A method, system and storage medium for unmanned aerial vehicles (UAVs) to automatically intrude into inspection areas.

Abstract

This invention provides an automatic intrusion inspection method for unmanned aerial vehicles (UAVs). The method includes: performing target detection on video frames acquired by the UAV to obtain image features for each frame; sequentially comparing the image features of each frame with preset target features; when an image feature in an image successfully matches a target feature, using the successfully matched image as the current frame; performing trajectory prediction on the successfully matched image features of the current frame to obtain a predicted image and a predicted trajectory; inputting the predicted image into a GhostNetV2 network for feature extraction to obtain predicted image features; inputting the next frame of the current frame into a GhostNetV2 network for feature extraction to obtain actual image features; calculating the similarity between the predicted image features and the actual image features; filtering and correcting the predicted trajectory based on the similarity; and performing route planning based on the predicted trajectory to generate a tracking route. This method reduces the computational load required for target object identification during feature extraction, lowers computing power requirements, and improves real-time tracking performance.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicles (UAVs), and in particular to a method, system, and storage medium for automatic intrusion inspection by UAVs. Background Technology

[0002] With the advancement of technology, drone technology is constantly developing. Currently, remote inspection of designated areas by controlling the gimbal camera of a drone has improved the efficiency of inspection. However, when real-time inspection of a target area is required, the drone operator needs to constantly operate the joystick to control the rotation of the gimbal camera and observe whether there are any abnormalities in the image. Although it is more efficient than traditional inspection, the efficiency is still relatively low. Moreover, long-term inspection will also cause the operator's attention to decrease. Therefore, automatic tracking algorithms are needed to assist in completing the inspection task.

[0003] Existing technologies utilize the BoT-SORT multi-target tracking method to achieve automatic tracking of targets. For example, the invention with publication number "CN118072013.A" improves the BoT-SORT multi-target tracking algorithm based on region partitioning, enabling target tracking in complex scenarios. It obtains the position of the detected target through a detection algorithm, and then uses an improved tracking algorithm to associate the target and trajectory, thus achieving automatic tracking of the specified target. However, these algorithms have shortcomings in real-time performance, often resulting in problems with untimely tracking. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a method and system for automatic intrusion inspection by unmanned aerial vehicles (UAVs), which enables the target tracking algorithm to reduce the amount of computation in the feature extraction process and reduce the computing power requirement when identifying target objects.

[0005] Specifically, the present invention is achieved through the following technical solution:

[0006] On the one hand, the present invention provides a method for automatic unmanned aerial vehicle (UAV) intrusion inspection, which includes:

[0007] Target detection is performed on the video frames acquired by the drone to obtain the original image features of each frame;

[0008] The original image features of each frame are compared with the preset target features in sequence. When the original image features and the target features are successfully matched, the image with the successful match is used as the image to be predicted.

[0009] Trajectory prediction is performed on the original image features of the image to be predicted to obtain the predicted image and the predicted trajectory.

[0010] The predicted image features are obtained by extracting features from the predicted image using the GhostNetV2 network;

[0011] The similarity of the predicted image features with the original image features of the next frame of the corresponding image to be predicted is compared, and the predicted trajectory is filtered and corrected based on the similarity.

[0012] Route planning is performed based on the predicted trajectory to generate real-time tracking routes.

[0013] The existing BoT-SORT-ReID algorithm's Re-ID module has been improved by replacing the Resnet50 network in ReID with the GhostNetV2 network, which effectively reduces the computational cost of Re-ID feature extraction and improves the real-time performance of target tracking.

[0014] Furthermore, the feature extraction of the predicted image using the GhostNetV2 network specifically includes:

[0015] The acquired predicted image data is convolved and separated, then batch normalized and ReLU activated to obtain the first intermediate feature;

[0016] The acquired predicted image data is processed by a fully connected attention mechanism to obtain the second intermediate feature;

[0017] The third intermediate feature is obtained by multiplying the first intermediate feature and the second intermediate feature.

[0018] The third intermediate feature is separated by convolution of the second Ghost module and then batch normalized to obtain the fourth intermediate feature.

[0019] The fourth intermediate feature is added to the acquired original image data to obtain the image features.

[0020] Furthermore, the GhostNetV2 network is trained through the following steps:

[0021] The acquired training image data is input into the Ghost Netv2 network to be trained to obtain predicted image features;

[0022] Global average pooling is performed on the predicted image features to obtain the feature map;

[0023] The feature map and its corresponding standard map are used to calculate the feature loss; the feature map is processed through a fully connected layer to calculate the predicted probability distribution, and the distribution loss is calculated based on the predicted probability distribution and the standard probability.

[0024] The parameters of the Ghost Netv2 network are adjusted based on the distribution loss and feature loss until the network converges.

[0025] Furthermore, the loss between the predicted image features and the standard image features is calculated using the following formula:

[0026]

[0027] in These represent the feature maps obtained after random samples, random positive samples, and random negative samples have passed through the network, respectively. α represents the threshold.

[0028] Furthermore, at least one Ghost module of the GhostNetV2 network is trained with structural reparameterization to reduce the model structure.

[0029] Furthermore, after the model converges, fine-tuning of the network model is performed using the following formula:

[0030]

[0031] Where ρ represents the pruning operation, r is the pruning rate, and θ represents the fine-tuning operation. Let represent the network after the t-th pruning.

[0032] On the other hand, the present invention also provides an automatic drone intrusion inspection system, which includes: a drone gimbal: used to track and acquire video frames of the inspection area;

[0033] Target detection module: used to perform target detection on video frames acquired by the UAV to obtain the original image features of each frame;

[0034] Target matching module: It is used to compare the original image features of each frame with the preset target features in sequence. When the original image features and the target features are successfully matched, the image with the successful match is used as the current frame.

[0035] Trajectory prediction module: Used to predict the trajectory of the original image features of the image to be predicted, so as to obtain the predicted image and the predicted trajectory;

[0036] GhostNetV2 network: used to extract features from the predicted image to obtain the predicted image features;

[0037] Feature matching module: It is used to compare the similarity of the features of the predicted image with the features of the original image of the next frame of the corresponding image to be predicted, and to filter and correct the predicted trajectory based on the similarity.

[0038] Route generation module: Used to plan routes based on predicted trajectories and generate real-time tracking routes.

[0039] Furthermore, the GhostNetV2 network includes:

[0040] The first Ghost module layer is used to separate the acquired predicted image data through the first Ghost module convolution, perform batch normalization, and then ReLU activation to obtain the first intermediate features.

[0041] Decomposed Fully Connected Attention Layer: The acquired prediction image data is processed by the decomposed fully connected attention mechanism to obtain the second intermediate feature;

[0042] Feature aggregation layer: used to multiply the first intermediate feature and the second intermediate feature to obtain the third intermediate feature;

[0043] Second Ghost module layer: The third intermediate feature is separated by convolution in the second Ghost module and then batch normalized to obtain the fourth intermediate feature;

[0044] Feature concatenation layer: The fourth intermediate feature is added to the acquired predicted image data to obtain the predicted image features.

[0045] Furthermore, the drone automatic intrusion inspection system also includes a network update module, which includes:

[0046] Feature map reconstruction submodule: used to perform global average pooling on the predicted image features to obtain the feature map;

[0047] Feature loss calculation submodule: used to calculate feature loss based on the feature map and its corresponding standard map;

[0048] Distribution loss calculation submodule: used to process the feature map through a fully connected layer to calculate the predicted probability distribution, and calculate the distribution loss based on the predicted probability distribution and the standard probability;

[0049] The network update submodule is used to adjust the parameters of the Ghost Netv2 network based on the distributed loss and feature loss until the network converges.

[0050] On the other hand, the present invention also provides a storage medium storing a computer-readable storage program, which, when executed by a processor, implements the steps of an automatic intrusion inspection method for unmanned aerial vehicles as described in any of the preceding claims.

[0051] This invention discloses an automatic intrusion inspection method and system for unmanned aerial vehicles (UAVs). Firstly, it improves the Re-ID module of the existing BoT-SORT-ReID algorithm by replacing the ResNet50 network in ReID with a GhostNetV2 network. This effectively reduces the computational load during Re-ID feature extraction, improves real-time performance in target tracking, and, after each prediction, adjusts relevant parameters in the trajectory prediction module based on the prediction and actual offset results to better match the predicted target's motion preferences, thus improving prediction accuracy. Furthermore, recognizing that the feature reuse technique commonly used in GhostNetV2 networks increases computational costs on network hardware, impacting execution efficiency, the Ghostmodule undergoes structural reparameterization training. This reduces the network deployment structure, and feature fusion is performed in the weight space rather than the feature space, improving feature fusion efficiency. Finally, the trained network model is pruned to further simplify the number of model parameters, reduce computational and storage requirements, and improve system real-time performance.

[0052] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0053] Figure 1 A structural block diagram of an exemplary unmanned aerial vehicle (UAV) inspection system provided by the present invention;

[0054] Figure 2 A flowchart of an automatic unmanned aerial vehicle (UAV) intrusion inspection method provided by the present invention;

[0055] Figure 3 A structural block diagram of an automatic drone intrusion inspection system provided by the present invention;

[0056] Figure 4 This is a flowchart illustrating feature extraction from an exemplary GhostNetV2 network according to the present invention.

[0057] Figure 5 A flowchart illustrating the training process of an exemplary GhostNetV2 network provided by the present invention;

[0058] Figure 6 A structural block diagram of an exemplary Ghost module provided by the present invention;

[0059] Figure 7 This is a schematic diagram illustrating the derivation of the structural reparameterization described in this invention. Detailed Implementation

[0060] Please see Figure 1Drone inspections require the prior deployment of hardware equipment, typically using a drone equipped with a gimbal camera to complete the inspection task. This invention uses a three-axis gimbal camera as an example, comprising a three-axis electronically stabilized gimbal, a thermal imaging and visible light camera, and a laser rangefinder module; these components are mounted on the drone via a connecting device. The drone is equipped with a chip that stores a target tracking algorithm, which performs target recognition, detection, and tracking on the image data collected by the gimbal camera based on preset target features. After being assembled, the drone travels to a preset flight area to complete the inspection task, collecting image data in real time and detecting the presence of a preset target within the flight area. If a target is found, it is located, and the drone is controlled to track the target; otherwise, the inspection proceeds according to the set program.

[0061] During the inspection process, large areas of image data need to be collected and processed and analyzed in real time to enable target identification, detection, and tracking. To meet the real-time requirements of UAVs for automated intrusion inspections, significant computing resources are needed to assist in image data processing. However, the computing power of UAVs is limited, typically making it difficult to meet the demands of traditional BoT-SORT multi-target tracking methods. Furthermore, if image data processing is performed using a cloud server, the transmission latency will result in slow and untimely issuance of control commands from the UAV.

[0062] Based on this, the present invention provides an automatic drone intrusion inspection method and system. It improves the Re-ID module of the existing BoT-SORT-ReID algorithm by replacing the ResNet50 network in ReID with a GhostNetV2 network. This reduces the computational load during feature extraction when the target tracking algorithm identifies the target object, thus lowering the computational power requirement. The automatic drone intrusion inspection system includes a drone gimbal, a target detection module 10, a trajectory prediction module 20, a GhostNetV2 network 30, a feature matching module 40, and a route generation module 50. The automatic drone intrusion inspection system controls the drone gimbal to execute the automatic drone intrusion inspection method. Please refer to [link to relevant documentation]. Figure 2 and Figure 3 The method for automatically intruding into inspection areas by drones specifically includes:

[0063] The target detection module is used to perform step S10: perform target detection on the video frames acquired by the UAV to obtain the original image features of each frame.

[0064] The target matching module is used to perform step S20: sequentially compare the original image features of each frame with the preset target features. When the original image features in the image successfully match the target features, the successfully matched image is used as the image to be predicted.

[0065] During drone inspections, images are continuously captured on the inspected area, forming a series of video frames. Object detection algorithms can identify objects of interest in each frame and label their location and category, providing crucial information for subsequent analysis, such as identifying people, vehicles, and other objects in the images. In this invention, the captured images are matched against the target features of the object to be tracked. When a matching feature appears in the image, the tracking task begins. At this point, the object is marked in the image using a bounding box, obtaining its location and status information.

[0066] The trajectory prediction module 30 executes step S30: performing trajectory prediction on the image features that are successfully matched in the current frame to obtain the predicted image and the predicted trajectory.

[0067] This invention uses a Kalman filter to model the motion of the object and predict its position in the next frame. The Kalman filter directly outputs the width and height of the bounding box, rather than a scaling factor, which helps to obtain more accurate bounding box predictions. It also employs Camera Motion Compensation (CMC) technology to estimate the global motion of the camera and correct the parameters of the Kalman filter, enabling the tracking algorithm to maintain tracking stability when the camera moves, thus solving the problem of the impact of camera motion on tracking accuracy. At this point, the tracking trajectory and the predicted image are generated based on the prediction results.

[0068] To ensure the accuracy of continuous tracking, it is necessary to determine the accuracy of the prediction results during the tracking process. When prediction deviations occur, the parameters of the Kalman filter or the camera motion compensation parameters need to be corrected. At this point, the accuracy of the prediction results is determined using IOU and ReID.

[0069] GhostNetV2 network execution step S40: Input the predicted image into the GhostNetV2 network for feature extraction to obtain the predicted image features.

[0070] To transform images into a form more suitable for analysis and understanding, feature extraction is required. This data transformation process involves data transformation, filtering, dimensionality reduction, and other processes. In this application, the GhostNetV2 network is used to perform feature extraction of image data to solve the problem of limited computing resources on the drone's payload, while simultaneously meeting the requirements of real-time tracking.

[0071] Please see Figure 4 The feature extraction process of the GhostNetV2 network includes:

[0072] The acquired predicted image data is convolved and separated, then batch normalized and ReLU activated to obtain the first intermediate feature;

[0073] The acquired predicted image data is processed by a fully connected attention mechanism to obtain the second intermediate feature;

[0074] The decomposed fully connected attention mechanism differs from conventional fully connected layers. It is based on two fully connected layers: one layer gathers features along the horizontal direction, and the other layer gathers features along the vertical direction. The feature gathering in both directions can be specifically characterized as follows:

[0075]

[0076] in, and Z represents the learnable weights, and Z represents the image data.

[0077] The two fully connected layers capture long-range correlations along two directions respectively. At this point, the fully connected attention mechanism is decomposed and used in conjunction with the Ghost module to capture long-distance spatial location dependencies.

[0078] To improve the efficiency and accuracy of UAVs performing automated intrusion inspection tasks, this invention also trains the existing Ghost NetV2 network to adapt it to each specific object, or task. Please refer to... Figure 5 The training process of the Ghost Netv2 network in this invention is as follows:

[0079] The acquired training image data is input into the Ghost Netv2 network to be trained to obtain predicted image features;

[0080] In one embodiment of the invention, the Ghost module is trained using the open-source ImageNet dataset. P identities and K images of each person are randomly selected as training samples, where P is 16 and K is 4. Considering that pedestrians are small targets from the drone's perspective, each image is resized to 64*32 pixels, padded with 10 pixels, and then randomly cropped into a 64*32 rectangular image. This rectangular image is input into the model, and the model outputs the ReID feature F.

[0081] Global average pooling is performed on the predicted image features to obtain the feature map;

[0082] The predicted image features, i.e. ReID features F, are input into a global average pooling layer to restore the feature map.

[0083] Calculate the feature loss based on the feature map and its corresponding standard map;

[0084] The predicted feature F uses triplet loss as the feature loss, and its mathematical formula is as follows:

[0085]

[0086] in The feature map is obtained after random samples are processed by the network. The feature maps obtained after random positive samples pass through the network. This represents the feature map obtained after random negative samples pass through the network. α represents the threshold.

[0087] When the value of the expression within the brackets is greater than 0, the result after the plus sign is the expression itself, and the loss is greater than 0, so the parameters are updated. When the value of the expression within the brackets is less than 0, the result after the plus sign is 0, so no gradient can be generated, and the parameters are not updated. In this embodiment, α is set to 0.3.

[0088] The feature map is processed through a fully connected layer to calculate the predicted probability distribution, and the distribution loss is calculated based on the predicted probability distribution and the standard probability.

[0089] The parameters of the Ghost Netv2 network are adjusted based on the distribution loss and feature loss until the network converges.

[0090] The predicted probability distribution uses the cross-entropy loss function, the mathematical formula of which is as follows:

[0091]

[0092] Where y represents the actual probability distribution. The predicted probability distribution is represented by the following expression for the total training loss:

[0093] L total =L feature +L p

[0094] Once the model converges, training is stopped and the latest model weights are saved for later use. During training, triplet loss is used to represent the image feature loss of the training set, which enables the trained Ghost Netv2 network to better distinguish the differences between similar images and improve the drone's ability to identify targets. The introduction of cross-entropy loss accelerates the model's convergence speed, reduces the training time required when switching between different targets, and enables faster and more effective responses when facing different tasks.

[0095] The feature matching module executes step S50: calculate the similarity between the predicted image features and the actual image features, and complete the filtering and correction of the predicted trajectory based on the similarity.

[0096] There are various ways to compare similarity, such as comparing cosine similarity, Euclidean distance, grayscale algorithms, etc. In this invention, the cosine similarity between image features and target features is calculated, and then it is determined whether the cosine similarity value exceeds a preset threshold. If it does, it is determined that the predicted image features and the actual image features are the same object, which indicates that the confidence of the trajectory prediction is high. However, if the cosine similarity is lower than the preset threshold, it is determined that the predicted trajectory of the Kalman filter has an error and needs to be corrected. The parameters of the Kalman filter are then adjusted according to the prediction difference.

[0097] The route generation module executes step S60: Based on the predicted trajectory, it performs route planning and generates a real-time tracking route.

[0098] By comparing the real-time tracking route with the actual position of the tracked object, the drone's gimbal and flight attitude are adjusted to center the tracked object in the camera's position. The actual movement path of the tracked object is then recorded, along with video footage, generating the object's trajectory. This data is used for subsequent analysis and investigation. It also provides important evidence and convenience in fields such as security and surveillance.

[0099] In GhostNetV2 networks, feature reuse is typically used in network design. The Ghost module reuses feature maps from other layers through concatenation operations to ensure a large number of channels and network capacity. However, this increases the computational cost on the hardware, thus affecting execution efficiency. Figure 6 As shown, existing Ghost modules perform depthwise separable convolution on the input feature map to obtain the intrinsic map. The intrinsic map is then processed by a ReLU activation function to obtain the Ghost map. Finally, the input feature map and the Ghost map are concatenated to output the final result. This feature reuse technique reduces the efficiency of feature matching by the camera in this invention, affecting the real-time tracking performance. Based on these shortcomings, this invention, referencing the concept of reparameterization, performs structural reparameterization training on the convolutional kernels used for dconv, enabling the final deployed Ghost module to reduce processing time and meet the requirements of real-time tracking.

[0100] Please see Figure 7 'a' represents the original Ghost module. 'b' replaces the 'cat' operation with the 'add' operation on the basis of 'a'. 'c' moves the ReLU operation after the 'add' operation on the basis of 'b'. 'd' adds a BN operation to the jump connection on the basis of 'c'. Based on the idea of ​​reparameterization, the operation within the box in 'd' can be equivalent to the 'dconv' operation, thus obtaining the structure in 'e'.

[0101] That is, according to Figure 7The structure shown in (d) is trained by reparameterization, and the value of the convolution kernel of dconv is continuously corrected so that the final Ghost module deployment only includes the trained dconv layer and ReLU layer. At this time, the feature fusion process is carried out in the weight space rather than in the feature space, which improves the efficiency of feature fusion.

[0102] Furthermore, after training the network model, to further streamline the model, this invention also performs pruning. By removing redundant neurons or connections, pruning can reduce the number of model parameters and lower computational and storage requirements without significantly affecting model performance. The iterative pruning strategy employed in this invention has the following mathematical formula:

[0103]

[0104] Where ρ represents the pruning operation, r is the pruning rate, and θ represents the fine-tuning operation. Let represent the network after the t-th pruning.

[0105] When the network is unable to make effective predictions, stop the pruning process and use the network obtained from the last effective pruning as the final network.

[0106] In summary, the UAV automatic intrusion inspection method and system described in this invention first improves the Re-ID module of the existing BoT-SORT-ReID algorithm by replacing the ResNet50 network in ReID with a GhostNetV2 network. This effectively reduces the computational load during Re-ID feature extraction and improves the real-time performance of target tracking. Furthermore, after each prediction, the relevant parameters in the trajectory prediction module can be corrected based on the prediction results and actual offset results to better match the predicted target's motion preferences and improve prediction accuracy. Further, recognizing that the feature reuse technique commonly used in GhostNetV2 networks increases the computational cost on network hardware, thus affecting execution efficiency, the Ghost module undergoes structural reparameterization training. This reduces the network deployment structure, and the feature fusion process is performed in the weight space rather than the feature space, improving feature fusion efficiency. Finally, the trained network model is pruned to further simplify the number of model parameters, reduce computational and storage requirements, and improve the system's real-time performance.

[0107] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the automatic intrusion inspection method for unmanned aerial vehicles described in any of the above embodiments.

[0108] This invention can take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0109] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A method for automatic unmanned aerial vehicle (UAV) intrusion inspection, characterized in that, include: Target detection is performed on the video frames acquired by the drone to obtain the original image features of each frame; The original image features of each frame are compared with the preset target features in sequence. When the original image features and the target features are successfully matched, the image with the successful match is used as the image to be predicted. Trajectory prediction is performed on the original image features of the image to be predicted to obtain the predicted image and the predicted trajectory. The predicted image features are obtained by extracting features from the predicted image using the GhostNetV2 network; The similarity of the predicted image features with the original image features of the next frame of the corresponding image to be predicted is compared, and the predicted trajectory is filtered and corrected based on the similarity. Route planning is performed based on the predicted trajectory to generate real-time tracking routes.

2. The method for automatic unmanned aerial vehicle (UAV) intrusion inspection according to claim 1, characterized in that, Feature extraction of the predicted image using the GhostNetV2 network specifically includes: The acquired predicted image data is convolved and separated, then batch normalized and ReLU activated to obtain the first intermediate feature; The acquired predicted image data is processed by a fully connected attention mechanism to obtain the second intermediate feature; The third intermediate feature is obtained by multiplying the first intermediate feature and the second intermediate feature. The third intermediate feature is separated by convolution using the second Ghost module, and then batch normalization is performed to obtain the fourth intermediate feature. The fourth intermediate feature is added to the obtained predicted image data to obtain the predicted image features.

3. The method for automatic unmanned aerial vehicle (UAV) intrusion inspection according to claim 2, characterized in that, The GhostNetV2 network is trained through the following steps: The acquired training image data is input into the Ghost Netv2 network to be trained to obtain predicted image features; Global average pooling is performed on the predicted image features to obtain the feature map; The feature map and its corresponding standard map are used to calculate the feature loss; the feature map is processed through a fully connected layer to calculate the predicted probability distribution, and the distribution loss is calculated based on the predicted probability distribution and the standard probability. The parameters of the Ghost Netv2 network are adjusted based on the distribution loss and feature loss until the network converges.

4. The method for automatic unmanned aerial vehicle (UAV) intrusion inspection according to claim 3, characterized in that, The loss between the predicted image features and the standard image features is calculated using the following formula: in α represents the feature maps obtained after random samples, random positive samples, and random negative samples are processed by the network, respectively; α represents the threshold.

5. A method for automatic unmanned aerial vehicle (UAV) intrusion inspection according to any one of claims 2-4, characterized in that, Perform structural reparameterization training on at least one Ghost module of the GhostNetV2 network shown to reduce the model structure.

6. The method for automatic unmanned aerial vehicle (UAV) intrusion inspection according to claim 5, characterized in that, After the model converges, fine-tuning of the network model is performed using the following formula: Where ρ represents the pruning operation, r is the pruning rate, and θ represents the fine-tuning operation. Let represent the network after the t-th pruning.

7. A drone automatic intrusion inspection system, characterized in that, include Drone gimbal: Used to track and acquire video frames of the inspection area; Target detection module: used to perform target detection on video frames acquired by the UAV to obtain the original image features of each frame; Target matching module: It is used to compare the original image features of each frame with the preset target features in sequence. When the original image features and the target features are successfully matched, the image with the successful match is used as the current frame. Trajectory prediction module: Used to predict the trajectory of the original image features of the image to be predicted, so as to obtain the predicted image and the predicted trajectory; GhostNetV2 network: used to extract features from the predicted image to obtain the predicted image features; Feature matching module: It is used to compare the similarity of the features of the predicted image with the features of the original image of the next frame of the corresponding image to be predicted, and to filter and correct the predicted trajectory based on the similarity. Route generation module: Used to plan routes based on predicted trajectories and generate real-time tracking routes.

8. The automatic intrusion inspection system for unmanned aerial vehicles according to claim 7, characterized in that, The GhostNetV2 network includes: The first Ghost module layer is used to perform convolutional separation on the acquired predicted image data, followed by batch normalization and ReLU activation to obtain the first intermediate features. Decomposed Fully Connected Attention Layer: The acquired prediction image data is processed by the decomposed fully connected attention mechanism to obtain the second intermediate feature; Feature aggregation layer: used to multiply the first intermediate feature and the second intermediate feature to obtain the third intermediate feature; Second Ghost module layer: The third intermediate feature is separated by convolution in the second Ghost module and then batch normalized to obtain the fourth intermediate feature; Feature stitching layer: The fourth intermediate feature is added to the acquired original image data to obtain the image features.

9. The automatic intrusion inspection system for unmanned aerial vehicles according to claim 8, characterized in that, It also includes a network update module, which includes: Feature map reconstruction submodule: Used to perform global average pooling on the predicted image features to obtain the feature map; Feature loss calculation submodule: used to calculate feature loss based on the feature map and its corresponding standard map; Distribution loss calculation submodule: used to process the feature map through a fully connected layer to calculate the predicted probability distribution, and calculate the distribution loss based on the predicted probability distribution and the standard probability; The network update submodule is used to adjust the parameters of the Ghost Netv2 network based on the distributed loss and feature loss until the network converges.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the automatic intrusion inspection method for unmanned aerial vehicles as described in any one of claims 1 to 6.

Images