resilience to intermittent sampling. Finally, leveraging the clearer physical meaning of the range Doppler images compared to the original radar echo signals, detection results can be obtained efficiently and accurately using the determined target prediction boxes. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0043] Figure 1 This is a flowchart illustrating a target detection method for resisting intermittent sampling interference provided in an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the main framework of a target detection method against intermittent sampling interference provided in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram illustrating the process of generating intermittent sampling and forwarding interference signals.

[0046] Figure 4This is a schematic diagram of the structure of the visual target detection network provided in an embodiment of the present invention;

[0047] Figure 5 A schematic diagram of the Focus module provided in an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of the Conv module provided in an embodiment of the present invention;

[0049] Figure 7 A schematic diagram of the DWConv module provided in an embodiment of the present invention;

[0050] Figure 8 This is a schematic diagram of a stacking module provided in an embodiment of the present invention;

[0051] Figure 9 A schematic diagram of the CSP1 module provided in an embodiment of the present invention;

[0052] Figure 10 This is a schematic diagram of the BottleNeck1 module provided in an embodiment of the present invention;

[0053] Figure 11 A schematic diagram of the CSP2 module provided in an embodiment of the present invention;

[0054] Figure 12 This is a schematic diagram of the BottleNeck2 module provided in an embodiment of the present invention;

[0055] Figure 13 A schematic diagram of the SPP module provided in an embodiment of the present invention;

[0056] Figure 14 This is a schematic diagram of the upsampling module provided in an embodiment of the present invention;

[0057] Figure 15 This is a schematic diagram of the detection head module provided in an embodiment of the present invention;

[0058] Figure 16 This is a schematic diagram of the label box and prediction box provided in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] Given the limitations of existing technologies, such as the large amount of training samples required, unclear physical meaning, numerous steps involved, inability to simultaneously measure range and velocity, and poor performance under extremely low signal-to-noise ratio conditions, this invention proposes a target detection method based on a visual target detection network to combat intermittent sampling interference. This method converts radar echo signals into range Doppler images, which are then input into a visual target detection network. This network detects the real target based on the difference between the target and the intermittently sampled interference signal in the range Doppler image, thereby obtaining the real target's range and velocity information. Figure 1 The flowchart of the method of the present invention is as follows: Figure 2 This is a schematic diagram of the main framework of the method of the present invention. In one aspect, embodiments of the present invention provide a target detection method resistant to intermittent sampling interference, comprising:

[0061] S100, Acquire radar echo signal;

[0062] Specifically, to facilitate understanding of the subsequent steps, a detailed explanation of the relevant radar signal model will be provided first:

[0063] 1) Transmitted signal model

[0064] Suppose that the radar transmits a linear frequency modulated (LFM) signal. LFM is a widely used radar modulation signal, possessing advantages such as ease of generation, simple processing, and the ability to achieve a high pulse compression ratio during pulse compression. The original LFM signal can be expressed as follows (1):

[0065]

[0066] In the formula, j represents the imaginary unit. π represents pi. This represents the frequency modulation slope of the linear frequency modulated signal. B represents the bandwidth of the signal. T p This represents the pulse width of the linear frequency modulated signal. t represents time. This represents a rectangular window function with a width of , and its expression is as follows (2):

[0067]

[0068] By adding a carrier frequency to the original linear frequency modulated signal, the radar's transmitted signal can be obtained. The radar transmitted signal can be expressed as follows (3):

[0069]

[0070] In the formula, f0 represents the carrier frequency of the signal.

[0071] 2) Intermittent sampling and forwarding interference signal model

[0072] The basic process of intermittent sampling and forwarding interference signal generation is as follows: Figure 3 As shown, the jammer first uses a pulse width of T w The repetition period is T s The radar's transmitted signal is sampled using a rectangular pulse sequence, and the sampled slice signal is then forwarded to obtain the intermittently sampled and forwarded jamming signal.

[0073] The rectangular pulse sequence used for sampling can be represented by the following equation (4):

[0074]

[0075] In the formula, * denotes the convolution operation. δ(·) represents the impulse response function. n represents the nth sampling pulse. N represents the number of sampling pulses. This indicates a width of T. w The rectangular window function. Its expression is as follows (5):

[0076]

[0077] The intermittent sampling interference signal received by the radar can be expressed as follows (6):

[0078]

[0079] In the formula, A J This indicates the amplitude of the interference signal. m represents the sequence number of the linear frequency modulated signal transmitted by the radar. This represents the echo delay caused by the jammer's distance corresponding to the m-th linear frequency modulated signal transmitted by the radar. The value of m ranges from [0, M-1]. M represents the number of pulses transmitted by the radar within one coherent processing interval. The expression is as follows (7):

[0080]

[0081] In the formula, R J This indicates the distance between the jammer and the radar when the jammer receives the first transmitted signal from the radar. V J This represents the radial velocity of the jammer relative to the radar. (T) r This indicates the pulse repetition interval. C represents the speed of light.

[0082] 3) Echo signal model

[0083] The target echo signal can be expressed as follows (8):

[0084]

[0085] In the formula, A TThis indicates the amplitude of the target echo signal. This represents the echo time delay caused by the target distance for the m-th linear frequency modulated signal transmitted by the radar. The expression is as follows (9):

[0086]

[0087] In the formula, R T This indicates the distance between the target and the radar when the first pulse signal emitted by the radar reaches the target location. V T This indicates the radial velocity of the target relative to the radar.

[0088] The radar echo signal received by the radar (i.e., the radar echo signal) consists of the target echo signal, the interference signal, and the noise signal. Its expression is as follows (10):

[0089] r(t) = x(t) + j(t) + n(t)

[0090] In the formula, n(t) represents the noise signal.

[0091] S200: Perform Doppler conversion processing on the radar echo signal to obtain the range Doppler image;

[0092] It should be noted that Doppler transformation processing includes fast-time pulse compression processing and slow-time discrete Fourier transform processing.

[0093] In some embodiments, the radar echo signal is demodulated, and the radar echo signal is convolved with a matched filter to complete the fast time dimension pulse compression processing of the radar echo signal; the radar echo signal after pulse compression is processed by the slow time dimension discrete Fourier transform to obtain the range Doppler spectrum; and the range Doppler spectrum is plotted to obtain the range Doppler map.

[0094] Specifically, the radar echo signal from multiple pulses received by the radar is converted into a range Doppler image. Taking a specific embodiment as an example, suppose the number of echo pulses received by the radar within one coherent processing interval is 32. After performing fast-time dimension pulse compression processing and slow-time dimension discrete Fourier transform processing on the data composed of these 32 echo pulses, the corresponding range Doppler image can be obtained. Let the dimension of the obtained range Doppler image be 343×434×3. To further describe the principle of obtaining the range Doppler image through Doppler conversion processing, the target echo signal in the radar echo signal in step S100 is used as an example to illustrate the principle of Doppler image generation:

[0095] First, the target echo signal needs to be demodulated. This can be achieved by multiplying the target echo signal by the complex form of the carrier signal, as shown in equation (11):

[0096] x r (t)=x(t)exp(-j·2·π·f0·t)

[0097] Substituting formula (8) into formula (11) yields the following formula (12):

[0098]

[0099] In the formula, the range of values ​​for t is:

[0100] The matched filter corresponding to the transmitted signal can be expressed as follows (13):

[0101] h(t)=conj(ζ(-t))

[0102] In the formula, the range of values ​​for t is: conj(·) represents the conjugate operation.

[0103] Then, pulse compression of the target echo signal can be achieved by convolving the demodulated target echo signal with a matched filter. This process can be expressed as follows (14):

[0104] In the formula, * represents the convolution operation. By substituting equations (12) and (13) into equation (14) and rearranging, we can obtain the following equation (15):

[0105]

[0106] In the formula, sinc(·) represents the Singer function. In this formula, y(t) has a value of t. The maximum time. And The value is approximately equal to By finding From the value of , we can calculate the distance R of the target. T . The value corresponds to the x-coordinate of the target on the distance-Doppler map, and their relationship is as follows: Where 'a' represents the x-coordinate of the target on the distance-Doppler map. RD This represents the width of the Doppler graph. T RD This represents the radar detection time corresponding to the width of the Doppler image, and this value is known.

[0107] After pulse compression processing of the fast time dimension of the target echo signal, a discrete Fourier transform is performed on its slow time dimension to obtain the range-Doppler spectrum of the target echo signal. The range-Doppler spectrum is then plotted to obtain the range-Doppler map. The process of calculating the range-Doppler spectrum is as follows. First, substituting formula (9) into formula (15) yields the following formula (16):

[0108]

[0109] In the formula, the value of m ranges from [0, M-1]. M represents the number of pulses transmitted by the radar within one coherent processing interval. Taking m in equation (16) as the independent variable and simplifying the equation, its discrete form can be obtained as follows: equation (17):

[0110]

[0111] Consider the following discrete Fourier transform pair: (18)

[0112]

[0113] Here, ω represents the digital domain angular frequency. X(ω) represents the Fourier transform.

[0114] The discrete Fourier transform of equation (17) can be expressed as follows: (19)

[0115]

[0116] As can be seen, the spectral center ω0 corresponding to equation (17) is as follows: equation (20):

[0117]

[0118] ω0 is the digital domain angular frequency, and it has the following relationship with the corresponding real frequency:

[0119]

[0120] In the formula, f d This represents the Doppler frequency. This indicates the pulse repetition frequency of the radar transmitted signal.

[0121] The relationship between the target velocity and the Doppler frequency can be obtained from formulas (20) and (21), as shown in formula (22). Once the Doppler frequency of the target is obtained, the target velocity can be calculated using formula (22).

[0122]

[0123] In summary, this invention uses the range Doppler image corresponding to the radar echo signal as the input to the subsequent detection network. This approach offers four advantages: First, it compresses the raw echo data, reducing the amount of data the network needs to process. Second, the range Doppler image incorporates more domain knowledge compared to the raw echo data, reducing the number of samples required for network training. Third, the physical meaning of the range Doppler image is clearer than that of the raw echo signal. The target's position in the range Doppler image corresponds to its range and velocity information. This allows the subsequent target detection network to obtain the target's range and velocity information immediately after detecting its position in the Doppler image, thus achieving integrated velocity and range measurement. Fourth, the difference between the target signal and interference signal in the range Doppler image is intuitive and easily distinguishable, which helps reduce the difficulty for the subsequent target detection network in detecting real targets under intermittent sampling interference conditions.

[0124] S300. Analyze the distance Doppler image using a visual target detection network to obtain the target feature map;

[0125] It should be noted that the target feature map includes a preset number of prediction boxes; the visual target detection network is generated by training on labeled samples based on the improved YOLOX-Nano network, which includes a single detection head structure and a focus loss function.

[0126] In some embodiments, the method further includes creating a visual object detection network based on an improved YOLOX-Nano network.

[0127] The process involves creating a visual object detection network based on an improved YOLOX-Nano network. This includes setting up the improved YOLOX-Nano network, training it using training samples and an overall loss function, and adjusting the network based on the training results to obtain the visual object detection network. The improved YOLOX-Nano network consists of a backbone, a connectivity component, and a detection head, with the detection head using a single-head structure. The overall loss function includes foreground prediction loss, category prediction loss, and bounding box localization loss.

[0128] In some embodiments, in the step of training the improved YOLOX-Nano network for prediction using the overall loss function, the focus loss function is used to calculate the foreground prediction loss, the cross-entropy loss function is used to calculate the category prediction loss, and the cross-union loss is used to calculate the bounding box localization loss.

[0129] In some embodiments, the visual target detection network includes a backbone, a connection part, and a detection head part. The visual target detection network is used to analyze the range Doppler image to obtain a target feature map, including: performing multi-level feature extraction on the range Doppler image through the backbone to obtain features at different levels; fusing the features at different levels through the connection part to obtain fused features; and performing regression and classification processing on the fused features through the detection head part to obtain the target feature map.

[0130] First, it's important to note that YOLOX has a total of seven models of different sizes. Sorted from largest to smallest parameter count, these seven models are YOLOX-x, YOLOX-Darknet53, YOLOX-l, YOLOX-m, YOLOX-s, YOLOX-Tiny, and YOLOX-Nano. The smallest model, YOLOX-Nano, has 0.9M parameters, which is only one-hundredth of the largest model, YOLOX-x, with 99.1M parameters. The original YOLOX-Nano network contains three detector heads. The width and height of the output predicted feature maps from these three detector heads are 1 / 8, 1 / 16, and 1 / 32 of the width and height of the input image, respectively. YOLOX uses this three-detector-head design to detect targets of different scales; specifically, detector heads with output predicted feature map sizes of 1 / 8, 1 / 16, and 1 / 32 of the input image size are responsible for predicting small, medium, and large targets, respectively. However, in the field of radar target detection, targets typically appear as small targets on range Doppler maps. The original YOLOX-Nano network's detection heads, responsible for predicting large and medium-sized targets, are unnecessary and redundant for the intermittent sampling-resistant target detection task in this embodiment of the invention. Excessive detection heads increase model size, thereby increasing memory and computational resource consumption. To address this issue, this embodiment of the invention improves upon YOLOX-Nano, proposing a single-detector YOLOX-Nano target detection network. This network primarily reduces the number of model parameters and computational cost by pruning the detection heads of the original YOLOX-Nano network, removing the 16x downsampling and 32x downsampling detection heads, thus maintaining a relatively constant detection accuracy. The visual target detection network structure based on the improved YOLOX-Nano network proposed in this embodiment of the invention is as follows: Figure 4 As shown, it includes a Backbone, a Neck, and a Head. The Backbone comprises Backbone1, Backbone2, and Backbone3. Specifically, as... Figure 4As shown, Backbone1 includes Focus (pixel focusing module), DWConv (depthwise separable convolution module), and CSP1 (cross stage partial network 1 module); Backbone2 includes DWConv and CSP1; Backbone3 includes DWConv, SPP (spatial pyramid pooling module), and CSP1; Neck includes Conv (convolution module), upsampling module, stacking module, and CSP2 (cross stage partial network 2 module).

[0131] Specifically, considering the limitations of computing resources in real-world applications, this embodiment of the invention uses the YOLOX-Nano model as the baseline network and improves upon it to achieve end-to-end target detection, ranging, and velocity measurement against ISRJ interference. Taking a specific embodiment as an example, to facilitate network data processing, the dimension of the range Doppler image is first scaled to 640×640×3. The scaled range Doppler image is then input into the trained, improved YOLOX-Nano network. First, the backbone of the network extracts features at different levels from the input range Doppler image (the input range Doppler image is processed sequentially through Backbone1, Backbone2, and Backbone3 to obtain features at different levels). Then, the Neck of the network fuses the features output from the backbone at different levels. Finally, the Head of the network performs regression and classification operations on the fused features output from the Neck, outputting a feature map with dimension 80×80, where 80×80 corresponds to 6400 predicted boxes. The value 6 corresponds to the width, height, x-coordinate and y-coordinate of the center point of the predicted bounding box, the confidence score that the predicted bounding box contains the object to be detected, and the class score of the object within the predicted bounding box. Specifically, the output predicted feature map (i.e., the target feature map) has a dimension of 80×80×(bbox+obj+cls), where 80×80 corresponds to 6400 predicted bounding boxes, bbox(bounding box, predicted box) has a value of 4, corresponding to the width, height, and x- and y-coordinates of the center point of the predicted bounding box, obj(object, target) has a value of 1, corresponding to the confidence score that the predicted bounding box contains the target. Since the predicted target has only one possible class, namely the true target, cls(class, category) has a value of 1, corresponding to the score that the predicted bounding box contains the true target.

[0132] To describe in detail the function of the visual target detection network, the following section explains the specific results of each module and the detection head in the improved YOLOX-Nano network:

[0133] Focus: A schematic diagram of the Focus module is shown below. Figure 5 As shown, its main operation is to extract every other pixel from the input image and then stack them. Inputting a 640×640×3 image into the Focus module yields a 320×320×12 feature map. The main function of the Focus module is to downsample the input image, reduce the number of model parameters and computational cost, and improve the model's running speed.

[0134] Conv: Conv is a regular convolutional module. A diagram of a regular convolutional module is shown below. Figure 6 As shown. Its characteristic is that the number of channels in the convolutional kernel is the same as the number of channels in the input feature map, and the number of channels in the output feature map is the same as the number of convolutional kernels. Its function is to extract features from the input data. Generally, a batch normalization layer and an activation layer are added after a regular convolutional layer. The purpose of the batch normalization layer is to standardize the features output from the previous layer, making the network easier to train. The purpose of the activation layer is to introduce non-linear factors, thereby enhancing the non-linear expressive power of the network.

[0135] DWConv: DWConv is a depthwise separable convolution. A schematic diagram of depthwise separable convolution is shown below. Figure 7 As shown. Compared to regular convolution, depthwise separable convolution has fewer parameters and lower computational cost, and is therefore generally used in lightweight networks. Depthwise separable convolution mainly consists of two parts: channel-wise convolution and pointwise convolution. Channel-wise convolution is characterized by having only one channel in the kernel, and the number of kernels is equal to the number of channels in the input feature map. After channel-wise convolution, the number of channels in the output feature map is equal to the number of channels in the input feature map. Each kernel in channel-wise convolution performs a convolution operation with each channel feature of the input feature map, thereby extracting features from each channel of the input feature map. Pointwise convolution is characterized by having a 1×1 kernel size, and the number of channels in the kernel is equal to the number of channels in the input feature map. After pointwise convolution, the number of channels in the output feature map is the same as the number of kernels in pointwise convolution. The main function of pointwise convolution is to fuse the feature maps output by channel-wise convolution layers.

[0136] Stacked Modules: A schematic diagram of stacked modules is shown below. Figure 8 As shown, the main function of the stacking module is to concatenate multiple input feature maps along the channel dimension.

[0137] CSP1 module: The main structure of CSP1 is as follows Figure 9 As shown. By Figure 9 As can be seen, the CSP1 module mainly consists of the Conv module, several BottleNeck1 modules, and stacked modules. The composition of the BottleNeck1 module is as follows: Figure 10 As shown, the BottleNeck1 module mainly consists of a Conv module, a DWConv module, and an addition module. After the input feature map passes through the first Conv module of BottleNeck1, its number of channels is halved. Then, after passing through the DWConv module, its number of channels is increased back to the original number. This method of first reducing and then increasing the number of channels in the input feature map helps reduce the number of network parameters and computational cost.

[0138] CSP2: The main structure of the CSP2 module is as follows Figure 11 As shown. The CSP2 module has a similar structure to the CSP1 module, the main difference being that the CSP1 module uses the BottleNeck1 structure, while the CSP2 module uses the BottleNeck2 structure. The structure of BottleNeck2 is as follows... Figure 12 As shown in the BottleNeck2 structure, the input feature map is first reduced in number of channels by a Conv module, and then restored in number of channels by a DWConv module.

[0139] SPP Module: A schematic diagram of the SPP module is shown below. Figure 13 As shown, the SPP module mainly consists of a regular convolutional module (Conv), a max-pooling layer module (MaxPool), and a stacking module. The SPP module uses pooling kernels of different sizes to perform max-pooling operations on the input feature map, thereby extracting features from different receptive fields.

[0140] Upsampling module: A schematic diagram of the upsampling module is shown below. Figure 14 As shown. This embodiment of the invention uses nearest neighbor interpolation upsampling. After passing through an upsampling module with a magnification factor of 2, the width and height of the input feature map will be increased by a factor of 1 each.

[0141] Detection Head: The original YOLOX detection head module contains three detection heads. These three detection heads output predicted feature maps with sizes of 20*20, 40*40, and 80*80, respectively, suitable for detecting large, medium, and small targets. In the range Doppler spectrum of radar echo signals, real targets typically appear as small-area targets. To address this characteristic, this invention removes the detection heads with output predicted feature maps of 20*20 and 40*40 from the original YOLOX, retaining only the detection head with an output predicted feature map size of 80*80. This improvement significantly reduces the number of network parameters and computational load, making the method of this embodiment more suitable for practical engineering applications. The detection head module used in this embodiment is as follows: Figure 15 As shown.

[0142] In the YOLOX loss function calculation, the foreground prediction uses the binary cross-entropy (BCE) loss function. The cross-entropy loss function has advantages such as fast convergence and simple function differentiation. However, the intermittent sampling-resistant target detection task based on range Doppler maps suffers from the problem of positive and negative sample imbalance, where the area occupied by the target in the range Doppler map is usually much smaller than the area occupied by the background. The cross-entropy loss function is not suitable for this problem because it assigns equal weights to the loss of a single positive sample and the loss of a single negative sample. When the number of negative samples is much greater than the number of positive samples, the loss corresponding to the negative samples will also be much greater than the loss corresponding to the positive samples. This causes the model to tilt towards negative samples during training, resulting in a decrease in the model's performance in detecting positive samples. To address this problem, this embodiment of the invention proposes to introduce a Focal loss function to replace the cross-entropy loss function used in the foreground prediction part of the original YOLOX loss function, thereby improving the performance of the intermittent sampling-resistant target detection model under the condition of positive and negative sample imbalance. The focal loss function is used to address the problem of poor detection accuracy in first-order visual object detection algorithms under conditions of extreme sample imbalance. Focal loss increases the loss weight of hard-to-classify samples by adding an adjustment factor to the cross-entropy loss function, thereby making the model more focused on training on hard-to-classify samples. The expression for the focal loss is as follows (23):

[0143] L Focal =-((1-p) γ ·y·log(p)+p γ ·(1-y)·log(1-p))

[0144] In the formula, γ represents the adjustable coefficient. When y takes the value of 1, formula (23) can be expressed as in formula (24). When the training sample is an easily classified positive sample, the predicted value p will approach the label value 1, and the corresponding loss value will approach 0. If the value of γ is increased, the corresponding loss value will further decrease, that is, the contribution of easily classified positive samples to the total loss value will decrease, as shown in formula (24) below:

[0145] L Focal =-(1-p) γ ·log(p)

[0146] Similarly, when y is 0, formula (23) can be expressed as in formula (25). When the training sample is an easily classifiable negative sample, the predicted value p will approach the label value 0, and the corresponding loss value will also approach 0. If the value of γ is increased, the corresponding loss value will further decrease, that is, the contribution of easily classifiable negative samples to the total loss will decrease, as shown in formula (25):

[0147] L Focal =-p γ ·log(1-p)

[0148] In intermittent sampling-resistant target detection tasks, the imbalance between positive and negative samples leads to relatively high classification accuracy for the majority of negative samples (background samples) and relatively low classification accuracy for the minority of positive samples (true target samples). Focal loss dynamically reduces the loss value corresponding to samples with low classification difficulty during training, allowing the model to focus more on training on samples with high classification difficulty. This alleviates the imbalance problem in intermittent sampling-resistant target detection tasks and effectively improves the model's detection performance.

[0149] The overall loss function for the improved YOLOX-Nano network training process is explained below:

[0150] The loss function used in training the improved YOLOX-Nano network proposed in this embodiment of the invention mainly consists of three parts: foreground prediction loss, category prediction loss, and bounding box localization loss. The foreground prediction loss is calculated using the focal loss function. The expression of the foreground prediction loss function is shown in the following equation (26):

[0151]

[0152] In the formula, W and H are the width and height of the predicted feature map output by the network, respectively. i represents the index of the predicted bounding box. This represents the probability that the i-th prediction box contains the target. This represents the label value corresponding to the foreground prediction of the i-th prediction box.

[0153] The cross-entropy loss function is used to calculate the category prediction loss. The category prediction loss function is shown in equation (27) below:

[0154]

[0155] In the formula, P represents the number of positive sample prediction boxes. This indicates the probability that the predicted bounding box belongs to the true target category. This indicates the category label of the predicted bounding box. When the category of the target contained in the predicted bounding box is the true target, The value is 1. Otherwise It is 0.

[0156] The bounding box localization loss is calculated using IOU loss (intersection over union loss). The expression for the bounding box localization loss is shown in equation (28):

[0157]

[0158] In the formula, IOU i This represents the intersection-union ratio (IOU) between the i-th predicted bounding box and its corresponding label box. i The calculation expression is shown in equation (29):

[0159]

[0160] In the formula, I i It represents the area of ​​the intersection between the i-th predicted bounding box and its corresponding label box, such as... Figure 16 The intersection of the label box and the prediction box is shown (top left represents the label box, bottom right represents the prediction box). U i It represents the combined area of ​​the i-th prediction box and its corresponding label box.

[0161] S400. Determine the target prediction box based on the prediction box and a preset detection threshold value;

[0162] It should be noted that the preset detection threshold is determined by the false alarm rate index.

[0163] In a specific embodiment, among the 6400 predicted bounding boxes output by the visual object detection network, only those with a detection score greater than the detection threshold are selected as the final detection result (i.e., the target prediction box). The detection score is equal to the confidence that the predicted box contains the object to be detected multiplied by the category score of the object in the predicted box. The detection threshold is adjusted and set according to the false alarm rate index.

[0164] S500: Determine the detection result based on the parameters of the target prediction box;

[0165] It should be noted that the parameters of the target prediction bounding box include the confidence score of the object to be detected contained within the bounding box, the object's class score, and the coordinates of the center point of the bounding box; the detection results include the detection score, target distance, and target velocity; in some specific embodiments, such as Figure 4As shown, the output predicted feature map (i.e., the target feature map) has a dimension of 80×80×(bbox+obj+cls), where 80×80 corresponds to 6400 predicted boxes. The parameters of each predicted box are as follows: the value of bbox(bounding box, predicted box) is 4, which corresponds to the width, height, and horizontal and vertical coordinates of the center point of the predicted box; the value of obj(object, target) is 1, which corresponds to the confidence score of the target contained in the predicted box; since there is only one possible category for the predicted target, namely the true target, the value of cls(class, category) is 1, which corresponds to the score of the predicted target contained in the predicted box being the true target.

[0166] In some embodiments, a detection score is calculated based on the confidence level of the object to be detected contained within the target prediction box and the object's category score; the target distance is calculated based on the x-coordinate of the center point of the target prediction box, combined with the width of the range Doppler graph and the radar range detection range; and the target velocity is calculated based on the y-coordinate of the center point of the target prediction box, combined with the height of the range Doppler graph and the radar velocity range. The x-coordinate and y-coordinate of the center point are determined using the center point coordinates, and the radar range detection range and radar velocity range are determined based on the radar corresponding to the radar echo signal.

[0167] Specifically, the number of target prediction boxes selected is the number of targets detected by the network. The x-coordinate of the center point of each prediction box corresponds to the target's distance, and the y-coordinate corresponds to the target's velocity. The target's distance is obtained by dividing the x-coordinate of the prediction box's center point by the width of the distance-Doppler image and then multiplying it by the radar's range detection range. The target's velocity is obtained by dividing the y-coordinate of the prediction box's center point by the height of the distance-Doppler image and then multiplying it by the radar's velocity range.

[0168] In summary, this invention transforms the intermittent sampling-resistant target detection problem into a range-Doppler image-based visual target detection problem, and utilizes an improved YOLOX-Nano network to train a visual target detection network to achieve end-to-end intermittent sampling-resistant target detection, ranging, and velocity measurement. This method has the advantages of clear physical meaning, simple steps, integrated velocity and ranging measurement, and fewer training samples required. Furthermore, addressing the poor performance of existing deep learning-based intermittent sampling-resistant target detection methods under extremely low signal-to-noise ratio (SNR) conditions, a focal loss function is introduced to improve the recognition loss weight of samples under low SNR conditions. This allows this invention to maintain a high target detection accuracy even under low SNR conditions. Simultaneously, considering the characteristic that target signals are typically small targets on the range-Doppler image, this paper proposes a single-detector-head YOLOX-Nano visual target detection network. This network is obtained by trimming the detector head of the original YOLOX-Nano network. Compared to the original YOLOX-Nano network, the network proposed in this invention has fewer parameters and lower computational cost, making it more suitable for practical engineering applications.

[0169] On the other hand, embodiments of the present invention provide a target detection system resistant to intermittent sampling interference, comprising: a first module for acquiring radar echo signals; a second module for performing Doppler conversion processing on the radar echo signals to obtain a range Doppler map; wherein the Doppler conversion processing includes fast-time pulse compression processing and slow-time discrete Fourier transform processing; and a third module for analyzing the range Doppler map using a visual target detection network to obtain a target feature map; wherein the target feature map includes a preset number of prediction boxes; and the visual target detection network is based on an improved YOLOX-Nan... The improved YOLOX-Nano network, trained on labeled samples, includes a single detection head structure and a focal loss function. The fourth module determines the target prediction box based on a preset detection threshold, where the preset detection threshold is determined by the false alarm rate. The fifth module determines the detection result based on the parameters of the target prediction box, including the confidence score of the object within the box, the object's class score, and the coordinates of the center point of the box. The detection result includes the detection score, target distance, and target velocity.

[0170] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0171] Another aspect of the present invention provides a target detection device resistant to intermittent sampling interference, including a processor and a memory;

[0172] Memory is used to store programs;

[0173] The processor executes the program as described above.

[0174] The content of the method embodiments of the present invention is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0175] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0176] The content of the method embodiments of the present invention is applicable to the computer-readable storage medium embodiments. The specific functions implemented by the computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0177] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0178] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0179] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0180] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution means, apparatus, or device (such as a computer-based device, a processor-including device, or other means that can fetch and execute instructions from, or in conjunction with, an instruction execution means, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution means, apparatus, or device.

[0182] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0183] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0184] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0185] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0186] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A target detection method resistant to intermittent sampling interference, characterized in that, include: Acquire radar echo signals; The radar echo signal is subjected to Doppler conversion processing to obtain a range Doppler map; wherein, the Doppler conversion processing includes fast time-dimensional pulse compression processing and slow time-dimensional discrete Fourier transform processing; The distance Doppler image is analyzed using a visual target detection network to obtain a target feature map; wherein, the target feature map includes a preset number of prediction boxes; the visual target detection network is generated by training a modified YOLOX-Nano network with labeled samples, and the modified YOLOX-Nano network includes a single detection head structure and a focus loss function; Based on the predicted bounding box, a target predicted bounding box is determined by a preset detection threshold; wherein, the preset detection threshold is determined by the false alarm rate index. Based on the parameters of the target prediction box, a detection result is determined; wherein, the parameters of the target prediction box include the confidence score of the object to be detected contained within the target prediction box, the object's category score, and the coordinates of the center point of the target prediction box; the detection result includes the detection score, target distance, and target velocity; The step of determining the detection result based on the parameters of the target prediction bounding box includes: The detection score is calculated based on the confidence level of the objects to be detected contained within the target prediction box and the category score of the objects. The target distance is calculated based on the x-coordinate of the center point of the target prediction box, combined with the width of the range Doppler image and the radar range detection range. The target velocity is calculated based on the ordinate of the center point of the target prediction box, combined with the height of the range Doppler image and the radar velocity range. The horizontal and vertical coordinates of the center point are determined by the center point coordinates, and the radar range and radar speed range are determined based on the radar corresponding to the radar echo signal.

2. The target detection method against intermittent sampling interference according to claim 1, characterized in that, The process of performing Doppler conversion processing on the radar echo signal to obtain a range Doppler image includes: The radar echo signal is demodulated, and the demodulated radar echo signal is convolved with a matched filter to complete the fast time dimension pulse compression processing of the radar echo signal. The radar echo signal after pulse compression is subjected to slow-time discrete Fourier transform processing to obtain the range Doppler spectrum. The distance-Doppler spectrum is plotted to obtain the distance-Doppler map.

3. The target detection method against intermittent sampling interference according to claim 1, characterized in that, Also includes: A visual object detection network was created based on the improved YOLOX-Nano network.

4. The target detection method against intermittent sampling interference according to claim 3, characterized in that, The visual object detection network, based on the improved YOLOX-Nano network, includes: An improved YOLOX-Nano network is set up. Based on training samples, the improved YOLOX-Nano network is trained for prediction using the overall loss function. Based on the training results, the improved YOLOX-Nano network is adjusted to obtain a visual object detection network. The improved YOLOX-Nano network includes a backbone, a connection part, and a detection head part, wherein the detection head part uses a single detection head structure; the overall loss function includes foreground prediction loss, category prediction loss, and bounding box localization loss.

5. The target detection method against intermittent sampling interference according to claim 4, characterized in that, In the step of training the improved YOLOX-Nano network for prediction using the combined global loss function, the focus loss function is used to calculate the foreground prediction loss, the cross-entropy loss function is used to calculate the category prediction loss, and the cross-union loss is used to calculate the predicted bounding box localization loss.

6. The target detection method against intermittent sampling interference according to claim 1, characterized in that, The visual target detection network includes a backbone, a connector, and a detection head. The step of analyzing the distance Doppler image using the visual target detection network to obtain a target feature map includes: Multi-level feature extraction is performed on the distance Doppler image through the main part to obtain features at different levels; The different levels of features are fused through the connection portion to obtain fused features; The target feature map is obtained by performing regression and classification processing on the fused features through the detection head.

7. A target detection system resistant to intermittent sampling interference, characterized in that, include: The first module is used to acquire radar echo signals; The second module is used to perform Doppler conversion processing on the radar echo signal to obtain a range Doppler map; wherein, the Doppler conversion processing includes fast time dimension pulse compression processing and slow time dimension discrete Fourier transform processing; The third module is used to analyze the distance Doppler image using a visual target detection network to obtain a target feature map; wherein, the target feature map includes a preset number of prediction boxes; the visual target detection network is generated by training a modified YOLOX-Nano network with labeled samples, and the modified YOLOX-Nano network includes a single detection head structure and a focus loss function; The fourth module is used to determine the target prediction box based on the prediction box and a preset detection threshold; wherein the preset detection threshold is determined by the false alarm rate index. The fifth module is used to determine the detection result based on the parameters of the target prediction box; wherein, the parameters of the target prediction box include the confidence score of the object to be detected contained in the target prediction box, the class score of the object, and the coordinates of the center point of the target prediction box; the detection result includes the detection score, target distance, and target velocity; The step of determining the detection result based on the parameters of the target prediction bounding box includes: The detection score is calculated based on the confidence level of the objects to be detected contained within the target prediction box and the category score of the objects. The target distance is calculated based on the x-coordinate of the center point of the target prediction box, combined with the width of the range Doppler image and the radar range detection range. The target velocity is calculated based on the ordinate of the center point of the target prediction box, combined with the height of the range Doppler image and the radar velocity range. The horizontal and vertical coordinates of the center point are determined by the center point coordinates, and the radar range and radar speed range are determined based on the radar corresponding to the radar echo signal.

8. A target detection device resistant to intermittent sampling interference, comprising a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 6.

Images