A method, device, electronic device and medium for optimizing target point of unpowered airdrop

Abstract

The present application provides a method, device, electronic device and medium for optimizing target points for unpowered airdrops. First, a target image is acquired, and pixel points in the target area and hollow area of ​​the target image are traversed and sampled to form a set of candidate points. Then, for each candidate point, the features of the candidate point in multiple directions are extracted to form a multi-dimensional vector feature corresponding to the candidate point. Afterwards, the vector feature corresponding to each candidate point is input into the delivery target point optimization model to predict the probability of successful delivery when each candidate point is used as a delivery target point. The delivery target point optimization model is obtained by training a deep neural network using a pre-constructed target point training sample set through supervised learning. Finally, the candidate point with the highest probability of successful delivery is selected as the optimal delivery target point corresponding to the target image. For airdrop areas of arbitrary shapes, the optimal delivery target point can be found quickly and accurately, thereby improving the success rate of material airdrops.

Description

Technical Field

[0001] The present application relates to the technical field of airdrop target point prediction, and in particular to a method, device, electronic equipment and medium for optimizing an unpowered airdrop target point. Background Art

[0002] Airdrop is a commonly used delivery method in modern military and civilian fields. Its purpose is to make the landing point of airdropped materials within a designated area. In practical applications, the airdrop process is often accompanied by complex random interference from the flight environment, including environmental interference such as wind fields and ocean currents, positioning errors of the drop point, initial alignment errors, initial orientation errors, and dynamic simulation model deviations. For unpowered airdrops, since there is no assistance from advanced guidance and control equipment, it is necessary to focus on improving the airdrop prediction algorithm. Under the condition of considering the final landing point position error, a more appropriate drop target point should be selected for the airdrop system to maximize the possibility of the materials falling into the delivery area.

[0003] Since there is an error between the actual landing point and the delivery target point, and the target area may be of any irregular shape and there may be unreachable areas inside it, it is very difficult to quickly optimize the airdrop target point. The existing airdrop target point optimization methods mainly include analytical methods, simulation methods and traditional optimization methods. Among them, the simulation method has strong adaptability to the target shape, but its calculation time is long; although the analytical method can quickly give the calculation results, it is only applicable to targets with relatively regular shapes; although the traditional optimization method improves the calculation efficiency, it also faces the problem of long optimization time. Therefore, there is an urgent need for an airdrop target point optimization method that combines accuracy and agility to optimize the quality of the solution to the greatest extent to reduce resource consumption. Summary of the invention

[0004] The present application provides a method, device, electronic equipment and medium for optimizing the target point of an unpowered airdrop, which can quickly and accurately find the optimal delivery target point for an airdrop area of ​​any shape, thereby improving the success rate of the airdrop of supplies.

[0005] A first aspect of an embodiment of the present application provides a method for optimizing a target point for an unpowered airdrop, the method comprising:

[0006] Obtain a target image, and perform traversal sampling on pixel points in a target area and a hollow area of ​​the target image to form a candidate point set;

[0007] For each candidate point in the above candidate point set, extract features of the candidate point in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point;

[0008] The vector features corresponding to each candidate point are input into the delivery target point optimization model to predict the probability of successful delivery when each candidate point is used as the delivery target point. The delivery target point optimization model is obtained by training the deep neural network through supervised learning using a pre-constructed target point training sample set;

[0009] The candidate point with the highest probability of successful delivery is selected as the optimal delivery target point corresponding to the above target image.

[0010] Optionally, the shape of the target area includes three types: circle, rectangle and polygon. When the target area is a circle, the target area is determined by the coordinates of the center point and the radius of the circle; when the target area is a rectangle, the target area is determined by the coordinates of the center point, length, width and the angle between the length direction and the X-axis of the rectangle; when the target area is a polygon, the target area is determined by the coordinates of each corner point of the polygon;

[0011] The shapes of the hollow area include three types: circle, rectangle and polygon. The hollow area is located within the boundary of the target area or partially overlaps with the target area.

[0012] Optionally, the construction of the target point training sample set includes:

[0013] Obtain several target images, randomly generate corresponding target areas and hollow areas, and obtain a sample image set;

[0014] Traverse the pixel points in the target area and the hollow area in the sample image to form a sample point set;

[0015] For each sample point in the sample point set, extract features of the sample point in multiple directions to form a multi-dimensional vector feature corresponding to the sample point;

[0016] According to the preset CEP error, each sample point is simulated for landing, the probability of successful delivery when the sample point is used as the delivery target point is obtained, and the corresponding probability label is generated;

[0017] The multidimensional vector features and probability labels corresponding to the sample points are combined to obtain training samples, and all training samples in each sample image are counted to form the above-mentioned target point training sample set.

[0018] Optionally, features of the sample points in multiple directions are extracted to form multi-dimensional vector features corresponding to the sample points, including:

[0019] With the above sample point as the center, draw several rays evenly outward along the circumference;

[0020] The sum of the lengths of the line segments of each ray falling into the target area is calculated respectively to obtain the features of the sample points in multiple directions, thereby forming a multi-dimensional vector feature corresponding to the sample points.

[0021] Optionally, the above-mentioned performing a landing point simulation on each sample point according to a preset CEP error, obtaining a probability of successful delivery when the sample point is used as a delivery target point, and generating a corresponding probability label, including:

[0022] Determine the CEP error according to the type of target image;

[0023] The sample point is used as the delivery target point, and the sample point is simulated M times according to the determined CEP error. The number of times the landing point is within the above target area is counted, and the probability of successful delivery when the sample point is used as the delivery target point is calculated:

[0024]

[0025] in, is the value of the probability label corresponding to the jth sample point in the kth sample image.

[0026] Optionally, the training of the above delivery target point optimization model includes:

[0027] Constructing a delivery target point optimization model based on a feedforward neural network, comprising an input layer, multiple hidden layers and an output layer, wherein the number of neuron nodes in the input layer is the same as the dimension of the multidimensional vector feature, and the number of neuron nodes in the output layer is the same as the length of the probability label;

[0028] A number of training samples are randomly selected from the target point training sample set, and the multi-dimensional vector features of the sample points in the training samples are input into the delivery target point optimization model to obtain a predicted value of the probability of successful delivery when the sample points are used as delivery target points;

[0029] Based on the above prediction values ​​and the probability labels of the sample points in the training samples, the loss function of the above delivery target point optimization model is calculated:

[0030]

[0031] Among them, y t is the predicted value, is the true value of the probability label, and N is the number of training samples;

[0032] The model is trained using the stochastic gradient descent method, including: obtaining the derivatives of the loss function with respect to the model parameters layer by layer through back propagation, and updating the model parameters until the loss function reaches a preset threshold, and then stopping the updating. The model parameters include the weights and biases corresponding to the neuron nodes of each layer.

[0033] A second aspect of an embodiment of the present application provides a device for optimizing a target point for unpowered airdrops, the device comprising:

[0034] The candidate point acquisition module is used to acquire the target image, traverse and sample the pixel points in the target area and the hollow area of ​​the target image, and form a candidate point set;

[0035] A vector feature extraction module is used to extract the features of each candidate point in the candidate point set in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point;

[0036] A model prediction module is used to input the vector features corresponding to each candidate point into the delivery target point optimization model to predict the probability of successful delivery when each candidate point is used as the delivery target point. The delivery target point optimization model is obtained by training the deep neural network through supervised learning using a pre-constructed target point training sample set;

[0037] The optimal point determination module is used to select the candidate point with the highest probability of successful delivery as the optimal delivery target point corresponding to the above target image.

[0038] Optionally, the above-mentioned vector feature extraction module includes:

[0039] The lead submodule is used to make a number of rays uniformly outward along the circumference direction with the above sample point as the center of the circle;

[0040] The feature calculation submodule is used to respectively calculate the sum of the lengths of the line segments of each ray falling into the above target area, obtain the features of the sample points in multiple directions, and form the multi-dimensional vector features corresponding to the above sample points.

[0041] A third aspect of an embodiment of the present application provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps in the method described above in the first aspect of the present application are implemented.

[0042] A fourth aspect of an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the steps of the method described above in the first aspect of the present application.

[0043] Compared with the prior art, this application has the following advantages:

[0044] The unpowered airdrop target point optimization method provided by the present application mainly includes an offline training stage and an online application stage. In the offline training stage, a pre-constructed target point training sample set is used to train the deep neural network through supervised learning to obtain a delivery target point optimization model. In the online application stage, first, the pixel points in the target area and hollow area of ​​the target image are extracted as candidate points, and the features of each candidate point in multiple directions are calculated to form a multi-dimensional vector feature corresponding to the candidate point, and then input into the delivery target point optimization model for prediction, and the probability of successful delivery when each candidate point is used as a delivery target point is obtained, and finally the candidate point with the largest probability of successful delivery is selected as the optimal delivery target point corresponding to the target image. Thus, the trained delivery target point optimization model is directly used to solve and optimize the unpowered airdrop target point, so as to select a more suitable delivery target point for the airdrop system, so as to maximize the possibility of materials falling into the target area. And through the training of a large number of samples, the adaptability of the delivery target point optimization model is continuously enhanced, and for airdrop areas of any shape, the optimal calculation result can be given more quickly to meet the timeliness requirements of the temporary delivery target point planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0046] Figure 1 is a flow chart of a method for optimizing a target point for unpowered airdrops proposed in one embodiment of the present application;

[0047] Figure 2 is a schematic diagram of a sample image when a hollow area is located within the boundary of a target area in a method for optimizing a target point for unpowered airdrop proposed in an embodiment of the present application;

[0048] Figure 3 It is a schematic diagram of a sample image when a hollow area and a target area partially overlap in the unpowered airdrop target point optimization method proposed in an embodiment of the present application;

[0049] Figure 4 It is a schematic diagram of multi-dimensional vector feature extraction in the unpowered airdrop target point optimization method proposed in one embodiment of the present application;

[0050] Figure 5 It is a schematic diagram of the result of predicting the optimal delivery target point using the unpowered airdrop target point optimization method proposed in one embodiment of the present application;

[0051] Figure 6This is a flow chart of constructing a target point training sample set in the unpowered airdrop target point optimization method proposed in one embodiment of the present application;

[0052] Figure 7 It is a training flow chart of a delivery target point optimization model in an unpowered airdrop target point optimization method proposed in an embodiment of the present application;

[0053] Figure 8 It is a structural schematic diagram of a delivery target point optimization model in a non-powered airdrop target point optimization method proposed in an embodiment of the present application;

[0054] Fig. 9 It is a structural schematic diagram of an unpowered airdrop target point optimization device proposed in one embodiment of the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are 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.

[0056] During the airdrop process, there is a difference between the actual landing point and the delivery target point. At the same time, the target area may be of any irregular shape and there may be unreachable areas inside it. Therefore, it is very difficult to quickly optimize the airdrop target point. The current mainstream airdrop target point optimization methods include simulation methods, analytical methods and traditional optimization methods.

[0057] The simulation method estimates the probability of each candidate target point falling into the delivery area through a large number of simulation experiments. Since it is an approximate calculation, the more simulations are performed, the closer the estimated result is to the true value. In order to obtain the optimal target point, it is necessary to traverse all candidate points, which will take a huge amount of computing time. In practical applications, the simulation method often terminates the calculation after only one feasible solution is obtained.

[0058] The analytical method directly obtains the optimal target point position through a calculation formula. This method takes very little time, but is only suitable for target areas with simple shapes and cannot solve irregular delivery areas with holes.

[0059] In order to make up for the shortcomings of the above two methods, a variety of delivery target point calculation methods based on traditional optimization technology have emerged, including genetic algorithms, particle swarm algorithms, etc. Compared with ergodic simulation, the calculation time of this type of method is shortened, but because each target must be recalculated online and there is a lack of experience to draw on, its optimization time is still difficult to meet the timeliness requirements of ad hoc delivery target point planning.

[0060] In view of this, this application adopts a supervised learning method to solve the unpowered airdrop target point, which is mainly divided into an offline training stage and an online application stage. In the offline training stage, an optimized sample library of unpowered airdrop target points is constructed, and a deep neural network is trained by supervised learning; in the online application stage, the multi-dimensional vector features extracted from the target image are input into the trained deep neural network to obtain the optimal airdrop target point calculation result. Under the supervised learning framework, offline training of neural networks takes a long time, but online applications only require inference calculations, and the time spent is extremely small. At the same time, through the training of a large number of samples, the adaptability of the neural network is continuously enhanced, and it can give the optimal calculation results for airdrop target areas of any shape.

[0061] For details, please refer to Figure 1 , Figure 1 Flowchart of the method for optimizing the target point of unpowered airdrop proposed in one embodiment of the present application. Figure 1 As shown, the method comprises the following steps:

[0062] Step S101: Acquire a target image, and perform traversal sampling on pixel points in a target area and a hollow area of ​​the target image to form a candidate point set.

[0063] In the above steps, the initial image is first obtained, such as a remote sensing satellite image corresponding to the target block where the materials are planned to be airdropped, or a color image obtained by taking a camera from a high altitude. Then, according to actual needs, the shape of the target area and the shape of the hollow area are marked in the initial image, and the target image is obtained after image preprocessing, such as normalization and grayscale processing. Because in actual applications, there may be some inaccessible areas within the boundaries of the designated target area, such as abandoned illegal factories and uninhabited mountaintops. If the materials are airdropped to these inaccessible areas, it will affect the speed of the rescue operation and cause immeasurable consequences. Therefore, when airdropping materials, it is necessary to avoid the materials falling into these inaccessible areas.

[0064] Exemplarily, the shape of the target area includes three types: circle, rectangle and polygon. When the target area is a circle, the target area is determined by the center point coordinates and radius of the circle; when the target area is a rectangle, the target area is determined by the center point coordinates, length, width and the angle between the length and the X-axis of the rectangle; when the target area is a polygon, the target area is determined by the coordinates of each corner point of the polygon. Similarly, the shape of the hollow area includes three types: circle, rectangle and polygon, and the hollow area is located within the boundary of the target area or partially overlaps with the target area.

[0065] Please refer to Figure 2When the hollow area is within the boundary of the target area, the image size is converted to a uniform standard size by normalizing and graying the initial image, and the pixel values ​​of the pixels in the hollow area are uniformly set to 128 by setting the local area pixel values, the pixel values ​​of the remaining pixels within the boundary of the target area after excluding the hollow area pixels are uniformly set to 255, and the pixel values ​​of the pixels outside the boundary of the target area are uniformly set to 0, so as to distinguish the hollow area, the target area and the non-target area. Figure 3 As shown, the black part is the non-target area, the white part with a circular boundary is the target area, and the gray part with a quadrilateral boundary is the hollow area.

[0066] Please refer to Figure 3 , when the hollow area partially overlaps with the target area, it means that the part of the designated hollow area that overlaps with the target area has an impact on the delivery of materials to the target area, and it is necessary to prevent the materials from falling into the overlapping area, so the overlapping area is regarded as the actual hollow area; and the other part has no impact on the delivery of materials to the target area, so it is regarded as the actual non-target area together with the non-target area. That is, the part where the designated hollow area overlaps with the target area is regarded as the actual hollow area, and the remaining part is included in the non-target area. Then, by normalizing and graying the initial image, the pixel values ​​of the pixels in the overlapping area are uniformly set to 128, the pixel values ​​of the remaining pixels within the boundary of the target area after excluding the pixels in the overlapping area are uniformly set to 255, and the pixel values ​​of the pixels outside the boundary of the target area are uniformly set to 0, thereby distinguishing the overlapping area (that is, the actual hollow area), the target area and the non-target area. As Figure 3 As shown, the black part is the actual non-target area, the white part with a rectangular boundary is the target area, and the gray part with a circular boundary is the hollow area.

[0067] After the above image processing is performed on the initial image, the target image is obtained, and the pixel points in the target area and the pixel points in the actual hollow area are taken as candidate points to obtain a candidate point set. For example, there are 6326 candidate points in total, and the candidate point set can be recorded as:

[0068] P t = {P t k} k＝1,...,6326

[0069] Among them, P t is the candidate point set of the target image, P t k is the kth candidate point position.

[0070] Step S102: for each candidate point in the above candidate point set, extract features of the candidate point in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point.

[0071] In the above steps, for any candidate point, take it as the center of the circle, make several rays uniformly outward along the circumferential direction, and then calculate the sum of the lengths of the line segments of each ray falling into the target area, so as to obtain the features of the candidate point in multiple directions and form the multi-dimensional vector features corresponding to the candidate point.

[0072] For example, taking any candidate point as the center, rays are drawn uniformly in 360 directions, and the angle between adjacent rays is 1°, so one candidate point corresponds to 360 rays. Then, in each of the 360 ​​directions, the sum of the lengths of the line segments where the rays fall into the target area is calculated. Figure 4 As shown in the figure, when the kth candidate point makes a ray in a certain direction, the ray is cut off by the hollow area of ​​the quadrilateral, and the line segment formed by the ray falling into the target area is divided into two segments. The sum of the lengths of the two segments is the feature of the candidate point in this direction. In the specific calculation, the expression of the straight line where the ray l1 is located can be calculated by the coordinates of the candidate point and the angle between the ray and the horizontal line, and the upper boundary l can be calculated according to the coordinates of the four corner points of the hollow area of ​​the quadrilateral. a and the lower boundary l b Then solve the equation for the line where ray l1 and upper boundary l a , lower boundary l b As well as the coordinates of the intersection of the circular target area boundary, the distance formula between the two points is used to calculate the lengths of the two line segments where ray l1 falls into the target area, and finally the characteristics of the candidate point in the direction of ray l1 are added together.

[0073] The features of each candidate point in 360 directions are counted to form the 360-dimensional vector features corresponding to the candidate point as follows:

[0074]

[0075] in, is the vector feature corresponding to the kth candidate point, It is the sum of the lengths of the line segments of the rays in each direction falling into the target area.

[0076] Finally, assuming that there are 6326 candidate points in total, the feature set formed by the vector features of all candidate points of the target image is:

[0077]

[0078] It should be noted that the higher the dimension of the vector features of the candidate points, the higher the accuracy of the subsequent optimal delivery target point prediction using the vector features, but at the same time the amount of calculation is also greater. Therefore, the dimension of the vector features of the candidate points is not restricted here and can be set according to actual conditions.

[0079] Step S103: Input the vector features corresponding to each candidate point into the delivery target point optimization model to predict the probability of successful delivery of each candidate point as the delivery target point. The above delivery target point optimization model is obtained by using a pre-constructed target point training sample set and training the deep neural network through supervised learning.

[0080] In the above steps, the multi-dimensional vector features corresponding to the calculated candidate points are directly input into the pre-trained delivery target point optimization model, and the predicted probability values ​​corresponding to each candidate point can be obtained through forward calculation. The predicted probability value refers to the probability that the final delivery point falls within the target area when a candidate point is used as a delivery target point for delivery, that is, the probability of successful delivery with the candidate point as the delivery target point.

[0081] Exemplarily, a feedforward neural network can be used to construct a delivery target point optimization model, and then several groups of training samples are randomly selected from the target point training sample set corresponding to the sample image, and the neural network model is trained by a stochastic gradient descent algorithm. Specifically, by inputting the vector features of the sample points into the neural network model, the output predicted probability value is obtained, and it is compared with the true value of the probability to calculate the loss function Loss. Then, the derivative of the loss function to each parameter of the model (weights W and bias b corresponding to each layer of neuron nodes, etc.) is obtained layer by layer through back propagation, and each parameter of the model is updated. This cycle is executed multiple times until the model converges, thereby obtaining a trained delivery target point optimization model.

[0082] Among them, for sample images, when delineating target areas and hollow areas, the shapes of target areas and hollow areas can be randomly generated based on circles, rectangles, and complex polygons, and then used for model training, so that the types of sample images are richer, thereby improving the accuracy of model prediction. Therefore, through the training of a large number of samples of different shapes, the adaptability of the delivery target point optimization model is continuously enhanced, and for airdrop areas of any shape, the optimal calculation results can be given more quickly to meet the timeliness requirements of the planning of ad hoc delivery target points.

[0083] Step S104: Select the candidate point with the highest probability of successful delivery as the optimal delivery target point corresponding to the target image.

[0084] In the above steps, after obtaining the predicted probability values ​​corresponding to all candidate points in the above target image, the predicted probability values ​​can be sorted in descending order, and the candidate point corresponding to the first predicted probability value can be selected as the optimal delivery target point corresponding to the target image, that is, the candidate point with the highest probability of successful delivery is selected as the optimal delivery target point corresponding to the target image, thereby maximizing the possibility of successful delivery of materials.

[0085] For example, assume that there are 6326 candidate points in total, among which the kth candidate point P t k The probability value predicted by the delivery target point optimization model is Then the probability set formed by the predicted probability values ​​of all candidate points of the target image is:

[0086]

[0087] Then, the candidate point with the highest probability of successful delivery is selected as the optimal delivery target point corresponding to the target image through the maximum value function:

[0088]

[0089] Among them, K is the serial number of the candidate point corresponding to the optimal delivery target point, is the predicted probability value corresponding to the kth candidate point.

[0090] In order to test and verify the performance of the delivery target point optimization model proposed in this application, the test samples are first obtained, and the target area and hollow area are randomly generated. Then, the unpowered airdrop target point optimization method of this application and the simulation method in the prior art are used to predict the optimal delivery target point of the test samples. The test results are as follows: Figure 5 shown.

[0091] Figure 5In the figure, the white part is the target area, which is circular in shape, the gray part is the hollow area, which is quadrilateral in shape, and the black part is the non-target area. The optimal delivery target point predicted by the unpowered airdrop target point optimization method of the present application is shown as the triangle point in the figure, whose coordinates are (38, 26), the corresponding prediction probability value is 0.918, and the calculation time is 0.146 seconds; the optimal delivery target point predicted by the simulation method in the prior art is shown as the circular point in the figure, whose coordinates are (40, 27), the corresponding prediction probability value is 0.924, and the calculation time is about 50 seconds. It can be found that the final positions of the optimal delivery target point predicted by the unpowered airdrop target point optimization method of the present application and the optimal delivery target point predicted by the simulation method are very close, but the calculation time is greatly shortened, indicating that the unpowered airdrop target point optimization method of the present application can not only achieve the effect of the existing method in terms of prediction accuracy and effectiveness, but also spends very little time, has faster calculation speed and higher agility. Compared with the existing technology, it can better meet the timeliness requirements of ad hoc delivery target point planning.

[0092] In another embodiment of the present application, the process of training a deep neural network by supervised learning using a pre-constructed target point training sample set to obtain a delivery target point optimization model is as follows:

[0093] Step S201: construct a target point training sample set.

[0094] Please refer to Figure 6 , constructing the target point training sample set specifically includes:

[0095] Step S201 - 1 : Acquire several target images, randomly generate corresponding target areas and hollow areas, and obtain a sample image set.

[0096] In the above steps, a target image set is formed by acquiring several target images:

[0097] I={I k} k＝1,...,N

[0098] Among them, I is the target image set, I k is a grayscale image with width and height L, and N is the number of images in the collection.

[0099] For each target image, a target area and a hollow area are randomly generated, and the target image is processed for local area pixel values. The pixel values ​​of the pixels in the hollow area are uniformly set to 128, the pixel values ​​of the remaining pixels within the boundary of the target area after excluding the pixels in the hollow area are uniformly set to 255, and the pixel values ​​of the pixels outside the boundary of the target area are uniformly set to 0, thereby distinguishing the hollow area, the target area and the non-target area to obtain a sample image.

[0100] Exemplarily, the shapes of the target area and the hollow area can be circular, rectangular, and complex polygonal. When generating a sample image set, the shape of the target area can be set in a ratio of circular: rectangular: complex polygon = 1:1:2 to improve the diversity of the sample. For example, 40 sample images are obtained, of which 10 have circular, 10 have rectangular, and 20 have complex polygonal shapes.

[0101] Step S201 - 2 : traverse the pixel points in the target area and the hollow area in the sample image to form a sample point set.

[0102] For each sample image, traverse all pixels in its target area and hollow area to form a sample point set:

[0103]

[0104] Among them, P k is the sample point set corresponding to the kth sample image, P j is the jth sample point that meets the regional requirements, M k is the number of pixels in the kth sample image that meet the region requirements.

[0105] Step S201 - 3 : For each sample point in the sample point set, extract features of the sample point in multiple directions to form a multi-dimensional vector feature corresponding to the sample point.

[0106] For each sample point P in a sample image j , take it as the center of the circle, make several rays uniformly outward along the circumference, and then calculate the sum of the lengths of the line segments of each ray falling into the target area, and get the sample point P j Features in multiple directions form sample points P j The corresponding multi-dimensional vector features.

[0107] For example, taking the sample point P j As the center, make rays evenly in 360 directions, and the angle between adjacent rays is 1°, so a total of 360 rays can be made. i ,i=1,...,360. Count each ray l i The sum of the lengths of the line segments falling into the target area d i , so the corresponding 360-dimensional vector features can be obtained as follows:

[0108]

[0109] in, is the sample point P in the sample image j The corresponding vector features.

[0110] Therefore, the set of vector features of each sample point extracted from the sample image is:

[0111]

[0112] Among them, X k is the feature set corresponding to the sample image.

[0113] Therefore, the set of vector features extracted from all sample images for each sample point is:

[0114]

[0115] Among them, X is the feature set corresponding to each sample point that meets the regional requirements in all sample images.

[0116] Step S201 - 4 : performing landing point simulation on each sample point according to the preset CEP error, obtaining the probability of successful delivery when the sample point is used as the delivery target point, and generating a corresponding probability label.

[0117] In the above steps, the CEP error is first determined according to the type of the original target image. CEP, also known as circular error probability, is an indicator used to describe the error distribution. It means that within a circular area, 50% of the measured values ​​fall within this area. The radius of this circular area is the CEP. For example, if the CEP of a navigation system is 10 meters, then 50% of the measured values ​​of this system will fall within a circular area with a radius of 10 meters centered on the user's position. If the target image is obtained based on this navigation system, the CEP error is set to 10.

[0118] Then, the sample point is used as the delivery target point, and the sample point is simulated M times according to the determined CEP error. The number of times the landing point is within the above target area is counted, and the probability of successful delivery when the sample point is used as the delivery target point is calculated:

[0119]

[0120] in, is the value of the probability label corresponding to the jth sample point in the kth sample image.

[0121] Therefore, the set of probability label values ​​formed by each sample point extracted from a sample image is:

[0122]

[0123] Among them, y k is the set of probability label values ​​corresponding to the sample image.

[0124] Therefore, the set of probability label values ​​of each sample point extracted from all sample images is:

[0125]

[0126] Among them, y is the set of probability label values ​​corresponding to the sample points of all sample images that meet the regional requirements.

[0127] Step S201 - 5 : combining the multidimensional vector features and probability labels corresponding to the sample points to obtain training samples, and counting all the training samples in each sample image to form the above-mentioned target point training sample set.

[0128] By combining the feature set X obtained in step S201-3 and the probability label value set y obtained in step S201-4, the features corresponding to the subscripts and subscripts and probability label value Combine them to get a set of training sample data. That is, the vector features and probability label values ​​corresponding to each sample point constitute a set of sample data for training. So all the training sample data can be expressed as:

[0129]

[0130] Where D is the target point training sample set, (X t ,y t ) is any set of training samples in the sample set, is the total number of training samples.

[0131] Step S202: construct a delivery target point optimization model and train it using a target point training sample set.

[0132] Please refer to Figure 7 ,The training process of the delivery target point optimization model specifically includes:

[0133] Step S202-1: Construct a delivery target point optimization model based on a feedforward neural network, including an input layer, multiple hidden layers and an output layer, wherein the number of neuron nodes in the input layer is the same as the dimension of the multidimensional vector feature, and the number of neuron nodes in the output layer is the same as the length of the probability label.

[0134] In the above steps, the structure of the delivery target point optimization model constructed is as follows: Figure 8 As shown in Figure 1, it includes 1 input layer, 2 hidden layers and 1 output layer. Assuming that the dimension of the sample point vector feature extracted above is 360, the number of neuron nodes in the input layer of the delivery target point optimization model is also 360; the number of neuron nodes in the output layer is the same as the probability label The length of is the same, which is 1; the number of neuron nodes in each hidden layer is 1024.

[0135] Step S202-2: randomly extract a number of training samples from the target point training sample set, input the multidimensional vector features of the sample points in the training samples into the delivery target point optimization model, and obtain the predicted value of the probability of successful delivery when the sample points are used as delivery target points.

[0136] Step S202-3: Based on the above prediction values ​​and the probability labels of the sample points in the training samples, the loss function of the above delivery target point optimization model is calculated:

[0137]

[0138] Among them, y t is the predicted value, is the true value of the probability label, and N is the number of training samples.

[0139] For example, assume that 1000 sample points are randomly selected from the sample image each time, and their features are input into the neural network model, and the output of the model is 1000 predicted values ​​y t ,t=1,...,1000, the true value of the corresponding probability label is The loss function is calculated as follows:

[0140]

[0141] Step S202-4: Use the stochastic gradient descent method to train the model, including: obtaining the derivatives of the loss function with respect to the model parameters layer by layer through back propagation, and updating the model parameters until the loss function reaches a preset threshold and then stops updating. The model parameters include the weights W and bias b corresponding to the neuron nodes of each layer.

[0142] By combining the stochastic gradient descent method and the error back propagation method, the gradient of the loss function is first calculated for the weights and other parameters of all neuron nodes in the neural network through forward propagation, and then fed back to the neuron nodes of each layer to update the weights and other parameters to minimize the loss function, thereby training and optimizing the model.

[0143] Based on the same inventive concept, the second aspect of the embodiment of the present application provides a non-powered airdrop target point optimization device, such as Figure 8 As shown, the device comprises:

[0144] The candidate point acquisition module 1 is used to acquire a target image, and perform traversal sampling on the pixel points in the target area and the hollow area of ​​the target image to form a candidate point set;

[0145] The vector feature extraction module 2 is used to extract the features of each candidate point in the candidate point set in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point;

[0146] Model prediction module 3, used to input the vector features corresponding to each candidate point into the delivery target point optimization model, and predict the probability of successful delivery when each candidate point is used as the delivery target point. The delivery target point optimization model is obtained by training the deep neural network through supervised learning using a pre-constructed target point training sample set;

[0147] The optimal point determination module 4 is used to select a candidate point with the highest probability of successful delivery as the optimal delivery target point corresponding to the target image.

[0148] Optionally, in the above device, the shape of the target area includes three types: circle, rectangle and polygon. When the target area is a circle, the target area is determined by the coordinates of the center point and the radius of the circle; when the target area is a rectangle, the target area is determined by the coordinates of the center point, length, width and the angle between the length direction and the X-axis of the rectangle; when the target area is a polygon, the target area is determined by the coordinates of each corner point of the polygon;

[0149] The shapes of the hollow area include three types: circle, rectangle and polygon. The hollow area is located within the boundary of the target area or partially overlaps with the target area.

[0150] Optionally, the above-mentioned vector feature extraction module includes:

[0151] The lead submodule is used to make a number of rays uniformly outward along the circumference direction with the above sample point as the center of the circle;

[0152] The feature calculation submodule is used to respectively calculate the sum of the lengths of the line segments of each ray falling into the above target area, obtain the features of the sample points in multiple directions, and form the multi-dimensional vector features corresponding to the above sample points.

[0153] Optionally, in the above device, the construction of the target point training sample set includes:

[0154] The sample image acquisition submodule is used to acquire several target images, randomly generate corresponding target areas and hollow areas, and obtain a sample image set;

[0155] A sample point acquisition submodule is used to traverse the pixel points in the target area and the hollow area in the sample image to form a sample point set;

[0156] A feature extraction submodule is used to extract the features of each sample point in the sample point set in multiple directions to form a multi-dimensional vector feature corresponding to the sample point;

[0157] The probability label generation submodule is used to simulate the landing point of each sample point according to the preset CEP error, obtain the probability of successful delivery when the sample point is used as the delivery target point, and generate the corresponding probability label;

[0158] The training sample set generation submodule is used to combine the multidimensional vector features and probability labels corresponding to the sample points to obtain training samples, and to count all training samples in each sample image to form the above-mentioned target point training sample set.

[0159] Optionally, the above-mentioned probability label generation submodule includes:

[0160] A CEP error determination unit, used to determine the CEP error according to the type of the target image;

[0161] The landing point simulation unit is used to use the sample point as the delivery target point, perform M landing point simulations on the sample point according to the determined CEP error, and count the number of times the landing point is within the above target area, and calculate the probability of successful delivery when the sample point is used as the delivery target point:

[0162]

[0163] in, is the value of the probability label corresponding to the jth sample point in the kth sample image.

[0164] Optionally, in the above device, the training of the delivery target point optimization model includes:

[0165] A model building submodule, used to build a delivery target point optimization model based on a feedforward neural network, including an input layer, multiple hidden layers and an output layer, wherein the number of neuron nodes in the input layer is the same as the dimension of the multidimensional vector feature, and the number of neuron nodes in the output layer is the same as the length of the probability label;

[0166] The sample prediction submodule is used to randomly extract a number of training samples from the target point training sample set, input the multi-dimensional vector features of the sample points in the training samples into the delivery target point optimization model, and obtain the predicted value of the probability of successful delivery when the sample points are used as delivery target points;

[0167] The loss calculation submodule is used to calculate the loss function of the above delivery target point optimization model based on the above prediction value and the probability label of the sample point in the training sample:

[0168]

[0169] Among them, y t is the predicted value, is the true value of the probability label, and N is the number of training samples;

[0170] The parameter updating submodule is used to train the model using the stochastic gradient descent method, including: obtaining the derivatives of the above loss function with respect to the model parameters layer by layer through back propagation, and updating the model parameters until the above loss function reaches a preset threshold and then stops updating. The above model parameters include the weights and biases corresponding to the neuron nodes of each layer.

[0171] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0172] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0174] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0175] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0176] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0177] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0178] The above is a detailed introduction to the unpowered airdrop target point optimization method, device, electronic device and medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for optimizing target points of unpowered airdrop, characterized in that: include: Acquire a target image, and perform traversal sampling on pixel points in a target area and a hollow area of ​​the target image to form a candidate point set; For each candidate point in the candidate point set, extracting features of the candidate point in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point, including: taking the candidate point as the center of a circle, uniformly making a number of rays outward along the circumference direction; respectively calculating the sum of the lengths of the line segments of each ray falling into the target area, obtaining features of the candidate point in multiple directions, and forming a multi-dimensional vector feature corresponding to the candidate point; Inputting the vector features corresponding to each candidate point into the delivery target point optimization model, and predicting the probability of successful delivery when each candidate point is used as the delivery target point, the delivery target point optimization model is obtained by training a deep neural network through supervised learning using a pre-constructed target point training sample set; The candidate point with the highest probability of successful delivery is selected as the optimal delivery target point corresponding to the target image.

2. The unpowered airdrop target point optimization method according to claim 1, characterized in that: The shape of the target area includes three types: circle, rectangle and polygon. When the target area is a circle, the target area is determined by the center point coordinates and radius of the circle; when the target area is a rectangle, the target area is determined by the center point coordinates, length, width and angle between the length direction and the X-axis of the rectangle; When the target area is a polygon, the target area is determined by the coordinates of each corner point of the polygon; The shapes of the hollow area include three types: circle, rectangle and polygon. The hollow area is located within the boundary of the target area or partially overlaps with the target area.

3. The unpowered airdrop target point optimization method according to claim 1 or 2, characterized in that: The construction of the target point training sample set includes: Obtain several target images, randomly generate corresponding target areas and hollow areas, and obtain a sample image set; Traversing the pixel points in the target area and the hollow area in the sample image to form a sample point set; For each sample point in the sample point set, extract features of the sample point in multiple directions to form a multi-dimensional vector feature corresponding to the sample point; According to the preset CEP error, each sample point is simulated for landing, the probability of successful delivery when the sample point is used as the delivery target point is obtained, and the corresponding probability label is generated; The multidimensional vector features and probability labels corresponding to the sample points are combined to obtain training samples, and all training samples in each sample image are counted to form the target point training sample set.

4. The unpowered airdrop target point optimization method according to claim 3 is characterized in that: The method of performing a drop point simulation on each sample point according to the preset CEP error to obtain the probability of successful delivery when the sample point is used as the delivery target point and generating a corresponding probability label includes: Determine the CEP error according to the type of target image; The sample point is used as the delivery target point, and the sample point is simulated M times according to the determined CEP error. The number of times the landing point is in the target area is counted, and the probability of successful delivery when the sample point is used as the delivery target point is calculated: in, is the value of the probability label corresponding to the jth sample point in the kth sample image.

5. The unpowered airdrop target point optimization method according to claim 3, characterized in that: The training of the delivery target point optimization model includes: Constructing a delivery target point optimization model based on a feedforward neural network, comprising an input layer, a plurality of hidden layers and an output layer, wherein the number of neuron nodes in the input layer is the same as the dimension of the multidimensional vector feature, and the number of neuron nodes in the output layer is the same as the length of the probability label; Randomly extracting a number of training samples from the target point training sample set, inputting the multidimensional vector features of the sample points in the training samples into the delivery target point optimization model, and obtaining a predicted value of the probability of successful delivery when the sample points are used as delivery target points; Based on the predicted value and the probability labels of the sample points in the training samples, the loss function of the delivery target point optimization model is calculated: Among them, y t is the predicted value, is the true value of the probability label, and N is the number of training samples; The model is trained using the stochastic gradient descent method, including: obtaining the derivatives of the loss function with respect to the model parameters layer by layer through back propagation, and updating the model parameters until the loss function reaches a preset threshold and then stops updating. The model parameters include the weights and biases corresponding to the neuron nodes of each layer.

6. A non-powered airdrop target point optimization device, characterized in that: include: A candidate point acquisition module is used to acquire a target image, traverse and sample pixel points in a target area and a hollow area of ​​the target image, and form a candidate point set; A vector feature extraction module, used to extract features of each candidate point in the candidate point set in multiple directions to form a multi-dimensional vector feature corresponding to the candidate point; A model prediction module is used to input the vector features corresponding to each candidate point into a delivery target point optimization model to predict the probability of successful delivery of each candidate point as a delivery target point. The delivery target point optimization model is obtained by training a deep neural network in a supervised learning manner using a pre-constructed target point training sample set; The optimal point determination module is used to select a candidate point with the highest probability of successful delivery as the optimal delivery target point corresponding to the target image.

7. The unpowered airdrop target point optimization device according to claim 6, characterized in that: The vector feature extraction module comprises: The lead submodule is used to make a number of rays uniformly outward along the circumference with the sample point as the center of the circle; The feature calculation submodule is used to respectively calculate the sum of the lengths of the line segments of each ray falling into the target area, obtain the features of the sample point in multiple directions, and form a multi-dimensional vector feature corresponding to the sample point.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 5 are implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

Images