A target recognition method and system from the perspective of a drone

Abstract

The present invention discloses a target recognition method and system from the perspective of an unmanned aerial vehicle (UAV), which relates to the field of UAV technology. The method uses multiple cameras carried by a UAV to simultaneously photograph a target area from different angles, performs preprocessing to obtain multi-perspective image data, registers the preprocessed multi-perspective images, and applies a convolutional neural network to extract image features from the image at each perspective, and establishes associations between different perspectives through feature matching. Based on the feature matching results, the multi-perspective images are fused, and then a fusion algorithm is used to supplement occluded areas. The present invention uses multiple cameras carried by a UAV to photograph a target area from different angles to obtain multi-perspective image data, helping to eliminate occlusion problems that exist under a single perspective. Secondly, the obtained comprehensive target view reduces information loss caused by single-perspective occlusion, ensures the stability and reliability of feature points, and improves the recognition accuracy and robustness of the entire system.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs), and in particular to a target recognition method and system from the perspective of an UAV. Background Art

[0002] Drones are widely used in many fields such as industry, agriculture, and environmental monitoring. Target recognition technology can help quickly locate problem areas and provide timely data support so that corresponding protective measures can be taken. It not only improves efficiency, but also changes the way traditional industries operate. With the continuous development of technology, drones have great application potential in more fields. Target recognition technology, as one of the core applications of drones, has promoted the advancement of drone intelligence and automation, bringing innovation and change to all walks of life.

[0003] During the flight of a drone, the target's perspective is constantly changing, and it is easy for the target to be partially occluded, which will make it difficult to maintain high accuracy in target recognition results. Therefore, how to fuse multi-perspective images to restore the occluded area and improve the accuracy of target recognition is the problem we need to solve. To this end, a target recognition method and system from the perspective of a drone are proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a target recognition method and system from the perspective of a drone to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] In a first aspect, a method for identifying a target from a drone's perspective comprises the following steps:

[0007] Step 1: Use multiple cameras onboard the drone to simultaneously photograph the target area from different angles, and perform preprocessing to obtain multi-view image data;

[0008] Step 2: Register the pre-processed multi-view images and apply convolutional neural networks to extract image features from each view, and establish associations between different views through feature matching;

[0009] Step 3: Based on the feature matching results, the multi-view images are fused, and then the fusion algorithm is used to supplement the occluded areas to generate a complete target view;

[0010] Step 4: Automatically detect and identify the target object by combining the fused image data;

[0011] Step 5: Verify the results of target detection and recognition and analyze the accuracy of target recognition.

[0012] A further improvement of the technical solution of the present invention is that in step 1, the process of acquiring multi-view image data includes:

[0013] Identify the target area to be photographed and the type of information required from different angles. Based on the shooting requirements, equip the drone with multiple cameras and adjust the camera angles and focal lengths to ensure that detailed information about the target area can be captured from different perspectives. Thoroughly inspect the drone and cameras to ensure they are in good working condition, including battery charge, signal connectivity, and camera clarity.

[0014] Plan the drone's flight path, altitude, and speed based on the target area's terrain, weather conditions, and filming requirements, ensuring coverage of all required shooting angles and locations, avoiding obstacles and no-fly zones, and complying with flight regulations and safety requirements for the target area;

[0015] Operate the drone to shoot according to the set flight plan, control the drone to fly along the predetermined flight route and altitude, and start multiple cameras for synchronous shooting;

[0016] The drone transmits the multi-view image data it captures in real time to the ground control station, which then performs preliminary sorting, deletes photos that clearly do not meet the requirements (such as overexposure and blur), and performs pre-processing operations such as denoising and color correction on the remaining images.

[0017] An image database is established to record the shooting time and geographic location information of each picture, and the acquired multi-view image data is stored in the image database for easy use in subsequent processing.

[0018] A further improvement of the technical solution of the present invention is that in step 2, the multi-view image registration process includes:

[0019] Select an image with the most central view from the multi-view images as the reference image, use the SIFT feature detection algorithm to detect key points in the reference image and each image to be registered, and generate a 128-dimensional descriptor vector for each key point. The descriptor vector describes the local image features around the key point.

[0020] Use the BFMatcher matching algorithm to find corresponding key point pairs between the reference image and the image to be registered, use the RANSAC algorithm to remove incorrect matching points, and estimate the geometric transformation model. Then, apply the geometric transformation model to align the image to be registered to the spatial coordinate system of the reference image;

[0021] Image features used for image fusion are extracted from the registered multi-view images, including color features, texture features, shape features, spatial relationship features, and multi-view consistency features. The registered images are then input into a convolutional neural network model. After passing through multiple convolutional and pooling layers of the convolutional neural network model, a series of feature maps are obtained.

[0022] For each image from each perspective, local feature vectors are extracted from the selected feature map, and the similarity between feature vectors from different perspectives is calculated using similarity measurement. Based on the similarity measurement results, feature matching relationships between images from different perspectives are established. A matching threshold is set to determine which pairs are considered matching pairs. Pairs above the matching threshold are considered valid matching relationships.

[0023] Save the feature matching results between each view image and the reference image, including matching point pairs and their corresponding feature descriptors.

[0024] A further improvement of the technical solution of the present invention is that the calculation formula of the similarity measure is:

[0025] ;

[0026] Where, represents the similarity measure, represents the dimension of the feature vector, Indicates the The first The local eigenvector dimensional component, Indicates the The first The local eigenvector dimensional component, The value range is between 0 and 1. When two eigenvectors are exactly the same, the similarity reaches the maximum value of 1.

[0027] A further improvement of the technical solution of the present invention is that in step 3, the process of generating the complete target view includes:

[0028] Based on the calculated similarity measurement results, confirm each pair of matching points and their corresponding feature descriptors, and organize all valid matching point pairs and their corresponding coordinate information for subsequent use;

[0029] Create a blank target view image of a size that can accommodate the contents of all input images. The target view image is a 16x16 pixel block. For each local area of ​​the 16x16 pixel block, determine the corresponding position of the area in the different view images based on the information of the matching point pairs. Apply the weighted averaging method to perform a weighted average of the pixel values ​​in the overlapping area to fuse the image contents of multiple view angles into the target view.

[0030] By analyzing matching point pairs and feature descriptors, the occluded area is identified. For the detected occluded area, its size is m×n pixels. For each pixel position in the occluded area, the new pixel value after filling is calculated and filled using an image restoration algorithm based on content-aware filling. The filling algorithm is inferred and generated based on the texture, color and structure information of the surrounding pixels;

[0031] The filled image is fused with the original image to ensure the natural transition and consistency of the fused area, and the fused image is post-processed including denoising, sharpening, color correction, etc., and the processed image is used as the final output to obtain a complete target view.

[0032] A further improvement of the technical solution of the present invention is that the calculation expression of the new pixel value after filling is:

[0033] ;

[0034] ;

[0035] ;

[0036] Where, Represents the new pixel value after filling, represents the center point, Indicates Radius around point The actual pixel value within the range, Represents the weight function, defining the importance of pixels in different directions, represents the energy difference from the center pixel, and Relative to the center point The vertical and horizontal offsets, represents the difference between two pixels, is a smaller than An integer that limits the size of the local area for comparison.

[0037] A further improvement of the technical solution of the present invention is that: in step 4, the process of identifying the target object includes:

[0038] Locate the target object in the fused image, detect feature points including texture features and shape features from the image, and then generate a descriptor for each detected feature point to describe the texture and shape information of its local area;

[0039] Use the sliding window method to traverse the entire image, use a fixed-size window to extract features at each position, match the extracted features with the features in the database, preset a matching threshold, classify the content in the window, and calculate the matching score to determine whether it is the target object;

[0040] The detection results are post-processed and the non-maximum suppression algorithm is applied to remove the target frames of repeated detections to improve the quality of the final output. The position of the identified target object is marked on the image, and the position of the identified target object and related information are output.

[0041] A further improvement of the technical solution of the present invention is that the calculation expression of the matching score is:

[0042] ;

[0043] Where, represents the matching score, Indicates the The matching score of feature points, Indicates the number of feature points extracted within the window, Represents a small constant (such as 0.01) to avoid the denominator being zero. Represents the preset matching threshold, represents the ideal upper limit of the matching score, and represents the ideal similarity between feature points. Indicates selection and The smaller one among them ensures that the matching score does not exceed the ideal upper limit, because The value range of is generally between [0,1]. Also within this range, so The maximum value of , The value range is between 0 and 1.

[0044] A further improvement of the technical solution of the present invention is that in step 5, the process of analyzing the target recognition accuracy includes:

[0045] Obtain a set of real scene images containing target objects, use bounding boxes to mark the location and size of the targets, and accurately label the target objects in the images;

[0046] Compare the target detection and recognition results with the target objects in the real scene image, analyze the color features, texture features, shape features and spatial relationship features of the two, determine the matching degree of each feature, and analyze the accuracy of target recognition;

[0047] Visualization tools are used to display the analysis results of target recognition accuracy, as well as the location and size of target objects.

[0048] In a second aspect, a target recognition system from the perspective of an unmanned aerial vehicle is used to implement the target recognition method from the perspective of an unmanned aerial vehicle, including an image data management platform, wherein the image data management platform is communicatively connected to an image acquisition module, a multi-view image registration and fusion module, a feature extraction and matching module, and a target detection and classification module, wherein the modules are electrically connected;

[0049] The image acquisition module uses drones to capture multi-view images of target objects, providing high-quality raw image data, laying the foundation for subsequent image processing and recognition;

[0050] The multi-view image registration and fusion module uses a feature detection algorithm to detect key points in the reference image and the image to be registered, and aligns the images to the same coordinate system, thereby fusing the image content from multiple perspectives into the target view and filling in the occluded areas;

[0051] The feature extraction and matching module is used to extract color features, texture features, shape features, spatial relationship features and multi-view consistency features from the registered images, generate feature maps through convolutional neural networks, calculate the similarity of feature vectors between different viewpoints, and establish feature matching relationships between images from different viewpoints;

[0052] The target detection and classification module uses a fixed-size window to traverse the entire image, extract features and match them with features in the database, and classify the content in the window to determine whether it is a target object.

[0053] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art:

[0054] 1. The present invention provides a target recognition method and system from the perspective of an unmanned aerial vehicle (UAV). The method uses multiple cameras carried by the UAV to capture the target area from different angles, obtaining multi-perspective image data, helping to eliminate the occlusion problem existing from a single perspective. Secondly, the obtained comprehensive target view reduces information loss caused by single-perspective occlusion, ensures the stability and reliability of feature points, and improves the recognition accuracy and robustness of the entire system.

[0055] 2. The present invention provides a method and system for target recognition from the perspective of a drone. By aligning and fusing images from multiple perspectives, the problem of information loss caused by occlusion can be effectively eliminated. The fused image not only contains more detailed information, but also has higher resolution and clarity, which helps to improve the accuracy of target detection and makes feature extraction more accurate. The complete target view generated by image fusion technology further expands the target recognition accuracy from the perspective of a drone. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0057] Figure 1 Schematic diagram of the method flow of the present invention;

[0058] Figure 2 Schematic diagram of the multi-view image registration workflow of the present invention;

[0059] Figure 3 A schematic diagram of a workflow for generating a complete target view of the present invention;

[0060] Figure 4 Schematic diagram of the system functional modules of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0062] Example 1, as Figure 1 、 Figure 2 As shown, the present invention provides a target recognition method from the perspective of a drone, comprising the following steps:

[0063] Step 1: Use multiple cameras on the drone to simultaneously shoot the target area from different angles, perform pre-processing to obtain multi-view image data, clarify the target area to be shot, and the type of information to be obtained from different angles. According to the shooting requirements, equip the drone with multiple cameras, adjust the angle and focal length of the camera to ensure that detailed information of the target area can be captured from different perspectives, and conduct a detailed inspection of the drone and camera to ensure that they are in good working condition, including battery power, signal connection, camera clarity, etc. According to the terrain, weather conditions and shooting requirements of the target area, plan the flight route, altitude and speed of the drone to ensure that all angles and positions required for shooting are covered. , ensure that the flight route avoids obstacles and no-fly zones, abides by the flight regulations and safety provisions of the target area, operates the drone to perform shooting operations according to the set flight plan, controls the drone to fly according to the predetermined flight route and altitude, and starts multiple cameras for synchronous shooting. The drone transmits the multi-view image data captured in real time to the ground control station for preliminary sorting, deletes photos that obviously do not meet the requirements (such as overexposure, blur), and performs pre-processing operations such as denoising and color correction on the retained images, establishes an image database, records the shooting time and geographic location information of each picture, and stores the acquired multi-view image data in the image database for use in subsequent processing;

[0064] Step 2: Register the preprocessed multi-view images, and apply convolutional neural networks to extract image features from each view, and establish associations between different viewpoints through feature matching. Select an image with the most central viewpoint from the multi-view images as the reference image, use the SIFT feature detection algorithm to detect key points in the reference image and each image to be registered, and generate a 128-dimensional descriptor vector for each key point. The descriptor vector describes the local image features around the key point, and uses the BFMatcher (Brute-Force Matcher) matching algorithm to find the corresponding key point pairs between the reference image and the image to be registered. Use the RANSAC (Random Sample Consensus) algorithm to remove incorrect matching points and estimate the geometric transformation model, and then apply Use the geometric transformation model to align the image to be registered to the spatial coordinate system of the reference image, extract image features for image fusion from the registered multi-view images, including color features, texture features, shape features, spatial relationship features and multi-view consistency features, and input the registered images into the convolutional neural network model. After passing through multiple convolutional layers and pooling layers of the convolutional neural network model, a series of feature maps are obtained. For each view of the image, local feature vectors are extracted from the selected feature maps, and the similarity of feature vectors between different viewpoints is calculated using a similarity metric. Based on the similarity metric results, a feature matching relationship is established between images of different viewpoints, and a matching threshold is set to determine whether they are considered matching pairs. Pairs above the matching threshold are considered to be valid matching relationships. The feature matching results between each view image and the reference image are saved, including matching point pairs and their corresponding feature descriptors.

[0065] Furthermore, the calculation formula of the similarity measure is:

[0066] ;

[0067] Where, represents the similarity measure, represents the dimension of the feature vector, Indicates the The first The local eigenvector dimensional component, Indicates the The first The local eigenvector dimensional component, The value range is between 0 and 1. When the two eigenvectors are exactly the same, the similarity reaches the maximum value of 1. When the two eigenvectors are completely different (that is, there is no common non-zero component), the similarity is close to 0.

[0068] Step 3: Based on the feature matching results, the multi-view images are fused, and then the fusion algorithm is used to supplement the occluded areas to generate a complete target view;

[0069] Step 4: Automatically detect and identify the target object by combining the fused image data;

[0070] Step 5: Verify the results of target detection and recognition and analyze the accuracy of target recognition.

[0071] Example 2, as Figure 3 As shown, based on Example 1, the present invention provides a technical solution: Preferably, in step 3, the process of generating the complete target view includes:

[0072] According to the calculated similarity measurement results, confirm each pair of matching points and their corresponding feature descriptors, and organize all valid matching point pairs and their corresponding coordinate information for subsequent use, create a blank target view image, the size of which should be able to accommodate the content of all input images. The size of the target view image is 16x16 pixel blocks. For each local area of ​​the 16x16 pixel block, determine the corresponding position of the area in the image of different view angles based on the information of the matching point pair, and apply the weighted averaging method to perform weighted averaging on the pixel values ​​in the overlapping area, and fuse the image content of multiple view angles into the target view. By analyzing the matching points Based on the feature descriptors, the occluded area is identified. For the detected occluded area, its size is m×n pixels. For each pixel position in the occluded area, the new pixel value after filling is calculated and filled using an image restoration algorithm based on content-aware filling. The filling algorithm is based on the texture, color and structure information of the surrounding pixels for inference and generation. The filled image is fused with the original image to ensure a natural transition and consistency in the fused area. The fused image is then post-processed, including denoising, sharpening, color correction, etc., and the processed image is used as the final output to obtain a complete target view.

[0073] Furthermore, the calculation expression of the new pixel value after filling is:

[0074] ;

[0075] ;

[0076] ;

[0077] Where, Represents the new pixel value after filling, represents the center point, Indicates Radius around point The actual pixel value within the range, Represents the weight function, defining the importance of pixels in different directions, represents the energy difference from the center pixel, and Relative to the center point The vertical and horizontal offsets, represents the difference between two pixels, is a smaller than An integer that limits the size of the local area to be compared. increases (i.e. the difference between the surrounding pixels and the central pixel increases), The value of gradually decreases, which means that the influence of pixels with large differences on the final result will be reduced accordingly, that is, the more similar pixels are, the greater their contribution to the filling effect;

[0078] In step 4, the process of identifying the target object includes:

[0079] The target object is located in the fused image, and feature points including texture features and shape features are detected from the image. A descriptor is generated for each detected feature point to describe the texture and shape information of its local area. The sliding window method is used to traverse the entire image, and a fixed-size window is used to extract features at each position. The extracted features are matched with the features in the database. A matching threshold is preset, and the content in the window is classified. The matching score is calculated to determine whether it is the target object. The detection results are post-processed, and the non-maximum suppression algorithm is applied to remove the target box that is repeatedly detected to improve the quality of the final output. The position of the identified target object is marked on the image, and the position of the identified target object and related information are output;

[0080] Furthermore, the calculation expression of the matching score is:

[0081] ;

[0082] Where, represents the matching score, Indicates the The matching score of feature points, Indicates the number of feature points extracted within the window, Represents a small constant (such as 0.01) to avoid the denominator being zero. Represents the preset matching threshold, represents the ideal upper limit of the matching score, and represents the ideal similarity between feature points. Indicates selection and The smaller one among them ensures that the matching score does not exceed the ideal upper limit, because The value range of is generally between [0,1]. Also within this range, so The maximum value of , The value range is between 0 and 1. are very close to hour, Close to 1, it means that the features in the window are highly matched with the target features, and it is likely to be the target object. Much smaller than ,but If it is much smaller than 1, it means the matching degree is not high and it is unlikely to be the target object. When the matching score is generally high, Tends to keep high value, on the contrary, if most matches have low scores, will decrease as the denominator increases;

[0083] In step 5, the process of analyzing target recognition accuracy includes:

[0084] Obtain a set of real-scene images containing target objects, use bounding boxes to mark the position and size of the targets to accurately label the target objects in the images, compare the target detection and recognition results with the target objects in the real-scene images, analyze the color features, texture features, shape features, and spatial relationship features of the two, clarify the matching degree of each feature, and analyze the accuracy of target recognition. Use visualization tools to display the analysis results of target recognition accuracy, as well as the position and size of the target objects.

[0085] Example 3, as Figure 4 As shown, based on Examples 1-2, the present invention further provides a target recognition system from the perspective of a drone, which is used to implement the target recognition method from the perspective of a drone, including an image data management platform, which is communicatively connected to an image acquisition module, a multi-view image registration and fusion module, a feature extraction and matching module, and a target detection and classification module, wherein the modules are electrically connected;

[0086] The image acquisition module uses drones to capture multi-view images of target objects, providing high-quality raw image data and laying the foundation for subsequent image processing and recognition;

[0087] The multi-view image registration and fusion module uses a feature detection algorithm to detect key points in the reference image and the image to be registered, and aligns the images to the same coordinate system. It then fuses the image content from multiple perspectives into the target view to fill in the occluded areas.

[0088] The feature extraction and matching module is used to extract color features, texture features, shape features, spatial relationship features, and multi-view consistency features from the registered images, generate feature maps through convolutional neural networks, calculate the similarity of feature vectors between different viewpoints, and establish feature matching relationships between images from different viewpoints;

[0089] The target detection and classification module uses a fixed-size window to traverse the entire image, extract features and match them with the features in the database, and classify the content in the window to determine whether it is a target object.

[0090] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A target recognition method from the perspective of a drone, characterized in that: The following steps are involved: Step 1: Use multiple cameras onboard the drone to simultaneously photograph the target area from different angles, and perform preprocessing to obtain multi-view image data; Step 2: Register the preprocessed multi-view images and apply a convolutional neural network to extract image features from each view. Then, establish the association between different views through feature matching. The multi-view image registration process includes: Select an image with the most central perspective from the multi-view images as the reference image, use the SIFT feature detection algorithm to detect key points in the reference image and each image to be registered, and generate a 128-dimensional descriptor vector for each key point; Use the BFMatcher matching algorithm to find corresponding key point pairs between the reference image and the image to be registered, use the RANSAC algorithm to remove incorrect matching points, and estimate the geometric transformation model. Then, apply the geometric transformation model to align the image to be registered to the spatial coordinate system of the reference image; Image features used for image fusion are extracted from the registered multi-view images, including color features, texture features, shape features, spatial relationship features, and multi-view consistency features. The registered images are then input into a convolutional neural network model. After passing through multiple convolutional and pooling layers of the convolutional neural network model, a series of feature maps are obtained. For each image from each perspective, local feature vectors are extracted from the selected feature map, and the similarity between feature vectors from different perspectives is calculated using similarity measurement. Based on the similarity measurement results, feature matching relationships between images from different perspectives are established. A matching threshold is set to determine which pairs are considered matching pairs. Pairs above the matching threshold are considered valid matching relationships. Save the feature matching results between each view image and the reference image, including matching point pairs and their corresponding feature descriptors; Step 3: Based on the feature matching results, the multi-view images are fused, and then the fusion algorithm is used to supplement the occluded areas to generate a complete target view; Step 4: Automatically detect and identify the target object by combining the fused image data; Step 5: Verify the results of target detection and recognition and analyze the accuracy of target recognition.

2. The target recognition method from the perspective of a drone according to claim 1, characterized in that: In step 1, the process of acquiring multi-view image data includes: Identify the target area to be photographed and the type of information to be obtained from different angles. Equip the drone with multiple cameras and adjust the camera angles and focal lengths based on the shooting requirements. Plan the drone's flight path, altitude, and speed based on the target area's terrain, weather conditions, and filming requirements; Operate the drone to shoot according to the set flight plan, control the drone to fly along the predetermined flight route and altitude, and start multiple cameras for synchronous shooting; The drone transmits the multi-view image data it captures in real time to the ground control station, which then performs preliminary sorting, deletes photos that clearly do not meet the requirements, and performs pre-processing operations such as denoising and color correction on the remaining images. An image database is established to record the shooting time and geographical location information of each picture, and the acquired multi-view image data is stored in the image database.

3. The target recognition method from the perspective of a drone according to claim 1, characterized in that: The calculation formula of the similarity metric is: ； Where, represents the similarity measure, represents the dimension of the feature vector, Indicates the The first The local eigenvector dimensional component, Indicates the The first The local eigenvector dimensional component, The value range is between 0 and 1. When two eigenvectors are exactly the same, the similarity reaches the maximum value of 1.

4. The target recognition method from the perspective of a drone according to claim 3, characterized in that: In step 3, the process of generating the complete target view includes: According to the calculated similarity measurement results, confirm each pair of matching points and their corresponding feature descriptors, and organize all valid matching point pairs and their corresponding coordinate information; Create a blank target view image with a size of 16x16 pixel blocks. For each local area of ​​the 16x16 pixel block, determine the corresponding position of the area in the different view images based on the information of the matching point pairs. Apply the weighted averaging method to perform a weighted average of the pixel values ​​in the overlapping area to fuse the image content of multiple view angles into the target view. By analyzing the matching point pairs and feature descriptors, the occluded area is identified. For the detected occluded area, its size is m×n pixels. For each pixel position in the occluded area, the new pixel value after filling is calculated and filled using the image restoration algorithm based on content-aware filling; The filled image is fused with the original image, and the fused image is subjected to post-processing operations including denoising, sharpening, and color correction. The processed image is then used as the final output to obtain a complete target view.

5. The target recognition method from the perspective of a drone according to claim 4, characterized in that: The calculation expression of the new pixel value after filling is: ； ； ； Where, Represents the new pixel value after filling, represents the center point, Indicates Radius around point The actual pixel value within the range, represents the weight function, represents the energy difference from the center pixel, and Relative to the center point The vertical and horizontal offsets, represents the difference between two pixels, is a smaller than An integer that limits the size of the local area for comparison.

6. The target recognition method from the perspective of a drone according to claim 5, characterized in that: In the step 4, the process of identifying the target object includes: Locate the target object in the fused image, detect feature points including texture features and shape features from the image, and then generate a descriptor for each detected feature point to describe the texture and shape information of its local area; Use the sliding window method to traverse the entire image, use a fixed-size window to extract features at each position, match the extracted features with the features in the database, preset a matching threshold, classify the content in the window, and calculate the matching score to determine whether it is the target object; The detection results are post-processed, and the non-maximum suppression algorithm is applied to remove the target frames of repeated detections, and the positions of the identified target objects are marked on the image, and the positions of the identified target objects and related information are output.

7. The target recognition method from the perspective of a drone according to claim 6, characterized in that: The calculation expression of the matching score is: ； Where, represents the matching score, Indicates the The matching score of feature points, Indicates the number of feature points extracted within the window, Represents a small constant to avoid the denominator being zero. Indicates the preset matching threshold, Indicates selection and The smaller one, The value range is between 0 and 1.

8. The target recognition method from the perspective of a drone according to claim 7, characterized in that: In step 5, the process of analyzing target recognition accuracy includes: Obtain a set of real scene images containing target objects, use bounding boxes to mark the location and size of the targets, and accurately label the target objects in the images; Compare the target detection and recognition results with the target objects in the real scene image, analyze the color features, texture features, shape features and spatial relationship features of the two, determine the matching degree of each feature, and analyze the accuracy of target recognition; Visualization tools are used to display the analysis results of target recognition accuracy, as well as the location and size of target objects.

9. A target recognition system from the perspective of an unmanned aerial vehicle, for implementing the target recognition method from the perspective of an unmanned aerial vehicle according to any one of claims 1 to 8, comprising an image data management platform, characterized in that: The image data management platform is communicatively connected to an image acquisition module, a multi-view image registration and fusion module, a feature extraction and matching module, and a target detection and classification module, wherein electrical signals are connected between the modules; The image acquisition module uses a drone to capture multi-view images of the target object; The multi-view image registration and fusion module uses a feature detection algorithm to detect key points in the reference image and the image to be registered, and aligns the images to the same coordinate system, thereby fusing the image content from multiple perspectives into the target view and filling in the occluded areas; The feature extraction and matching module is used to extract color features, texture features, shape features, spatial relationship features and multi-view consistency features from the registered images, generate feature maps through convolutional neural networks, calculate the similarity of feature vectors between different viewpoints, and establish feature matching relationships between images from different viewpoints; The target detection and classification module uses a fixed-size window to traverse the entire image, extract features and match them with features in the database, and classify the content in the window to determine whether it is a target object.

Images