Unmanned aerial vehicle deicing method based on image recognition

Abstract

The invention discloses an unmanned aerial vehicle deicing method based on image recognition, and relates to the technical field of unmanned aerial vehicle deicing, and the method comprises the steps: calling an image collection assembly of a target unmanned aerial vehicle to carry out the image collection of a target deicing region, obtaining a deicing region image, carrying out the feature recognition of the deicing region image through an image feature recognizer, and obtaining a deicing region image; obtaining an icing feature set; obtaining a regional environment state parameter group sequence; determining a plurality of environment state parameter feature groups; obtaining an environment state parameter feature group sequence; determining an iterative association environment state parameter feature group; and obtaining a deicing parameter set of the target unmanned aerial vehicle, and controlling the target unmanned aerial vehicle to deice the target deicing area. The technical problem that in the prior art, unmanned aerial vehicle deicing lacks correlation analysis on the environment state, the actual icing condition of the deicing area is inaccurately mastered, and the deicing quality is poor is solved, and the technical effect of improving the unmanned aerial vehicle deicing reliability is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle deicing, and in particular to an unmanned aerial vehicle deicing method based on image recognition. Background Art

[0002] At present, drones often rely on fixed paths and preset deicing parameters when performing deicing operations, which can easily cause the deicing strategy to be out of touch with the actual situation. At the same time, the icing conditions in the deicing area are not accurately understood. For example, when faced with uneven ice thickness, the mode and power of the deicing device are not adjusted according to the thickness of the ice layer, resulting in excessive energy waste in thin ice areas and incomplete deicing in thick ice areas. Ultimately, the quality of deicing is affected, and untreated frozen areas may be left, and even lead to serious consequences such as excessive energy consumption and low operating efficiency.

[0003] The existing technology of drone deicing has the technical problems of lack of correlation analysis of environmental conditions and inaccurate grasp of the actual icing conditions in the deicing area, resulting in poor deicing quality. Summary of the invention

[0004] The present application provides a UAV deicing method based on image recognition, which is used to solve the technical problems in the prior art that UAV deicing lacks correlation analysis of environmental conditions, has inaccurate understanding of the actual icing conditions in the deicing area, and leads to poor deicing quality.

[0005] In view of the above problems, the present application provides a method for deicing a drone based on image recognition, the method comprising:

[0006] Calling the image acquisition component of the target UAV to acquire images of the target deicing area to obtain an image of the deicing area, and using an image feature identifier to perform feature recognition on the deicing area image to obtain an icing feature set;

[0007] Calling the environment acquisition component of the target UAV to continuously monitor the environmental status of the target deicing area within a preset monitoring window to obtain a sequence of regional environmental status parameter groups, wherein each regional environmental status parameter group includes a plurality of regional environmental status parameters;

[0008] Constructing multiple asynchronous recognition network layers to perform parameter feature analysis on the regional environmental state parameter group sequence to determine multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields;

[0009] Serializing the plurality of environmental state parameter feature groups based on the sizes of the plurality of asynchronous receptive fields to obtain an environmental state parameter feature group sequence;

[0010] Performing iterative correlation interaction analysis on the environmental state parameter feature group sequence in order from front to back to determine an iterative correlation environmental state parameter feature group;

[0011] Based on the icing feature set and the iteratively associated environmental state parameter feature group, UAV deicing parameter identification is performed, a target UAV deicing parameter set is obtained, and the target UAV is controlled to de-ice the target deicing area according to the target UAV deicing parameter set.

[0012] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0013] The present application acquires images of the target deicing area by calling the image acquisition component of the target drone, obtains the deicing area image, uses the image feature identifier to perform feature recognition on the deicing area image, obtains the icing feature set, and then calls the environment acquisition component of the target drone to continuously monitor the environmental state of the target deicing area within the preset monitoring window, and obtains the regional environmental state parameter group sequence, wherein each regional environmental state parameter group includes multiple regional environmental state parameters, and then constructs multiple asynchronous recognition network layers to perform parameter feature analysis on the regional environmental state parameter group sequence, and determines multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields, and the multiple environmental state parameter feature groups are serialized based on the size of the multiple asynchronous receptive fields to obtain the environmental state parameter feature group sequence, and then the environmental state parameter feature group sequence is sequentially iterated and associated interactively analyzed in order from front to back, and the iteratively associated environmental state parameter feature group is determined, and the drone deicing parameter recognition is performed based on the icing feature set and the iteratively associated environmental state parameter feature group, and the target drone is controlled to de-ice the target deicing area according to the target drone deicing parameter set. The technical effect of improving the deicing quality of drones is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0015] Figure 1 A schematic diagram of a flow chart of a method for deicing a drone based on image recognition provided in an embodiment of the present application;

[0016] Figure 2 A schematic diagram of a process for determining an iteratively associated environmental state parameter feature group in an image recognition-based UAV deicing method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application provides a UAV deicing method based on image recognition, which is used to solve the technical problems in the prior art that UAV deicing lacks correlation analysis of environmental conditions, has inaccurate understanding of the actual icing conditions in the deicing area, and leads to poor deicing quality.

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0019] It should be noted that the terms "including" and "having" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or inherent to these processes, methods, products or devices.

[0020] Examples, such as Figure 1 As shown, the present application provides a method for deicing a drone based on image recognition, wherein the method comprises:

[0021] S100: calling the image acquisition component of the target UAV to acquire images of the target deicing area to obtain an image of the deicing area, and using an image feature identifier to perform feature recognition on the deicing area image to obtain an icing feature set;

[0022] Further, the image acquisition component of the target drone is called to acquire an image of the target deicing area to obtain an image of the deicing area, and an image feature identifier is used to perform feature recognition on the deicing area image to obtain an icing feature set. Step S100 of the embodiment of the present application further includes:

[0023] Obtaining a preset flight path of the target UAV;

[0024] The target UAV collects images of the target deicing area according to the preset flight altitude and the preset flight speed and the preset flight path to obtain an initial deicing area image;

[0025] The initial deicing area image is subjected to histogram equalization processing, and the icing sub-area of ​​the processed initial deicing area image is identified using a segmentation algorithm to obtain the deicing area image.

[0026] Furthermore, step S100 in the embodiment of the present application further includes:

[0027] Counting the number of pixels of each gray value in the initial deicing area image to obtain a gray histogram;

[0028] The pixel values ​​of the initial deicing area image are redistributed based on the grayscale histogram to obtain the deicing area image.

[0029] In a possible embodiment, the target UAV is any type of UAV used for deicing operations, and the deicing is performed by using hot air blowing, infrared heating, low-temperature spray deicing agent or mechanical vibration deicing. The target deicing area is any area that needs to be deiced after ice formation. The image acquisition component is a device installed on the target UAV for stable image acquisition, including a high-definition camera, an RGB camera, an infrared camera and a stabilization system (such as a three-axis gimbal).

[0030] The image acquisition component is used to acquire images of the target deicing area to obtain a deicing area image that can reflect the icing conditions of the target deicing area. Then, the trained image feature identifier is used to perform feature recognition on the deicing area image to extract features reflecting the degree of icing, such as the thickness, shape, texture and area of ​​the ice layer, to obtain the icing feature set. By acquiring the icing feature set, a technical effect of providing data support for subsequent reliable and accurate deicing of drones is achieved.

[0031] In one embodiment, the geographical scope of the target deicing area, the distribution of obstacles (such as buildings, trees, etc.), the performance parameters of the UAV (such as endurance time, flight altitude) and other information are obtained, and the technical personnel in this field determine the path for the target UAV to collect images of the target deicing area based on the above information to obtain the preset flight path. Preferably, a grid coverage path planning method is used to divide the target area into several sub-areas, each of which corresponds to a flight path segment. In addition, overlapping areas are set in the path design to avoid blind areas for image acquisition.

[0032] Preferably, the technicians in this field determine the flight altitude and flight speed of the target drone when collecting images based on the above information. Exemplarily, the preset flight altitude is adjusted by the technicians in this field based on the camera field of view and the size of the target area, such as 20 to 50 meters. The preset flight speed is set by the technicians in this field after balancing the image clarity and collection efficiency, such as 1 to 2 m / s.

[0033] Optionally, the planned preset flight path is uploaded to the UAV navigation system in the form of a GPS coordinate sequence, and the target UAV uses the image acquisition component to collect images of the target deicing area according to the navigation path to obtain the initial set of images of the deicing area to be spliced. Preferably, the three-axis gimbal is started to ensure the stability of the camera and avoid flight jitter affecting the image clarity. Each initial image of the deicing area to be spliced ​​is associated with corresponding geographic coordinates (GPS information), timestamp and flight altitude. The initial set of images of the deicing area to be spliced ​​is spliced ​​according to the geographic coordinates to obtain the initial deicing area image.

[0034] Furthermore, in order to improve the contrast of the image, the initial deicing area image is subjected to histogram equalization processing, the contrast of the low-brightness and high-brightness areas is stretched, and the visibility of the iced area is enhanced. Optionally, the input format is an RGB image or a grayscale image, and the number of pixels of each grayscale value in the image is counted to generate a grayscale histogram. The grayscale histogram is used to analyze the image brightness distribution. Exemplarily, if the values ​​of the histogram are concentrated in the low grayscale range, it means that the image is dark as a whole; if the values ​​of the histogram are concentrated in the high grayscale range, it means that the image is bright as a whole; if the histogram distribution is uneven, the image contrast is low.

[0035] According to the grayscale histogram, the number of pixels of each grayscale value is gradually accumulated to calculate the cumulative distribution function (CDF). The cumulative distribution function represents the cumulative number of pixels from the lowest grayscale value to the current grayscale value. Then, the CDF is normalized to the range of 0 to 1 for subsequent calculations. Based on the cumulative distribution function, the pixel values ​​are redistributed to make the distribution of grayscale values ​​more uniform. Thus, the cumulative distribution value corresponding to each grayscale value is linearly mapped to the range of 0 to 255 to obtain a new grayscale value. This mapping relationship is recorded to generate a mapping table of pixel values. Each pixel in the image is traversed, its original grayscale value is read, and then the original grayscale value is replaced with the new grayscale value according to the pixel value mapping table. As a result, the distribution range of grayscale values ​​in the processed image is wider and the contrast is enhanced. The redistributed pixel values ​​are saved as a new image to obtain the deicing area image. The technical effect of significantly enhancing the icing features that may be masked by insufficient illumination or contrast in the initial deicing area image is achieved, and providing reliable data support for subsequent icing area identification.

[0036] In one embodiment, based on a segmentation algorithm, such as a threshold segmentation method, the optimal global grayscale threshold of the initial deicing area image is automatically calculated, and the initial deicing area image is divided into an icing sub-area and a non-icing sub-area, so that the image of the position of the icing sub-area in the initial deicing area image is obtained as the deicing area image.

[0037] Optionally, multiple sample deicing area images and multiple sample icing feature sets are obtained as training data, and the training data is used to perform supervised training on a framework built based on a feedforward neural network to learn the mapping relationship between the deicing area images and the icing feature sets, and obtain the icing feature set. The icing feature set includes multidimensional feature data that characterizes the icing area conditions, including position features, area features, thickness features, shape features, etc. By obtaining the icing feature set, a technical effect of laying the foundation for reliable deicing of subsequent drones is achieved.

[0038] S200: calling the environment acquisition component of the target UAV to continuously monitor the environmental status of the target deicing area within a preset monitoring window to obtain a sequence of regional environmental status parameter groups, wherein each regional environmental status parameter group includes a plurality of regional environmental status parameters;

[0039] In one possible embodiment, first, ensure that the environmental acquisition components (such as temperature sensors, humidity sensors, anemometers, infrared cameras) on the drone are working properly. Then, call the initialization program of the environmental acquisition component to set the sampling frequency and range of each acquisition device. When the communication between the environmental acquisition component and the drone main control system and the ground station meets the requirements, collect the data collected by each sensor in the preset monitoring window to obtain the regional environmental state parameter group sequence. Among them, the regional environmental state parameter group sequence reflects the environmental state of the target deicing area in the preset monitoring window. Changes over time.

[0040] Optionally, each regional environmental status parameter group reflects the environmental conditions of the target deicing area at a time point within a preset monitoring window, including parameters such as temperature, wind speed and humidity. By obtaining the regional environmental status parameter group sequence, a technical effect of providing data support for subsequent refined analysis of regional environmental characteristics is achieved.

[0041] S300: construct multiple asynchronous recognition network layers to perform parameter feature analysis on the regional environmental state parameter group sequence to determine multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields;

[0042] Further, multiple asynchronous recognition network layers are constructed to perform parameter feature analysis on the regional environmental state parameter group sequence to determine multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields. Step S300 of the embodiment of the present application also includes:

[0043] Retrieving historical environmental change information of the target deicing area using the data types of the multiple regional environmental status parameters as an index to determine multiple parameter change period sets;

[0044] Calculating the mean of each of the plurality of parameter change period sets to obtain the mean of the plurality of parameter change period sets;

[0045] Calculate the sum of the mean values ​​of the multiple parameter change periods divided by the mean values ​​of the multiple parameter change periods to obtain multiple change coefficients;

[0046] Multiplying the multiple variation coefficients by a preset receptive field to obtain multiple asynchronous receptive fields, wherein the asynchronous receptive field is a feature extraction bandwidth of the asynchronous recognition network layer for the regional environment state parameter group sequence;

[0047] The multiple asynchronous recognition network layers are constructed based on the multiple asynchronous receptive fields.

[0048] Furthermore, step S300 in the embodiment of the present application further includes:

[0049] According to the multiple asynchronous receptive fields, convolution kernel parameters and time windows are set for the initial convolutional neural network layers respectively to obtain multiple initial asynchronous recognition network layers with completed parameter configurations;

[0050] A plurality of sample data sets are respectively obtained to perform supervised training on the plurality of initial asynchronous recognition network layers until the training converges, thereby obtaining the plurality of asynchronous recognition network layers that have completed the training.

[0051] In a possible embodiment, the regional environmental state parameter group sequence reflects the environmental state changes of the target deicing area within a preset monitoring window. By constructing multiple asynchronous recognition network layers, parameter analysis of different scales is performed on the regional environmental state parameter group sequence, and the environmental state changes are deeply analyzed to obtain the multiple environmental state parameter feature groups. The multiple asynchronous recognition network layers include multiple asynchronous receptive fields. The network layers constructed with different receptive fields have different data extraction ranges. The larger the receptive field, the greater the breadth of data that can be identified. The smaller the receptive field, the more sensitive the identification of data changes over time. By obtaining the multiple environmental state parameter feature groups, the technical effect of providing reliable data for the subsequent interactive analysis of features obtained from different receptive fields is achieved.

[0052] Optionally, the historical environmental change information is the environmental change of the target deicing area in the historical time, including the change time of different environmental state parameters. Exemplarily, a temperature change time set is obtained according to the change time when the temperature change amplitude exceeds the preset temperature threshold in the historical environmental change information. The historical environmental change information of the target deicing area is retrieved by using the data type of the environmental state parameters of the multiple areas as an index to obtain the multiple parameter change period sets. The multiple parameter change period sets reflect the change period of multiple environmental state parameters of the target deicing area in the historical time.

[0053] Then, the mean values ​​of the sets of the multiple parameter change cycles are calculated respectively to obtain the mean values ​​of the multiple parameter change cycles, wherein the mean values ​​of the multiple parameter change cycles reflect the average change cycles of the multiple environmental state parameters. The ratios of the mean values ​​of the multiple parameter change cycles to the sum of the mean values ​​of the multiple parameter change cycles are taken as the multiple change coefficients. The multiple change coefficients reflect the fluctuations of the multiple environmental state parameters in the historical time. The larger the change coefficient, the greater the fluctuation degree of the corresponding environmental state parameter.

[0054] Optionally, a preset receptive field pre-set by a person skilled in the art is obtained. The preset receptive field is a preset maximum feature extraction bandwidth. Then, a plurality of variation coefficients are multiplied by the preset receptive field to obtain a plurality of asynchronous receptive fields, each of which corresponds to an environmental state parameter. The greater the degree of parameter fluctuation, the larger the data extraction bandwidth is required to analyze the overall data situation, and the smaller the degree of parameter fluctuation, the smaller the data extraction bandwidth is required to capture the fine fluctuation of the data. The asynchronous receptive field is the feature extraction bandwidth of the asynchronous recognition network layer for the sequence of environmental state parameter groups of the region. In other words, the asynchronous receptive field refers to the data range or time window covered by the asynchronous recognition network layer when performing feature extraction on the sequence of environmental state parameter groups within a specific time period.

[0055] Therefore, by obtaining different asynchronous receptive fields, when extracting features from the regional environmental state parameter group sequence, the feature extraction conditions of different environmental state parameters can be considered respectively, thereby ensuring the comprehensiveness of the features of the multiple environmental state parameter feature groups extracted. Each environmental state parameter feature group corresponds to an asynchronous recognition network layer constructed by an asynchronous receptive field.

[0056] In one embodiment, the asynchronous receptive field defines the time window and data range covered when the network layer processes the sequence of environmental state parameter groups. In convolutional neural networks (CNNs), these receptive fields directly affect the time window setting and the size of the convolution kernel. The time window refers to the input data range (i.e., sequence length) processed by the network layer at one time. The time window size is consistent with the asynchronous receptive field, reflecting the range of attention of the network layer to changes in input parameters. The size (length or width) of the convolution kernel matches the receptive field size to ensure that key features are extracted within a specific time range.

[0057] A person skilled in the art sets the time window and the convolution kernel parameters of the initial convolutional neural network layer according to the sizes of the multiple asynchronous receptive fields. Preferably, different time windows are assigned to each network layer according to the sizes of the multiple asynchronous receptive fields, and the size of each time window is directly equal to the value of an asynchronous receptive field. Exemplarily, if the asynchronous receptive field is 16, the time window is 16 time points. If the asynchronous receptive field is 8, the time window is 8 time points. The length of the convolution kernel is equal to or slightly smaller than the receptive field to capture the feature pattern within the receptive field. According to the parameter complexity and model requirements, multiple convolution kernels are set to extract different features.

[0058] A plurality of sample data sets are extracted from the historical data for training the initial network layer, wherein each sample data set includes a plurality of sample area environment state parameter group sequences and a plurality of sample environment state parameter feature groups of corresponding types.

[0059] Then, multiple sample data sets are divided into training sets, validation sets, and test sets. This ensures that the data covers a variety of environmental change scenarios and improves the generalization ability of the model. Each initial asynchronous recognition network layer is trained using a supervised learning method.

[0060] Input multiple sample data sets into the corresponding multiple initial asynchronous recognition network layers respectively. Extract features based on convolution operation and calculate output values. Adjust network layer parameters (such as convolution kernel weights) through error back propagation algorithm (such as gradient descent). After repeated iterations, gradually optimize the feature extraction ability of the network layer. Select the mean square error as the loss function. When the loss function value drops to a preset threshold, or the training error remains stable in several consecutive iterations, the network layer training is considered to converge, and the multiple asynchronous recognition network layers that have been trained are obtained. Preferably, each asynchronous recognition network layer is trained separately without interfering with each other.

[0061] Optionally, the parameter features of the regional environmental state parameter group sequence are extracted using the multiple asynchronous recognition network layers that have been trained to obtain the multiple environmental state parameter feature groups. Each environmental state parameter feature group includes the change features of multiple environmental state parameters obtained by analyzing an asynchronous recognition network layer constructed in an asynchronous receptive field within a preset monitoring window corresponding to the regional environmental state parameter group sequence. By using the multiple asynchronous recognition network layers to extract features of multiple environmental state parameters at different scales, high-quality feature data is provided for subsequent analysis.

[0062] S400: Serializing the plurality of environment state parameter feature groups based on the sizes of the plurality of asynchronous receptive fields to obtain an environment state parameter feature group sequence;

[0063] S500: performing iterative correlation interaction analysis on the environment state parameter feature group sequence in order from front to back to determine an iterative correlation environment state parameter feature group;

[0064] Further, such as Figure 2 As shown, the environmental state parameter feature group sequence is subjected to iterative correlation interaction analysis in order from front to back to determine the iterative correlation environmental state parameter feature group. Step S500 of the embodiment of the present application further includes:

[0065] Extracting a first environmental state parameter feature group and a second environmental state feature group in sequence from the environmental state parameter feature group sequence;

[0066] Performing iterative correlation and interaction analysis on the first environmental state parameter feature group and the second environmental state parameter feature group to determine a first correlated environmental state parameter feature group;

[0067] Performing iterative correlation and interactive analysis on the first associated environmental state parameter feature group and the third environmental state parameter feature group of the environmental state parameter feature group sequence to determine a second associated environmental state parameter feature group;

[0068] An iterative correlation interaction analysis is performed on the n-1th associated environmental state parameter feature group and the nth environmental state parameter feature group of the environmental state parameter feature group sequence to determine the iteratively associated environmental state parameter feature group, wherein n is the number of environmental state parameter feature group sequences in the environmental state parameter feature group sequence, and n is an integer greater than or equal to 1.

[0069] In a possible embodiment, each environmental state parameter feature group corresponds to an asynchronous receptive field, and according to the size of the corresponding asynchronous receptive field, the corresponding environmental state parameter feature groups are sorted in order from small to large of the asynchronous receptive field to obtain the environmental state parameter feature group sequence.

[0070] In a possible embodiment, in order to integrate the environmental state parameter features extracted at different scales, the environmental state parameter feature group sequence is subjected to iterative correlation interaction analysis in order from front to back, the interaction relationship between adjacent environmental state parameter feature groups is analyzed, and an iterative correlation environmental state parameter feature group is determined. The iterative correlation environmental state parameter feature group integrates the environmental conditions reflected by the environmental state parameter feature groups at each scale, and reflects the environmental state of the target UAV when performing the UAV de-icing operation within the preset monitoring window.

[0071] In one embodiment, the first environmental state parameter feature group and the second environmental state parameter feature group are sequentially extracted from the environmental state parameter feature group sequence. Exemplarily, the first environmental state parameter feature group may be (temperature: -5°, humidity: 80, wind speed: 5), and the second environmental state parameter feature group may be (temperature: -4.8°, humidity: 81, wind speed: 4.8). The first environmental state parameter feature group and the second environmental state parameter feature group are subjected to iterative correlation interaction analysis to analyze the similarity between the same environmental state parameters, and then determine the first correlation environmental state parameter feature group containing different environmental state parameter development trends and different environmental state parameter characteristics.

[0072] Based on the same principle of obtaining the first associated environmental state parameter characteristic group, the first associated environmental state parameter characteristic group and the third environmental state parameter characteristic group of the environmental state parameter characteristic group sequence are subjected to iterative association interaction analysis to determine the second associated environmental state parameter characteristic group. Based on the same principle of obtaining the first associated environmental state parameter characteristic group, the n-1th associated environmental state parameter characteristic group and the nth environmental state parameter characteristic group of the environmental state parameter characteristic group sequence are subjected to iterative association interaction analysis to determine the iteratively associated environmental state parameter characteristic group, wherein n is the number of environmental state parameter characteristic group sequences in the environmental state parameter characteristic group sequence, and n is an integer greater than or equal to 1.

[0073] Furthermore, step S500 of the embodiment of the present application further includes:

[0074] Calculating the similarity of the features corresponding to the same environmental state parameters in the first environmental state parameter feature group and the second environmental state parameter feature group to obtain a first associated similarity group;

[0075] Performing intra-group interaction matrix identification on the first association similarity group to determine a first association interaction matrix;

[0076] A convolution calculation is performed on the first association interaction matrix and the second environmental state feature group to determine a first association environmental state parameter feature group.

[0077] Furthermore, step S500 of the embodiment of the present application further includes:

[0078] Obtain an intra-group interaction normalization processing formula, wherein the intra-group interaction normalization processing formula is:

[0079]

[0080] Among them, S[f(x i ,y i )] is the normalized value of the i-th association similarity in the first association similarity group, f(x i,y i ) is the i-th association similarity in the first association similarity group, x i is the i-th first environmental state parameter feature in the first environmental state parameter feature group, y i is the i-th second environment state parameter feature of the same type as the i-th first environment state parameter in the second environment state parameter group, e is the base of the natural logarithm, m is the total number of associated similarities in the first associated similarity group, and m is an integer greater than or equal to 1;

[0081] The association similarities in the first association similarity group are respectively input into the intra-group interaction normalization processing formula for normalization processing, and the processing result is embedded in a matrix to obtain the first association interaction matrix.

[0082] In one embodiment, the cosine similarity calculation formula is used to calculate the similarity of the corresponding features of the same environmental state parameters in the first environmental state parameter feature group and the second environmental state parameter feature group to obtain a first associated similarity group. The first associated similarity group reflects the similarity between the first environmental state parameter feature group and the second environmental state parameter feature group. Further, the intra-group interaction normalization processing formula is used to normalize the associated similarities in the first associated similarity group, and the processed normalized values ​​are embedded in the matrix to obtain the first associated interaction matrix. The first associated interaction matrix reflects the strength of association between the first environmental state parameter feature group and the second environmental state parameter feature group. The first associated interaction matrix can be used to avoid data redundancy when performing associated interaction analysis on adjacent environmental state parameter feature groups, thereby improving data processing efficiency.

[0083] Then, the first association interaction matrix and the second environmental state feature group are input into the graph convolution network for convolution calculation, so as to fuse information of different scales and output the first association environmental state parameter feature group. Through the interaction of information of different scales, an enhanced environmental state parameter feature group containing more trend information is obtained, so as to better capture the change of environmental state, and achieve the technical effect of providing an analytical data basis for effectively avoiding the influence of the environment on the de-icing of UAVs.

[0084] S600: Identify UAV deicing parameters based on the icing feature set and the iteratively associated environmental state parameter feature group, obtain a target UAV deicing parameter set, and control the target UAV to de-ice the target deicing area according to the target UAV deicing parameter set.

[0085] In an embodiment of the present application, the icing feature set reflects the degree of icing in the target deicing area, and the iteratively associated environmental state parameter feature group reflects the environmental trend change of the deicing of the UAV. By using a parameter identifier to identify the icing feature set and the iteratively associated environmental state parameter feature group, the target UAV deicing parameter set is obtained. Then, the target UAV deicing parameter set is sent to the control unit of the target UAV to control the target UAV to perform deicing operations on the target deicing area. The technical effect of improving the reliability of UAV deicing is achieved.

[0086] Optionally, multiple historical icing feature sets, multiple historical iteratively associated environmental state parameter feature groups, and multiple historical target UAV deicing parameter sets are obtained as training data sets, and the training data are used to supervise the training of the framework built based on the feedforward neural network, and the mapping relationship between the icing feature set and the iteratively associated environmental state parameter feature group and the target UAV deicing parameter set is learned during the training until the training converges to obtain the trained parameter identifier. The target UAV deicing parameter set includes parameters such as heating intensity, heating time, mechanical vibration frequency, mechanical vibration amplitude, injection pressure, and injection flow rate.

[0087] In summary, the embodiments of the present application have at least the following technical effects:

[0088] 1. This application uses an image acquisition component to acquire high-resolution images of the target deicing area, and uses an image feature identifier to accurately identify the location, area, thickness and other features of the icing area, and generates an icing feature set. This achieves the technical effect of enabling the drone to perform precise operations on the actual icing area.

[0089] 2. This application combines environmental status monitoring data and uses multiple asynchronous recognition network layers to perform multi-dimensional analysis of the data to grasp the trend of environmental changes, thereby achieving the technical effect of reducing the impact of environmental changes on the de-icing capability of the drone and ensuring the accuracy and reliability of the de-icing effect.

[0090] 3. By identifying UAV de-icing parameters according to the icing feature set and the iteratively associated environmental state parameter feature group, obtaining the target UAV de-icing parameter set, and controlling the target UAV to de-ice the target de-icing area according to the target UAV de-icing parameter set, the actual de-icing situation is grasped, and then the working mode of the UAV is adjusted, avoiding the technical effect of excessive de-icing or inefficient de-icing.

[0091] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0092] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0093] This specification and the drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A method for deicing an unmanned aerial vehicle based on image recognition, characterized in that: The method comprises: Calling the image acquisition component of the target UAV to acquire images of the target deicing area to obtain an image of the deicing area, and using an image feature identifier to perform feature recognition on the deicing area image to obtain an icing feature set; Calling the environment acquisition component of the target UAV to continuously monitor the environmental status of the target deicing area within a preset monitoring window to obtain a sequence of regional environmental status parameter groups, wherein each regional environmental status parameter group includes a plurality of regional environmental status parameters; Constructing multiple asynchronous recognition network layers to perform parameter feature analysis on the regional environmental state parameter group sequence to determine multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields; Serializing the plurality of environmental state parameter feature groups based on the sizes of the plurality of asynchronous receptive fields to obtain an environmental state parameter feature group sequence; Performing iterative correlation interaction analysis on the environmental state parameter feature group sequence in order from front to back to determine an iterative correlation environmental state parameter feature group; Based on the icing feature set and the iteratively associated environmental state parameter feature group, UAV deicing parameter identification is performed, a target UAV deicing parameter set is obtained, and the target UAV is controlled to de-ice the target deicing area according to the target UAV deicing parameter set.

2. The method for deicing a drone based on image recognition as claimed in claim 1, characterized in that: Calling the image acquisition component of the target UAV to acquire images of the target deicing area to obtain an image of the deicing area, and using an image feature identifier to perform feature recognition on the deicing area image to obtain an icing feature set, including: Obtaining a preset flight path of the target UAV; The target UAV collects images of the target deicing area according to the preset flight altitude and the preset flight speed and the preset flight path to obtain an initial deicing area image; The initial deicing area image is subjected to histogram equalization processing, and the icing sub-area of ​​the processed initial deicing area image is identified using a segmentation algorithm to obtain the deicing area image.

3. The method for deicing a drone based on image recognition as claimed in claim 2, characterized in that: include: Counting the number of pixels of each gray value in the initial deicing area image to obtain a gray histogram; The pixel values ​​of the initial deicing area image are redistributed based on the grayscale histogram to obtain the deicing area image.

4. The method for deicing a drone based on image recognition as claimed in claim 1, characterized in that: Constructing multiple asynchronous recognition network layers to perform parameter feature analysis on the regional environmental state parameter group sequence to determine multiple environmental state parameter feature groups, wherein the multiple asynchronous recognition network layers include multiple asynchronous receptive fields, including: Retrieving historical environmental change information of the target deicing area using the data types of the multiple regional environmental status parameters as an index to determine multiple parameter change period sets; Calculating the mean of each of the plurality of parameter change period sets to obtain the mean of the plurality of parameter change period sets; Calculate the sum of the mean values ​​of the multiple parameter change periods divided by the mean values ​​of the multiple parameter change periods to obtain multiple change coefficients; Multiplying the multiple variation coefficients by a preset receptive field to obtain multiple asynchronous receptive fields, wherein the asynchronous receptive field is a feature extraction bandwidth of the asynchronous recognition network layer for the regional environment state parameter group sequence; The multiple asynchronous recognition network layers are constructed based on the multiple asynchronous receptive fields.

5. The method for deicing a drone based on image recognition as claimed in claim 1, characterized in that: include: According to the multiple asynchronous receptive fields, convolution kernel parameters and time windows are set for the initial convolutional neural network layers respectively to obtain multiple initial asynchronous recognition network layers with completed parameter configurations; A plurality of sample data sets are respectively obtained to perform supervised training on the plurality of initial asynchronous recognition network layers until the training converges, thereby obtaining the plurality of asynchronous recognition network layers that have completed the training.

6. The method for deicing a drone based on image recognition as claimed in claim 4, characterized in that: The environmental state parameter feature group sequence is subjected to iterative correlation interaction analysis in order from front to back to determine an iterative correlation environmental state parameter feature group, including: Extracting a first environmental state parameter feature group and a second environmental state feature group in sequence from the environmental state parameter feature group sequence; Performing iterative correlation and interaction analysis on the first environmental state parameter feature group and the second environmental state parameter feature group to determine a first correlated environmental state parameter feature group; Performing iterative correlation and interactive analysis on the first associated environmental state parameter feature group and the third environmental state parameter feature group of the environmental state parameter feature group sequence to determine a second associated environmental state parameter feature group; An iterative correlation interaction analysis is performed on the n-1th associated environmental state parameter feature group and the nth environmental state parameter feature group of the environmental state parameter feature group sequence to determine the iteratively associated environmental state parameter feature group, wherein n is the number of environmental state parameter feature group sequences in the environmental state parameter feature group sequence, and n is an integer greater than or equal to 1.

7. The method for deicing a drone based on image recognition as claimed in claim 6, characterized in that: include: Calculating the similarity of the features corresponding to the same environmental state parameters in the first environmental state parameter feature group and the second environmental state parameter feature group to obtain a first associated similarity group; Performing intra-group interaction matrix identification on the first association similarity group to determine a first association interaction matrix; A convolution calculation is performed on the first association interaction matrix and the second environmental state feature group to determine a first association environmental state parameter feature group.

8. The method for deicing a drone based on image recognition as claimed in claim 7, characterized in that: include: Obtain an intra-group interaction normalization processing formula, wherein the intra-group interaction normalization processing formula is: Among them, S[f(x i ,y i )] is the normalized value of the i-th association similarity in the first association similarity group, f(x i ,y i ) is the i-th association similarity in the first association similarity group, x i is the i-th first environmental state parameter feature in the first environmental state parameter feature group, y i is the i-th second environment state parameter feature of the same type as the i-th first environment state parameter in the second environment state parameter group, e is the base of the natural logarithm, m is the total number of associated similarities in the first associated similarity group, and m is an integer greater than or equal to 1; The association similarities in the first association similarity group are respectively input into the intra-group interaction normalization processing formula for normalization processing, and the processing result is embedded in a matrix to obtain the first association interaction matrix.

Images