A UAV target recognition and automatic tracking algorithm platform

Abstract

The present invention relates to the technical field of image processing, and discloses an unmanned aerial vehicle target recognition and automatic tracking algorithm platform, an image acquisition module, which is used for collecting characteristic images of a target based on multiple unmanned aerial vehicles at fixed time intervals; an image scoring module, which is used for evaluating the characteristic images based on preset rules and calculating the tracking score of each characteristic image; a dynamic optimization module, which is used for determining a first observation point of a target by combining multiple unmanned aerial vehicles and an optimization algorithm, and assigning a unmanned aerial vehicle to the first observation point; a tracking planning module, which is used for determining second to Nth observation points according to the first observation point, and assigning the second to Nth observation points to the remaining unmanned aerial vehicles one by one to obtain a tracking circle of the target; and a dynamic adjustment module, which is used for acquiring characteristic images based on time series, and generating a predicted position of the target at the next moment in combination with a preset prediction model.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to an unmanned aerial vehicle target recognition and automatic tracking algorithm platform. Background Art

[0002] With the development of UAV technology and the expansion of its application areas, the application of UAV in target recognition and automatic tracking tasks has become increasingly important. These tasks usually involve the recognition and tracking of specific targets.

[0003] However, in the existing technology, the coordination strategy of multiple drones is based on static predefined distribution, and it is impossible to dynamically adjust the position distribution of drones according to the motion state of the target and environmental changes, resulting in low tracking accuracy. In addition, the information sharing mechanism between drones limits the coordination of the multi-drone system. Summary of the invention

[0004] The present invention provides an unmanned aerial vehicle target recognition and automatic tracking algorithm platform to solve the technical problems raised in the background technology.

[0005] The present invention provides a UAV target recognition and automatic tracking algorithm platform, comprising:

[0006] An image acquisition module is used to collect characteristic images of the target based on multiple drones at fixed time intervals;

[0007] An image scoring module, used to evaluate feature images based on preset rules and calculate the tracking score of each feature image;

[0008] A dynamic optimization module, for determining a first observation point of a target by combining multiple drones and an optimization algorithm, and assigning a drone to the first observation point;

[0009] A tracking planning module is used to determine the second to Nth observation points according to the first observation point, and to assign the second to Nth observation points to the remaining UAVs one by one to obtain a tracking circle of the target;

[0010] The dynamic adjustment module is used to obtain the feature image based on the time series and generate the predicted position of the target at the next moment in combination with the preset prediction model;

[0011] If the prediction is successful, the target is continuously tracked using the target’s tracking circle;

[0012] If the prediction fails, the prediction model is updated in reverse based on the predicted position and the actual position of the target, and returns to the dynamic optimization module.

[0013] Furthermore, preset rules include:

[0014] The feature image is semantically segmented based on the semantic tool; the semantic segmentation includes outputting the pixels of the target in the feature image as 1 and outputting the pixels of the non-target in the feature image as 0, thereby obtaining a semantic matrix;

[0015] Divide the semantic matrix into central area, sub-central area and peripheral area;

[0016] The number of elements with element values ​​of 1 in the central area, sub-central area, and edge area of ​​the semantic matrix is ​​counted respectively, and the tracking score of the feature image is calculated. The calculation formula of the tracking score is as follows:

[0017] ;

[0018] in, represents the tracking score of the feature image, Indicates the number of elements whose value in the center area is 1. represents the center weight, Indicates the number of elements whose element value is 1 in the secondary center area. represents the secondary center weight, Indicates the number of elements whose element value is 1 in the edge area. represents the edge weight, represents the target ideal weight, Indicates the number of elements whose target ideal element value is 1. It means to find the absolute value. , and All are not 0, , and The sum is 1, .

[0019] Furthermore, multiple drones and optimization algorithms are combined to determine the first observation point of the target, including:

[0020] Step 1, The drones are initialized based on the constraints to obtain the optimized population.

[0021] The individual code is represented as: ;in, represents the flight speed of the i-th UAV, represents the flight angle of the i-th UAV, represents the shooting angle of the i-th drone, ;

[0022] The constraints include: ;in, Indicates the minimum flight speed of the drone. Indicates the maximum flight speed of the drone;

[0023] , ;in, Indicates the maximum flight angle of the drone on the left side of the horizontal direction. Indicates the maximum flight angle of the drone on the right side in the horizontal direction. Indicates the maximum upward flight angle of the drone in the vertical direction. Indicates the maximum flight angle of the drone in the vertical direction;

[0024] , ;in, Indicates that in the horizontal direction, the drone is Adjust the shooting angle. Indicates that in the vertical direction, the drone is Adjust the shooting angle. Indicates the maximum shooting angle of the drone in the vertical direction. Indicates the maximum shooting angle of the drone in the vertical direction;

[0025] Step 2, the tracking score of the feature image obtained by each UAV based on the corresponding individual coding scheme is used as the individual objective function;

[0026] Step 3, calculate the objective function value of the individual, and sort each individual from large to small based on the objective function value to obtain the individual feature sorting;

[0027] Step 4: sort the features from front to back and retain a preset number of individuals, and use the remaining individuals as parents for update iteration;

[0028] Step 5, after looping steps 3 and 4 for a preset number of update iterations, stop the update iteration, output the individual with the largest objective function value, and take the position of the drone corresponding to the individual with the largest objective function value as the first observation point.

[0029] Furthermore, the remaining individuals are used as parents for update iterations, including:

[0030] Divide the parents into two groups, each including parents, including a first update group and a second update group;

[0031] The calculation formula for the first update group is as follows:

[0032] ;

[0033] in, Indicates the first update group Parent The output individual encoding of the update iteration is Indicates the first update group Parent The output individual encoding of the update iteration is represents the first updated weight, represents the mean individual code of the first update group, , represents the first mean deviation weight, represents the correction bias parameter, , represents the random perturbation individual code, , Represents a random number;

[0034] The calculation formula for the second update group is as follows:

[0035] ;

[0036] in, Indicates the first Parent The output individual encoding of the update iteration is Indicates the first Parent The output individual encoding of the update iteration is represents the second updated weight, represents the second mean deviation weight, Indicates the preset number of update iterations, Indicates Update iterations, Indicates The individual encoding that ranks first in the feature ranking of the update iteration.

[0037] Further, determining the second to Nth observation points according to the first observation point includes:

[0038] Based on the topological structure constructed by the elements with element values ​​of 1 in the semantic matrix, the elements with element values ​​of 1 are mapped to nodes of the topological structure, and edges are constructed between adjacent nodes to obtain the topological structure;

[0039] Step 6, constructing an image coordinate system according to the feature image corresponding to the first observation point; wherein the origin of the image coordinate system is the central element of the semantic matrix corresponding to the feature image;

[0040] Map the target to the image coordinate system, and the calculation formula is as follows:

[0041] , , ;

[0042] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, and They represent the horizontal focal length and vertical focal length of the drone respectively. Represents the horizontal coordinate of the element corresponding to the central node in the topological structure, Indicates the ordinate of the element corresponding to the central node in the topological structure, Indicates the actual height of the target. Represents the Euclidean distance between any two nodes in the topological structure. represents the abscissa of the central element of the semantic matrix, The ordinate represents the central element of the semantic matrix;

[0043] Step 7: Based on the first observation point, construct the drone coordinate system and , and They are mapped to the drone coordinate system respectively; the origin of the drone coordinate system represents the drone corresponding to the first observation point; the calculation formula is as follows:

[0044] , ;

[0045] , ;

[0046] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, represents the rotation matrix, represents the yaw angle matrix, represents the pitch angle matrix, Indicates the flight angle of the drone corresponding to the first observation point, Indicates the shooting angle of the drone corresponding to the first observation point;

[0047] Step 8: Based on the GPS coordinates of the drone, , and They are mapped in the geographic coordinate system respectively, and the calculation formula is as follows:

[0048] , , ;

[0049] , ;

[0050] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system. , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the drone. and Represent the horizontal axis increment and the vertical axis increment respectively. represents the radius of the Earth;

[0051] Step 9, taking the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system as the center of the circle, the Euclidean distance between the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the UAV corresponding to the target and the first observation point as the radius, and the angle between the UAV corresponding to the target and the first observation point and the horizontal plane as the rotation angle, to obtain the characteristic circle; divide the characteristic circle into N equal parts, and take the first observation point as the starting point to obtain the second to Nth observation points, and divide the remaining UAVs into the second to Nth observation points one by one.

[0052] Furthermore, the preset prediction models include:

[0053] Step 10, based on steps 6 to 8, the feature image of the historical time period is processed to obtain the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the target in the geographic coordinate system in the historical time period;

[0054] Step 11, constructing a training parameter set according to the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the time series; extracting a number of fixed-length training parameters from the training sample set through a sliding time window of length h to obtain a number of training samples, and using the training parameters of the corresponding training samples at the next moment as sample labels;

[0055] Step 12: A prediction model is obtained based on training samples and sample labels, and a hidden layer of the prediction model is constructed based on a recurrent neural network.

[0056] Furthermore, the target is continuously tracked using the target's tracking circle, including:

[0057] The horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system from the gth moment to the g+hth moment are constructed as feature vectors and input into the prediction model to obtain the predicted position of the target at the g+h+1th moment;

[0058] With the predicted target position at the g+h+1th moment as the center, a characteristic circle is constructed, and N UAVs are assigned to the characteristic circle at the g+h+1th moment to obtain the target's tracking circle for continuous tracking of the target.

[0059] Furthermore, the prediction model is updated in reverse, including:

[0060] The mean square error between the target predicted position and the actual target position at the g+h+1th moment is used as the loss function of the hidden layer of the prediction model, and the weight parameters and bias parameters of the hidden layer of the prediction model are updated through back propagation.

[0061] The beneficial effects of the present invention are as follows: by introducing semantic segmentation technology to process the target image, the characteristic pixels of the target are classified and weighted, thereby realizing intelligent scoring and optimization of the image; combining with the dynamic optimization algorithm, based on the image segmentation results and UAV parameters, the optimal observation point and distribution scheme are calculated in real time; at the same time, the time series images are analyzed and predicted through the recurrent neural network to generate the motion trajectory and next position of the target, ensuring that the multi-UAV system can accurately track the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a module diagram of a UAV target recognition and automatic tracking algorithm platform of the present invention;

[0063] Figure 2 It is a schematic diagram of a tracking circle of the object of the present invention. DETAILED DESCRIPTION

[0064] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0065] like Figure 1 and Figure 2 As shown, a UAV target recognition and automatic tracking algorithm platform includes:

[0066] An image acquisition module is used to collect characteristic images of the target based on multiple drones at fixed time intervals;

[0067] An image scoring module, used to evaluate feature images based on preset rules and calculate the tracking score of each feature image;

[0068] A dynamic optimization module, for determining a first observation point of a target by combining multiple drones and an optimization algorithm, and assigning a drone to the first observation point;

[0069] A tracking planning module is used to determine the second to Nth observation points according to the first observation point, and to assign the second to Nth observation points to the remaining UAVs one by one to obtain a tracking circle of the target;

[0070] The dynamic adjustment module is used to obtain the feature image based on the time series and generate the predicted position of the target at the next moment in combination with the preset prediction model;

[0071] If the prediction is successful, the target is continuously tracked using the target’s tracking circle;

[0072] If the prediction fails, the prediction model is updated in reverse based on the predicted position and the actual position of the target, and returns to the dynamic optimization module.

[0073] In one embodiment of the present invention, the preset rules include:

[0074] The feature image is semantically segmented based on the semantic tool; the semantic segmentation includes outputting the pixels of the target in the feature image as 1 and outputting the pixels of the non-target in the feature image as 0, thereby obtaining a semantic matrix;

[0075] Divide the semantic matrix into central area, sub-central area and peripheral area;

[0076] The number of elements with element values ​​of 1 in the central area, sub-central area, and edge area of ​​the semantic matrix is ​​counted respectively, and the tracking score of the feature image is calculated. The calculation formula of the tracking score is as follows:

[0077] ;

[0078] in, represents the tracking score of the feature image, Indicates the number of elements whose value in the center area is 1. represents the center weight, Indicates the number of elements whose element value is 1 in the secondary center area. represents the secondary center weight, Indicates the number of elements whose element value is 1 in the edge area. represents the edge weight, represents the target ideal weight, Indicates the number of elements whose target ideal element value is 1. It means to find the absolute value. , and All are not 0, , and The sum is 1, .

[0079] Specifically, the characteristic image of the target is obtained by the drone, and the characteristic image is finely divided and scored to quantify the effect of the drone on target tracking. Through scientific scoring models and weight design, the quality of the characteristic image can be dynamically evaluated, thereby guiding the drone to track the target with higher precision and achieve continuous observation of the target. The semantic segmentation of the characteristic image is the basis for achieving tracking scoring. Specifically, the target pixels (such as the target body) in the characteristic image are marked as 1 by the semantic segmentation tool; non-target pixels (such as the background area) are marked as 0, thereby generating a semantic matrix. In the semantic matrix, the image is divided into the central area, sub-central area and edge area according to the positional relationship: the central area represents the pixel set of the target in the center of the image, reflecting whether the target is in the ideal observation position. The sub-central area represents the pixel set of the target slightly off the center but still in the suboptimal position. The edge area represents the pixel set of the target close to the edge of the image, indicating that there may be problems with the drone's observation of the target.

[0080] In one embodiment of the present invention, a first observation point of a target is determined by combining multiple drones and an optimization algorithm, including:

[0081] Step 1, The drones are initialized based on the constraints to obtain the optimized population.

[0082] The individual code is represented as: ;in, represents the flight speed of the i-th UAV, represents the flight angle of the i-th UAV, represents the shooting angle of the i-th drone, ;

[0083] The constraints include: ;in, Indicates the minimum flight speed of the drone. Indicates the maximum flight speed of the drone;

[0084] , ;in, Indicates the maximum flight angle of the drone on the left side of the horizontal direction. Indicates the maximum flight angle of the drone on the right side in the horizontal direction. Indicates the maximum upward flight angle of the drone in the vertical direction. Indicates the maximum flight angle of the drone in the vertical direction;

[0085] , ;in, Indicates that in the horizontal direction, the drone is Adjust the shooting angle. Indicates that in the vertical direction, the drone is Adjust the shooting angle. Indicates the maximum shooting angle of the drone in the vertical direction. Indicates the maximum shooting angle of the drone in the vertical direction;

[0086] Step 2, the tracking score of the feature image obtained by each UAV based on the corresponding individual coding scheme is used as the individual objective function;

[0087] Step 3, calculate the objective function value of the individual, and sort each individual from large to small based on the objective function value to obtain the individual feature sorting;

[0088] Step 4: sort the features from front to back and retain a preset number of individuals, and use the remaining individuals as parents for update iteration;

[0089] Step 5, after looping steps 3 and 4 for a preset number of update iterations, stop the update iteration, output the individual with the largest objective function value, and take the position of the drone corresponding to the individual with the largest objective function value as the first observation point.

[0090] The flight parameters of each drone are encoded through the optimization algorithm, and multiple drones are used to collaborate to collect feature images of the target. The flight speed, flight angle and shooting direction of the drone are screened by the optimization algorithm, and the scoring results of the feature images are integrated to dynamically optimize the observation position of the drone, so as to determine a high-quality target observation point that meets the requirements. The drone observation position corresponding to the individual with the highest objective function value is determined as the first observation point. This point is the initial key position for multi-drone collaboration and can provide the best target observation effect. The scoring model not only considers the location distribution of the target (such as whether it is close to the center), but also combines the deviation of the number of target pixels from the ideal value to comprehensively quantify the observation effect of the drone. By using the scoring value as the objective function of the optimization algorithm, it is ensured that the optimization result directly serves the drone target tracking task.

[0091] It should be noted that the semantic segmentation tool can be Google's DeepLab.

[0092] In one embodiment of the present invention, the remaining individuals are used as parents to perform update iterations, including:

[0093] Divide the parents into two groups, each including parents, including a first update group and a second update group;

[0094] The calculation formula for the first update group is as follows:

[0095] ;

[0096] in, Indicates the first update group Parent The output individual encoding of the update iteration is Indicates the first update group Parent The output individual encoding of the update iteration is represents the first updated weight, represents the mean individual code of the first update group, , represents the first mean deviation weight, represents the correction bias parameter, , represents the random perturbation individual code, , Represents a random number;

[0097] The calculation formula for the second update group is as follows:

[0098] ;

[0099] in, Indicates the first Parent The output individual encoding of the update iteration is Indicates the first Parent The output individual encoding of the update iteration is represents the second updated weight, represents the second mean deviation weight, Indicates the preset number of update iterations, Indicates Update iterations, Indicates The individual encoding that ranks first in the feature ranking of the update iteration.

[0100] Specifically, the group update mechanism fully combines the advantages of global exploration and local convergence through the independent optimization of two groups of parents, and has significant advantages in improving algorithm efficiency, enhancing population diversity and the quality of optimization results. This innovative design not only improves the robustness of the optimization algorithm, but also has a wide range of practical application value. Compared with the traditional overall population update method, the group update design can better handle multi-objective optimization problems. The first update group focuses on global exploration ability, and guides the population to search in a larger parameter space through mean adjustment and random perturbation to avoid falling into the local optimum. The second update group focuses on local convergence ability, and quickly approaches the optimal solution with the current optimal individual as the guide. The division of labor and cooperation between the two groups ensures that the algorithm achieves a dynamic balance between global exploration and local optimization. The first update group ensures that the individuals form an overall optimization trend within the group by adjusting the mean; the second update group further improves the convergence efficiency of the algorithm by guiding the optimal individual. The introduction of random perturbations ensures that the population still has the ability to explore in the later iteration process to avoid premature phenomenon.

[0101] In one embodiment of the present invention, determining the second to Nth observation points according to the first observation point includes:

[0102] Based on the topological structure constructed by the elements with element values ​​of 1 in the semantic matrix, the elements with element values ​​of 1 are mapped to nodes of the topological structure, and edges are constructed between adjacent nodes to obtain the topological structure;

[0103] Step 6, constructing an image coordinate system according to the feature image corresponding to the first observation point; wherein the origin of the image coordinate system is the central element of the semantic matrix corresponding to the feature image;

[0104] Map the target to the image coordinate system, and the calculation formula is as follows:

[0105] , , ;

[0106] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, and They represent the horizontal focal length and vertical focal length of the drone respectively. Represents the horizontal coordinate of the element corresponding to the central node in the topological structure, Indicates the ordinate of the element corresponding to the central node in the topological structure, Indicates the actual height of the target. Represents the Euclidean distance between any two nodes in the topological structure. represents the abscissa of the central element of the semantic matrix, The ordinate represents the central element of the semantic matrix;

[0107] Step 7: Based on the first observation point, construct the drone coordinate system and , and They are mapped to the drone coordinate system respectively; the origin of the drone coordinate system represents the drone corresponding to the first observation point; the calculation formula is as follows:

[0108] , ;

[0109] , ;

[0110] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, represents the rotation matrix, represents the yaw angle matrix, represents the pitch angle matrix, Indicates the flight angle of the drone corresponding to the first observation point, Indicates the shooting angle of the drone corresponding to the first observation point;

[0111] Step 8: Based on the GPS coordinates of the drone, , and They are mapped in the geographic coordinate system respectively, and the calculation formula is as follows:

[0112] , , ;

[0113] , ;

[0114] in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system. , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the drone. and Represent the horizontal axis increment and the vertical axis increment respectively. represents the radius of the Earth;

[0115] Step 9, taking the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system as the center of the circle, the Euclidean distance between the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the UAV corresponding to the target and the first observation point as the radius, and the angle between the UAV corresponding to the target and the first observation point and the horizontal plane as the rotation angle, to obtain the characteristic circle; divide the characteristic circle into N equal parts, and take the first observation point as the starting point to obtain the second to Nth observation points, and divide the remaining UAVs into the second to Nth observation points one by one.

[0116] Specifically, target information is extracted through feature images: first, the target pixel information is extracted through the semantic matrix in the feature image, and these pixel points are mapped into a topological structure, where the elements with a value of 1 in the semantic matrix are mapped to the nodes of the topological structure, and the edges between the nodes are constructed through the adjacent relationship to form the topological structure of the target. Establish an image coordinate system and an image coordinate system: Take the central element of the semantic matrix as the origin to establish an image coordinate system. Map the node position of the target in the semantic matrix to the image coordinate system to obtain the three-dimensional coordinate information of the target. Take the drone corresponding to the first observation point as the origin to establish a drone coordinate system, and map the shooting coordinates of the target to the drone coordinate system through the rotation matrix. Combined with the GPS information of the drone, the target is mapped from the drone coordinate system to the geographic coordinate system to obtain the geographic location coordinates of the target. According to the geographic location of the target and the position of the first observation point, a feature circle is generated with the target as the center and the distance between the target and the first observation point as the radius.

[0117] Specifically, the key to the dynamic target monitoring scheme through multi-UAV collaboration is to dynamically determine the second to Nth observation points based on the characteristic image information and the position of the first observation point, so that all UAVs can be efficiently distributed around the target to achieve the best multi-angle observation. This design does not rely solely on a single observation point, but rather reasonably allocates multiple UAVs by constructing a characteristic circle to give full play to the monitoring capabilities of the UAV swarm. Through the position of the first observation point and the geographic coordinates of the target, a characteristic circle with the target as the center and the distance between the first observation point and the target as the radius is determined. The essence of the characteristic circle is that each point on its circumference can be regarded as an "optimal observation point" with the same observation effect, because these points all meet the equality of the target distance. A single optimal observation point (i.e., the first observation point) cannot cover all the potential information around the target. In fact, in the geometric relationship between the target and the UAV, each equally divided point on the characteristic circle is the optimal observation point. Therefore, evenly distributing UAVs to multiple observation points on the characteristic circle can maximize the use of the "optimal area" around the target and avoid the simplification of information or the bias of observation results. Through the multi-observation point allocation mechanism, each drone can obtain the characteristic information of the target from different angles, ensuring the consistency and comprehensiveness of the overall monitoring results.

[0118] In one embodiment of the present invention, the preset prediction model includes:

[0119] Step 10, based on steps 6 to 8, the feature image of the historical time period is processed to obtain the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the target in the geographic coordinate system in the historical time period;

[0120] Step 11, constructing a training parameter set according to the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the time series; extracting a number of fixed-length training parameters from the training sample set through a sliding time window of length h to obtain a number of training samples, and using the training parameters of the corresponding training samples at the next moment as sample labels;

[0121] Step 12: A prediction model is obtained based on training samples and sample labels, and a hidden layer of the prediction model is constructed based on a recurrent neural network.

[0122] Specifically, from the processing results of step 6 to step 8, the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the target in the geographic coordinate system are extracted. These coordinate information are calculated based on the feature image in the historical time period. The extracted time series coordinate data is converted into a training parameter set, and the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of each time point are combined into a training parameter. The sliding time window is a fixed time period of length h, which is used to extract training samples with a fixed length from the time series data. For each training sample, the position of the target at the next moment is used as the sample label. A recurrent neural network (RNN) is used as the hidden layer structure of the prediction model. The characteristic of RNN is that it can process time series data, especially tasks with time dependence. Its hidden layer memorizes the information of historical time points through a cyclic structure, and combines this information with the current time point data to predict the future state. The training samples generated from the sliding time window are used as the input of the model, and the position of the target at the next moment is used as the label of the model. The model is supervised by the training samples and labels, and the weight parameters of the neural network are adjusted so that it can accurately predict the position of the target at the next moment. The output of the prediction model can help the UAV system make autonomous decisions and dynamically adjust the observation position of the UAV according to the prediction results, thereby achieving efficient and continuous target monitoring.

[0123] In one embodiment of the present invention, continuously tracking a target using a tracking circle of the target includes:

[0124] The horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system from the gth moment to the g+hth moment are constructed as feature vectors and input into the prediction model to obtain the predicted position of the target at the g+h+1th moment;

[0125] With the predicted target position at the g+h+1th moment as the center, a characteristic circle is constructed, and N UAVs are assigned to the characteristic circle at the g+h+1th moment to obtain the target's tracking circle for continuous tracking of the target.

[0126] In one embodiment of the present invention, the prediction model is reversely updated, including:

[0127] The mean square error between the target predicted position and the actual target position at the g+h+1th moment is used as the loss function of the hidden layer of the prediction model, and the weight parameters and bias parameters of the hidden layer of the prediction model are updated through back propagation.

[0128] Specifically, by constructing a tracking circle of the target, dynamic and continuous tracking of the target is achieved. First, based on the geographic coordinates (horizontal axis, vertical axis and vertical coordinates) of the target in the historical time period, a time series feature vector containing the target's motion trajectory is generated and input into the trained prediction model to accurately predict the position of the target at the future moment. Subsequently, with the predicted position as the center of the circle, and the Euclidean distance obtained in step 9 as the radius and horizontal angle (that is, only the center of the feature circle is changed), a feature circle is constructed, and the feature circle is divided into observation points equal to the number of drones. Drones are dynamically assigned to these observation points to form a drone tracking circle around the target, and the target is monitored from multiple optimal positions. In this way, the drone system can adjust the position and distribution of the tracking circle in real time when the target changes its position, realize continuous observation and efficient coordination of the target, and ensure the stability and accuracy of monitoring.

[0129] In one embodiment of the present invention, after finding the first observation point by assigning an optimization algorithm to N drones, the remaining N-1 drones are immediately matched with characteristic circles to obtain a tracking circle, and characteristic images of the target at N angles are obtained. Then, the predicted position of the target is continuously output based on the prediction model, so that a characteristic circle is constructed in advance at the predicted position, and then a characteristic image of the target is obtained. Therefore, the present invention can achieve refined tracking of the target by processing the characteristic image. The advantage over single drone tracking is that the target can be tracked based on multiple drones gradually approaching a better position, and multi-angle tracking of the target can also be achieved, so that the tracking accuracy and tracking shooting effect are greatly improved. Compared with the existing multi-drone tracking, the advantage is that the tracking effect of some drones will not be poor, and the ineffective work of some drones is avoided.

[0130] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. A drone target recognition and automatic tracking algorithm system, characterized in that: include: An image acquisition module is used to collect characteristic images of the target based on multiple drones at fixed time intervals; An image scoring module, used to evaluate feature images based on preset rules and calculate the tracking score of each feature image; The dynamic optimization module is used to combine multiple drones and the optimization algorithm to determine the first observation point of the target and assign a drone to the first observation point, including: Step 1, The drones are initialized based on the constraints to obtain the optimized population. The individual code is represented as: ;in, represents the flight speed of the i-th UAV, represents the flight angle of the i-th UAV, represents the shooting angle of the i-th drone, ; The constraints include: ;in, Indicates the minimum flight speed of the drone. Indicates the maximum flight speed of the drone; , ;in, Indicates the maximum flight angle of the drone on the left side of the horizontal direction. Indicates the maximum flight angle of the drone on the right side in the horizontal direction. Indicates the maximum upward flight angle of the drone in the vertical direction. Indicates the maximum flight angle of the drone in the vertical direction; , ;in, Indicates that in the horizontal direction, the drone is Adjust the shooting angle. Indicates that in the vertical direction, the drone is Adjust the shooting angle. Indicates the maximum shooting angle of the drone in the vertical direction. Indicates the maximum shooting angle of the drone in the vertical direction; Step 2, the tracking score of the feature image obtained by each UAV based on the corresponding individual coding scheme is used as the individual objective function; Step 3, calculate the objective function value of the individual, and sort each individual from large to small based on the objective function value to obtain the individual feature sorting; Step 4: sort the features from front to back and retain a preset number of individuals, and use the remaining individuals as parents for update iteration, specifically including: Divide the parents into two groups, each including parents, including a first update group and a second update group; The calculation formula for the first update group is as follows: ； in, Indicates the first update group Parent The output individual encoding of the update iteration is Indicates the first update group Parent The output individual encoding of the update iteration is represents the first updated weight, represents the mean individual code of the first update group, , represents the first mean deviation weight, represents the correction bias parameter, , represents the random perturbation individual code, , Represents a random number; The calculation formula for the second update group is as follows: ； in, Indicates the first Parent The output individual encoding of the update iteration is Indicates the first Parent The output individual encoding of the update iteration is represents the second updated weight, represents the second mean deviation weight, Indicates the preset number of update iterations, Indicates Update iterations, Indicates The individual encoding that ranks first in the feature ranking of the update iteration; Step 5, after looping step 3 and step 4 for a preset number of update iterations, the update iteration is stopped, the individual with the largest objective function value is output, and the position of the UAV corresponding to the individual with the largest objective function value is taken as the first observation point; the tracking planning module is used to determine the second to Nth observation points according to the first observation point, and assign the second to Nth observation points to the remaining UAVs one by one to obtain the tracking circle of the target; The dynamic adjustment module is used to obtain the feature image based on the time series and generate the predicted position of the target at the next moment in combination with the preset prediction model; If the prediction is successful, the target is continuously tracked using the target’s tracking circle; If the prediction fails, the prediction model is updated in reverse based on the predicted position and the actual position of the target, and returns to the dynamic optimization module.

2. The unmanned aerial vehicle target recognition and automatic tracking algorithm system according to claim 1 is characterized in that: Preset rules include: The feature image is semantically segmented based on the semantic tool; the semantic segmentation includes outputting the pixels of the target in the feature image as 1 and outputting the pixels of the non-target in the feature image as 0, thereby obtaining a semantic matrix; Divide the semantic matrix into central area, sub-central area and peripheral area; The number of elements with element values ​​of 1 in the central area, sub-central area, and edge area of ​​the semantic matrix is ​​counted respectively, and the tracking score of the feature image is calculated. The calculation formula of the tracking score is as follows: ； in, represents the tracking score of the feature image, Indicates the number of elements whose value in the center area is 1. represents the center weight, Indicates the number of elements whose element value is 1 in the secondary center area. represents the secondary center weight, Indicates the number of elements whose element value is 1 in the edge area. represents the edge weight, represents the target ideal weight, Indicates the number of elements whose target ideal element value is 1. It means to find the absolute value. , and All are not 0, , and The sum is 1, .

3. The unmanned aerial vehicle target recognition and automatic tracking algorithm system according to claim 1 is characterized in that: Determining second to Nth observation points according to the first observation point includes: Based on the topological structure constructed by the elements with element values ​​of 1 in the semantic matrix, the elements with element values ​​of 1 are mapped to nodes of the topological structure, and edges are constructed between adjacent nodes to obtain the topological structure; Step 6, constructing an image coordinate system according to the feature image corresponding to the first observation point; wherein the origin of the image coordinate system is the central element of the semantic matrix corresponding to the feature image; Map the target to the image coordinate system, and the calculation formula is as follows: ， ， ； in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, and They represent the horizontal focal length and vertical focal length of the drone respectively. Represents the horizontal coordinate of the element corresponding to the central node in the topological structure, Indicates the ordinate of the element corresponding to the central node in the topological structure, Indicates the actual height of the target. Represents the Euclidean distance between any two nodes in the topological structure. represents the abscissa of the central element of the semantic matrix, The ordinate represents the central element of the semantic matrix; Step 7: Based on the first observation point, construct the drone coordinate system and , and They are mapped to the drone coordinate system respectively; the origin of the drone coordinate system represents the drone corresponding to the first observation point; the calculation formula is as follows: ， ； ， ； in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the image coordinate system, represents the rotation matrix, represents the yaw angle matrix, represents the pitch angle matrix, Indicates the flight angle of the drone corresponding to the first observation point, Indicates the shooting angle of the drone corresponding to the first observation point; Step 8: Based on the GPS coordinates of the drone, , and They are mapped in the geographic coordinate system respectively, and the calculation formula is as follows: ， ， ； ， ； in, , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system. , and Respectively represent the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the drone. and Represent the horizontal axis increment and the vertical axis increment respectively. represents the radius of the Earth; Step 9, taking the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system as the center of the circle, the Euclidean distance between the horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the UAV corresponding to the target and the first observation point as the radius, and the angle between the UAV corresponding to the target and the first observation point and the horizontal plane as the rotation angle, to obtain the characteristic circle; divide the characteristic circle into N equal parts, and take the first observation point as the starting point to obtain the second to Nth observation points, and divide the remaining UAVs into the second to Nth observation points one by one.

4. The unmanned aerial vehicle target recognition and automatic tracking algorithm system according to claim 3 is characterized in that: Preset forecasting models, including: Step 10, based on steps 6 to 8, the feature image of the historical time period is processed to obtain the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the target in the geographic coordinate system in the historical time period; Step 11, constructing a training parameter set according to the horizontal axis coordinate, the vertical axis coordinate and the vertical coordinate of the time series; extracting a number of fixed-length training parameters from the training sample set through a sliding time window of length h to obtain a number of training samples, and using the training parameters of the corresponding training samples at the next moment as sample labels; Step 12: A prediction model is obtained based on training samples and sample labels, and a hidden layer of the prediction model is constructed based on a recurrent neural network.

5. The unmanned aerial vehicle target recognition and automatic tracking algorithm system according to claim 4 is characterized in that: Continuously track the target using the target's tracking circle, including: The horizontal axis coordinate, vertical axis coordinate and vertical coordinate of the target in the geographic coordinate system from the gth moment to the g+hth moment are constructed as feature vectors and input into the prediction model to obtain the predicted position of the target at the g+h+1th moment; With the predicted target position at the g+h+1th moment as the center, a characteristic circle is constructed, and N UAVs are assigned to the characteristic circle at the g+h+1th moment to obtain the target's tracking circle for continuous tracking of the target.

6. The unmanned aerial vehicle target recognition and automatic tracking algorithm system according to claim 5 is characterized in that: Update the prediction model in reverse order, including: The mean square error between the target predicted position and the actual target position at the g+h+1th moment is used as the loss function of the hidden layer of the prediction model, and the weight parameters and bias parameters of the hidden layer of the prediction model are updated through back propagation.

Images