Four-rotor unmanned aerial vehicle 3D target detection method based on detection motor characteristics

Abstract

The invention discloses a four-rotor unmanned aerial vehicle 3D target detection method based on motor feature detection. The method comprises the following steps: constructing a four-rotor unmanned aerial vehicle motor data set according to an image collected by a camera; performing target labeling on the motor data set of the quadrotor unmanned aerial vehicle so as to construct a target detection data set; constructing a quad-rotor unmanned aerial vehicle motor recognition model based on an efficient layer aggregation network fused with an attention mechanism, and performing training based on the target detection data set; and based on the trained quadrotor unmanned aerial vehicle motor identification model, realizing unmanned aerial vehicle 3D target detection. And the detection accuracy and robustness of the unmanned aerial vehicle under high-altitude and small-scale target conditions are effectively improved. According to the invention, based on the detected motor characteristics and the geometric characteristics of the quadrotor unmanned aerial vehicle structure, 3D detection of the unmanned aerial vehicle is realized, and reliable technical support is provided for robust detection in a complex background, shielding interference and multi-target scene.

Description

Technical Field

[0001] This invention belongs to the field of image target detection technology, and in particular relates to a 3D target detection method for quadcopter UAVs based on the detection of motor features. Background Technology

[0002] In the fields of intelligent sensing and aerial image processing, UAV recognition and structural analysis technologies are gradually becoming a research hotspot, especially in applications such as low-altitude target management, air traffic monitoring, border patrol, and target monitoring in complex environments, where the demand for 3D detection and recognition technology for quadcopter UAVs is becoming increasingly urgent. With the frequent increase in low-altitude flight activities, how to achieve robust detection of quadcopter UAVs in complex backgrounds, occlusion interference, and multi-target scenarios has become a key problem that urgently needs to be solved.

[0003] However, most existing UAV detection technologies are based on two-dimensional image recognition methods, which have significant limitations in acquiring spatial information about targets. Specifically, two-dimensional image recognition methods can only provide planar position information of the target, making it difficult to accurately reconstruct the three-dimensional structure of the UAV and achieve high-precision spatial positioning. In complex scenarios, such as those with occlusion, changes in lighting, or interference from multiple targets, the detection accuracy and robustness of traditional two-dimensional image recognition methods will drop significantly, making it difficult to meet the high-precision requirements for UAV structure reconstruction and spatial positioning in practical applications.

[0004] Furthermore, existing technologies also have shortcomings in handling small target detection. When quadcopter drones fly at high altitudes, their size in images is relatively small, making it difficult for traditional two-dimensional image recognition methods to effectively extract features of small targets, resulting in low detection accuracy. This limitation is particularly prominent in practical application scenarios such as low-altitude target management, because drones often appear as small-scale targets in these scenarios, with complex backgrounds and many interfering factors.

[0005] In summary, existing technologies lack robust detection capabilities in complex backgrounds, occlusion interference, and multi-target scenarios, particularly exhibiting significant shortcomings in acquiring target spatial information and processing small target detection. Therefore, a novel UAV 3D target detection method is urgently needed. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a 3D target detection method for quadcopter UAVs based on motor characteristics. This method is based on the YOLOv8x backbone network and combines the spatial distribution characteristics of UAV motors to design a recognition method suitable for small target detection scenarios, effectively improving the detection accuracy and robustness of UAVs under high-altitude and small-scale target conditions.

[0007] This invention proposes a 3D target detection method for quadcopter UAVs based on detecting motor features, comprising the following steps:

[0008] A dataset of quadcopter drone motors was constructed based on images captured by the camera.

[0009] Target annotation is performed on the quadcopter drone motor dataset to construct a target detection dataset;

[0010] A high-efficiency layer aggregation network based on the fusion attention mechanism is used to construct a quadcopter UAV motor recognition model, which is then trained on the target detection dataset.

[0011] Based on the trained quadcopter drone motor recognition model, 3D target detection of drones is achieved.

[0012] Optionally, the quadcopter drone motor recognition model includes a backbone network, a neck network, and a head network, wherein the neck network incorporates a Swing Transformer module and an efficient layer aggregation network that integrates attention mechanisms, and the head network incorporates a small target detection layer.

[0013] Optionally, the backbone network includes a CBS module, a C2F module, and an SPPF module, wherein:

[0014] The CBS module includes lightweight convolution, normalization, and SiLU activation functions for extracting primary features;

[0015] The C2F module serves as the core structure for extracting deep semantic information.

[0016] The SPPF module employs a residual design to extract multi-scale features through three consecutive 5×5 max pooling operations, and then concatenates the features before and after pooling in the output stage, preserving local details while compressing the feature map size.

[0017] Optionally, the basic units of the Swin Transformer module introduced by the neck network include window multi-head self-attention, shift window multi-head self-attention, multilayer perceptron, and layer normalization.

[0018] Optionally, the efficient layer aggregation network with fusion attention mechanism introduced by the neck network effectively integrates features from different levels through a skip connection strategy, and uses coordinate attention, channel attention and spatial attention to assign weights to multi-layer input features in different dimensions.

[0019] Optionally, the process of assigning weights to multi-layer input features using coordinate attention, channel attention, and spatial attention in different dimensions includes:

[0020] First, coordinate attention and channel attention are combined to obtain the joint weight representation of position and channel dimensions. Then, coordinate attention and spatial attention are fused to obtain the joint weight representation of position and spatial dimensions. Finally, based on the efficient layer aggregation strategy, the obtained joint weight representation of position and channel dimensions and the joint weight representation of position and spatial dimensions are fused to obtain the output features.

[0021] Optionally, the process of realizing 3D target detection for a drone based on the trained quadcopter drone motor recognition model includes:

[0022] Based on the detected features of the quadcopter drone motor, extract the bounding rectangle of the motor target;

[0023] By representing the motor with the center point of the top border of each circumscribed rectangle, the top surface coordinates of the four motors are obtained.

[0024] Based on the known relationship between the side length and height of the quadcopter drone, calculate the length of the base and the pixel length corresponding to the height in the image;

[0025] Assuming the height direction is parallel to the downward direction of the image's y-axis, the offset vector of the quadcopter drone to the y-axis is obtained based on the pixel length corresponding to the height in the image;

[0026] Subtract the offset vector from the top surface position coordinates of the four motors to obtain the corresponding bottom surface position coordinates of the four motors;

[0027] Based on the top and bottom position coordinates of the four motors, the 3D detection coordinates of the quadcopter UAV motors are finally obtained.

[0028] Optionally, the quadcopter UAV motor recognition model adopts the SIoU loss function, which includes angle cost, distance cost, shape cost, and intersection-union ratio cost.

[0029] Optionally, the detection and evaluation indicators of the quadcopter UAV motor identification model include accuracy, recall, average precision, and mean average precision.

[0030] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] This invention addresses the shortcomings of the YOLOv8 target detection algorithm in small target applications by proposing an improved YOLOv8-based quadcopter drone motor recognition model. Using the YOLOv8x model as the base network, a high-efficiency aggregation network with a fusion attention mechanism is introduced into the neck network, effectively enhancing feature fusion capabilities through positional, channel, and spatial attention mechanisms. Then, a Swin Transformer is introduced into the neck network, and a small target detection layer is added to the head network to enhance the extraction capability for small targets. Finally, the SIoU loss function is used to improve the regression accuracy between predicted and ground truth bounding boxes in the target detection algorithm. This method significantly improves the problems of insufficient feature representation and low detection accuracy of quadcopter drone motors. Based on the detected motor features and the geometric characteristics of the quadcopter drone structure, 3D detection of the drone is also achieved, providing reliable technical support for robust detection in complex backgrounds, occlusion interference, and multi-target scenarios. Attached Figure Description

[0033] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0034] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the quadcopter drone motor detection network structure based on the improved YOLOv8 according to an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the basic CBS convolutional structure according to an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the C2F structure according to an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the SPPF structure according to an embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram of the ELAN-A module network structure according to an embodiment of the present invention;

[0040] Figure 7 This is a schematic diagram of the Swing Transformer structure according to an embodiment of the present invention. Detailed Implementation

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0042] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0043] Example 1

[0044] This embodiment provides a 3D target detection method for quadcopter UAVs based on detecting motor features, such as Figure 1 As shown, the specific steps include:

[0045] Step 1: Construct a quadcopter drone motor dataset based on images acquired by the camera;

[0046] Step 2: Label image features. Label the quadcopter motors in each image in the quadcopter motor dataset to construct the target detection dataset, and divide the training set and test set in an 8:2 ratio.

[0047] Step 3: Construct a quadcopter UAV motor recognition model based on an efficient layer aggregation network with a fusion attention mechanism;

[0048] Step 4: Train the quadcopter drone motor recognition model built in Step 3;

[0049] Step 5: Use the quadcopter drone motor recognition model trained in Step 4 to test the test set and verify it through the test evaluation indicators.

[0050] Step 6: Based on the detection motor features and the geometric characteristics of the quadcopter UAV structure, realize 3D inspection of the UAV.

[0051] This embodiment proposes a reliable identification method for quadcopter drone motors based on the improved YOLOv8x framework, such as... Figure 2 As shown. The improved YOLOv8x framework includes a backbone network, a neck network, and a head network. The neck network introduces a Swin Transformer module and an efficient layer aggregation network (ELAN-A module) that incorporates an attention mechanism; the head network introduces a small object detection layer; the backbone network includes CBS, C2F, and SPPF modules, with the structure shown below. Figure 3 , Figure 4 and Figure 5 As shown in the figure. Finally, based on the detection motor features and the geometric characteristics of the quadcopter UAV structure, 3D inspection of the UAV is achieved.

[0052] To improve the recognition accuracy of quadcopter drones without increasing the complexity of the YOLOv8 network, this embodiment proposes an efficient layer aggregation network structure based on an attention mechanism, namely the ELAN-A module, whose structure is as follows: Figure 6 As shown, this module enhances feature extraction and fusion capabilities by introducing multiple attention mechanisms. Specifically, the ELAN-A module effectively integrates features from different levels through a skip connection strategy and uses coordinate attention (CA), channel attention (ECA), and spatial attention (ESA) to assign weights to multi-layer input features in different dimensions, thereby improving the network's ability to perceive key information. The three attention mechanisms complement each other, jointly enhancing the network's ability to model complex features.

[0053] In the overall architecture, the ELAN-A modules are deployed at three key locations in the neck network to enhance the aggregation and propagation of multi-scale features. This structure effectively improves the model's accuracy and robustness in small object detection tasks. Its input-output relationship and computational flow are as follows: Figure 3 As shown, the relevant operation definitions are as follows:

[0054]

[0055] F y =CBS1(F x-1 )+CBS2(F x-1 )+F x +F x+1 ,

[0056] in F represents multiplication. x ,F x+1 ∈R C×H×W This is an intermediate state of the data. The position weight W is obtained during the CA stage. h1 W h2 ∈R 1×H×1 and W w1 W w2 ∈R 1×1×W Obtain the channel weight W in ECA. c ∈R C×1×1 Obtain spatial weights W in ESA s ∈R 1×H×W This embodiment effectively enhances the network's ability to focus on input features by introducing a multi-dimensional attention mechanism, and achieves efficient fusion of features from different layers, thereby enhancing the overall feature representation capability.

[0057] The specific process is as follows: First, the joint weight representation of position and channel dimensions is obtained by combining the CA and ECA modules. Then, the joint weight representation of position and spatial dimensions is obtained by fusing the CA and ESA modules. The ESA module is a lightweight spatial attention structure extended from the ECA mechanism in this embodiment. Finally, based on the efficient layer aggregation strategy of the ELAN module, the multi-dimensional features guided by the above weights are fused to obtain the output features. This method fully integrates the advantages of the ELAN structure and multiple attention mechanisms, effectively enhancing the feature integration capability of the neck network and providing support for improving overall detection performance.

[0058] The feasible overall architecture of the Swin Transformer consists of four phases, such as... Figure 7 As shown, each stage downsamples the feature map layer by layer through Patch Merging, gradually reducing the spatial resolution while increasing the channel dimension to construct a hierarchical feature representation. The input image is first divided into several non-overlapping image patches. Each patch is transformed into a corresponding feature vector through linear mapping, achieving preliminary feature embedding. This process is similar to patch embedding in Vision Transformer (ViT), but Swin Transformer uses a finer-grained partitioning strategy, which helps to preserve richer local structural information.

[0059] Furthermore, its basic units consist of Windowed Multi-Head Self-Attention (W-MSA), Shifted Windowed Multi-Head Self-Attention (SW-MSA), Multilayer Perceptron (MLP), and Layer Normalization (LN). W-MSA performs self-attention computation within a local window to reduce overall computational overhead; SW-MSA enhances cross-window information interaction by introducing a sliding window mechanism; the MLP module is used to perform nonlinear feature transformations; and LN provides normalization processing in each layer to promote stable model training and accelerate convergence.

[0060] The feasible backbone network integrates three key structures: CBS, C2F, and SPPF. The CBS module consists of lightweight convolution, normalization, and SiLU activation functions, used for primary feature extraction. The C2F module, as the core structure, is responsible for extracting deep semantic information. The SPPF module adopts a residual design, extracting multi-scale features through three consecutive 5×5 max pooling operations, and concatenating the features before and after pooling in the output stage. This preserves local details while compressing the feature map size, thereby enhancing the ability to perceive targets at different scales.

[0061] In practice, the YOLOv8 proposed in this embodiment originally used the CIoU loss function, which obtains the target position by dividing the Euclidean distance from the center of the predicted bounding box to the center of the ground truth bounding box by the diagonal length of the smallest bounding box. However, when the target size is small, the center point deviation contributes little to the IoU, and CIoU has a weak optimization effect on small target detection. For example, airport scenes are characterized by a large number of small-scale targets, making the use of CIoU limited. Therefore, in order to obtain the optimal model parameters, the model proposed in this embodiment uses the SIoU loss function. SIoU further considers the vector angle between the ground truth bounding box and the predicted bounding box, improving the model's ability to detect small targets. SIoU specifically includes four parts: angle cost, distance cost, shape cost, and intersection-over-union ratio (IoU) cost.

[0062] In SIoU, the angle loss function calculates the loss based on the angle between the ground truth bounding box and the predicted bounding box. During training, this allows the predicted bounding box to quickly drift towards the nearest axis, and then regress to the other coordinate. This mechanism effectively improves convergence speed. When the angle is equal to the predicted angle, minimize α first; otherwise, minimize β first, where α is the angle between the ground truth bounding box and the predicted bounding box, and β is the complementary angle of α. Angle cost is defined as follows:

[0063]

[0064] Where Λ is the angle loss function, and These are the coordinates of the center point of the true bounding box. and These are the coordinates of the center point of the prediction box. C h σ represents the vertical distance between the ground truth bounding box and the predicted bounding box, and σ represents the straight-line center distance between the ground truth bounding box and the predicted bounding box.

[0065] Based on the consideration of angle cost, the distance cost is redefined; the closer the angle α is to the distance cost, the greater the distance cost.

[0066] The greater the contribution of distance cost to the loss function, the closer the predicted bounding box is to the ground truth bounding box. Shape cost considers the shape difference between the ground truth bounding box and the predicted bounding box, constraining the aspect ratio of the box to make the predicted box closer to the ground truth box in shape. The SIoU loss function is defined as:

[0067]

[0068] Here, Δ represents the distance cost, and Ω represents the shape cost. Through the synergistic effect of these components, the SIoU loss function can align bounding boxes with higher accuracy. For example, in airport target detection tasks, SIoU can better perform its function in accurately locating targets, whether they are small people and vehicles or large aircraft, providing strong support for improving the model's detection accuracy and training efficiency, thus enabling accurate target detection in complex airport scenarios.

[0069] Experimental environment and evaluation index verification:

[0070] The experimental environment for this embodiment is the SCNet cloud platform environment, with the server equipped with a Hygon C86718532C CPU and an NVIDIA GeForce RTX3080 GPU. The input dataset images were resized to 800x800 pixels. An Adam optimizer with an initial learning rate of 0.001 and a weight decay rate of 0.0005 was used. Four evaluation metrics were used: precision, recall, average precision (AP), and mean average precision (mAP). In this embodiment, the IoU threshold for AP and mAP was 0.5 to determine whether the target was detected correctly.

[0071] Accuracy is the ratio of the number of correctly predicted samples to the total number of positive samples detected by the model. The formula is:

[0072]

[0073] Recall is the ratio of the number of correctly predicted samples to the actual number of positive samples. The formula is:

[0074]

[0075] Average precision is the area under the PR curve, calculated using the following formula:

[0076]

[0077] Therefore, when the number of all classes tested is n, the formula for calculating the overall class mAP is:

[0078]

[0079] The feasible process for 3D target detection of quadrotor drones based on a trained quadrotor drone motor recognition model includes:

[0080] This embodiment considers 3D target detection of a quadcopter drone as an approximation of a cuboid mapping in an image. Based on known information, this embodiment transforms 3D target detection into a cuboid modeling problem in the image. It is assumed that the height projection direction of the cuboid is vertically upward. Based on the detection results of the drone motors, the bounding rectangles of each motor target are extracted. To simplify the feature point description, the center point of the upper bounding box of each detection box is used to approximate the motor, ultimately obtaining the two-dimensional position coordinates of the four motors: p0 = (x0, y0), p1 = (x1, y1), p2 = (x2, y2), p3 = (x3, y3), which are used for subsequent structure reconstruction and pose analysis. Given that the relationship between the drone's side length and height is r, the base length L is calculated, taking the side length from p0 to p1 as an example:

[0081]

[0082] Next, calculate the pixel length H corresponding to the height in the image based on the proportion:

[0083] H = r·L;

[0084] Then, assuming the height direction is parallel to the downward direction of the image's y-axis, the offset vector along the y-axis is obtained.

[0085] Finally, since the acquisition motor is located at the top of the drone, the height offset vector above is subtracted from the four points on the top surface to obtain the coordinates of the bottom vertex: p4 = (x4, y4), p5 = (x5, y5), p6 = (x6, y6), p7 = (x7, y7).

[0086]

[0087] The final 3D detection coordinates of the quadcopter UAV are [p0, p1, p2, p3, p4, p5, p6, p7]. Compared with other multi-sensor 3D detection methods, the method proposed in this embodiment, based on a pure vision approach, has the advantages of low cost and simplicity.

[0088] Analysis of experimental results:

[0089] To verify the effectiveness of the improved target detection algorithm proposed in this embodiment, the experimental results are shown in Table 1 below:

[0090] Table 1

[0091] category P(％) R(％) mAP(%) YOLOv3 88.3 79.3 82 YOLOv4 90.8 83.6 85.7 YOLOv5s 93.4 77.9 83.6 YOLOv8 (anchor) 87.9 55.2 86.1 Ours 93.4 88.8 90.9

[0092] Experimental results were compared with those of YOLOv3, YOLOv4, YOLOv5s, YOLOv8x, and the proposed method. Experimental results show that the proposed method for small target detection in airport scenarios achieves a 4.8% improvement in mAP compared to the original network, and exhibits the best performance across all evaluation metrics. This embodiment demonstrates an 8.9% improvement in mAP compared to YOLOv3, a 5.2% improvement compared to YOLOv4, and a 7.3% improvement compared to YOLOv5s.

[0093] This embodiment, through the integrated design of the above-mentioned network structure and processing strategy, not only enhances the model's ability to perceive small-scale features, but also improves the modeling accuracy of the quadcopter's structural consistency, providing reliable technical support for subsequent 3D scene reconstruction and aircraft positioning.

[0094] Example 2

[0095] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0096] Example 3

[0097] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0098] Example 4

[0099] This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0100] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A 3D target detection method for a quadcopter UAV based on detecting motor features, characterized in that, Includes the following steps: A dataset of quadcopter drone motors was constructed based on images captured by the camera. Target annotation is performed on the quadcopter drone motor dataset to construct a target detection dataset; A high-efficiency layer aggregation network based on the fusion attention mechanism is used to construct a quadcopter UAV motor recognition model, which is then trained on the target detection dataset. Based on the trained quadcopter drone motor recognition model, 3D target detection of drones is achieved.

2. The method according to claim 1, characterized in that, The quadcopter UAV motor recognition model includes a backbone network, a neck network, and a head network. The neck network incorporates a Swing Transformer module and an efficient layer aggregation network that integrates attention mechanisms, while the head network incorporates a small target detection layer.

3. The method according to claim 2, characterized in that, The backbone network includes a CBS module, a C2F module, and an SPPF module, wherein: The CBS module includes lightweight convolution, normalization, and SiLU activation functions for extracting primary features; The C2F module serves as the core structure for extracting deep semantic information. The SPPF module employs a residual design to extract multi-scale features through three consecutive 5×5 max pooling operations, and then concatenates the features before and after pooling in the output stage, preserving local details while compressing the feature map size.

4. The method according to claim 2, characterized in that, The basic units of the Swin Transformer module introduced by the neck network include window multi-head self-attention, shifted window multi-head self-attention, multilayer perceptron, and layer normalization.

5. The method according to claim 2, characterized in that, The efficient layer aggregation network with fusion attention mechanism introduced in the neck network effectively integrates features from different levels through a skip connection strategy, and uses coordinate attention, channel attention, and spatial attention to assign weights to multi-layer input features in different dimensions.

6. The method according to claim 5, characterized in that, The process of assigning weights to multi-layer input features using coordinate attention, channel attention, and spatial attention includes: First, coordinate attention and channel attention are combined to obtain the joint weight representation of position and channel dimensions. Then, coordinate attention and spatial attention are fused to obtain the joint weight representation of position and spatial dimensions. Finally, based on the efficient layer aggregation strategy, the obtained joint weight representation of position and channel dimensions and the joint weight representation of position and spatial dimensions are fused to obtain the output features.

7. The method according to claim 1, characterized in that, The process of realizing 3D target detection for drones based on the trained quadcopter drone motor recognition model includes: Based on the detected features of the quadcopter drone motor, extract the bounding rectangle of the motor target; By representing the motor with the center point of the top border of each circumscribed rectangle, the top surface coordinates of the four motors are obtained. Based on the known relationship between the side length and height of the quadcopter drone, calculate the length of the base and the pixel length corresponding to the height in the image; Assuming the height direction is parallel to the downward direction of the image's y-axis, the offset vector of the quadcopter drone to the y-axis is obtained based on the pixel length corresponding to the height in the image; Subtract the offset vector from the top surface position coordinates of the four motors to obtain the corresponding bottom surface position coordinates of the four motors; Based on the top and bottom position coordinates of the four motors, the 3D detection coordinates of the quadcopter UAV motors are finally obtained.

8. The method according to claim 1, characterized in that, The quadcopter UAV motor recognition model uses the SIoU loss function, which includes angle cost, distance cost, shape cost, and intersection-union ratio cost.

9. The method according to claim 1, characterized in that, The detection and evaluation metrics for the quadcopter drone motor identification model include accuracy, recall, average precision, and mean average precision.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images