Target detection method and device, computer device, storage medium and product

Abstract

The application relates to a target detection method and device, computer equipment, a storage medium and a product. The method comprises the following steps: acquiring a range-Doppler image of a target echo of multiple scatterers in an interference background; performing deep extraction processing on the range-Doppler image based on a pre-constructed target detection model; and determining a target detection result according to an extraction result. The application directly performs target detection on the range-Doppler image through the constructed target detection model, without needing to perform anti-interference processing before target detection. Moreover, the target detection model can perform deep extraction on the range-Doppler image, improves the features of image data, and improves the accuracy of target detection through the extraction result.

Description

Technical Field

[0001] This application relates to the field of radar anti-jamming technology, and in particular to a target detection method, apparatus, computer equipment, storage medium, and product. Background Technology

[0002] Radar is an indispensable electronic device in modern remote sensing systems. It can be used for target detection by emitting electromagnetic waves to illuminate a target and receiving its echo, thereby obtaining information such as the distance, rate of change of distance, azimuth, and altitude from the target to the electromagnetic wave emission point. The modern electromagnetic environment is complex and ever-changing, and radar jamming techniques are constantly evolving. Faced with increasingly complex jamming scenarios, radar anti-jamming capabilities, especially in the field of radar signal processing, are receiving increasing attention from researchers.

[0003] Traditional radar anti-jamming methods rely on filtering to remove or suppress interference. However, in complex electromagnetic environments, the diverse and complex patterns of interference can lead to poor target detection performance. Summary of the Invention

[0004] Therefore, it is necessary to provide a target detection method, apparatus, computer equipment, storage medium, and product that can improve the detection effect in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a target detection method, which includes:

[0006] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0007] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0008] In one embodiment, the target detection model includes a backbone network, a feature pyramid network, and an output prediction layer. The backbone network includes multiple convolutional layers connected by residual connections. Based on the pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined based on the extraction result, including:

[0009] The distance-Doppler image is input into the backbone network, and the depth-deepening extraction process is performed through multiple convolutional layers in the backbone network to obtain the extraction result.

[0010] The extracted results are input into a feature pyramid network for multi-level fusion processing to obtain fused data;

[0011] The fused data is input to the prediction layer and then processed to obtain the target detection result.

[0012] In one embodiment, the extraction result includes high-dimensional features and low-dimensional features. The feature pyramid includes a convolutional network. The extraction result is input into the feature pyramid network for multi-level fusion processing to obtain fused data, including:

[0013] High-dimensional and low-dimensional features are input into a convolutional network. The convolutional network then upsamples the high-dimensional features layer by layer to obtain the upsampling result.

[0014] The upsampling results and low-dimensional features are fused using a convolutional network to obtain fused data.

[0015] In one embodiment, the output layer includes a decoupling head layer and a fusion transpose layer, which inputs the fused data into the output prediction layer for prediction processing to obtain the target detection result, including:

[0016] The fused data is input into the decoupling head layer for decoupling processing to obtain the decoupling result;

[0017] The decoupling results are input into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection results, which include target category output, target foreground output and anchor box output.

[0018] In one embodiment, the method further includes:

[0019] Obtain the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds.

[0020] The object detection model is obtained by training the training dataset and the initial model.

[0021] In one embodiment, the above-mentioned training based on the training dataset and the initial model to obtain the object detection model includes:

[0022] Input the sample distance-Doppler images from the training dataset into the initial model to obtain the predicted bounding box information;

[0023] By correlating and comparing the predicted bounding box information with the ground truth labeled bounding boxes, multiple positive sample predicted anchor boxes are obtained.

[0024] The target detection model is determined based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function.

[0025] In one embodiment, the above-mentioned determination of the target detection model based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function includes:

[0026] Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes;

[0027] Calculate the target loss value between the predicted anchor box and the true labeled box for each target positive sample according to the preset loss function;

[0028] If the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0029] In one embodiment, the preset loss function includes the intersection-union loss function and the binary cross-entropy loss function; the target loss value includes a first loss value, a second loss value, and a third loss value; the predicted bounding box information includes category output, foreground output, and anchor box output; and the target loss value between the predicted anchor box and the ground truth bounding box of each target positive sample is calculated according to the preset loss function, including:

[0030] Based on the binary classification cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined;

[0031] Based on the binary classification cross-entropy loss function, a second loss value is determined between the foreground output and the foreground output corresponding to the ground truth bounding box.

[0032] Based on the cross-union loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

[0033] Secondly, this application also provides a target detection device. The device includes:

[0034] The image acquisition module is used to acquire the range-Doppler image of the target echo from a multi-scatterer under interference background;

[0035] The result determination module is used to perform depth enhancement extraction processing on the range-Doppler image based on a pre-built target detection model, and determine the target detection result based on the extraction result; the target detection result is used to characterize the type of target echo.

[0036] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0037] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0038] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0039] Fourthly, this application also provides a computer-readable storage medium. This computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0040] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0041] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0042] Fifthly, this application also provides a computer program product. This computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0043] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0044] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0045] The aforementioned target detection methods, apparatus, computer equipment, storage media, and products first acquire range-Doppler images of target echoes from multiple scatterers under interference backgrounds. Then, based on a pre-constructed target detection model, depth-deepening extraction is performed on the range-Doppler images, and the target detection result is determined based on the extraction result. In this application, the target detection model directly performs target detection on the range-Doppler images without requiring anti-interference processing before target detection. Furthermore, the target detection model of this application can perform depth-deepening extraction on the range-Doppler images, improving the features of the image data, and thus enhancing the accuracy of target detection. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the application environment of a target detection method in one embodiment;

[0047] Figure 2 This is a flowchart illustrating a target detection method in one embodiment;

[0048] Figure 3 This is a flowchart illustrating the process of determining the target detection result in one embodiment;

[0049] Figure 4 This is a schematic diagram of a feature pyramid in one embodiment;

[0050] Figure 5 This is a flowchart illustrating the process of obtaining target detection results in one embodiment;

[0051] Figure 6 This is a flowchart illustrating the target detection method in another embodiment;

[0052] Figure 7 This is a flowchart illustrating the process of obtaining the target detection model in one embodiment;

[0053] Figure 8 This is a flowchart illustrating the process of determining a target detection model in one embodiment;

[0054] Figure 9 This is a flowchart illustrating the process of determining the target loss value in one embodiment;

[0055] Figure 10 This is a structural block diagram of a target detection device in one embodiment;

[0056] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] First, before introducing the technical solutions of the embodiments of this application in detail, the technical background on which the embodiments of this application are based will be introduced.

[0059] Radar is an indispensable electronic device in modern remote sensing systems. The modern electromagnetic environment is complex and ever-changing, and radar jamming techniques are constantly evolving. Faced with increasingly complex interference scenarios, radar anti-jamming capabilities, especially in the field of radar signal processing, are receiving increasing attention from researchers.

[0060] Traditional radar anti-jamming methods are based on filtering. These methods first require interference filtering or suppression. In complex electromagnetic environments, interference patterns are diverse and complex, posing a serious threat and challenge to interference filtering and suppression. Poor interference filtering and suppression will directly affect the subsequent CFAR (Constant False Alarm Rate) target detection performance.

[0061] Based on this, this application provides a target detection method, apparatus, computer equipment, storage medium, and product, aiming to solve the above-mentioned technical problems.

[0062] The target detection method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 acquires range-Doppler images of target echoes from multiple scatterers under interference backgrounds; based on a pre-built target detection model, it performs depth-deepening extraction processing on the range-Doppler images and determines the target detection result based on the extraction results. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0063] In one exemplary embodiment, such as Figure 2 As shown, this application provides a target detection method, which is applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps S201 to S202. Wherein:

[0064] S201, Obtain the range-Doppler image of the target echo from the multi-scatterer under interference background.

[0065] The target echo of the multi-scatterer can be either simulated data or measured data.

[0066] In this embodiment of the application, the terminal can first acquire the target echo of the multi-scatterer under the interference background, and then input the target echo into the range-Doppler (RD image) obtained after preprocessing such as pulse compression and coherent accumulation. Alternatively, it can acquire the range-Doppler image directly input by the user.

[0067] S202, based on a pre-built target detection model, performs depth-deepening extraction on the range-Doppler image and determines the target detection result based on the extraction result.

[0068] The target detection results are used to characterize the type of target echo. The target detection model can employ networks such as YOLOX, Fsater R-CNN, YOLOv4, and SSD. Depth enhancement is implemented based on the specific structure of the target detection model; however, this application does not specifically limit the structure of the target detection model.

[0069] In this embodiment, based on the range-Doppler image obtained in the above embodiments, the terminal can input the range-Doppler image into a pre-built target detection model, perform depth-deepening extraction processing, and determine the target detection result based on the extraction result. The target detection result can be obtained by inputting the extraction result into the detection model, or it can be obtained based on the correspondence between the extraction result and the detection result.

[0070] The aforementioned target detection method acquires a range-Doppler image of the target echo from a multi-scattering object under interference background; performs depth-deepening extraction on the range-Doppler image based on a pre-constructed target detection model; and determines the target detection result based on the extraction result. This application embodiment directly detects targets in the range-Doppler image using a constructed target detection model, without requiring anti-interference processing before target detection. Furthermore, the target detection model of this application can perform depth-deepening extraction on the range-Doppler image, improving the features of the image data, and thus enhancing the accuracy of target detection.

[0071] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 3 The target detection model in this application embodiment includes a backbone network, a feature pyramid network, and an output prediction layer. The backbone network includes multiple convolutional layers, which are connected by residual connections. This application embodiment relates to a process of performing depth-deepening extraction on a range-Doppler image based on a pre-built target detection model, and determining the target detection result based on the extraction result, including the following steps S301 to S302. Wherein:

[0072] S301, the distance-Doppler image is input into the backbone network, and the depth-deepening extraction process is performed through multiple convolutional layers in the backbone network to obtain the extraction result.

[0073] The backbone network can be either the Darknet53 backbone network of YOLOv3 or the backbone network of YOLOv4. This application does not specifically limit the type of backbone network. Darknet53 contains 53 convolutional layers, each layer's convolutional kernel consists of 1x1 and 3x3 convolutional kernels, and it also uses a residual connection method.

[0074] In this embodiment, based on the range-Doppler image obtained in the above embodiment, the range-Doppler image is input into the backbone network, and multiple convolutional layers with residual connections in the backbone network perform depth-deepening extraction on the range-Doppler image to obtain the extraction result.

[0075] S302, the extracted results are input into the feature pyramid network for multi-level fusion processing to obtain fused data.

[0076] In this context, the Feature Pyramid Network (FPN) is also known as the bottleneck layer, such as... Figure 4 As shown.

[0077] In this embodiment of the application, based on the extraction results obtained from the above embodiments, the extraction results can be input into the feature pyramid network for multi-level fusion processing. The multi-level fusion processing can be to fuse the extraction results with the acquired feature data, or to fuse the upsampled data and downsampled data of itself.

[0078] S302, the fused data input and output prediction layer is processed to obtain the target detection result.

[0079] In this embodiment of the application, based on the fusion data obtained in the above embodiments, the fusion data can be input into the prediction layer and pre-built prediction model for prediction processing, or the fusion data can be input into the prediction layer, the fusion data can be pre-processed, and the pre-processed fusion data can be predicted to obtain the target detection result.

[0080] In this embodiment, the backbone network structure allows for increased model depth without gradient explosion. Deep models enhance the ability to extract input features, improving model learning efficiency. The feature pyramid network enables multi-level fusion processing, enhancing the detection of small targets and resulting in high accuracy of the output target detection results.

[0081] In an exemplary embodiment, based on the above embodiments, the extraction result of this application embodiment includes high-dimensional features and low-dimensional features, and the feature pyramid includes a convolutional network. This application embodiment involves inputting the extraction result into the feature pyramid network for multi-level fusion processing to obtain fused data, specifically including:

[0082] High-dimensional and low-dimensional features are input into a convolutional network. The high-dimensional features are upsampled layer by layer through the convolutional network to obtain the upsampled results. The upsampled results and low-dimensional features are then fused through the convolutional network to obtain fused data.

[0083] In this embodiment, the Feature Pyramid Network (FPN), also known as the bottleneck layer, performs multi-scale fusion of features extracted by the backbone network, improving the efficiency of target feature extraction (especially for small targets). This network structure is also a convolutional network, fusing high-dimensional features from the backbone network's output layer with features extracted from the corresponding downsampling layers (low-dimensional features). Since high-dimensional features are rich in semantic information and have low resolution, they are advantageous for detecting large targets; low-dimensional features have less semantic information but higher resolution, making them advantageous for detecting small targets. Therefore, the FPN network, by combining high-dimensional and low-dimensional features, enhances the detection of small targets.

[0084] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 5The output layer of this application embodiment includes a decoupling head layer and a fusion transpose layer. This application embodiment relates to the process of inputting fused data into the output prediction layer for prediction processing to obtain target detection results, including the following S401 to S402. Wherein:

[0085] S401 inputs the fused data into the decoupling head layer for decoupling processing to obtain the decoupling result.

[0086] In this embodiment, the decoupled head, compared to the coupled head, can improve object detection accuracy and network convergence speed. The entire output prediction layer contains three decoupled heads, representing the feature information at three scales of the preceding FPN output. Each decoupled head contains three branches: object class output (cls_output), object foreground output (obj_out_put), and anchor box output (reg_output). Finally, the three branches are concatenated and fused to output the feature information of the decoupled head, which is the decoupling result.

[0087] S402, the decoupling result is input into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection result, which includes target category output, target foreground output and anchor box output.

[0088] In this embodiment, the decoupling results output by the three decoupling heads are concatenated, fused, and transposed again to obtain the two-dimensional prediction box information for target detection (number of prediction boxes * prediction box information (target category output, target foreground output, and anchor box output)).

[0089] In this embodiment of the application, setting a decoupling head can effectively improve the performance evaluation value of the output result.

[0090] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 6 The method in this application embodiment further includes the following S501 to S502. Wherein:

[0091] S501, Obtain the training dataset.

[0092] The training dataset includes sample range-Doppler images of multiple scattering target echoes under backgrounds of active suppression, active deception interference, and composite interference.

[0093] In this embodiment, the terminal first collects the dataset and divides it into a test dataset and a training dataset. The dataset is divided into a training sample set and a test sample set, with the number of samples in the training sample set being greater than that in the test set. The ratio of the training sample set to the test set is 9:1.

[0094] S502, train the object detection model based on the training dataset and the initial model to obtain the object detection model.

[0095] In this embodiment, the training dataset is input into the initial model to obtain the output result. Alternatively, the training data can be preprocessed before being input into the initial model. Then, based on the output result and a preset loss function, the network parameters of the model are updated to obtain the object detection model.

[0096] The embodiments of this application train the detection model to obtain a target detection model with higher accuracy.

[0097] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 7 This application embodiment relates to the process of training an object detection model based on a training dataset and an initial model, including the following steps S601 to S603. Wherein:

[0098] S601, input the sample distance-Doppler image from the training dataset into the initial model to obtain the prediction box information.

[0099] In this embodiment of the application, the sample distance-Doppler image in the training dataset is input into the initial model. The initial model first performs further depth extraction processing on the sample distance-Doppler image based on its own network results, and then generates prediction box information based on the extraction results.

[0100] S602, the predicted bounding box information and the ground truth labeled bounding box are correlated and compared to obtain multiple positive sample predicted anchor boxes;

[0101] In this embodiment, based on the predicted bounding box information obtained in the above embodiments, the predicted bounding box information and the ground truth bounding boxes can be correlated and compared to obtain multiple positive sample predicted anchor boxes. The ground truth bounding boxes can be pre-acquired or calculated on-site.

[0102] S603, determine the target detection model based on multiple positive sample predicted anchor boxes, ground truth labeled boxes and preset loss function.

[0103] In this embodiment, based on the positive sample predicted anchor boxes and ground truth labeled boxes obtained in the above embodiments, the positive sample predicted anchor boxes and ground truth labeled boxes can be compared according to a preset loss function to determine the target detection model. Alternatively, multiple positive sample predicted anchor boxes can be selected, and the suitable positive sample predicted anchor boxes can be compared with the ground truth labeled boxes to determine the target detection model.

[0104] In this embodiment, the detection performance of the target detection model is greatly improved by using multiple positive sample predictions of anchor boxes, ground truth labeled boxes, and a preset loss function.

[0105] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 8 This application embodiment relates to the process of determining an object detection model based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function, including the following steps S701 to S703. Wherein:

[0106] S701, Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes.

[0107] In this embodiment, the predicted bounding box information output by the network is obtained through the decoupling head during network training.

[0108] Anchors are needed to associate predicted bounding boxes with ground truth bounding boxes to calculate the loss function and update network parameters. Currently, the object detection industry mainly uses two methods: anchor-based and anchor-free. The first method, anchor-based, involves setting anchor generation rules based on the target feature map output by the coupling head. Each feature pixel generates multiple candidate anchor boxes, and each generated anchor is compared with the ground truth to determine the difference. Clearly, anchor-based methods have high computational overhead due to the need to generate a large number of predicted target anchor boxes. The second method, anchor-free, generates only one anchor box per feature pixel. Taking YOLOv3 as an example, each FPN output feature pixel generates at least three anchor boxes; therefore, it reduces the computational cost by two-thirds compared to the anchor-based method. To improve real-time performance and network lightweighting, the YOLOX network in this embodiment uses the anchor-free method for computation.

[0109] Anchor-free processing yields predicted anchor boxes for all outputs. A label assignment network is then used to correlate and compare all predicted anchor boxes with the input ground truth, selecting the positive sample predicted anchor boxes.

[0110] Since most predicted anchor boxes are negative samples, with only a small fraction being positive samples, the label assignment network employs two main methods to select positive predicted anchor boxes: preliminary screening and fine screening based on the Optimal Transport Assignment (SimOTA) method. The preliminary screening primarily includes the following two methods: 1. Center Point Method: The coordinates of the top-left and bottom-right corners of the ground truth are calculated to determine the ground truth range. Then, the sum of the distances (four values ​​in total) between the center point of the predicted anchor box and the two sides of the ground truth box is used to determine if it is greater than 0, thus extracting anchor boxes within the ground truth range. If it is greater than 0, the predicted anchor box is considered a positive sample. Any predicted anchor box whose center point is within the ground truth range is preliminarily identified as a positive anchor box. 2. Target Box Method: Using the center point of the ground truth as a reference, a square with a side length of 5 is set, and all anchor boxes within the square are selected.

[0111] The second step is fine-tuning, which mainly consists of three parts: loss function calculation, cost calculation, and SimOTA solution. For each selected positive sample anchor box, the loss (i.e., cls_loss, obj_loss, and reg_loss) is calculated based on the ground truth of the input target class output (cls_output), target foreground output (obj_out_put), and anchor box output (reg_output). The cost loss is obtained by weighted summing of cls_loss and reg_loss, as shown in formula (1).

[0112]

[0113] Among them, C ij , These represent the cost between the positive sample anchor box and the ground truth of the j-th target, the target class loss, and the target predicted anchor box size loss, respectively. λ is a weighting coefficient, typically set to 3. Positive sample predicted anchor boxes are then finely selected based on the cost loss and the SimOTA method for label assignment.

[0114] S702, calculate the target loss value between the predicted anchor box and the true labeled box of each target positive sample according to the preset loss function.

[0115] In this embodiment, based on the target positive sample predicted anchor boxes obtained from the above embodiments, the target loss value between each target positive sample predicted anchor box and the real labeled box can be calculated based on a preset loss function.

[0116] S703, if the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0117] In this embodiment, real-time performance and network lightweighting are improved by selecting a target positive sample prediction anchor box from multiple positive sample prediction anchor boxes.

[0118] In one exemplary embodiment, based on the above embodiments, please refer to... Figure 9 The preset loss function in this application embodiment includes the intersection-union loss function and the binary classification cross-entropy loss function. The target loss value includes a first loss value, a second loss value, and a third loss value. The predicted box information includes category output, foreground output, and anchor box output. This application embodiment involves the process of calculating the target loss value between the predicted anchor box of each target positive sample and the ground truth labeled box according to the preset loss function, including the following S801 to S803. Wherein:

[0119] S801, based on the binary classification cross-entropy loss function, determines the first loss value between the category output and the category output corresponding to the true bounding box.

[0120] In this embodiment, based on the category output of the target positive sample predicted anchor box obtained in the above embodiment, the category output and the ground truth labeled box are input into the binary cross-entropy loss function to obtain the first loss value between the category output and the category output corresponding to the ground truth labeled box.

[0121] S802, based on the binary classification cross-entropy loss function, determines the second loss value between the foreground output and the foreground output corresponding to the true bounding box.

[0122] In this embodiment, based on the foreground output of the predicted anchor box of the target positive sample obtained in the above embodiment, the foreground output of the class and the foreground output corresponding to the ground truth box are input into the binary cross-entropy loss function to obtain the second loss value between the foreground output and the foreground output corresponding to the ground truth box.

[0123] S803, based on the cross-union ratio loss function, determines the third loss value between the anchor box output and the anchor box output corresponding to the true annotation box.

[0124] In this embodiment, based on the anchor box output of the target positive sample prediction anchor box obtained in the above embodiment, the anchor box output and the anchor box output corresponding to the real labeled box are input into the binary classification cross-entropy loss function to obtain the third loss value between the anchor box output and the anchor box output corresponding to the real labeled box.

[0125] This application embodiment minimizes the loss by measuring the difference between the predicted results of the model and the actual labels, that is, the predicted results are as close as possible to the actual labels, thereby completing the task more accurately.

[0126] In an exemplary embodiment, based on the above embodiments, the method of this application embodiment further includes the following steps:

[0127] Step 1: Obtain the training dataset; input the sample distance-Doppler image from the training dataset into the initial model to obtain the predicted bounding box information; compare and correlate the predicted bounding box information with the ground truth labeled bounding boxes to obtain multiple positive sample predicted anchor boxes;

[0128] Step 2: Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes; based on the binary cross-entropy loss function, determine the first loss value between the class output of the positive sample prediction anchor box and the class output corresponding to the ground truth box; based on the binary cross-entropy loss function, determine the second loss value between the foreground output of the positive sample prediction anchor box and the foreground output corresponding to the ground truth box; based on the cross-union loss function, determine the third loss value between the anchor box output of the positive sample prediction anchor box and the anchor box output corresponding to the ground truth box.

[0129] Step 3: If the first loss value, the second loss value, or the third loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0130] Step 4: Obtain the range-Doppler image of the target echo from the multi-scatterer against the background of interference;

[0131] Step 5: Input the distance-Doppler image into the backbone network, and perform depth-deepening extraction through multiple convolutional layers in the backbone network to obtain high-dimensional and low-dimensional features;

[0132] Step 6: Input the high-dimensional features and low-dimensional features into the convolutional network, and perform upsampling on the high-dimensional features layer by layer through the convolutional network to obtain the upsampling results; and then fuse the upsampling results and low-dimensional features through the convolutional network to obtain fused data;

[0133] Step 7: Input the fused data into the decoupling head layer for decoupling processing to obtain the decoupling result; input the decoupling result into the fusion transpose layer for connection fusion and transpose processing to obtain the target detection result.

[0134] This application embodiment can perform high-accuracy, high-real-time end-to-end target detection based on range-Doppler (RD) images in complex interference backgrounds, improving target detection efficiency. Compared with several benchmark models such as Fsater R-CNN, YOLOv4, and SSD, this application has outstanding overall performance advantages, including average accuracy, recall, and precision. Moreover, it replaces the traditional target detection and interference suppression functions, improving target detection performance in interference backgrounds.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a target detection apparatus for implementing the target detection method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more target detection apparatus embodiments provided below can be found in the limitations of the target detection method described above, and will not be repeated here.

[0137] In one embodiment, such as Figure 10 As shown, a target detection device 900 is provided, comprising:

[0138] Image acquisition module 901 is used to acquire range-Doppler images of target echoes from multiple scatterers under interference background;

[0139] The result determination module 902 is used to perform depth enhancement extraction processing on the range-Doppler image based on a pre-built target detection model, and determine the target detection result based on the extraction result; the target detection result is used to characterize the type of target echo.

[0140] In one embodiment, the target detection model includes a backbone network, a feature pyramid network, and an output prediction layer. The backbone network includes multiple convolutional layers connected by residual connections. The result determination module 902 includes:

[0141] The extraction result determination unit is used to input the distance-Doppler image into the backbone network, and perform depth-deepening extraction processing through multiple convolutional layers in the backbone network to obtain the extraction result.

[0142] The data fusion determination unit is used to input the extracted results into the feature pyramid network for multi-level fusion processing to obtain fused data.

[0143] The result determination unit is used to input and output the fused data to the prediction layer for prediction processing to obtain the target detection result.

[0144] In one embodiment, the extraction result includes high-dimensional features and low-dimensional features, the feature pyramid includes a convolutional network, and the fused data determination unit includes:

[0145] The first result determination subunit is used to input high-dimensional and low-dimensional features into the convolutional network, and the convolutional network performs layer-by-layer upsampling on the high-dimensional features to obtain the upsampling result; and

[0146] The data fusion determination sub-unit is used to fuse the upsampling results and low-dimensional features through a convolutional network to obtain fused data.

[0147] In one embodiment, the output layer includes a decoupling header layer and a fusion transpose layer, and the result determination unit includes:

[0148] The second result determination sub-unit is used to input the fused data into the decoupling head layer for decoupling processing to obtain the decoupling result.

[0149] The target result determination subunit is used to input the decoupling result into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection result, which includes target category output, target foreground output and anchor box output.

[0150] In one embodiment, the above-mentioned apparatus further includes:

[0151] The training set acquisition module is used to acquire the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds.

[0152] The model determination module is used to train the object detection model based on the training dataset and the initial model.

[0153] In one embodiment, the above model determination module includes:

[0154] The information determination unit is used to input the sample distance-Doppler image from the training dataset into the initial model to obtain the prediction box information;

[0155] The positive anchor box determination unit is used to correlate and compare the predicted box information with the ground truth labeled box to obtain multiple positive sample predicted anchor boxes;

[0156] The model determination unit is used to determine the target detection model based on multiple positive samples, predicted anchor boxes, ground truth bounding boxes, and a preset loss function.

[0157] In one embodiment, the model determination unit includes:

[0158] The target anchor box determination sub-unit is used to select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes;

[0159] The target value determination subunit is used to calculate the target loss value between the predicted anchor box and the true labeled box of each target positive sample according to the preset loss function.

[0160] The model determines the sub-unit, which is used to adjust the parameters in the initial model if the target loss value does not meet the preset convergence condition, and continue training based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0161] In one embodiment, the preset loss function includes the intersection-union loss function and the binary cross-entropy loss function; the target loss value includes a first loss value, a second loss value, and a third loss value; the predicted bounding box information includes category output, foreground output, and anchor box output; and the target value determination subunit is specifically used for:

[0162] Based on the binary classification cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined;

[0163] Based on the binary classification cross-entropy loss function, a second loss value is determined between the foreground output and the foreground output corresponding to the ground truth bounding box.

[0164] Based on the cross-union loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

[0165] Each module in the aforementioned target detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0166] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a target detection method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0167] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0168] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0169] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0170] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0171] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0172] The distance-Doppler image is input into the backbone network, and the depth-deepening extraction process is performed through multiple convolutional layers in the backbone network to obtain the extraction result.

[0173] The extracted results are input into a feature pyramid network for multi-level fusion processing to obtain fused data;

[0174] The fused data is input to the prediction layer and then processed to obtain the target detection result.

[0175] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0176] High-dimensional and low-dimensional features are input into a convolutional network. The convolutional network then upsamples the high-dimensional features layer by layer to obtain the upsampling result.

[0177] The upsampling results and low-dimensional features are fused using a convolutional network to obtain fused data.

[0178] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0179] The fused data is input into the decoupling head layer for decoupling processing to obtain the decoupling result;

[0180] The decoupling results are input into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection results, which include target category output, target foreground output and anchor box output.

[0181] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0182] Obtain the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds.

[0183] The object detection model is obtained by training the training dataset and the initial model.

[0184] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0185] Input the sample distance-Doppler images from the training dataset into the initial model to obtain the predicted bounding box information;

[0186] By correlating and comparing the predicted bounding box information with the ground truth labeled bounding boxes, multiple positive sample predicted anchor boxes are obtained.

[0187] The target detection model is determined based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function.

[0188] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0189] Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes;

[0190] Calculate the target loss value between the predicted anchor box and the true labeled box for each target positive sample according to the preset loss function;

[0191] If the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0192] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0193] Based on the binary classification cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined;

[0194] Based on the binary classification cross-entropy loss function, a second loss value is determined between the foreground output and the foreground output corresponding to the ground truth bounding box.

[0195] Based on the cross-union loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0197] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0198] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0199] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0200] The distance-Doppler image is input into the backbone network, and the depth-deepening extraction process is performed through multiple convolutional layers in the backbone network to obtain the extraction result.

[0201] The extracted results are input into a feature pyramid network for multi-level fusion processing to obtain fused data;

[0202] The fused data is input to the prediction layer and then processed to obtain the target detection result.

[0203] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0204] High-dimensional and low-dimensional features are input into a convolutional network. The convolutional network then upsamples the high-dimensional features layer by layer to obtain the upsampling result.

[0205] The upsampling results and low-dimensional features are fused using a convolutional network to obtain fused data.

[0206] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0207] The fused data is input into the decoupling head layer for decoupling processing to obtain the decoupling result;

[0208] The decoupling results are input into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection results, which include target category output, target foreground output and anchor box output.

[0209] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0210] Obtain the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds.

[0211] The object detection model is obtained by training the training dataset and the initial model.

[0212] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0213] Input the sample distance-Doppler images from the training dataset into the initial model to obtain the predicted bounding box information;

[0214] By correlating and comparing the predicted bounding box information with the ground truth labeled bounding boxes, multiple positive sample predicted anchor boxes are obtained.

[0215] The target detection model is determined based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function.

[0216] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0217] Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes;

[0218] Calculate the target loss value between the predicted anchor box and the true labeled box for each target positive sample according to the preset loss function;

[0219] If the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0220] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0221] Based on the binary classification cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined;

[0222] Based on the binary classification cross-entropy loss function, a second loss value is determined between the foreground output and the foreground output corresponding to the ground truth bounding box.

[0223] Based on the cross-union loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

[0224] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0225] Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference;

[0226] Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo.

[0227] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0228] The distance-Doppler image is input into the backbone network, and the depth-deepening extraction process is performed through multiple convolutional layers in the backbone network to obtain the extraction result.

[0229] The extracted results are input into a feature pyramid network for multi-level fusion processing to obtain fused data;

[0230] The fused data is input to the prediction layer and then processed to obtain the target detection result.

[0231] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0232] High-dimensional and low-dimensional features are input into a convolutional network. The convolutional network then upsamples the high-dimensional features layer by layer to obtain the upsampling result.

[0233] The upsampling results and low-dimensional features are fused using a convolutional network to obtain fused data.

[0234] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0235] The fused data is input into the decoupling head layer for decoupling processing to obtain the decoupling result;

[0236] The decoupling results are input into the fusion and transpose layer for connection fusion and transpose processing to obtain the target detection results, which include target category output, target foreground output and anchor box output.

[0237] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0238] Obtain the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds.

[0239] The object detection model is obtained by training the training dataset and the initial model.

[0240] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0241] Input the sample distance-Doppler images from the training dataset into the initial model to obtain the predicted bounding box information;

[0242] By correlating and comparing the predicted bounding box information with the ground truth labeled bounding boxes, multiple positive sample predicted anchor boxes are obtained.

[0243] The target detection model is determined based on multiple positive sample predicted anchor boxes, ground truth labeled boxes, and a preset loss function.

[0244] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0245] Select the target positive sample prediction anchor box from multiple positive sample prediction anchor boxes;

[0246] Calculate the target loss value between the predicted anchor box and the true labeled box for each target positive sample according to the preset loss function;

[0247] If the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point the training ends and the target detection model is obtained.

[0248] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0249] Based on the binary classification cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined;

[0250] Based on the binary classification cross-entropy loss function, a second loss value is determined between the foreground output and the foreground output corresponding to the ground truth bounding box.

[0251] Based on the cross-union loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

[0252] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0253] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0254] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0255] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A target detection method, characterized in that, The method includes: Acquire range-Doppler images of target echoes from multiple scatterers against a background of interference; Based on a pre-built target detection model, depth-deepening extraction is performed on the range-Doppler image, and the target detection result is determined according to the extraction result; the target detection result is used to characterize the type of target echo. The target detection model includes a backbone network, a feature pyramid network, and an output prediction layer. The backbone network includes multiple convolutional layers connected by residual connections. The process of performing depth-deepening extraction on the range-Doppler image based on the pre-built target detection model and determining the target detection result based on the extraction result includes: inputting the range-Doppler image into the backbone network and performing depth-deepening extraction through multiple convolutional layers in the backbone network to obtain the extraction result; inputting the extraction result into the feature pyramid network for multi-level fusion processing to obtain fused data; and inputting the fused data into the output prediction layer for prediction processing to obtain the target detection result. The output prediction layer includes a decoupling head layer and a fusion transpose layer. The step of inputting the fused data into the output prediction layer for prediction processing to obtain the target detection result includes: inputting the fused data into the decoupling head layer for decoupling processing to obtain a decoupling result; and inputting the decoupling result into the fusion transpose layer for connection fusion and transpose processing to obtain the target detection result. The target detection result includes target category output, target foreground output, and anchor box output.

2. The method according to claim 1, characterized in that, The extraction results include high-dimensional features and low-dimensional features. The feature pyramid includes a convolutional network. The extraction results are input into the feature pyramid network for multi-level fusion processing to obtain fused data, including: The high-dimensional features and the low-dimensional features are input into the convolutional network, and the high-dimensional features are upsampled layer by layer through the convolutional network to obtain the upsampling result; and The upsampling result and the low-dimensional features are fused using the convolutional network to obtain the fused data.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Obtain the training dataset; the training dataset includes sample range-Doppler images of multiple scattering target echoes under active suppression, active deception interference, and composite interference backgrounds; The target detection model is obtained by training the training dataset and the initial model.

4. The method according to claim 3, characterized in that, The step of training the target detection model based on the training dataset and the initial model includes: Input the sample distance-Doppler images from the training dataset into the initial model to obtain the predicted bounding box information; The predicted bounding box information is correlated and compared with the ground truth labeled bounding boxes to obtain multiple positive sample predicted anchor boxes; The target detection model is determined based on multiple positive sample predicted anchor boxes, the ground truth labeled boxes, and a preset loss function.

5. The method according to claim 4, characterized in that, The step of determining the target detection model based on multiple predicted anchor boxes from positive samples, the ground truth bounding boxes, and a preset loss function includes: Select the target positive sample prediction anchor box from the plurality of positive sample prediction anchor boxes; Calculate the target loss value between the predicted anchor box of each target positive sample and the ground truth labeled box according to the preset loss function; If the target loss value does not meet the preset convergence condition, the parameters in the initial model are adjusted, and training continues based on the training dataset until the target loss value meets the preset convergence condition, at which point training ends and the target detection model is obtained.

6. The method according to claim 5, characterized in that, The preset loss function includes the intersection-over-union loss function and the binary cross-entropy loss function; the target loss value includes a first loss value, a second loss value, and a third loss value; the predicted bounding box information includes category output, foreground output, and anchor box output; and the step of calculating the target loss value between each predicted anchor box of the target positive sample and the ground truth labeled box according to the preset loss function includes: Based on the binary cross-entropy loss function, the first loss value between the category output and the category output corresponding to the true bounding box is determined; Based on the binary cross-entropy loss function, the second loss value between the foreground output and the foreground output corresponding to the ground truth bounding box is determined; Based on the cross-union ratio loss function, a third loss value is determined between the anchor box output and the anchor box output corresponding to the true annotation box.

7. A target detection device, characterized in that, The device includes: The image acquisition module is used to acquire the range-Doppler image of the target echo from a multi-scatterer under interference background; The result determination module is used to perform depth-deepening extraction processing on the range-Doppler image based on a pre-built target detection model, and determine the target detection result based on the extraction result; the target detection result is used to characterize the type of target echo. The target detection model includes a backbone network, a feature pyramid network, and an output prediction layer. The backbone network comprises multiple convolutional layers connected by residual connections. The process of performing depth-deepening extraction on the range-Doppler image based on the pre-built target detection model and determining the target detection result based on the extraction result includes: inputting the range-Doppler image into the backbone network, performing depth-deepening extraction through multiple convolutional layers in the backbone network to obtain the extraction result; and inputting the extraction result into the feature pyramid network for multi-layer processing. The process involves a fusion process to obtain fused data; the fused data is then input into the output prediction layer for prediction processing to obtain the target detection result; the output prediction layer includes a decoupling head layer and a fusion transpose layer; the process of inputting the fused data into the output prediction layer for prediction processing to obtain the target detection result includes: inputting the fused data into the decoupling head layer for decoupling processing to obtain a decoupling result; and inputting the decoupling result into the fusion transpose layer for connection fusion and transpose processing to obtain the target detection result, which includes target category output, target foreground output, and anchor box output.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements 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 computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images