Air-ground cooperative target identification method based on fuzzy integral decision level fusion

Abstract

The invention belongs to the technical field of air-ground collaborative target recognition, and particularly relates to an air-ground collaborative target recognition method based on fuzzy integral decision-level fusion, which comprises the following steps: S0, acquiring acquired data; forming a set Gkcv by the data of each source for the same possible result in the same aspect of the same target; s1, setting a fuzzy density interval gi of the unmanned vehicle and the unmanned aerial vehicle; s2, solving a value of a correlation parameter lambda by using a Newton iteration method; s3, defining a subset Ai as a subset composed of the first i elements of the set Gkcv; calculating a fuzzy density interval w'i of the subset Ai; s4, sorting the confidence values in the set Gkcv from small to large; s5, performing data fusion by combining the sorted confidence values and the fuzzy density interval w'i of each subset Ai to obtain a corresponding fusion confidence interval; and S6, performing defuzzification on the fusion confidence interval to obtain corresponding information. According to the method, air-ground collaborative multi-dimensional heterogeneous target information can be effectively fused, and a high-confidence identification decision is output in a complex environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of air-ground collaborative target recognition, and in particular relates to an air-ground collaborative target recognition method based on fuzzy integral decision-level fusion. Background Art

[0002] In complex scenarios like outdoor search and disaster relief, accurate target identification is a crucial prerequisite for mission success. Unmanned vehicles and drones, representative of intelligent equipment, have been widely used in field detection.

[0003] However, in complex terrain conditions (such as hidden areas such as mountains and basins, or obstructions from buildings and obstacles in urban environments), target information acquisition faces multiple challenges. For example, environmental complexity: undulating terrain and artificial obstacles lead to detection blind spots, and information integrity is easily destroyed. Interference factors: Natural obstructions or artificial camouflage may cause misidentification. Equipment limitations: Unmanned vehicles have the advantages of strong endurance and close reconnaissance, but have poor adaptability to special terrain (such as steep slopes and swamps) and limited perception range; UAVs have wide-area search and rapid response capabilities, but due to payload limitations, the sensor accuracy is insufficient and the target false detection rate is high.

[0004] Furthermore, the information obtained by a single device is fragmented and unreliable, making it difficult to support decision-making in complex scenarios. To overcome this bottleneck, current research focuses on air-ground collaborative detection: leveraging the complementary advantages of unmanned vehicles and drones (e.g., unmanned vehicle ground-level scanning combined with drone high-altitude wide-area coverage) to achieve multi-dimensional information fusion. However, existing technologies still face two core challenges:

[0005] Problem 1: Multi-source information fusion is highly uncertain. The recognition accuracy and false alarm rates for different targets vary significantly (e.g., vehicles vs. personnel), and the target's damage status and threat level are complex. Traditional single-value fusion models (such as weighted average) cannot quantify the dynamic reliability of sensors, making the fusion results sensitive to abnormal data.

[0006] Problem 2: Loss of confidence in decisions with heterogeneous data. Environmental interference causes fluctuations in confidence intervals (e.g., fog reduces the confidence of optical equipment). Existing methods output a single decision value, losing information about uncertainty. Addressing this issue requires modeling the correlation between device reliability intervals and target features simultaneously, which increases mathematical complexity exponentially.

[0007] Therefore, how to effectively integrate multi-dimensional heterogeneous target information of air-ground collaboration and output high-confidence recognition decisions in complex environments has become a problem that needs to be solved urgently. Summary of the Invention

[0008] In response to the above-mentioned deficiencies in the existing technology, the present invention provides an air-ground collaborative target recognition method based on fuzzy integral decision-level fusion, which can effectively fuse multi-dimensional heterogeneous target information of air-ground collaboration and output high-confidence recognition decisions in complex environments.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion includes the following steps:

[0011] S0. Take a single UAV or unmanned vehicle as a source for collecting information, obtain target detection data from each source, and perform spatial and temporal calibration to obtain information data of each target collected by each source; the information data of a target includes multiple aspects of information, including the type and safety level of the target; the data of each aspect includes the probability of occurrence of multiple results, and the probability of occurrence is recorded as a confidence value; the confidence value of the cth aspect of the i-th source for the k-th target is the vth result is μ ikcv ; and μ ikc1 +…μ ikcv +μ ikcN =1, N is the number of possible outcomes of aspect c;

[0012] The confidence values ​​of the v-th result of each source on the c-th aspect of target k are grouped into a set G kcv , G kcv ={μ 1kcv , μ 2kcv ..., μ ikcv ..., μ nkcv};

[0013] S1. Set the fuzzy density interval g of each source for the vth result of the cth aspect i , the fuzzy density interval satisfies the regularity constraint;

[0014] S2, combined with fuzzy density interval g i , use Newton iteration method to solve the value of the associated parameter λ;

[0015] S3. Define subset A i For the set G kcv The subset consisting of the first i elements of , 1≤i≤n; based on the fuzzy density interval g i and the associated parameter λ, using the Newton-Raphson recursion method to calculate the subset A i The fuzzy density interval w' i ; Calculate A n times from i=1 to i=n i The fuzzy density interval w' i ; get w'1,w'2,…,w' n;

[0016] S4, for set G kcv The confidence values ​​in are sorted from small to large, and the confidence value of the i-th one after sorting is f(x' i );

[0017] S5, combined with the confidence values ​​after sorting f(x' i ), and each subset A i The fuzzy density interval w' i , use the improved fuzzy integral algorithm to perform data fusion and obtain the fusion confidence interval of the corresponding target information. The calculation formula includes:

[0018]

[0019] Where a and b represent the lower limit and upper limit of the confidence interval respectively;

[0020] S6, defuzzifying the fusion confidence interval obtained in S5, and obtaining defuzzified information of the cth aspect of the kth target as the vth result;

[0021] S7. Repeat the above process to obtain the defuzzification of various results of the cth aspect of the kth target, and comprehensively judge the recognition result of the cth aspect of the kth target; until the recognition results of all aspects of all targets are judged.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. Solve the problem of high uncertainty in multi-source information fusion. Existing methods cannot quantify the dynamic reliability of equipment due to fixed weights (such as the sudden drop in the confidence of drones in foggy weather). In this method, S1 dynamically sets the equipment fuzzy density interval g i , S3 recursively deduces the subset fuzzy density w by the Newton-Raphson method i This approach captures fluctuations in the reliability of multiple devices. By establishing a dynamic interval-based evaluation model for device reliability, it replaces the fixed-weight weighted average method used in existing technologies. When complex terrain obstructs or device status fluctuates, the fusion results automatically reduce the contribution of low-reliability devices, significantly suppressing interference from abnormal data and avoiding misjudgments.

[0024] 2. Break through the bottleneck of lack of confidence in heterogeneous data decision making. Existing methods (such as DS evidence theory) output a single judgment value and lose uncertainty information. In this method, S5 outputs a confidence interval (such as vehicle identification: [0.72, 0.85]) by improving the fuzzy integral, and S6 uses the maximum membership method to defuzzify. In this way, the fusion result retains the confidence uncertainty in the form of an interval (the upper limit represents an optimistic estimate, and the lower limit represents a conservative estimate). Compared with existing probabilistic fusion methods (such as Bayesian output 0.8), this solution can quantify the credible range of the recognition result (such as "the vehicle confidence level is between 0.72 and 0.85"), providing a buffer basis for high-risk decisions.

[0025] 3. Improve the ability to resolve target features in complex environments, addressing the issue of wide variations in recognition accuracy due to features such as target damage status and camouflage interference. In this method, S4 sorts the multi-source confidence scores by value (ensuring the monotonicity of the fuzzy integral), and S5 improves the integral algorithm to strengthen the nonlinear coupling calculation of the sorted confidence scores. This allows for adaptive fusion of complementary information through confidence sorting and fuzzy measure correlation modeling (S2λ parameter) to address heterogeneous data feature differences (e.g., partial vehicle damage and personnel camouflage) between unmanned vehicles (near-range high-definition) and drones (wide-area low-definition) in air-ground collaboration. Compared to direct average fusion, this method significantly improves the recognition rate for occluded and partially damaged targets.

[0026] In summary, this method can effectively integrate multi-dimensional heterogeneous target information of air-ground collaboration and output high-confidence recognition decisions in complex environments.

[0027] Preferably, in S3, the subset A is calculated according to the following formula i The fuzzy density interval w' i :

[0028] g λ (x1∪x2)=g λ (x1)+g λ (x2)+λg λ (x1)g λ (x2);

[0029] w' i =g λ (x1∪x2…∪x i )=g λ (x1∪x2…∪x i-1 )+g λ (x i )+λg λ (x1∪x2…∪x i-1 )g λ (x i );

[0030] Among them, g λ (xi )=g i ;x i =μ ikc ; x1∪x2…∪x i =A i .

[0031] Such a setting can 1. break through the combination paradox limitation of traditional fusion models. Traditional confidence fusion (such as weighted averaging and DS evidence theory) assumes that devices are independent of each other and ignores the nonlinear effect of reliability superposition (for example, when devices A and B work together, the overall reliability ≠ A+B). In scenarios with high device correlation (such as drones and unmanned vehicles jointly observing the same target), it is easy to cause overfitting of the fusion results. In this method, the complementarity or redundancy between devices is explicitly modeled through λ (λ>0 indicates complementarity, λ<0 indicates redundancy), solving the combination paradox problem of "1+1>2" or "1+1<1". Compared with the independence assumption model, it more accurately portrays the device interaction effect in actual collaborative scenarios.

[0032] 2. Achieve efficient recursive calculation of high-dimensional reliability intervals. The existing fuzzy integral needs to calculate 2n-1 subset measures (n is the number of devices). When n ≥ 5, the computational complexity increases exponentially, making it difficult to implement in engineering. This method uses a recursive formula (to reduce the computational complexity from O(2 n ) is reduced to O(n). By iteratively reusing the results of the preceding subset, repeated calculations are avoided, significantly improving the real-time performance of large-scale air-ground collaborative systems (e.g., 10+ devices).

[0033] In summary, the λ fuzzy measure recursive mechanism solves the common problems of device interaction effect distortion and high-dimensional computation in traditional fusion methods through mathematical innovation (combinatorial correlation modeling) and engineering optimization (linear complexity calculation), providing a high-precision, low-cost reliability foundation for subsequent confidence interval fusion (S5).

[0034] Preferably, in S2, the Newton iteration method is used to solve the following equation to calculate the associated parameter λ:

[0035]

[0036] Such a setting can significantly improve the solution efficiency and iterative convergence speed. Traditional methods (such as bisection method and simple iteration method) are not very effective in solving equations. When the equation is highly nonlinear (including n consecutive products), multiple iterations of the solution space must be traversed, resulting in slow convergence and high computational overhead. In this method, the Newton iteration method leverages local linear approximations and derivative information (first-order Taylor expansion) to transform the nonlinear problem into an iterative approximation process, exponentially reducing the number of iterations required for convergence. This method, especially in multi-device scenarios (n ≥ 3), can quickly approximate the true λ value, significantly reducing computational time.

[0037] 2. Enhance the robustness and accuracy of solutions in complex scenarios. The existing analytical methods (such as polynomial decomposition) have no closed-form solution when n>2; the gradient descent method is prone to fall into local optimality due to improper selection of initial values, resulting in deviation in the λ solution. In this method, the Newton iteration method adjusts the step size adaptively (Hessian matrix participates in the iteration) to improve the reliability of the equipment (such as partial g i It has strong adaptability to complex situations such as multimodal equations and can stably converge to the global optimal solution, avoiding distortion of fusion results caused by parameter sensitivity problems.

[0038] In summary, in the process of solving the associated parameter λ, the Newton iteration method solves the dual bottlenecks of low computational efficiency and poor parameter accuracy in traditional fusion methods through innovative applications at the mathematical level (derivative drive + second-order convergence), laying a high reliability foundation for subsequent confidence interval fusion (S3-S5).

[0039] Preferably, in S0, the starting point of the detection task is used as the calibration point of the data space information to perform spatial calibration; and the time calibration is performed using the least square method.

[0040] This setting, the combined design of spatial calibration (dynamic starting point benchmark) and temporal calibration (least squares fitting), specifically solves the pain points of spatial benchmark drift and temporal asynchronous error accumulation in air-ground collaboration, significantly optimizes the quality of original data, and provides high-confidence spatiotemporal consistency guarantees for subsequent interval fusion (S3-S5) and decision defuzzification (S6).

[0041] Preferably, in S1, the fuzzy density ranges of the unmanned vehicles and drones are set by conducting recognition experiments in different environments and weather conditions and combining expert experience.

[0042] With this setup, S1 sets the fuzzy density range by coupling environmental experiments with expert knowledge, breaking through the bottleneck of rigid device weights in traditional fusion. This provides a reliable foundation for environmental adaptation and anti-interference for subsequent fuzzy measurement recursion (S3) and confidence fusion (S5), comprehensively improving the robustness of target recognition in complex scenarios.

[0043] Preferably, in S6, the maximum membership method is used for defuzzification.

[0044] With this setting, the maximum membership method solves the problems of decision risk imbalance and information inheritance disconnection in traditional defuzzification methods with the mean principle (balance) and full-link compatibility (inheriting confidence interval information), and ultimately achieves the core goal of the air-ground collaborative system: to output stable and reliable target classification decisions in an uncertain information environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0046] Figure 1 Flowchart of this method;

[0047] Figure 2 Schematic diagram of the spatiotemporal alignment of target information in the embodiment. DETAILED DESCRIPTION

[0048] The following is a further detailed description through specific implementation methods:

[0049] Example:

[0050] like Figure 1 As shown, this embodiment discloses an air-ground collaborative target recognition method based on fuzzy integral decision-level fusion, including the following steps:

[0051] S0. Take a single UAV or unmanned vehicle as a source for collecting information, obtain target detection data from each source, and perform spatial and temporal calibration to obtain information data of each target collected by each source; the information data of a target includes multiple aspects of information, including the type and safety level of the target; the data of each aspect includes the probability of occurrence of multiple results, and the probability of occurrence is recorded as a confidence value; the confidence value of the cth aspect of the i-th source for the k-th target is the vth result is μ ikcv ; and μ ikc1 +…μ ikcv +μ ikcN =1, N is the number of possible outcomes of aspect c;

[0052] The confidence values ​​of the v-th result of each source on the c-th aspect of target k are grouped into a set G kcv , G kcv ={μ 1kcv , μ 2kcv ..., μ ikcv ..., μ nkcv};

[0053] In specific implementation, the starting point of the detection task is used as the calibration point of the data space information to perform spatial calibration; the time calibration is performed by least squares. To facilitate those skilled in the art to better understand spatial calibration and time calibration, the following description is provided.

[0054] Data information spatial alignment processing

[0055] During a mission, the drone and unmanned vehicle select the starting point of the reconnaissance mission as the calibration point for their spatial data information. Their latitude, longitude, and elevation information are converted to the same inertial coordinate system. After the mission begins, the drone and unmanned vehicle, already spatially aligned, begin reconnaissance and search for the target. Once the target is identified, laser ranging is used to determine the distance. The drone uses the angle and distance values, while the unmanned vehicle uses the distance value to convert the target's latitude and longitude coordinates, respectively. This results in spatially unified data from both systems.

[0056] Data information time alignment processing

[0057] The time alignment of data is to synchronize the collected target data to the same time using the algorithm used in a certain time period. There are many methods for time alignment, and the commonly used ones are the least squares criterion and interpolation method. Because the interpolation method requires a large amount of calculation, and when too many interpolation points are required, the required order of the interpolation function will increase accordingly, and the final value will be unstable. The least squares criterion can overcome these shortcomings. Therefore, this method uses the least squares method for time alignment processing. Assume that there are n groups of data: (t k ,F), where k=1,2,…,n, then the least squares expression is: F(t)=a·t k +b.

[0058] Based on the spatial alignment of the unmanned vehicle and the drone, the following is obtained: Figure 2 The time alignment diagram shown in the figure shows two different types of detection equipment with different sampling rates. O1A2 and O1A3 are the elevation measurement vectors of the unmanned vehicle, while O2A1 and O2A4 are the depression measurement vectors of the drone. O2A2 and O2A3 are the depression direction vectors of the drone after time alignment, and O2T is the time-calibrated vector.

[0059] The image acquisition rates of drones and unmanned vehicles are different. The time consumed for collecting each image is 0.08s and 0.0015s respectively. In order to improve the accuracy of time alignment, the high-precision collected data is fitted to the low-precision collected data. That is, the data collected by the unmanned vehicle is fitted to the drone, and the unmanned vehicle data is time-aligned to the drone, with the initial time being 0s and the end time being 0.08s.

[0060] Tables 1 and 2 respectively show the pitch angles when the UAV collects the target data and the pitch angles when the unmanned vehicle collects the target data information within a period of time.

[0061] Table 1 UAV collected data information

[0062]

[0063] Table 2 Data collected by unmanned vehicles

[0064]

[0065] The least squares method is used for alignment, and we have:

[0066]

[0067] Among them, F is the least square multi-order curve fitting function of the unmanned vehicle data. According to the fitting curve of the unmanned vehicle and the above formula, the angle value a of the drone after time alignment is calculated. x Among them, t x Corresponding to the middle period of the unmanned vehicle, we can get: a x =75.68°.

[0068] By fitting high-precision time data to low-precision time data, the time alignment problem of the two types of detection equipment is solved. The calculation and fitting method for the azimuth angle is the same as that for the elevation angle, so I will not elaborate on it here.

[0069] S1. Set the fuzzy density interval g of each source for the vth result of the cth aspect i , the fuzzy density interval satisfies the regularity constraint.

[0070] During implementation, fuzzy density ranges for unmanned vehicles and drones were set by conducting recognition experiments in various environments and weather conditions, combined with expert experience. This approach, coupled with environmental experiments and expert knowledge, overcomes the bottleneck of rigid device weights in traditional fusion. This provides a reliable foundation for environmental adaptability and interference resistance in subsequent fuzzy measure recursion (S3) and confidence fusion (S5), comprehensively improving the robustness of target recognition in complex scenarios.

[0071] Fuzzy measure describes the importance of each attribute and introduces the concepts of fuzzy density and correlation parameters, which are defined as follows:

[0072] Let X represent all elements in the set, each X n All elements in the subset form a set F, g λ is a fuzzy measure on the set F. If X1 and X2 belong to F, the intersection of X1 and X2 is an empty set, and λ>-1, then:

[0073] g λ (X1∪X2)=g λ (X1)+g λ (X2)+λg λ (X1)g λ (X2);

[0074] And g λ The conditions are: monotonically continuous and bounded, in a single point set (gλ {X i}) Fuzzy measure value, which is the fuzzy density g i , and g λ (X) = 1, then it is called a regular fuzzy measure (λ∈(-1,∞)).

[0075] Among them, the regular g λ The method for calculating the associated parameter λ in fuzzy measurement is:

[0076]

[0077] When λ=0, it indicates that the correlation between the two detectors is 0, that is, they do not affect each other, and the superposition method can be used for accumulation; when -1<λ<0, it indicates that there is redundant correlation between the attributes; when 0<λ, it indicates that there is complementary correlation between the attributes.

[0078] For a finite set F, g λ The fuzzy measure can be determined by its fuzzy density, that is:

[0079]

[0080] Fuzzy integrals have the ability to fuse the importance of multi-source information (fuzzy measures) with the objective evidence provided by each source (f-function). In practical applications, many problems cannot be solved with simple linear superposition, and fuzzy integrals also have a similar non-additive property. Using nonlinear superposition to fuse multi-source information can complement the strengths and weaknesses of unmanned vehicles and drones in reconnaissance and identification.

[0081] It should be noted that there are two forms of data information fusion: hard decision and soft decision. For hard decision, the information of the same target only has two cases of "yes" or "no". Therefore, the membership degree of one target n1 is:

[0082]

[0083] For soft decision making, there may be multiple possibilities for identifying the same target, but the probabilities of different possibilities are also different. This solution adopts the soft decision method. For the convenience of explanation, it is assumed that there is only one target and only one aspect of information is identified. In this case, the membership degree of source i to the target is:

[0084]

[0085] Among them, μ i1 +μ i2 +,...,+μ iN =1.

[0086] Since unmanned vehicles or drone detection equipment need to establish a matching relationship with N possible results when scouting and identifying a certain aspect of a target, and there is a correlation between each result, a soft decision-making method is used.

[0087] The probability P of unmanned vehicle or drone detection equipment obtaining target information mn , can be expressed by fuzzy measurement. The greater the possibility of obtaining target information, the greater the probability and the greater the fuzzy measurement value. Conversely, the smaller the possibility of obtaining target information, the smaller the probability and the smaller the fuzzy measurement value. Since the detection equipment is affected by different factors such as weather conditions, topography, and background complexity, the probability of obtaining target information in different situations will also be different. However, due to the stability of detection equipment recognition, the probability value will fluctuate within a certain range. Using a certain value as the fuzzy density of a detection device for a certain target recognition is inaccurate. Therefore, an interval W is selected as its fuzzy density interval, so that

[0088] S2, combined with fuzzy density interval g i , use Newton iteration method to solve the following equation to calculate the associated parameter λ:

[0089]

[0090] Traditional methods (such as bisection method, simple iteration method) are not very effective in solving equations. , because the equation is highly nonlinear (containing n consecutive products), it is necessary to traverse the search solution space multiple times, convergence is slow, and the computational overhead is high. In this method, the Newton iteration method uses local linear approximation and derivative information (first-order Taylor expansion) to transform the nonlinear problem into an iterative approximation process, exponentially reducing the number of iterations required for convergence. Especially in multi-device scenarios (n≥3), it can quickly approximate the true λ value, greatly reducing the computational time. In addition, the analytical methods of the prior art (such as polynomial decomposition) have no closed-form solutions when n>2; the gradient descent method is prone to fall into local optimality due to improper selection of initial values, resulting in deviations in the λ solution. In this method, the Newton iteration method adjusts the step size adaptively (the Hessian matrix participates in the iteration) to deal with uneven distribution of device reliability (such as some g i The Newton iteration method is highly adaptable to complex situations such as those involving λ close to 0 or 1 and multimodal equations, and can stably converge to the global optimal solution, avoiding distortion in fusion results caused by parameter sensitivity. Therefore, in solving the associated parameter λ, the Newton iteration method uses innovative mathematical applications (derivative-driven + second-order convergence) to overcome the dual bottlenecks of low computational efficiency and poor parameter accuracy in traditional fusion methods, laying a high-reliability foundation for subsequent confidence interval fusion (S3-S5).

[0091] S3. Define subset A iis a subset of the first i elements of the set G, 1≤i≤n; based on the fuzzy density interval g i and the associated parameter λ, using the Newton-Raphson recursion method to calculate the subset A i The fuzzy density interval w' i ; Calculate A n times from i=1 to i=n i The fuzzy density interval w' i ; get w'1,w'2,…,w' n ;

[0092] In specific implementation, the subset A is calculated according to the following formula i The fuzzy density interval w' i :

[0093] g λ (x1∪x2)=g λ (x1)+g λ (x2)+λg λ (x1)g λ (x2);

[0094] w' i =g λ (x1∪x2…∪x i )=g λ (x1∪x2…∪x i-1 )+g λ (x i )+λg λ (x1∪x2…∪x i-1 )g λ (x i );

[0095] Among them, g λ (x i )=g i ;x i =μ ikcv ; x1∪x2…∪x i =A i .

[0096] Traditional confidence fusion (such as weighted average and DS evidence theory) assumes that devices are independent of each other and ignores the nonlinear effect of reliability superposition (such as when devices A and B work together, the overall reliability ≠ A+B). In scenarios with high device correlation (such as drones and unmanned vehicles jointly observing the same target), it is easy to cause overfitting of the fusion results. In this method, the complementarity or redundancy between devices is explicitly modeled through λ (λ>0 indicates complementarity, λ<0 indicates redundancy), solving the combination paradox of "1+1>2" or "1+1<1". Compared with the independent assumption model, it more accurately depicts the device interaction effect in actual collaborative scenarios. In addition, the existing fuzzy integral needs to calculate 2n-1 subset measures (n is the number of devices). When n≥5, the computational complexity increases exponentially, making it difficult to implement in engineering. This method uses a recursive formula (to reduce the computational complexity from O(2 n ) is reduced to O(n). By iteratively reusing the results of the preceding subset, repeated calculations are avoided, significantly improving the real-time performance of large-scale air-ground collaborative systems (e.g., 10+ devices).

[0097] S4. Sort the confidence values ​​in the set G from small to large. The confidence value of the i-th one after sorting is f(x' i );

[0098] S5, combined with the confidence values ​​after sorting f(x' i ), and each subset A i The fuzzy density interval w' i , use the improved fuzzy integral algorithm to perform data fusion and obtain the fusion confidence interval of the corresponding target information. The calculation formula includes:

[0099]

[0100] Where a and b represent the lower limit and upper limit of the confidence interval respectively;

[0101] S6. Defuzzify the fusion confidence interval obtained in S5 to obtain the defuzzified information of the kth target, with the cth aspect being the vth result. In specific implementation, the maximum membership method is used for defuzzification. In this way, the maximum membership method, based on the mean principle (balance) and full-link compatibility (inheriting confidence interval information), solves the problems of decision risk imbalance and information inheritance failure in traditional defuzzification methods, ultimately achieving the core goal of the air-ground collaborative system: outputting stable and reliable target classification decisions in an uncertain information environment.

[0102] S7. Repeat the above process to obtain the defuzzification of various results of the cth aspect of the kth target, and comprehensively judge the recognition result of the cth aspect of the kth target; until the recognition results of all aspects of all targets are judged.

[0103] The existing methods cannot quantify the dynamic reliability of equipment due to fixed weights (e.g., the confidence of drones drops sharply in foggy weather). In this method, S1 dynamically sets the equipment fuzzy density interval g i , S3 recursively deduces the subset fuzzy density w by the Newton-Raphson method i (Capturing fluctuations in the reliability of multiple devices). This approach establishes an interval-based dynamic evaluation model for device reliability, replacing the fixed-weight weighted average method used in existing technologies. When faced with complex terrain obstructions or fluctuating device states, the fusion results automatically reduce the contribution of low-reliability devices, significantly suppressing anomalous data interference and avoiding misjudgments. Furthermore, existing methods (such as DS evidence theory) output a single decision value, losing uncertainty information. In this method, S5 uses a modified fuzzy integral to output a confidence interval (e.g., [0.72, 0.85] for vehicle identification), and S6 uses the maximum membership method to defuzzify. This preserves confidence uncertainty in the fusion result as an interval (the upper limit represents an optimistic estimate, the lower limit represents a conservative estimate). Compared to existing probabilistic fusion methods (such as Bayesian output of 0.8), this solution can quantify the confidence range of the recognition result (e.g., "vehicle confidence is between 0.72 and 0.85"), providing a buffer for high-risk decision-making. Furthermore, this method can improve target feature resolution in complex environments, addressing the problem of wide variations in recognition accuracy due to factors such as target damage status and camouflage interference. In this method, S4 sorts the confidence scores of multiple sources by value (ensuring the monotonicity of the fuzzy integral), and S5 improves the integral algorithm to strengthen the nonlinear coupled calculation of the sorted confidence scores. This method adaptively fuses complementary information by combining confidence sorting with fuzzy measure correlation modeling (S2λ parameter) to address the heterogeneous data characteristics of unmanned vehicles (near-range high-definition) and drones (wide-area low-definition) in air-ground collaboration, such as partially damaged vehicles and disguised personnel. Compared to direct average fusion, the recognition rate of occluded and partially damaged targets is significantly improved.

[0104] This method can effectively integrate multi-dimensional heterogeneous target information of air-ground collaboration and output high-confidence recognition decisions in complex environments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the technical solutions. Those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present invention that do not depart from the purpose and scope of the technical solutions of the present invention should be included in the scope of the claims of the present invention.

Claims

1. An air-ground collaborative target recognition method based on fuzzy integral decision-level fusion is characterized by: The following steps are involved: S0. Use a single UAV or unmanned vehicle as a source for collecting information, obtain target detection data from each source, and perform spatial and temporal calibration to obtain information data of each target collected by each source; The information data of a target includes multiple aspects of information, including the type and security level of the target; The data of each aspect includes the probability of occurrence of multiple results, and the probability of occurrence is recorded as the confidence value; the confidence value of the cth aspect of the i-th source to the k-th target is the v-th result is μ ikc ; and μ ikc1 +…μ ikc +μ ikc =1, N is the number of possible outcomes of aspect c; The confidence values ​​of the v-th result of each source on the c-th aspect of target k are grouped into a set G kcv , G kcv ={μ 1kcv , μ 2kcv ..., μ ikc ..., μ nkcv }; S1. Set the fuzzy density interval g of each source for the vth result of the cth aspect i , the fuzzy density interval satisfies the regularity constraint; S2, combined with fuzzy density interval g i , use Newton iteration method to solve the value of the associated parameter λ; S3. Define subset A i For the set G kcv The subset consisting of the first i elements of , 1≤i≤n; based on the fuzzy density interval g i and the associated parameter λ, using the Newton-Raphson recursion method to calculate the subset A i The fuzzy density interval w′ i ; Calculate A n times from i=1 to i=n i The fuzzy density interval w′ i , we get w′1,w′2,…,w′ n ; S4, for set G kcv The confidence values ​​in are sorted from small to large, and the confidence value of the i-th one after sorting is f(x′ i ); S5, combined with the confidence values ​​f(x′ i ), and each subset A i The fuzzy density interval w′ i , use the improved fuzzy integral algorithm to perform data fusion and obtain the fusion confidence interval of the corresponding target information. The calculation formula of the fuzzy integral algorithm includes: f(x′0)=0; Where a and b represent the lower limit and upper limit of the confidence interval respectively; S6, defuzzifying the fusion confidence interval obtained in S5, and obtaining defuzzified information of the cth aspect of the kth target as the vth result; S7. Repeat the above process to obtain the defuzzification of various results of the cth aspect of the kth target, and comprehensively judge the recognition result of the cth aspect of the kth target; until the recognition results of all aspects of all targets are judged.

2. The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion according to claim 1 is characterized by: In S3, the subset A is calculated according to the following formula i The fuzzy density interval w′ i : g λ (x1∪x2)=g λ (x1)+g λ (x2)+λg λ (x1)g λ (x2); w′ i =g λ (x1∪x2…∪x i )=g λ (x1∪x2…∪x i-1 )+g λ (x i )+λg λ (x1∪x2…∪x i-1 )g λ (x i ); among them,g λ (x i )=g i ;x i =μ ikcv ;x1∪x2…∪x i =A i 。 3. The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion according to claim 1 is characterized by: In S2, the Newton iteration method is used to solve the following equation to calculate the associated parameter λ:

4. The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion according to claim 1 is characterized by: In S0, the starting point of the detection task is used as the calibration point of the data space information to perform spatial calibration; and the least square method is used to perform temporal calibration.

5. The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion according to claim 1 is characterized by: In S1, by conducting recognition experiments in different environments and weather conditions and combining expert experience, the fuzzy density ranges of unmanned vehicles and drones were set.

6. The air-ground collaborative target recognition method based on fuzzy integral decision-level fusion according to claim 1 is characterized by: In S6, the maximum membership method is used for defuzzification.

Images