Wildlife monitoring method using unmanned aerial vehicle

Abstract

The present application relates to the field of wild animal monitoring, and particularly relates to a wild animal monitoring method using a UAV, the present application divides geographical blocks, extracts typical visual images, carries out interference superposition verification on the typical visual images of each geographical block, determines the visual hiding relationship of each type of wild animal in different geographical blocks, and subsequently uses a monitoring atlas to adaptively process the image information collected by the UAV in different geographical blocks, especially for the geographical blocks without visual hiding relationship, the form of sliding search of a search window is used to extract sensitive monitoring dimension visual features to quickly analyze the target wild animals existing in the image information. The present application uses a UAV to monitor wild animals, fully considers the visual hiding relationship of different wild animals in different geographical blocks, constructs a monitoring atlas to adaptively use different image processing methods, and improves the efficiency of monitoring a large protected area.

Description

Technical Field

[0001] This invention relates to the field of wildlife monitoring, and more particularly to a method for wildlife monitoring using unmanned aerial vehicles (UAVs). Background Technology

[0002] Wildlife monitoring is a crucial tool for ecological conservation and biodiversity research. In vast protected areas like Aksai in Subei County, Northwest my country, the complex terrain and harsh climate make traditional methods like manual patrols and fixed camera monitoring limited by accessibility and coverage, insufficient to meet the needs of large-scale, continuous monitoring of rare and endangered species. With the development of drone technology, using drones equipped with visible light or infrared payloads for aerial photography has become a new approach for wildlife surveys in protected areas, enabling rapid acquisition of large-area image information and improving monitoring efficiency.

[0003] For example, Chinese Patent Publication No. CN119649254A discloses a deep learning-based intelligent target recognition and perception system and method for unmanned aerial vehicles (UAVs), relating to the field of artificial intelligence technology. The system extracts the importance features of each target area and calculates the area importance score to generate a target area monitoring plan and acquire continuous frame images. Using an improved YOLOv5 model, the system takes continuous frame images as input, outputs bounding boxes of suspected target wild animals, and maps them to the actual geographic coordinate system to obtain their actual location coordinates. Keypoint detection is performed on the t-th frame image, keypoint coordinates are extracted, and visibility scores and confidence scores are calculated to form feature descriptors. Target detection and keypoint detection are performed on the (t+1)-th frame image to obtain the actual location coordinates and feature descriptors of candidate targets. Trajectory matching is then performed; if a match is found, the actual location coordinates and feature descriptors of the suspected target wild animal are updated; otherwise, a new target animal is defined. The total number of target wild animals is obtained by counting based on the trajectory matching results.

[0004] However, the following problems still exist in the existing technology. Existing drone monitoring typically uses a uniform identification model or manual interpretation standard for the entire protected area, ignoring the impact of typical visual background differences in different geographical areas on the camouflage effect of animals. This leads to serious missed detections of species that blend well with the background, such as snow leopards, and the computational cost of traversing and processing images collected over large protected areas is high. Summary of the Invention

[0005] To address these issues, this invention provides a method for wildlife monitoring using unmanned aerial vehicles (UAVs). This method overcomes the problems of low detection accuracy for concealed species due to the failure to consider typical visual background differences in different geographical areas and the visual concealment patterns of wild animals in each area, the lack of differentiated processing mechanisms based on geographical area differences, and the high computational consumption caused by indiscriminate traversal processing of large-area aerial images.

[0006] To achieve the above objectives, the present invention provides a method for wildlife monitoring using unmanned aerial vehicles (UAVs), comprising: A comprehensive survey of the protected area was conducted, and clustering was performed based on the clustering image features of the surveyed images to identify several geographical blocks and extract typical visual images of the geographical blocks. Extracting cutout samples of various types of wild animals, and performing interference overlay verification based on typical visual images of each geographic block, including placing the cutout samples within the typical visual images, constructing a retrieval window to perform feature retrieval on the typical visual images for a predetermined retrieval path to retrieve visual features in various dimensions, wherein the retrieval path must pass through the cutout samples. Based on the changes in visual features of each dimension during the feature retrieval process, the visual camouflage relationships of various types of wild animals in different geographical areas are determined, and the sensitive monitoring dimensions and the corresponding fluctuation constraints of the sensitive monitoring dimensions are determined in order to construct the monitoring map corresponding to each type of wild animal. The process involves identifying the species of wild animals to be monitored, controlling a drone to collect image information within the protected area, and processing the drone's image information by accessing monitoring maps based on the geographic region to which the image information belongs. A retrieval window is constructed to retrieve image information by sliding. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. Alternatively, image information can be identified using an image processing model to determine the corresponding type of wild animal present in the image information, thereby tagging the corresponding image information.

[0007] It should be noted that the process of conducting a comprehensive survey of the protected area, clustering the surveyed images based on their clustering characteristics to identify several geographic blocks, and extracting typical visual images of these geographic blocks includes the following steps: Acquire survey images of the protected area in advance; Based on the clustering image features of the mapped images, each image patch is clustered to determine several clusters; Each cluster is defined as a geographic block, and the image data collected by the UAV in the geographic block is defined as a typical visual image. The clustered image features include color dimension visual features, texture discrete dimension visual features, and texture direction dimension visual features.

[0008] It should be noted that the process of constructing a search window to perform feature retrieval on the typical visual images according to a predetermined search path, in order to retrieve visual features in various dimensions, includes: The cutout sample is placed within the typical visual image, and a coordinate system is established to determine the center coordinates of the cutout sample. Within the typical visual image, several retrieval paths are determined that pass through the center coordinates in different directions; A search window is constructed and moves along the search path with a predetermined step size to determine the local image retrieved by the search window, and the visual features of the local image in various dimensions are extracted. The visual features in each dimension include color, brightness, texture direction, texture dispersion, and shape. The retrieval window is a local area of ​​a predetermined size.

[0009] It should be noted that the process of determining the visual camouflage relationships of various types of wild animals in different geographic areas based on the changes in visual features across various dimensions during feature retrieval includes, Based on the visual features retrieved during the feature retrieval process, construct the change curves of the visual features in each dimension; For the visual feature change curves of each dimension, determine the target curve segment corresponding to the retrieval window containing the cutout sample, so as to determine the curve feature difference ratio between the target curve segment and the non-target curve segment, as well as the peak dispersion of the non-target curve segment; If the visual feature change curves of each dimension all meet the visual camouflage conditions, then it is determined that there is a visual camouflage relationship between the corresponding type of wild animal and the geographic block. If there is a visual feature change curve in any dimension that does not meet the visual hiding condition, then it is determined that there is no visual hiding relationship between the corresponding type of wild animal and the geographic block. The curve features include frequency, average amplitude, and peak dispersion.

[0010] It should be noted that verifying that the visual feature change curves of each dimension satisfy the corresponding visual hiding conditions includes, The curve feature difference ratio does not exceed the preset comprehensive difference ratio threshold, and the peak dispersion exceeds the preset peak dispersion threshold.

[0011] It should be noted that the process of determining the sensitive monitoring dimensions and the corresponding fluctuation constraints for these dimensions includes, If the dimension corresponding to the visual feature change curve that does not meet the visual hiding condition is determined, then the dimension is determined as the sensitive monitoring dimension. Determine the normal distribution of the amplitude of the target curve segment and the non-target curve segment corresponding to the visual feature change curve of the sensitive monitoring dimension, and determine the confidence interval for each. The confidence intervals corresponding to non-target curve segments are defined as background fluctuation constraints, and the confidence intervals corresponding to target curve segments are defined as target fluctuation constraints.

[0012] It should be noted that the process of constructing monitoring maps corresponding to various types of wild animals includes, The monitoring map includes the visual camouflage relationship between various types of wild animals and various geographical blocks, the sensitive monitoring dimensions of various types of wild animals in various geographical blocks, and the fluctuation constraints of the corresponding sensitive monitoring dimensions.

[0013] It should be noted that when processing UAV image information by calling the monitoring map based on the geographical block to which the image information belongs, Based on the species of wild animals, determine the visual camouflage relationships between the corresponding types of wild animals and each geographical area; For geographic blocks that do not have visual hiding relationships, a retrieval window is constructed to slide and retrieve image information. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. For geographic blocks with visually hidden relationships, image information is identified using an image processing model to determine the corresponding types of wild animals present in the image information, and the corresponding image information is then labeled.

[0014] It should be noted that the process of constructing a sliding search window to retrieve image information and extracting sensitive monitoring dimensions based on the monitoring map includes, The image information of individual geographic blocks is stitched together to form a new stitched image; A retrieval path is constructed that is distributed in parallel within the stitched image, and several retrieval windows are constructed to move along the corresponding retrieval path in order to extract the visual features of the sensitive monitoring dimensions corresponding to the stitched image. Compare the visual features of sensitive monitoring dimensions with the corresponding fluctuation constraints; Mark the image information that meets the marking conditions; The marking conditions include that the visual features of the corresponding sensitive monitoring dimension within the retrieval window are free from the corresponding background fluctuation constraints and the visual features belong to the target fluctuation constraints.

[0015] It should be noted that this also includes the acquisition time and acquisition location corresponding to the recorded and marked image information.

[0016] Compared with existing technologies, this invention divides geographical areas, extracts typical visual images, performs interferometric overlay verification on the typical visual images of each geographical area, determines visual features in each dimension, and, based on the changes in visual features in each dimension during feature retrieval, determines the visual camouflage relationships of different types of wild animals in different geographical areas. It also determines sensitive monitoring dimensions and corresponding fluctuation constraints for these sensitive monitoring dimensions to construct monitoring maps for each type of wild animal. Subsequently, the monitoring maps are used to adaptively process image information collected by drones in different geographical areas. Especially for geographical areas without visual camouflage relationships, a sliding search window is used to extract visual features of sensitive monitoring dimensions for rapid analysis of target wild animals present in the image information. This invention utilizes drones for wildlife monitoring, fully considers the visual camouflage relationships of different wild animals in different geographical areas, and adaptively employs different image processing methods to construct monitoring maps, ensuring the accuracy of wildlife monitoring and improving the efficiency of monitoring large protected areas.

[0017] The advantages of this invention lie in its ability to identify typical visual images of various geographical blocks within a protected area and extract matted samples of various wild animals. These matted samples are then placed within the typical visual images for interferometric overlay verification. During this verification process, the retrieval window moves along multiple retrieval paths, passing through the matted samples, to extract visual features across various dimensions. In practical applications, the land cover of some geographical blocks within a protected area is relatively uniform and stable, and the background visual features themselves exhibit relatively stable fluctuations. Furthermore, the retrieval window only needs to perform visual feature extraction, a simple process that does not involve complex visual computations or deep learning models, resulting in extremely low computational resource consumption. Leveraging the aforementioned characteristics, this invention fully considers the varying visual concealment characteristics of different wild animals in different geographical areas due to their physical differences. When the search window slides over the local image area where the wild animal is located, the visual features in various dimensions will exhibit different changes. Some of these visual features are sensitive to the presence of animals and will produce significant fluctuations, thus enabling the rapid and sensitive detection of suspected targets with low computational cost. Based on this, this invention obtains prior data through interferometric superposition verification, providing data support for providing monitoring maps. Subsequently, different image processing methods are used to analyze the image information collected by the UAV in various geographical areas, ensuring the accuracy of wild animal monitoring and improving the efficiency of monitoring large-area protected areas.

[0018] The advantages of this invention are that it determines the visual camouflage relationships of different types of wild animals in different geographical areas and identifies sensitive monitoring dimensions, thereby constructing a monitoring atlas. This atlas includes the visual camouflage relationships of various types of wild animals in different geographical areas, the corresponding sensitive monitoring dimensions, and the fluctuation constraints of these sensitive monitoring dimensions. These fluctuation constraints fully consider background and target fluctuation constraints, thus fully reflecting the fluctuation range of the sensitive monitoring dimensions extracted from the background and wild animal target areas during the movement of the retrieval window. This allows for subsequent monitoring of various wild animals and is also applicable to the analysis of image information from different geographical areas, ensuring the accuracy of wild animal monitoring and improving the efficiency of monitoring large protected areas.

[0019] The advantages of this invention are that, when determining the species of wild animals to be monitored, it utilizes the geographic blocks described in the images collected by drones, calls up monitoring maps, and processes the drone image information in a targeted manner. During actual monitoring, when there is no visual camouflage relationship between wild animal species and geographic blocks, the monitoring maps are fully utilized to construct parallel retrieval paths. Image information is quickly extracted through a retrieval window. During extraction, only sensitive monitoring dimensions suitable for the corresponding type of wild animal and geographic block are extracted, improving extraction efficiency and reducing interference caused by the introduction of redundant visual features. This allows for the rapid identification of suspected targets without relying on complex visual analysis or model integration. Furthermore, when there is a visual camouflage relationship between wild animal species and geographic blocks, image processing models are invoked to improve the accuracy of processing complex image information. Therefore, the monitoring maps are used to process different geographic blocks in a targeted manner, making it suitable for processing large amounts of image information, ensuring the accuracy of wild animal monitoring, and improving the efficiency of monitoring large protected areas. Attached Figure Description

[0020] Figure 1 A schematic diagram illustrating the steps of a wildlife monitoring method using a drone, as described in an embodiment of the invention. Figure 2 This is a schematic diagram illustrating feature retrieval using a constructed retrieval window, as described in an embodiment of the invention. Figure 3 A logic block diagram illustrating the determination of visual camouflage relationships of various types of wild animals in different geographical areas, as described in an embodiment of the invention. Figure 4 This is a logic block diagram illustrating the process of calling monitoring maps to process image information of a UAV, as described in an embodiment of the invention. In the image, 1: search window, 2: cutout sample, 3: search path. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Please see Figure 1 The diagram illustrates the steps of a wildlife monitoring method using a drone according to an embodiment of the present invention. The wildlife monitoring method using a drone according to an embodiment of the present invention includes: Step S1: Conduct a comprehensive survey of the protected area, and cluster the images based on their clustering features to identify several geographic blocks and extract typical visual images of the geographic blocks. Step S2: Extract cutout samples 2 of various types of wild animals, and perform interference overlay verification based on typical visual images of each geographic block. This includes placing the cutout sample 2 within the typical visual image, constructing a retrieval window 1 to perform feature retrieval on the typical visual image for a predetermined retrieval path 3 to retrieve visual features in various dimensions. The retrieval path 3 must pass through the cutout sample 2. Step S3: Based on the changes in visual features of each dimension during the feature retrieval process, determine the visual camouflage relationships of each type of wild animal in different geographical areas, and determine the sensitive monitoring dimensions and the fluctuation constraints of the corresponding sensitive monitoring dimensions in order to construct the monitoring map corresponding to each type of wild animal. Step S4: Determine the species of wild animals to be monitored, control the drone to collect image information within the protected area, and process the drone's image information by calling the monitoring map according to the geographical block to which the image information belongs, including... A retrieval window 1 is constructed to retrieve image information by sliding. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. Alternatively, image information can be identified using an image processing model to determine the corresponding type of wild animal present in the image information, thereby tagging the corresponding image information.

[0024] In practice, there are no specific limitations on the method of conducting traversal mapping. As an optional implementation method, drones can be used to perform flight missions within the protected area and collect surface image information in advance. To improve mapping efficiency and obtain a wider area of ​​surface coverage, the drones can fly at a higher altitude, so that the geographical area covered by a single image is wider, reducing the amount of data processing required for subsequent image stitching and geographical block division. After preprocessing such as stitching and correction, the collected image data can be used as mapping images for subsequent geographical block identification and division based on clustering image features, which will not be elaborated further.

[0025] In practice, there are no restrictions on the model of the drone. The drone must be equipped with a positioning system to know the location corresponding to the captured image. This is existing technology and will not be elaborated further.

[0026] In specific implementation, there is no limitation on the form of obtaining wildlife matting samples 2. Those skilled in the art can obtain image frames containing corresponding types of wildlife in advance, determine the wildlife outlines in the image frames, and perform matting processing on the local images within the wildlife outlines to obtain the corresponding matting samples.

[0027] In practice, no specific restrictions are placed on the classification of wild animal species. For different monitoring stations, the classification can be based on the species of wild animals to be monitored in the corresponding protected area, and the corresponding wild animal image sample 2 can be determined. This will not be elaborated further.

[0028] In practical implementation, the specific structure of the image processing model is not limited. For example, the YOLO series of target detection models can be used as the basic architecture, and a dedicated image processing model can be trained using pre-labeled corresponding wildlife image samples; existing multimodal large models can also be used to perform semantic understanding and target recognition processing on image information; or open-source image processing models with animal species recognition capabilities can be directly selected for deployment and application. Any model that can effectively identify and label the corresponding type of wildlife in the input image information can be applied to this solution, and the specific selection can be flexibly determined according to the computing power and accuracy requirements of the actual monitoring scenario.

[0029] To elaborate, the process of conducting a comprehensive survey of the protected area, clustering the surveyed images based on their clustering characteristics to identify several geographic blocks, and extracting typical visual images of these geographic blocks includes... Acquire survey images of the protected area in advance; Based on the clustering image features of the mapped images, each image patch is clustered to determine several clusters; Each cluster is defined as a geographic block, and the image data collected by the UAV in the geographic block is defined as a typical visual image. The clustered image features include color dimension visual features, texture discrete dimension visual features, and texture direction dimension visual features.

[0030] During implementation, The color dimension visual characteristics are the hue H component in the HSV color space, which is used to distinguish the differences in color expression of various terrain blocks. It is less affected by light intensity, and a single value can distinguish the color differences of different land cover.

[0031] The discrete dimension visual features of texture and the direction dimension visual features are statistical measures based on the gray-level co-occurrence matrix in existing technologies. The discrete-dimensional visual feature of texture is the entropy value of the gray-level co-occurrence matrix, which measures the complexity of texture in an image patch; The texture orientation dimension visual feature is the inverse difference moment of the gray-level co-occurrence matrix, which measures the uniformity of texture in image patches; In practice, after clustering is completed, it is necessary to filter the area of ​​each cluster, remove clusters with an area less than the lower limit, and retain only clusters with an area greater than or equal to the lower limit as valid geographic blocks.

[0032] The purpose of setting an area threshold is to reduce the interference of local anomalies on the clustering results. In real-world scenarios, the surface of protected areas may contain scattered patches of discolored bare rock, small areas of water accumulation, or temporary vegetation patches. These local anomalies are prone to forming small, fragmented clusters during clustering. If they are classified as separate geographic blocks, it would result in an excessive number of blocks that lack actual ecological significance.

[0033] The lower limit of the area needs to be calculated in conjunction with the spatial resolution of the image, taking 5m in actual space as an example. 2 The lower limit of the area is obtained by converting the area to the corresponding area of ​​the image.

[0034] Clustering conditions can be set to cluster image patches. Each image patch is a local area of ​​a mapped image, and its size is determined based on the actual space. (The last part, "1 m," appears to be an unrelated fragment and is omitted from the translation.) 2 The area is converted to the corresponding area of ​​the image and used as the area of ​​the image block.

[0035] The clustering conditions include that the visual feature difference ratio of the color dimension between any image patch in the cluster and its neighboring image patches is less than a predetermined limit value for the color feature difference ratio, the visual feature difference ratio of the texture direction dimension is less than a predetermined limit value for the inverse moment difference ratio, and the visual feature difference ratio of the texture discrete dimension is less than a predetermined limit value for the entropy value difference ratio.

[0036] The process for determining the color feature difference ratio limit, the inverse moment difference ratio limit, and the entropy value difference ratio limit includes: The mapping images are annotated by those skilled in the art, marking the boundaries of individual terrain features, such as grasslands, rock walls, deserts, rocky beaches, lakes, and other single terrain areas; The color dimension visual features, texture discrete dimension visual features, and texture direction dimension visual features of image patches in a single terrain are determined. The difference ratio of color dimension visual features between each image patch is calculated, and the statistical distribution of each difference ratio is determined to determine the 25th percentile. This reflects the difference in color dimension visual features between image patches when the homogeneity of a single terrain is slightly higher than the average level. Therefore, the 25th percentile is used as the limit value for color feature difference ratio to facilitate subsequent clustering screening.

[0037] Calculate the difference ratio of visual features in the discrete dimension of texture and the difference ratio of visual features in the direction of texture among each image patch. Similarly, determine the statistical distribution of the difference ratio of visual features in the discrete dimension of texture and the difference ratio of visual features in the direction of texture, and determine the corresponding 25th percentile to reflect the difference level of visual features in the discrete dimension of texture and the direction of texture. Then, take the 25th percentile corresponding to the statistical distribution of the difference ratio of visual features in the direction of texture as the limit value of the inverse moment difference ratio, and take the 25th percentile corresponding to the statistical distribution of the difference ratio of visual features in the discrete dimension of texture as the limit value of the entropy difference ratio.

[0038] In practice, the difference ratio is the ratio of the absolute difference between values ​​to the mean of the values.

[0039] Please see Figure 2 As shown, this is a schematic diagram of constructing a search window for feature retrieval according to an embodiment of the invention. The process of constructing search window 1 to perform feature retrieval on the typical visual image for a predetermined search path 3, in order to retrieve visual features in various dimensions, includes: Place the cutout sample 2 within the typical visual image and establish a coordinate system to determine the center coordinates of the cutout sample 2. Within the typical visual image, several retrieval paths 3 that pass through the center coordinates in different directions are determined; A search window 1 is constructed and moves along the search path 3 with a predetermined step size. The local image retrieved by the search window 1 is determined, and the visual features of the local image in various dimensions are extracted. The visual features in each dimension include color dimension visual features, brightness dimension visual features, texture direction dimension visual features, texture discrete dimension visual features, and shape dimension visual features. The retrieval window 1 is a local area of ​​a predetermined size.

[0040] In practice, two search paths 3 can be set with an angle difference of 90° to uniformly examine the changes in visual features of each dimension when the search window 1 passes through the local image from all directions. Of course, those skilled in the art can also increase the number of search paths 3 to improve accuracy.

[0041] In implementation, the length of retrieval window 1 is set to 1 / 10 of the length of a typical visual image, and the width is 1 / 10 of the width of a typical visual image. This is to quickly examine the entire image and avoid retrieval window 1 being too large and covering too wide an area, which could easily mix in multiple textures or color regions, resulting in a decrease in the fineness of feature extraction. The step size of retrieval window 1 can be set by the user. For example, if retrieval window 1 moves in the width direction, the step size of retrieval window 1 cannot exceed the width of retrieval window 1 in order to traverse retrieval path 3. In implementation, it can be set to half of the corresponding width.

[0042] In practice, the matted sample 2 can be enlarged to a total area of ​​1 / 4 of the total area of ​​a typical visual image, so that the retrieval window 1 can undergo retrieval in multiple steps when passing through the matted sample 2.

[0043] In practice, the visual characteristics of color dimension are the hue H component in the HSV color space; The brightness dimension feature is the brightness V component in the HSV color space; The discrete-dimensional visual feature of texture is the entropy value of the gray-level co-occurrence matrix, which measures the complexity of texture in an image patch; The texture orientation dimension visual feature is the inverse difference moment of the gray-level co-occurrence matrix, which measures the uniformity of texture in image patches; Shape dimension features are the proportion of edge pixels in a local image to the total number of pixels in that region. The Canny edge detection operator is applied to the local image within search window 1 to determine the number of edge pixels, and then the shape dimension features are calculated. In practice, the introduction of animal body outlines and patch edges can alter the edge density of local areas. For example, the edge density of a rocky background is low; when search window 1 passes over a snow leopard's body, the outline and patch edges increase the edge density, with the specific value depending on the snow leopard's texture.

[0044] This invention identifies typical visual images of various geographical blocks within a protected area and extracts matted samples 2 of various wild animals. These matted samples 2 are then placed within the typical visual images for interferometric overlay verification. During this verification process, the retrieval window 1 moves along retrieval paths 3 in multiple directions, passing through the matted samples 2, to extract visual features across various dimensions. In practical applications, the land cover of some geographical blocks within the protected area is relatively uniform and stable, and the background visual features themselves exhibit relatively stable fluctuations. Furthermore, the retrieval window 1 only needs to perform visual feature extraction, which is simple and does not involve complex visual computations or deep learning models, resulting in extremely low computational resource consumption. Leveraging the aforementioned characteristics, this invention fully considers the varying visual concealment characteristics of different wild animals in different geographical areas due to their physical differences. When the retrieval window 1 slides over the local image area where the wild animal is located, the visual features in various dimensions will exhibit different changes. Some of these visual features are sensitive to the presence of animals and will produce significant fluctuations, thus enabling the rapid and sensitive detection of suspected targets with low computational cost. Based on this, this invention obtains prior data through interferometric superposition verification, providing data support for providing monitoring maps. Subsequently, different image processing methods are used to analyze the image information collected by the UAV in various geographical areas, ensuring the accuracy of wild animal monitoring and improving the efficiency of monitoring large protected areas.

[0045] To elaborate, please refer to Figure 3 As shown, Figure 3 The following is a logical block diagram illustrating the process of determining the visual camouflage relationships of various types of wild animals in different geographic areas, as described in this embodiment of the invention. The process of determining the visual camouflage relationships of various types of wild animals in different geographic areas based on the changes in visual features across various dimensions during feature retrieval includes: Based on the visual features retrieved during the feature retrieval process, construct the change curves of the visual features in each dimension; For the visual feature change curves of each dimension, determine the target curve segment corresponding to when the search window 1 contains the cutout sample 2, so as to determine the curve feature difference ratio between the target curve segment and the non-target curve segment, as well as the peak dispersion of the non-target curve segment. If the visual feature change curves of each dimension all meet the visual camouflage conditions, then it is determined that there is a visual camouflage relationship between the corresponding type of wild animal and the geographic block. If there is a visual feature change curve in any dimension that does not meet the visual hiding condition, then it is determined that there is no visual hiding relationship between the corresponding type of wild animal and the geographic block. The curve features include frequency, average amplitude, and peak dispersion.

[0046] It is understandable that the significance of considering the target curve segment and the non-target curve segment is that the non-target curve segment reflects the visual characteristics of the corresponding dimension of the background area. The present invention intends to use the peak dispersion to consider the fluctuation interference of the background itself, while the target curve segment reflects the visual characteristics of the corresponding dimension of the area where the animal is located. The present invention intends to use the curve feature difference ratio to reflect the data representation when the retrieval window 1 passes through the area where the animal is located.

[0047] The curve feature difference ratio is the ratio of the absolute difference between the curve features corresponding to the non-target curve segment and the target curve segment to the mean.

[0048] The peak discrete value is the variance of the amplitude at each time step.

[0049] To elaborate, verifying that the visual feature change curves of each dimension satisfy the corresponding visual hiding conditions includes: The curve feature difference ratio does not exceed the preset comprehensive difference ratio threshold, and the peak dispersion exceeds the preset peak dispersion threshold.

[0050] In implementation, the comprehensive difference ratio threshold aims to reflect the boundary that, compared to the sample 2 after being cut out from the retrieval window 1, must have a relatively obvious difference in visual features to meet the recognition requirements. Furthermore, since there are multiple curve features, a corresponding comprehensive difference ratio threshold is set for each curve feature, which is then selected appropriately during comparison. The comprehensive difference ratio threshold can be determined based on the curve features of non-target curve segments. The determination process includes... The non-target curve segment is further subdivided into several sub-curve segments. The curve characteristics of the sub-curve segments are solved, and the difference ratio between the corresponding curve characteristics of the sub-curve segments is determined. After statistics, the 90th quantile of the difference ratio is determined. Since the non-target curve segment actually corresponds to the background area, its overall difference ratio is at a relatively stable level. Therefore, the 90th quantile is set as the comprehensive difference ratio threshold to reflect the situation where the difference level is relatively high. Of course, the above is only an example, and those skilled in the art can also adjust the comprehensive difference ratio threshold.

[0051] The peak dispersion threshold aims to distinguish whether the fluctuation of the corresponding dimension curve feature when the retrieval window 1 passes through the region where the non-cutout sample 2 is located masks the performance of the corresponding dimension curve feature itself when passing through the cutout sample 2. Therefore, it can be determined based on the peak dispersion of the corresponding dimension curve feature when passing through the cutout sample 2. When the statistical retrieval window 1 passes through the cutout sample 2, the peak dispersion of the target curve segment is used. 80% of the peak dispersion is used as the upper limit of the peak dispersion threshold of the non-target curve segment. This limits the upper limit of the peak dispersion of the non-target curve segment and avoids the peak dispersion from masking the curve features of the target curve segment and affecting the judgment accuracy.

[0052] To elaborate, the process of determining sensitive monitoring dimensions and corresponding fluctuation constraints for these dimensions includes: If the dimension corresponding to the visual feature change curve that does not meet the visual hiding condition is determined, then the dimension is determined as the sensitive monitoring dimension. Determine the normal distribution of the amplitude of the target curve segment and the non-target curve segment corresponding to the visual feature change curve of the sensitive monitoring dimension, and determine the confidence interval for each. The confidence intervals corresponding to non-target curve segments are defined as background fluctuation constraints, and the confidence intervals corresponding to target curve segments are defined as target fluctuation constraints.

[0053] In practice, the confidence intervals are all 95% confidence intervals of a normal distribution, to reflect the typical intervals of the target curve segment and non-target curve segment of the corresponding sensitive detection dimension when the corresponding type of wild animals are in the geographic block.

[0054] To elaborate, the process of constructing monitoring maps corresponding to various types of wild animals includes, The monitoring map includes the visual camouflage relationship between various types of wild animals and various geographical blocks, the sensitive monitoring dimensions of various types of wild animals in various geographical blocks, and the fluctuation constraints of the corresponding sensitive monitoring dimensions.

[0055] During implementation, the visual hiding relationships, sensitive monitoring dimensions, and fluctuation constraints of the corresponding sensitive monitoring dimensions are determined and recorded to construct a monitoring map.

[0056] This invention identifies the visual camouflage relationships of different types of wild animals in different geographical areas and determines sensitive monitoring dimensions, thereby constructing a monitoring atlas. The monitoring atlas includes the visual camouflage relationships of various types of wild animals in different geographical areas, the corresponding sensitive monitoring dimensions, and the fluctuation constraints of these sensitive monitoring dimensions. These fluctuation constraints fully consider background fluctuation constraints and target fluctuation constraints, thus fully reflecting the fluctuation range of the sensitive monitoring dimensions extracted from the background and wild animal target areas during the movement of the retrieval window 1. This allows for subsequent monitoring of various wild animals and is also applicable to the analysis of image information from different geographical areas, ensuring the accuracy of wild animal monitoring and improving the efficiency of monitoring large protected areas.

[0057] Please see Figure 4 As shown, this is a logical block diagram of the process of calling the monitoring map to process the image information of the UAV according to an embodiment of the invention. When calling the monitoring map to process the image information of the UAV based on the geographical block to which the image information belongs, Based on the species of wild animals, determine the visual camouflage relationships between the corresponding types of wild animals and each geographical area; For geographic blocks that do not have visual hiding relationships, a retrieval window 1 is constructed to retrieve image information by sliding. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. For geographic blocks with visually hidden relationships, image information is identified using an image processing model to determine the corresponding types of wild animals present in the image information, and the corresponding image information is then labeled.

[0058] In practice, the fluctuation range is a closed interval formed by the maximum and minimum values ​​of the corresponding values ​​of the sensitive detection dimension.

[0059] To elaborate, the process of constructing a retrieval window 1 to retrieve image information and extracting sensitive monitoring dimensions based on the monitoring map includes: The image information of individual geographic blocks is stitched together to form a new stitched image; A retrieval path 3 is constructed and distributed in parallel within the stitched image. Several retrieval windows 1 are constructed and moved in accordance with the retrieval path 3 to extract the visual features of the sensitive monitoring dimensions corresponding to the stitched image. Compare the visual features of sensitive monitoring dimensions with the corresponding fluctuation constraints; Mark the image information that meets the marking conditions; The marking conditions include that the visual features of the corresponding sensitive monitoring dimension within the retrieval window 1 are free from the corresponding background fluctuation constraints and the visual features belong to the target fluctuation constraints.

[0060] It is understandable that fluctuation constraints are essentially a range. When the fluctuation range of a sensitive detection dimension does not belong to the corresponding fluctuation constraint, the surface is free from the fluctuation constraint, reflecting that the image information may contain the corresponding type of wild animal.

[0061] In practice, retrieval paths 3 can be arranged horizontally in parallel, with each retrieval path 3 adjacent to the others, to cover complete image information.

[0062] When stitching together image information for a single geographic area, the overlapping areas between adjacent images acquired by the UAV can be utilized. Based on feature matching or geographic coordinate information, multiple images can be combined into a single, complete stitched image covering the geographic area. Alternatively, since the UAV simultaneously records high-precision GPS coordinates and flight attitude data during image acquisition, the images can be directly projected onto a unified coordinate system based on the geographic reference coordinates corresponding to each image, completing the geographic location-based stitching. After stitching, the originally scattered single images within the geographic area are transformed into a spatially continuous panoramic image. This facilitates seamless sliding retrieval of the search window 1 within a unified reference system, avoiding feature breaks across image boundaries and improving the completeness and efficiency of visual feature extraction for sensitive monitoring dimensions.

[0063] To elaborate, this also includes recording the acquisition time and location corresponding to the marked image information.

[0064] The tagged images, collection time, and collection location are recorded and kept for easy reference by the reserve staff.

[0065] When determining the species of wild animals to be monitored, this invention utilizes the geographic blocks described in the images collected by drones, calls up monitoring maps, and processes the drone image information in a targeted manner. During actual monitoring, when there is no visual camouflage relationship between wild animal species and geographic blocks, the monitoring maps are fully utilized to construct parallel retrieval paths 3. Image information is rapidly extracted through retrieval window 1. During extraction, only sensitive monitoring dimensions suitable for the corresponding type of wild animal and geographic block are extracted, improving extraction efficiency and reducing interference caused by the introduction of redundant visual features. This allows for the rapid identification of suspected targets without relying on complex visual analysis or model integration. Furthermore, when there is a visual camouflage relationship between wild animal species and geographic blocks, image processing models are invoked to improve the accuracy of processing complex image information. Thus, the monitoring maps are used to process different geographic blocks in a targeted manner, making it suitable for processing large amounts of image information, ensuring the accuracy of wild animal monitoring, and improving the efficiency of monitoring large protected areas.

[0066] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for wildlife monitoring using unmanned aerial vehicles (UAVs), characterized in that, include: A comprehensive survey of the protected area was conducted, and clustering was performed based on the clustering image features of the surveyed images to identify several geographical blocks and extract typical visual images of the geographical blocks. Extracting cutout samples of various types of wild animals, and performing interference overlay verification based on typical visual images of each geographic block, including placing the cutout samples within the typical visual images, constructing a retrieval window to perform feature retrieval on the typical visual images for a predetermined retrieval path to retrieve visual features in various dimensions, wherein the retrieval path must pass through the cutout samples. Based on the changes in visual features of each dimension during the feature retrieval process, the visual camouflage relationships of various types of wild animals in different geographical areas are determined, and the sensitive monitoring dimensions and the corresponding fluctuation constraints of the sensitive monitoring dimensions are determined in order to construct the monitoring map corresponding to each type of wild animal. The process involves identifying the species of wild animals to be monitored, controlling a drone to collect image information within the protected area, and processing the drone's image information by accessing monitoring maps based on the geographic region to which the image information belongs. A retrieval window is constructed to retrieve image information by sliding. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. Alternatively, image information can be identified using an image processing model to determine the corresponding type of wild animal present in the image information, thereby tagging the corresponding image information.

2. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, The process of conducting a comprehensive survey of the protected area, clustering the surveyed images based on their clustering features to identify several geographic blocks, and extracting typical visual images of these geographic blocks includes... Acquire survey images of the protected area in advance; Based on the clustering image features of the mapped images, each image patch is clustered to determine several clusters; Each cluster is defined as a geographic block, and the image data collected by the UAV in the geographic block is defined as a typical visual image. The clustered image features include color dimension visual features, texture discrete dimension visual features, and texture direction dimension visual features.

3. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, The process of constructing a search window to perform feature retrieval on the typical visual images according to a predetermined search path, in order to retrieve visual features in various dimensions, includes: The cutout sample is placed within the typical visual image, and a coordinate system is established to determine the center coordinates of the cutout sample. Within the typical visual image, several retrieval paths are determined that pass through the center coordinates in different directions; A search window is constructed and moves along the search path with a predetermined step size to determine the local image retrieved by the search window, and the visual features of the local image in various dimensions are extracted. The visual features in each dimension include color, brightness, texture direction, texture dispersion, and shape. The retrieval window is a local area of ​​a predetermined size.

4. The method for wildlife monitoring using unmanned aerial vehicles according to claim 3, characterized in that, The process of determining the visual camouflage relationships of various types of wild animals in different geographic areas based on the changes in visual features across various dimensions during feature retrieval includes: Based on the visual features retrieved during the feature retrieval process, construct the change curves of the visual features in each dimension; For the visual feature change curves of each dimension, determine the target curve segment corresponding to the retrieval window containing the cutout sample, so as to determine the curve feature difference ratio between the target curve segment and the non-target curve segment, as well as the peak dispersion of the non-target curve segment; If the visual feature change curves of each dimension all meet the visual camouflage conditions, then it is determined that there is a visual camouflage relationship between the corresponding type of wild animal and the geographic block. If there is a visual feature change curve in any dimension that does not meet the visual hiding condition, then it is determined that there is no visual hiding relationship between the corresponding type of wild animal and the geographic block. The curve features include frequency, average amplitude, and peak dispersion.

5. The method for wildlife monitoring using unmanned aerial vehicles according to claim 4, characterized in that, Verify that the visual feature change curves of each dimension satisfy the corresponding visual hiding conditions, including... The curve feature difference ratio does not exceed the preset comprehensive difference ratio threshold, and the peak dispersion exceeds the preset peak dispersion threshold.

6. The method for wildlife monitoring using unmanned aerial vehicles according to claim 5, characterized in that, The process of determining sensitive monitoring dimensions and corresponding fluctuation constraints for these dimensions includes: If the dimension corresponding to the visual feature change curve that does not meet the visual hiding condition is determined, then the dimension is determined as the sensitive monitoring dimension. Determine the normal distribution of the amplitude of the target curve segment and the non-target curve segment corresponding to the visual feature change curve of the sensitive monitoring dimension, and determine the confidence interval for each. The confidence intervals corresponding to non-target curve segments are defined as background fluctuation constraints, and the confidence intervals corresponding to target curve segments are defined as target fluctuation constraints.

7. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, The process of constructing monitoring maps for various types of wild animals includes, The monitoring map includes the visual camouflage relationship between various types of wild animals and various geographical blocks, the sensitive monitoring dimensions of various types of wild animals in various geographical blocks, and the fluctuation constraints of the corresponding sensitive monitoring dimensions.

8. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, When processing UAV image information by calling the monitoring map based on the geographical block to which the image information belongs, Based on the species of wild animals, determine the visual camouflage relationships between the corresponding types of wild animals and each geographical area; For geographic blocks that do not have visual hiding relationships, a retrieval window is constructed to slide and retrieve image information. Sensitive monitoring dimensions are extracted based on the monitoring map, and the fluctuation range is determined. Based on the relationship between the fluctuation range and the corresponding fluctuation constraint, the image information is marked. For geographic blocks with visually hidden relationships, image information is identified using an image processing model to determine the corresponding types of wild animals present in the image information, and the corresponding image information is then labeled.

9. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, The process of constructing a retrieval window for sliding image information retrieval and extracting sensitive monitoring dimensions based on the monitoring map includes the following steps: The image information of individual geographic blocks is stitched together to form a new stitched image; A retrieval path is constructed that is distributed in parallel within the stitched image, and several retrieval windows are constructed to move along the corresponding retrieval path in order to extract the visual features of the sensitive monitoring dimensions corresponding to the stitched image. Compare the visual features of sensitive monitoring dimensions with the corresponding fluctuation constraints; Mark the image information that meets the marking conditions; The marking conditions include that the visual features of the corresponding sensitive monitoring dimension within the retrieval window are free from the corresponding background fluctuation constraints and the visual features belong to the target fluctuation constraints.

10. The method for wildlife monitoring using unmanned aerial vehicles according to claim 1, characterized in that, It also includes recording the acquisition time and acquisition location corresponding to the marked image information.

Images