Typical ground small target real-time detection and positioning method

Abstract

The invention discloses a real-time detection and positioning method for a typical small ground target, belongs to the technical field of target detection in deep learning, solves the problem of how to quickly realize detection, recognition and positioning of the small ground target, and adopts a plurality of convolution layers and a convolution attention fusion module. For a noisy hyperspectral image model, a convolutional layer is used for obtaining low-level feature information and channel adjustment, so that hyperspectral image features are effectively utilized; obtaining a noise residual image by using a U-shaped network in jump connection with a plurality of convolution attention fusion modules, wherein the noise residual image and the input noise image have the same shape; constructing a total loss function constrained by a global gradient regularization operator as an optimization target of a training model to train a model network structure; converting a pixel position of a model output result into latitude and longitude coordinates; and detection, identification and positioning of small ground targets can be rapidly realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection in deep learning and relates to a real-time detection and positioning method for typical small ground targets. Background Art

[0002] Object detection can be divided into two stages: feature extraction and object detection. The main purpose of the feature extraction stage is to extract the feature information of the object from the image or video, such as color, texture, shape, etc. The object detection stage detects the object based on the extracted feature information and obtains information such as the object detection box and category.

[0003] At present, there are still many challenges in detecting small targets. The main problem faced by using drones for intelligent reconnaissance is how to effectively identify and locate small targets, especially under complex backgrounds and diverse environmental conditions, such as small target size, changeable shapes, and easy confusion with the background. Summary of the invention

[0004] The technical problem to be solved by the present invention is how to quickly detect, identify and locate small ground targets.

[0005] A typical ground small target real-time detection and positioning method, the model structure includes multiple convolutional layers and convolutional attention fusion modules; for noisy hyperspectral image I∈R H×W×B The model uses convolutional layers to obtain low-level feature information and channel adjustment, which effectively utilizes the features of hyperspectral images; a U-shaped network with jump connections of multiple convolutional attention fusion modules is used to obtain the noise residual image I N ∈R H×W×B , the noise residual map has the same shape as the input noise image; construct the total loss function constrained by the global gradient regularization operator as the optimization target of the training model to train the model network structure; convert the pixel position of the model output result into longitude and latitude coordinates; where H×W represents the image size resolution and B represents the channel dimension.

[0006] Furthermore, the convolutional attention fusion module includes a local branch and a global branch;

[0007] The local branch is expressed as:

[0008] F conv =W 3×3×3 (CS(W 1×1 (Y))) (1)

[0009] Among them, F conv Represents the output information of the local branch, W 3×3×3 is a 3×3×3 convolution operation, CS represents the channel reorganization operation, W 1×1represents a 1×1 convolution operation, and Y represents the input feature information;

[0010] The global branch is expressed as:

[0011] F att =W 1×1 Attention(Q,K,V)+Y (2)

[0012] Attention(Q,K,V)=VSoftmax(KQ / α) (3)

[0013] Among them, α is a learnable scaling parameter, F att It represents the output information of the global branch, Attention() represents the feature information generated after the added attention mechanism operation, Softmax() represents the activation function, Q, K, and V represent the feature information formed by using 3×3 deep convolution operations on the input respectively.

[0014] Furthermore, the final output of the convolutional attention fusion module is:

[0015] F out =F att +F conv (4)

[0016] Among them, F conv Indicates the output information of the output local branch.

[0017] Furthermore, the method for constructing the total loss function constrained by the global gradient regularization operator is as follows:

[0018] Construct the loss function L S as follows:

[0019]

[0020] In formula (5), represents the estimated noise-free hyperspectral image, and I represents the noisy hyperspectral image;

[0021] In order to improve the quality of denoising and reduce redundancy, a global gradient regularization operator is used to constrain , as follows:

[0022]

[0023] In formula (6), and denote the gradient operators applied along the horizontal, vertical, and spectral axes, respectively;

[0024] The total loss function is as follows:

[0025] Lloss =L s +λL g (7)

[0026] In formula (7), λ is the control L g The weight parameter of the loss.

[0027] Furthermore, the method for converting the pixel position of the model output result into longitude and latitude coordinates is as follows:

[0028] The actual latitude and longitude coordinates corresponding to the pixel point (x, y) are unique and are set as (xLng, yLat). The data provided by the drone is used to calibrate the map:

[0029] (1) According to the height of the drone and the camera angle of view, the length Height1 and width Width1 of the image on the screen corresponding to the actual field are calculated as follows:

[0030]

[0031] Height1=deviceHeight*Width1 / deviceWidth (9)

[0032] (2) Calculate the normalized position (tmpX, tmpY) of the target pixel on the screen relative to the center point of the screen, as follows:

[0033] tmpX=droneHeight*2 / Width (10)

[0034] tmpY=-y / x*tmpX (11)

[0035] (3) According to the current position and nose direction of the UAV, the vertex coordinates of the actual scene are calculated as follows:

[0036] xLng=droneLng+tmpX*cos(droneHeading)

[0037] -tmpY*sin(droneHeading)*M(12)

[0038] yLat=droneLat+tmpY*cos(droneHeading)

[0039] -tmpX*sin(droneHeading)*M(13)

[0040] Among them, droneHeight represents the height of the drone relative to the ground, θ represents the camera viewing angle, deviceHeight and deviceWidth represent the graphics resolution on the screen, Width1 represents the width of the image on the screen corresponding to the real world, Height1 represents the length of the image on the screen corresponding to the real world, tmpX and tmpY represent the normalized position information of the target pixel on the screen relative to the center point of the screen, droneHeading represents the direction of the drone head, droneLat and droneLng represent the latitude and longitude of the drone, and M represents the conversion coefficient between angle and meter.

[0041] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above-mentioned typical ground small target real-time detection and positioning method, and the processor is configured to execute the program stored in the memory.

[0042] A storage medium stores a computer program, which, when executed by a processor, executes the steps of the above-mentioned typical small ground target real-time detection and positioning method.

[0043] The advantages of the present invention are:

[0044] The invention adopts multiple convolutional layers and convolutional attention fusion modules; for a noisy hyperspectral image model, a convolutional layer is used to obtain low-level feature information and channel adjustment, so that the hyperspectral image features are effectively utilized; a U-shaped network with jump connections of multiple convolutional attention fusion modules is used to obtain a noise residual map, and the noise residual map has the same shape as the input noise image; a total loss function constrained by a global gradient regularization operator is constructed as an optimization target of the training model to train the model network structure; the pixel position of the model output result is converted into longitude and latitude coordinates; the convolutional attention fusion module of the invention utilizes global context information and local features, learns a large amount of labeled data, continuously updates and optimizes the parameters of the network model, obtains the optimal model during training, and then inputs unlabeled data into the model, thereby quickly realizing target detection and recognition, realizing real-time transmission of drone video and GPS positioning, altitude, orientation and other parameters, and on the basis of target detection, mutually converts its pixel points and longitude and latitude, thereby realizing its positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a model structure diagram of a typical small ground target real-time detection and positioning method according to an embodiment of the present invention;

[0046] Figure 2 is a network structure diagram of a convolutional attention fusion module of a typical small ground target real-time detection and positioning method according to an embodiment of the present invention;

[0047] Figure 3 It is a flowchart of the unmanned aerial vehicle target positioning of the typical ground small target real-time detection and positioning method of the embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of the coordinate solution of the unmanned aerial vehicle target positioning of the typical ground small target real-time detection and positioning method of the embodiment of the present invention;

[0049] Figure 5 is a system development flow chart of a typical small ground target real-time detection and positioning method according to an embodiment of the present invention;

[0050] Figure 6 This is a system interface diagram of a typical small ground target real-time detection and positioning method according to an embodiment of the present invention;

[0051] Figure 7 This is a drone real-time detection verification effect diagram of a typical ground small target real-time detection and positioning method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0054] Embodiment 1

[0055] like Figure 1 As shown, the model structure of a typical small ground target real-time detection and positioning method of an embodiment of the present invention includes using multiple convolutional layers and convolutional attention fusion modules; for a noisy hyperspectral image I∈R H×W×B The model uses convolutional layers to obtain low-level feature information and channel adjustment, which effectively utilizes the features of hyperspectral images; a U-shaped network with jump connections of multiple convolutional attention fusion modules is used to obtain the noise residual image I N ∈R H×W×B , the noise residual map has the same shape as the input noise image; construct the total loss function constrained by the global gradient regularization operator as the optimization target of the training model to train the model network structure; convert the pixel position of the model output result into longitude and latitude coordinates; where H×W represents the image size resolution and B represents the channel dimension.

[0056] like Figure 2The figure shows the network structure of the convolutional attention fusion module, including local branches and global branches. In the local branch, in order to enhance cross-channel interaction and promote information fusion, 1×1 convolution is first used to adjust the channel size. After that, the channel denoising operation is performed and the input tensor is divided into multiple groups along the channel dimension.

[0057] Subsequently, a 3×3×3 convolution is used to extract features; the local branch is expressed as:

[0058] F conv =W 3×3×3 (CS(W 1×1 (Y))) (1)

[0059] In formula (1), F conv Represents the output information of the local branch, W 3×3×3 is a 3×3×3 convolution operation, CS represents the channel reorganization operation, W 1×1 represents a 1×1 convolution operation, and Y represents the input feature information.

[0060] Similarly, the global branch is expressed as:

[0061] F att =W 1×1 Attention(Q,K,V)+Y (2)

[0062] Attention(Q,K,V)=VSoftmax(KQ / α) (3)

[0063] In formula (3), α is a learnable scaling parameter used to control the size of the matrix product of K and Q before applying the Softmax function; F att It represents the output information of the global branch, Attention() represents the feature information generated after the added attention mechanism operation, Softmax() represents the activation function, Q, K, and V represent the feature information formed by using 3×3 deep convolution operations on the input respectively.

[0064] Therefore, the final output of the network structure of the convolutional attention fusion module is:

[0065] F out =F att +F conv (4)

[0066] Construct the loss function L S As the optimization goal of the training network, the loss function is expressed as follows:

[0067]

[0068] In formula (5), represents the estimated noise-free hyperspectral image, and I represents the noisy hyperspectral image.

[0069] In order to improve the quality of denoising and reduce redundancy, a global gradient regularization operator is used to constrain

[0070]

[0071] In formula (6), and denote the gradient operators applied along the horizontal, vertical, and spectral axes, respectively.

[0072] The total loss function is as follows:

[0073] L loss =L s +λL g (7)

[0074] In formula (7), λ is the control L g The weight parameter of the loss. Based on experimental experience, λ is set to 0.01 to maintain a balance between various loss terms.

[0075] If the output of the target detection model is only the location information of the target in the image (pixel position), it is not very practical. This information must be converted into geographic location (latitude and longitude) to provide accurate geographic coordinates. To this end, it is necessary to solve the problem of target positioning from the perspective of the drone, that is, to realize the conversion between pixel points and longitude and latitude.

[0076] like Figure 3 As shown in the figure, it is a flowchart of UAV target positioning. The conversion and calculation parameters of pixel position and geographic location in the UAV reconnaissance image include parameters such as the UAV height, orientation, and camera pitch angle. Figure 4 The figure shows the coordinate solution of the target positioning of the drone. From the perspective of the drone, the ground is a horizontal plane, the upper left corner of the screen is the origin (0,0), and the right and downward are positive. The actual latitude and longitude coordinates corresponding to the pixel point (x, y) on any screen are unique, set as (xLng, yLat), so that the data provided by the drone can be used to calibrate the map.

[0077] The calculation process is as follows:

[0078] (1) According to the height of the drone and the camera angle of view, calculate the length Height1 and width Width1 of the image on the screen corresponding to the actual field.

[0079]

[0080] Height1=deviceHeight*Width1 / deviceWidth (9)

[0081] (2) Calculate the normalized position (tmpX, tmpY) of the target pixel on the screen relative to the center point of the screen.

[0082] tmpX=droneHeight*2 / Width (10)

[0083] tmpY=-y / x*tmpX (11)

[0084] (3) Calculate the vertex coordinates on the ground based on the current position and nose direction of the drone.

[0085] xLng=droneLng+tmpX*cos(droneHeading)

[0086] -tmpY*sin(droneHeading)*M(12)

[0087] yLat=droneLat+tmpY*cos(droneHeading)

[0088] -tmpX*sin(droneHeading)*M(13)

[0089] Among them, droneHeight represents the height of the drone relative to the ground, θ represents the camera viewing angle, deviceHeight and deviceWidth represent the image resolution on the screen, Width1 represents the width of the image on the screen corresponding to the real world, Height1 represents the length of the image on the screen corresponding to the real world, tmpX and tmpY represent the normalized position information of the target pixel on the screen relative to the center point of the screen, droneHeading represents the direction of the drone head, droneLat and droneLng represent the latitude and longitude of the drone, and M represents the conversion coefficient between angle and meter.

[0090] like Figure 5 As shown in the figure, it is a development flow chart of the target detection and positioning system of the present invention. After selecting the input source, the system will call the media processor such as MediaHandler to configure the camera, read the video file or load the image file. The system will handle all necessary media encoding and decoding and format conversion to ensure that data in different formats can smoothly enter the next processing stage. The user only needs to click the corresponding button to automatically complete the whole process.

[0091] The next step in processing the media input source is a continuous frame processing loop. In this loop, the system first preprocesses each frame of the image, including adjusting the image size, converting the color space, and normalizing it to ensure that they meet the input standards of the model. Subsequently, the preprocessed images will be fed into the model for detection and recognition. When performing multi-target detection, counting and tracking, the system will display the detection results in real time, realize target tracking, and store the detection records in the database. As the model outputs the detection results, the user interface will be updated in real time. This includes displaying the detection bounding box, classification label, and displaying the detection statistics in the display elements of the interface.

[0092] like Figure 6 As shown, it is the interface diagram of the target detection and positioning system of the present invention. A user-friendly UI interface is designed using the visualization tool Qt Designer using Python-based PySide6. In the process of designing an interactive typical ground target real-time detection system based on HCAFM-YOLOv8, a modular and hierarchical approach is adopted to integrate the functions of user interface, media processing and deep learning model. The system design idea is built around user experience, aiming to provide an intuitive and responsive user interface while performing complex image processing and target detection tasks in the background. When the user starts this application, the MainWindow class is first instantiated to initialize the application interface, set the necessary parameters, and prepare to receive user input. Users can select their input source through this interface. The choice of input source is very flexible. It can be a real-time video stream, a pre-recorded video file, or even a single static image, or a drone Rtsp push stream, which makes the system widely used in different usage scenarios.

[0093] Based on the YOLOv8 network structure, the method of the present invention makes targeted improvements based on the characteristics of the data set, and proposes a hybrid convolution and attention network (HCANet) for HSI denoising. By utilizing global context information and local features, a convolution and attention fusion module (CAFM) is designed to enhance the modeling of global and local features, thereby quickly realizing target detection. A detection system with user-friendly functions is designed based on PySide6, which realizes target detection using multi-source input such as images, videos, network cameras, and Rtsp real-time streaming.

[0094] The YOLOv8 network structure is built on the cutting-edge progress of deep learning and computer vision, providing unparalleled performance in terms of speed and accuracy. With the continuous optimization and improvement of the YOLO series network structure for target detection, the recognition accuracy and real-time detection are continuously improved. On this basis, the YOLOv8 network structure solves many challenges of small targets in terms of accuracy and speed. Therefore, on the basis of YOLOv8, the present invention adopts the hybrid convolution and attention network (HCANet) denoising of hyperspectral image (HSI), and utilizes global context information and local features to achieve high-precision and fast target detection. It learns a large amount of labeled data, continuously updates and optimizes the parameters of the network model, obtains the optimal model during training, and then inputs the unlabeled data into the model, thereby quickly realizing the detection and recognition of the target. The real-time transmission of drone video and GPS positioning, altitude, orientation and other parameters is realized through the wireless module. On the basis of target detection, its pixel points are converted into longitude and latitude to achieve its positioning.

[0095] Experimental verification

[0096] Taking H-type equipment as an example to verify the method of the present invention, the entire data set contains 5360 images, including 4341 training images, 483 verification images, and 536 test images. The professional labeling tool LabelImg is used for manual labeling. Each image has a corresponding XML file, which mainly contains the labeling information, labeling box and other information of the image. The model trained with this data set performs better in recognition and positioning.

[0097] The experimental indicators adopt the evaluation index parameters commonly used in the target detection algorithm, such as precision, recall, and average precision. After 1000 iterations of training, the model precision and recall indicators tend to be stable, and the model training is completed. In order to compare and analyze the experimental results, Faster-RCNN, YOLOv5, and YOLOv8 models were used for comparison, and the experimental results are shown in Table 1. The hyperparameter adjustment remains consistent. Compared with Faster-RCNN and YOLOv5 and YOLOv8 models, the model of the present invention has a certain improvement in accuracy, recall, and average accuracy. Among them, the mAP50-95 indicators increased by 6.8 percentage points, 4 percentage points, and 2.1 percentage points, respectively.

[0098] Table 1 Comparison of experimental results of each model

[0099] Prediction Model precision recall mAP50(B) mAP50-95 Faster-RCNN 87.99 96.81 95.79 85.4 YOLOv5 99.3 98.74 99.43 88.2 YOLOv8 99.4 99.2 99.5 90.1 HCAFM-YOLOv8 99.5 99.4 99.5 92.2

[0100] In addition, an independent test set is used for testing, with 100 pictures in the test set, and 10 pictures are randomly selected, as shown in Table 2. Through the prediction result experiment of the model, it can be found that for the same detection target, the recognition score of the model of the present invention is higher, that is, the judgment is more accurate.

[0101] Table 2 Comparison of prediction results of different models for the same target

[0102] Test image HCAFM-YOLOv8 Faster-RCNN Yolov5 Yolov8 1.jpg 95,95 89,91 90,93 88,91 6.jpg 89 88 88 88 18.jpg 96 91 93 91 24.jpg 93,95,96 81,89,87 92,91,88 91,89,87 36.jpg 95,94,94 88,88,89 90,91,91 88,88,89 44.jpg 95 89 90 88 60.jpg 95,96,95 90,90,89 91,92,92 90,90,87 78.jpg 95 88 91 88 90.jpg 93,94,95 89,90,91 90,91,91 89,90,91 100.jpg 94,95 88,89 89,91 88,89

[0103] In summary, through the comparison of experimental results, it can be seen that the model algorithm used in the present invention has achieved the best effect in training, including the accuracy, recall rate and average accuracy have all achieved the optimal results.

[0104] like Figure 7 As shown in the figure, taking the real-time access of DJI drone as an example, the upper left corner is the returned video, and the right side is the real-time detection screen. First, the trained model algorithm is loaded into the reconnaissance drone. Secondly, the video shot by the DJI drone is transmitted back to the reconnaissance drone in real time. Finally, the model algorithm performs real-time target detection and positioning on the returned video. Through experimental verification, the model has achieved good real-time detection effect; this achievement can be promoted and applied to the reconnaissance and calibration drone system to provide accurate information for the detection and perception of typical ground targets; it can greatly improve the effectiveness of intelligence reconnaissance, promote the formation of a system of drone systems, and play a reliable and effective role.

[0105] Embodiment 2

[0106] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute a typical ground small target real-time detection and positioning method in Example 1, and the processor is configured to execute the program stored in the memory.

[0107] Embodiment 3

[0108] A storage medium stores a computer program, which, when executed by a processor, executes the steps of a typical small ground target real-time detection and positioning method in embodiment 1.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A typical small ground target real-time detection and positioning method, characterized in that: The model structure includes multiple convolutional layers and convolutional attention fusion modules; for noisy hyperspectral images I∈R H×W×B The model uses convolutional layers to obtain low-level feature information and channel adjustment, which effectively utilizes hyperspectral image features; Use a U-shaped network with multiple convolutional attention fusion modules and skip connections to obtain the noise residual image I N ∈R H×W×B , the noise residual map has the same shape as the input noise image; The total loss function constrained by the global gradient regularization operator is constructed as the optimization target of the training model to train the model network structure; the pixel position of the model output result is converted into longitude and latitude coordinates; where H×W represents the image size resolution and B represents the channel dimension.

2. A typical small ground target real-time detection and positioning method according to claim 1, characterized in that: The convolutional attention fusion module includes a local branch and a global branch; The local branch is expressed as: F conv =W 3×3×3 (CS(W 1×1 (Y))) (1) Among them, F conv Represents the output information of the local branch, W 3×3×3 is a 3×3×3 convolution operation, CS represents the channel reorganization operation, W 1×1 represents a 1×1 convolution operation, and Y represents the input feature information; The global branch is expressed as: F att =W 1×1 Attention(Q,K,V)+Y (2) Attention(Q,K,V)=VSoftmax(KQ / α) (3) Among them, α is a learnable scaling parameter, F att It represents the output information of the global branch, Attention() represents the feature information generated after the added attention mechanism operation, Softmax() represents the activation function, Q, K, and V represent the feature information formed by using 3×3 deep convolution operations on the input respectively.

3. A typical ground small target real-time detection and positioning method according to claim 2, characterized in that: The final output of the convolutional attention fusion module is: F out =F att +F conv (4) Among them, F conv Indicates the output information of the output local branch.

4. A typical small ground target real-time detection and positioning method according to claim 3, characterized in that: The method for constructing the total loss function constrained by the global gradient regularization operator is as follows: Construct the loss function L s as follows: In formula (5), represents the estimated noise-free hyperspectral image, and I represents the noisy hyperspectral image; In order to improve the quality of denoising and reduce redundancy, a global gradient regularization operator is used to constrain The details are as follows: In formula (6), and denote the gradient operators applied along the horizontal, vertical, and spectral axes, respectively; The total loss function is as follows: THE loss =L s +λL g (7) In formula (7), λ is the control L g The weight parameter of the loss.

5. A typical small ground target real-time detection and positioning method according to claim 4, characterized in that: The method for converting the pixel position of the model output result into longitude and latitude coordinates is as follows: The actual latitude and longitude coordinates corresponding to the pixel point (x, y) are unique and are set as (xLng, yLat). The data provided by the drone is used to calibrate the map: (1) According to the height of the drone and the camera angle of view, the length Height1 and width Width1 of the image on the screen corresponding to the actual field are calculated as follows: Height1=deviceHeight*Width1 / deviceWidth (9) (2) Calculate the normalized position (tmpX, tmpY) of the target pixel on the screen relative to the center point of the screen, as follows: tmpX=droneHeight*2 / Width (10) tmpY=-y / x*tmpX (11) (3) According to the current position and nose direction of the UAV, the vertex coordinates of the actual scene are calculated as follows: xLng=droneLng+tmpX*cos(droneHeading) -tmpY*sin(droneHeading)*M(12)yLat=droneLat+tmpY*cos(droneHeading) -tmpX*sin(droneHeading)*M(13) Among them, droneHeight represents the height of the drone relative to the ground, θ represents the camera viewing angle, deviceHeight and deviceWidth represent the graphics resolution on the screen, Width1 represents the width of the image on the screen corresponding to the real world, Height1 represents the length of the image on the screen corresponding to the real world, tmpX and tmpY represent the normalized position information of the target pixel on the screen relative to the center point of the screen, droneHeading represents the direction of the drone head, droneLat and droneLng represent the latitude and longitude of the drone, and M represents the conversion coefficient between angle and meter.

6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the typical ground small target real-time detection and positioning method as described in any one of claims 1 to 5, and the processor is configured to execute the program stored in the memory.

7. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the typical small ground target real-time detection and positioning method according to any one of claims 1 to 5 are executed.

Images