Unmanned aerial vehicle low-altitude target identification method based on multi-source perception

Abstract

The invention discloses an unmanned aerial vehicle low-altitude target identification method based on multi-source perception, and belongs to the field of target identification, and the method comprises the steps: obtaining multi-source perception data, carrying out the preprocessing and dimension reduction of the multi-source perception data, and obtaining a dimension reduction feature set; fusing the dimension reduction feature set through a deep learning model to obtain a fused feature set; performing dimension judgment on the fused feature set, and obtaining a simplified feature set based on a judgment result; recognizing the simplified feature set through a deep learning model to obtain a preliminary recognition result; acquiring spatial data, and performing region classification according to the preliminary recognition result and the spatial data to obtain an environment classification result; and according to an environment classification result, the simplified feature set is identified again to obtain a target classification result, the target classification results of different unmanned aerial vehicles are integrated, and dynamic information of target identification is obtained according to the integration result. The accuracy and real-time performance of low-altitude target identification of the unmanned aerial vehicle can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition for unmanned aerial vehicles (UAVs), and in particular relates to a low-altitude target recognition method for UAVs based on multi-source perception. Background Art

[0002] Low-altitude target recognition by drones is a key research area in modern aviation safety and defense. Its importance lies in ensuring the safety and efficient management of low-altitude airspace. With the widespread adoption of drone technology, the need for target recognition in low-altitude environments is becoming increasingly urgent, especially in complex scenarios such as cities and airports. Accurate target identification is crucial for preventing unauthorized intrusions and collisions. However, existing methods for low-altitude target recognition often rely on a single sensor, making it difficult to cope with the diversity and dynamic changes of targets in complex environments. This results in insufficient recognition accuracy, especially in inclement weather or occlusion conditions, where traditional methods are prone to failure.

[0003] In this field, the core challenges stem from the effective integration of multi-source perception data and the spatial span problem in complex low-altitude environments. Multi-source perception involves the data fusion of multiple sensors such as optical, infrared, and radar. Each sensor captures information of different dimensions, and the heterogeneity of the data complicates feature extraction and fusion. If these data cannot be effectively integrated, it will be difficult for the system to accurately distinguish the shape, thermal characteristics, or motion trajectory of the target, which in turn affects the recognition performance. Furthermore, this problem directly leads to the system's lack of adaptability in environments with different spatial spans. Low-altitude environments include densely occluded areas in urban canyons and wide areas in open areas. The difference in spatial span requires the system to have flexible parameter adjustment capabilities, otherwise it will be difficult to maintain stable recognition effects over a large area. On a deeper level, this lack of adaptability limits the realization of multi-UAV collaborative perception, making it difficult for the system to expand its spatial coverage through a distributed network.

[0004] Therefore, how to design a UAV low-altitude target recognition method that can effectively extract multi-source perception data features and adaptively adjust to adapt to different spatial span environments has become a key issue in this study. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a UAV low-altitude target recognition method based on multi-source perception to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a method for identifying low-altitude targets of unmanned aerial vehicles based on multi-source perception, comprising:

[0007] Acquire multi-source perception data, pre-process and reduce the dimensionality of the multi-source perception data to obtain a reduced dimensionality feature set; fuse the reduced dimensionality feature set through a deep learning model to obtain a fused feature set; perform dimensionality judgment on the fused feature set, and obtain a simplified feature set based on the judgment result; identify the simplified feature set through a deep learning model to obtain a preliminary recognition result; acquire spatial data, perform regional classification based on the preliminary recognition result and the spatial data, and obtain an environmental classification result; based on the environmental classification result, re-identify the simplified feature set to obtain a target classification result; integrate the target classification results of different UAVs, and obtain dynamic information of target recognition based on the integration result.

[0008] Optionally, the process of obtaining the dimensionality reduction feature set includes:

[0009] Preprocessing is performed on multi-source perception data, wherein the multi-source perception data includes optical data, infrared data, and radar data; the preprocessing includes denoising and calibration processing; and dimensionality reduction is performed on the multi-source perception data through principal component analysis to obtain a reduced-dimensionality feature set.

[0010] Optionally, the process of obtaining the fused feature set includes:

[0011] A deep belief network is constructed, the restricted Boltzmann machine in the deep belief network is pre-trained, and the fusion feature set is fused through the trained deep belief network to obtain the fusion feature set.

[0012] Optionally, the process of obtaining the reduced feature set includes:

[0013] The dimensionality of the fused feature set is judged. When the dimension of the fused feature set is greater than a preset threshold, the dimensionality is reduced by a recursive feature elimination method. When the dimensionality reduction result of the recursive feature elimination method is greater than the preset threshold, redundant features are removed again by correlation analysis to obtain a streamlined feature set. Otherwise, the fused feature set is directly used as the streamlined feature set.

[0014] Optionally, the process of obtaining the environmental classification results includes:

[0015] The spatial data and dimensionality reduction feature set are classified through the machine learning model, and the preliminary recognition results are input into the machine learning model as auxiliary information to obtain the environmental classification results.

[0016] Optionally, the process of obtaining the target classification result includes:

[0017] According to the environmental classification results, an adaptive model is constructed, wherein the adaptive model is a deep learning model. During the construction of the adaptive model, the model parameters of the adaptive model are adjusted according to the corresponding relationship between the preset environmental classification results and the model parameters of the adaptive model; the simplified feature set is identified through the adaptive model to obtain the target classification result.

[0018] Optionally, the adaptive model is trained by transfer learning.

[0019] Optionally, the process of integrating the target classification results of different drones includes:

[0020] The integrity of the target classification result is judged. When the target classification result is complete, the target classification result is distributed through the consistent hashing algorithm to obtain a data mapping table between different drones. According to the data mapping table, the target classification result is transmitted through the data sharing mechanism between drone nodes to obtain a coverage-optimized data set. The coverage rate of the coverage-optimized data set is judged. When the coverage rate is greater than a threshold, the target classification results identified by the drone nodes are integrated through a distributed fusion algorithm to obtain a preliminary collaborative perception result. The preliminary collaborative perception structure is subjected to consistency verification and global optimization to obtain the final global target recognition result, i.e., the integration result.

[0021] Optionally, based on the integration result, the target motion trajectory in the integration result is predicted by a Kalman filtering method to obtain dynamic information of target recognition.

[0022] On the other hand, the present application provides a UAV low-altitude target recognition system based on multi-source perception, which is used to execute the above method.

[0023] Compared with the prior art, the present invention has the following advantages and technical effects:

[0024] The present invention discloses a method for low-altitude target recognition of unmanned aerial vehicles (UAVs) based on multi-source perception. The method first preprocesses and extracts features from optical, infrared and radar data, and then generates unified fusion features through principal component analysis dimensionality reduction and deep belief network fusion. A convolutional neural network is then used for preliminary target classification, and the network parameters are adaptively adjusted according to environmental characteristics. Finally, transfer learning is used to improve adaptability to complex environments, and global recognition results are obtained through distributed fusion of multiple UAV nodes. The present invention also combines Kalman filtering to achieve target trajectory prediction, effectively solving key problems such as multi-source heterogeneous data fusion, complex environment adaptive recognition, and multi-node collaborative perception, significantly improving the accuracy, robustness and real-time performance of target recognition, and providing an effective solution for intelligent perception in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0026] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0027] Figure 2 Schematic diagram of the system architecture of an embodiment of the present invention. DETAILED DESCRIPTION

[0028] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0030] like Figure 1 As shown, a method for identifying low-altitude targets of a drone based on multi-source perception in this embodiment may specifically include:

[0031] Step S101: Acquire multi-source perception data collected by optical, infrared, and radar sensors, use a preprocessing algorithm to denoise the optical data to generate a first optical feature set, perform thermal imaging correction on the infrared data to generate a first infrared feature set, and perform signal filtering on the radar data to generate a first radar feature set, thereby obtaining a multi-source feature set.

[0032] Multi-source data collected by optical, infrared, and radar sensors are obtained. Preprocessing algorithms are used to denoise the optical data, perform thermal imaging correction on the infrared data, and perform signal filtering on the radar data to obtain the first optical feature set, the first infrared feature set, and the first radar feature set. The processed feature sets are then integrated to generate a multi-source feature set.

[0033] For example, in practical applications, multi-source data collected by optical, infrared, and radar sensors requires preprocessing to generate high-quality feature sets, providing a reliable foundation for subsequent analysis. For example, optical data denoising aims to eliminate noise in images and improve target recognition accuracy. Images collected by optical sensors are often affected by lighting variations and lens smudges, resulting in salt-and-pepper or Gaussian noise. Median filtering can be used to replace the central pixel value with the median of a pixel's neighborhood within a sliding window to suppress salt-and-pepper noise. Specifically, thermal image correction of infrared data is used to eliminate sensor non-uniformity and ambient temperature interference. Thermal images collected by infrared sensors may exhibit streaks or dark spots due to variations in detector response. A non-uniformity correction algorithm can be used to generate correction coefficients based on a calibrated blackbody radiator. In one possible implementation, signal filtering of radar data is used to remove clutter and multipath interference, improving target detection accuracy. Radar sensors are often affected by ground reflections or nearby objects, resulting in high-frequency noise in the signal. A Kalman filter can be used to predict the position and correct the measured values ​​based on a target motion model. It should be noted that the integration process requires consideration of sensor temporal synchronization and spatial registration. For example, optical and infrared data require pixel coordinate alignment through camera calibration, while radar data requires coordinate transformation to match the image plane. Once synchronized, the fused features effectively handle harsh environments like rain and fog, ensuring the stability of the drone system's collected data.

[0034] Understandably, the advantages of the aforementioned preprocessing and fusion methods lie in improving data quality and information completeness. Denoising, correction, and filtering each optimize the data quality of a single sensor, providing a solid foundation for safe driving.

[0035] Step S102: For the multi-source feature set, a principal component analysis algorithm is used to perform dimensionality reduction processing on the first optical feature set, the first infrared feature set, and the first radar feature set to generate a second optical feature set, a second infrared feature set, and a second radar feature set to obtain a unified dimensional feature set.

[0036] The principal component analysis algorithm is used to reduce the dimensionality of the first optical feature set, the first infrared feature set, and the first radar feature set to generate the second optical feature set, the second infrared feature set, and the second radar feature set, resulting in a unified dimensional feature set. The principal component analysis algorithm is used to reduce the dimensionality of the first optical feature set, the first infrared feature set, and the first radar feature set to generate a unified dimensional feature set. The core of this process is to reduce data redundancy through dimensionality reduction while retaining key information. The basic principle of principal component analysis is to project high-dimensional data into a low-dimensional space and identify the principal component direction of the data, that is, the direction with the largest variance.

[0037] The optical feature set may contain information such as pixel values ​​and color distribution from high-resolution images, and its dimensions may reach thousands of dimensions. Principal component analysis analyzes the correlations among these features, extracting the principal components that contribute most to the variance and generating a second optical feature set. For example, if the original optical feature set has 1000 dimensions, after principal component analysis, the first 200 dimensions may be retained, as these dimensions represent over 90% of the data variance.

[0038] Specifically, the dimensionality reduction process for infrared feature sets is similar to that for optical feature sets, but their data characteristics are different. Infrared data usually reflects thermal imaging information and may contain features such as temperature distribution and thermal signal intensity.

[0039] For example, a set of infrared signatures might record the thermal signature values ​​of a target area and have a dimensionality of 800. Principal component analysis identifies the directions most correlated with thermal signature variations, generating a second set of infrared signatures with a dimensionality reduction of perhaps 150. This dimensionality reduction method highlights key thermal imaging features and facilitates subsequent processing. Principal component analysis is also applicable to radar signatures. Radar data typically includes information such as distance, velocity, and angle, and the feature dimensions may be high due to the high sampling frequency.

[0040] For example, a radar feature set might contain 600-dimensional information about signal strength and Doppler shift. Principal component analysis calculates the covariance matrix and extracts the principal components that contribute most to the variance, generating a second radar feature set with dimensions down to 100. This approach effectively reduces noise interference and highlights the target's motion characteristics.

[0041] For example, in the object detection scenario, a unified dimensionality feature set can be used as input to provide a more consistent feature representation.

[0042] It should be noted that the implementation of principal component analysis requires data standardization to eliminate dimensional differences.

[0043] For example, the pixel values ​​of an optical feature may range from 0 to 255, while the temperature values ​​of an infrared feature may range from -20 to 100 degrees Celsius. Before processing, all features must be normalized to a distribution with a mean of 0 and a standard deviation of 1 to ensure fairness in principal component analysis. This preprocessing can improve the reliability of the dimensionality reduction effect. For example, in practical applications, the reduced second optical feature set, second infrared feature set, and second radar feature set can be used for target classification tasks. Unifying the dimensionality of feature sets can reduce computational complexity while preserving the core information of multi-source sensor data. This approach is particularly important when processing large-scale multi-source data and can improve system efficiency and robustness.

[0044] In step S103, based on the unified dimensional feature set, a deep belief network algorithm is used to perform feature fusion on the second optical feature set, the second infrared feature set, and the second radar feature set to generate a fused feature vector, wherein the deep belief network extracts cross-modal features through a pre-trained restricted Boltzmann machine to obtain a fused feature set.

[0045] A pre-trained restricted Boltzmann machine is used to perform initial feature extraction on the second optical feature set, the second infrared feature set, and the second radar feature set to obtain a multimodal feature subset. A deep belief network is used to perform cross-modal feature fusion on the multimodal feature subset to generate a fused feature vector.

[0046] In one possible implementation, a pre-trained restricted Boltzmann machine (RBM) is used to perform initial feature extraction on the second optical, infrared, and radar feature sets, effectively capturing the deep features of each modal data. The RBM is an unsupervised learning model that models data distribution through an energy function and is suitable for processing high-dimensional, complex, and multi-source features.

[0047] For example, in a target recognition scenario, the second optical feature set might contain the target's texture and color information, with a dimension of 1000; the second infrared feature set records the target's thermal radiation characteristics, with a dimension of 800; and the second radar feature set reflects the target's distance and speed information, with a dimension of 600. A pre-trained restricted Boltzmann machine can model these feature sets separately and learn the hidden layer representation for each modality.

[0048] For example, optical features may be compressed into a 200-dimensional subset to preserve texture and edge information; infrared features may be refined into a 150-dimensional subset to highlight thermal radiation distribution; and radar features may be generated into a 100-dimensional subset to focus on motion trajectories. This initial extraction preserves the core information of each modality and provides high-quality input for subsequent fusion.

[0049] For example, cross-modal feature fusion of multimodal feature subsets using a deep belief network can achieve a unified representation of features from different modalities. Deep belief networks consist of stacked layers of restricted Boltzmann machines, which learn the correlations between cross-modal features through layer-by-layer training and fine-tuning.

[0050] In one embodiment, assuming the multimodal feature subset includes 200-dimensional, 150-dimensional, and 100-dimensional features for optical, infrared, and radar, the deep belief network can input these features into the first layer to generate a 500-dimensional joint representation, which is then gradually compressed into a 300-dimensional fused feature vector through hidden layers. This fusion process can capture the correspondence between the target shape in the optical features and the thermal signature in the infrared features.

[0051] For example, the metal part of a target may appear as a highly reflective area in an optical image and a high-temperature area in an infrared image. Radar signatures complement the target's dynamic information; for example, a moving target at a speed of 5 meters per second may be associated with specific patterns in both optical and infrared signatures.

[0052] It should be noted that the generation of fused feature vectors can provide a more comprehensive description for subsequent tasks.

[0053] For example, in target recognition tasks, the fused feature vector can simultaneously reflect the appearance, thermal characteristics, and motion state of the target, which is more robust than single modality features.

[0054] After integrating this information through a deep belief network, a comprehensive feature vector can be generated to support more accurate classification decisions. This cross-modal fusion method fully utilizes the complementarity of multi-source data and provides a reliable basis for target analysis in complex environments.

[0055] Step S104: If the dimension of the fused feature set is greater than the maximum allowed dimension, a recursive feature elimination algorithm is used to perform feature selection on the fused feature set, delete redundant features, and generate a streamlined feature set.

[0056] If the dimensionality of the fused feature set exceeds the maximum allowed dimensionality, the feature set dimensionality information is obtained and a recursive feature elimination algorithm is used for feature selection to remove redundant features and obtain a first reduced feature set. Based on this first reduced feature set, the correlation matrix between features is calculated. If the correlation coefficient exceeds a preset threshold, highly correlated features are removed to obtain a reduced feature set.

[0057] When the dimensionality of the fused feature set exceeds the maximum allowed dimensionality, obtaining the dimensionality information of the feature set is the first step. The fused feature set is usually composed of multimodal data such as optical, infrared, and radar features, and the dimensionality may be as high as hundreds or even thousands.

[0058] For example, in a target recognition system, the fused feature vector is 300-dimensional, while the system's maximum allowed dimension is 200. Obtaining dimensional information provides a clear understanding of the extent of the excess dimension, providing a basis for subsequent feature selection.

[0059] For example, when using a recursive feature elimination algorithm for feature selection, the core goal is to retain critical information by iteratively removing features that contribute least to the model. Suppose, in the aforementioned object recognition system, a random forest model is used to assess the importance of each feature. The recursive feature elimination algorithm repeatedly calculates the contribution of features to the classification task. For example, texture features within an optical feature set may be highly important, while certain redundant brightness features may be less important. Each iteration, the 10% of features with the lowest importance are removed until the dimensionality is reduced to 200, resulting in the first reduced feature set. This process ensures that the most valuable features for object recognition are retained while reducing computational complexity.

[0060] Specifically, a correlation matrix between features is calculated based on the first reduced feature set to further eliminate highly correlated features. The correlation matrix reflects the linear dependency between features by calculating the Pearson correlation coefficient between features.

[0061] For example, in a target recognition system, among the 300 features in the first reduced feature set, it may be found that the edge intensity of the optical features and the thermal intensity of the infrared features have a correlation coefficient of 0.9, exceeding the preset threshold of 0.8. In this case, the edge intensity features, which contribute more to the model, can be retained while the thermal intensity features are removed, ultimately resulting in a reduced feature set. This approach avoids information redundancy and improves the independence of the feature set.

[0062] In one possible implementation, the construction and analysis of the correlation matrix can be optimized based on the business scenario. For example, if a target recognition system is used for nighttime target detection, infrared features may be more important than optical features. When calculating correlation, key infrared features, such as thermal distribution, can be prioritized while removing portions of optical features that are highly correlated with infrared. This strategy ensures a streamlined feature set that better meets the business needs of nighttime detection.

[0063] For example, the combination of recursive feature elimination and correlation analysis can form an efficient feature selection process for multimodal data processing. In a target recognition system, the initial 300-dimensional feature set is reduced to 250 dimensions through recursive feature elimination and then further reduced to 200 dimensions through correlation analysis. This streamlined feature set not only meets the system's dimensionality constraints but also retains key information from cross-modal features, such as shape information from optical features, thermal information from infrared features, and distance information from radar features. This information supports each other and jointly improves the robustness of target recognition.

[0064] In one possible implementation, the generation of a streamlined feature set can also aid decision-making by visualizing a correlation matrix. The correlation matrix can be plotted as a heat map to intuitively display areas of high correlation between features.

[0065] For example, if there are large areas of high correlation between optical and infrared features in the heat map, the contribution of features in these areas to the model can be prioritized to guide the elimination decision. This visualization method improves the transparency and explainability of feature selection.

[0066] As you can understand, the above method progressively optimizes the feature set through recursive feature elimination and correlation analysis. From the initial high-dimensional feature set to the streamlined feature set, each step focuses on the core requirements of object recognition, ensuring a concise and efficient feature set. This multi-level feature selection strategy effectively addresses the high-dimensionality challenges of multimodal data and provides high-quality input for subsequent model training.

[0067] Step S105: If the dimension is less than or equal to the maximum allowed dimension, the fused feature set is directly determined to be a reduced feature set.

[0068] If the feature dimension is less than or equal to the maximum allowed dimension, the simplified feature set is directly determined as the fused feature set to obtain the fused feature set.

[0069] For example, in the field of feature selection, when the dimensionality of a fused feature set is less than or equal to the maximum allowed dimensionality, it can be directly determined as a reduced feature set. The key to this approach is to avoid unnecessary feature removal and preserve the information integrity of the original features. In one embodiment, diversity analysis of the feature set is considered. Even if the dimensionality does not exceed the limit, the diversity between features can still be checked to ensure low information redundancy.

[0070] For example, the information contribution of features can be assessed by calculating their variance. For example, suppose the variance of optical features is 20, the variance of infrared features is 15, and the variance of radar features is only 5. Low-variance features may contribute less, but because the dimensionality is within the limit, retaining them does not increase the computational burden. Instead, they may provide supplementary information in specific scenarios, such as differences in emotional expression in noisy environments.

[0071] Step S106: Based on the streamlined feature set, a convolutional neural network algorithm is used to classify the target and generate a target recognition result. The convolutional neural network extracts the spatial features of the target through multi-layer convolution and pooling to obtain a preliminary recognition result.

[0072] A convolutional neural network is used to process the reduced feature set, extracting spatial features through multi-layer convolution operations to produce a feature map. Pooling is used to reduce the dimensionality of the feature map to generate a compressed feature set. If the resolution of the compressed feature set falls below a preset threshold, the feature map resolution is adjusted using an upsampling module to produce an adjusted feature set. A fully connected layer is used to classify the adjusted feature set to produce a preliminary classification result. The preliminary classification result is then nonlinearly mapped using an activation function to produce the final classification result. The target recognition result is determined based on the degree of match between the final classification result and the preset category label.

[0073] For example, when processing a reduced feature set, a convolutional neural network (CNN) extracts spatial features from an image through a convolution operation. This operation slides a kernel across the input feature set, capturing patterns in local regions, such as edges and textures.

[0074] In one possible implementation, assuming a 3×3 convolution kernel size and a stride of 1 when processing fused features, the first layer of the network uses 32 convolution kernels, generating 32 feature maps. Each feature map retains the spatial structure of the original image. This approach effectively extracts local features of the target object and, in target recognition scenarios, can capture key contour information such as object shape.

[0075] Specifically, the pooling operation performs dimensionality reduction on the feature map, aiming to reduce the amount of computation and retain the main features. Maximum pooling is a commonly used method, which usually uses a 2×2 pooling window with a stride of 2. For example, after the above-mentioned feature map is pooled, the amount of data is reduced, but significant spatial features are still retained. This dimensionality reduction method can highlight key areas and reduce the interference of background noise in target recognition. In one embodiment, if the resolution of the compressed feature set is lower than a preset threshold, the feature map resolution needs to be adjusted through an upsampling module. Upsampling can use transposed convolution to gradually increase the resolution of the feature map. This operation can restore details lost due to pooling in target recognition, such as subtle local textures of the target, thereby improving the accuracy of subsequent classification.

[0076] Preferably, the fully connected layer maps the adjusted feature set to the classification space. The adjusted feature set is input into the fully connected layer, and a 10-dimensional vector is output, corresponding to 10 preset target categories. The fully connected layer generates preliminary classification results by learning the relationship between features and categories. This method can convert complex features into category probabilities in target recognition. For example, the activation function performs nonlinear mapping on the preliminary classification results, and ReLU or Softmax functions are commonly used. Softmax can convert a 10-dimensional vector into a probability distribution. For example, the probability that a target is identified as "A" is 0.9. This nonlinear mapping enhances the model's adaptability to complex patterns and makes the classification results more reliable.

[0077] It is understood that the target recognition result is determined by the degree of match between the final classification result and the preset category label. For example, if the system has 10 known target labels, and the classification result shows that the match between a target and the label "A" is 95%, then the target is judged as "A".

[0078] It should be noted that the layered approach from feature extraction to classification ensures the accuracy and efficiency of recognition results. Each module supports each other, forming a complete technology chain that is suitable for efficient processing in a single business scenario.

[0079] Step S107: Obtain the preliminary recognition result and the spatial span parameter of the current environment, and use the support vector machine algorithm to classify the urban occlusion environment and the open area environment. If the spatial span parameter indicates that it is an urban occlusion environment, adjust the convolution kernel size of the convolutional neural network to S1 (S1 is a positive integer, indicating a small convolution kernel).

[0080] The spatial span parameters of the environment and target recognition data are obtained, and an input feature set is generated through a feature extraction process. Using the support vector machine algorithm, the input feature set is classified based on the classification decision boundary to determine whether the environment is urban occlusion or open area. If the classification result is urban occlusion, the convolutional neural network's convolution kernel size is adjusted to a positive integer value S1, and the adjusted convolution kernel parameters are determined.

[0081] For example, obtaining the environment's spatial span parameters and target recognition data is the basis for urban environment classification. Spatial span parameters refer to the spatial scale of the environment, such as the length, width, and height of an area or the distance between key landmarks. Target recognition data includes the target recognition results obtained by fusion of sensor-collected images, point clouds, and infrared data. The target recognition results serve as reference data, while the dimensionality-reduced data and the aforementioned spatial scale parameters serve as input data for a support vector machine (SVM). This provides reliable input for SVM classification. It should be noted that the SVM algorithm constructs a classification decision boundary to classify the input feature set into either an urban occluded environment or an open area environment. The classification boundary is determined by determining the optimal hyperplane based on sample points in the training data. For example, during the training phase, 1,000 feature sets are input: 500 from occluded environments and 500 from open areas. The SVM calculates the boundary. Using dimensionality-reduced radar data as an example, the feature vectors of urban occluded environments tend to have high-density point clouds, while the point clouds of open areas have lower density. The classification results show that the point cloud density of the test area is 600 points per cubic meter, making it classified as an occluded environment. This classification method has clear logic and clear classification boundaries. Specifically, if the classification result is an urban occluded environment, the convolution kernel size of the convolutional neural network needs to be adjusted to a positive integer value S1. The size of the convolution kernel affects the receptive field of feature extraction. The occluded environment is densely populated with buildings, so a smaller convolution kernel is required to capture details. Preferably, S1 is set to 3 to generate a 3x3 convolution kernel. Compared with the 5x5 convolution kernel that may be used in open areas, the 3x3 convolution kernel is more suitable for extracting local features of dense buildings. For example, when processing image data of an occluded environment, the 3x3 convolution kernel can focus on the edges and textures of buildings to avoid excessive background noise interference.

[0082] In one embodiment, the adjusted convolution kernel parameters need to be further verified for their applicability. For example, a set of fused features is processed using the adjusted 3x3 convolution kernel. After extracting the feature map, it is observed whether the buildings and roads can be clearly distinguished. The verification results show that the sharpness of the building edges in the feature map has increased by 10%, indicating that the adjusted parameters are more suitable for occluded environments. This dynamic adjustment method can flexibly adapt to different environmental characteristics. Each link of the above method is closely connected, from data collection to classification to parameter adjustment, to form a complete urban environment analysis chain. For example, the precise acquisition of spatial span parameters provides reliable input for feature extraction, and the classification results of the support vector machine directly guide the parameter optimization of the convolutional neural network. This logically rigorous process ensures the accuracy of environmental classification and the targetedness of subsequent processing, providing efficient technical support for urban environment analysis.

[0083] Step S108: If it is an open area, adjust it to S2 (S2 is a positive integer, S2 is greater than S1) to generate an adaptive recognition model.

[0084] If the environment type is open area, parameters S1 and S2 are adjusted according to preset rules to obtain adjusted parameter values. Based on the adjusted parameter values, the aforementioned convolutional neural network is used to construct an adaptive recognition model to generate an initial model structure. Using this initial model structure, a training dataset is obtained and the model parameters are optimized to obtain an optimized model. Based on this optimized model, the recognition threshold is dynamically adjusted to generate a final adaptive recognition model. The final adaptive recognition model processes the input data to obtain the recognition result.

[0085] In one possible implementation, for parameter adjustment in open area environments, preset rules can be based on the spatial span characteristics of the environment. For example, open areas usually have a larger field of view and a lower object density, so the convolution kernel size can be adjusted from S1 to S2, where S2 is a larger positive integer value to capture feature information in a wider area. For example, if S1 is 3, which is suitable for small-scale feature extraction in urban occluded environments, then S2 can be adjusted to 5 to meet the global feature requirements of open areas. The adjusted parameter values ​​are verified experimentally to ensure the model's sensitivity to environmental characteristics. Specifically, when constructing an adaptive recognition model based on the adjusted S2 parameters, a dynamic hierarchical design of a convolutional neural network can be adopted. Preferably, the network structure includes multiple convolution layers and pooling layers, combined with a convolution kernel of size S2 to generate an initial model structure. For example, in an open area scenario, the initial model can be configured with 3 convolution layers, each layer using a convolution kernel of S2=5, with a stride of 1 to enhance the ability to extract large-scale features.

[0086] It should be noted that the initial model structure needs to flexibly adjust the number of layers or convolution kernels according to the actual task requirements. In one embodiment, when obtaining a training data set, image data of open areas and fusion data of two other aspects of data can be used, such as suburban roads or field scenes, which include diverse lighting and weather conditions. The data set may include 10,000 annotated images and corresponding fusion data after infrared and point cloud processing, covering scenes such as daytime, evening, and cloudy days. When optimizing the model parameters, the gradient descent method is used to adjust the weights through multiple iterations. For example, during the training process, the learning rate can be set to 0.001 and iterated 200 times to ensure model convergence. The optimized model can capture the characteristics of open areas more accurately.

[0087] It is understandable that dynamically adjusting the recognition threshold is the key to improving the adaptability of the model. For example, in an open area, the edges of the target object may be blurred due to changes in lighting. Therefore, the threshold range can be dynamically set according to the feature distribution of the training data set, such as adjusting the confidence threshold from 0.7 to 0.6 to improve the recognition rate of weak feature targets. The final adaptive recognition model is optimized through this threshold and can cope with complex environments more flexibly. For example, when the final model processes input data, it can input a set of real-time image data of an open area, such as a highway monitoring video frame. Based on the adjusted S2 parameters and the optimized threshold, the model identifies vehicles and pedestrians in the picture, and outputs classification results and location information. This approach can significantly improve the robustness of recognition and adapt to diverse scenarios in open areas.

[0088] In step S109, based on the adaptive recognition model, a transfer learning algorithm is used to reclassify the simplified feature set to generate a final recognition result. Transfer learning improves the model's adaptability to complex environments by fine-tuning the pre-trained model to obtain the target classification result.

[0089] A feature extraction algorithm is used to generate feature vectors for the reduced feature set. Using a transfer learning algorithm, the pretrained adaptive model is loaded and fine-tuned. If the dimension of the feature vector meets the preset threshold, the adaptive model is used to classify the feature vector, generating the target classification result. Based on the target classification result, the result optimization algorithm is used to adjust the classification boundaries and generate an optimized classification result. By comparing the optimized classification result with the target recognition result, it is determined whether the classification accuracy meets the preset threshold and the target recognition result is obtained. If the classification accuracy of the target recognition result falls below the preset threshold, the model fine-tuning parameters are adjusted and the adaptive model is regenerated.

[0090] For example, when processing open area environmental data, the core of the feature extraction algorithm lies in extracting key information from complex inputs to generate a streamlined feature set. The input is a fusion of drone low-altitude aerial imagery, infrared, and radar data. The feature extraction algorithm can extract terrain contours and vegetation distribution characteristics through edge detection and texture analysis, generating a feature vector containing information such as color mean and texture density. Preferably, the feature vector dimension is controlled within 200 dimensions to reduce the complexity of subsequent calculations. This approach can effectively retain key environmental information and facilitate model processing.

[0091] In one possible implementation, the transfer learning algorithm achieves efficient modeling by loading a pre-trained model. For example, a convolutional neural network model pre-trained on an urban environment dataset is used to fine-tune it for open area scenarios. During fine-tuning, the weights of the last two fully connected layers can be adjusted to adapt to the feature distribution of the target scene. It should be noted that the choice of pre-trained model must match the target domain, such as selecting a model that is good at processing natural environments to ensure the transfer effect. Specifically, the dimensionality of the feature vector is a key step before classification. Assuming the preset threshold is 60 dimensions, if the feature vector is 50 dimensions, which meets the requirements, it is directly input into the adaptive model for classification.

[0092] For example, the model can classify input features as "forest", "grassland" or "water area". The classification process relies on the model's learning of feature weights to ensure that the classification results accurately reflect the type of environment. For example, the result optimization algorithm improves accuracy by adjusting the classification boundaries. Assuming that the initial classification mistakenly classifies an area as "grassland", the optimization algorithm can analyze the texture density of the feature vector and dynamically adjust the boundaries based on the training data to correct it to "forest". This adjustment is based on dynamic analysis of data distribution and can effectively reduce misclassification. In one embodiment, the classification accuracy is determined by comparing the optimized classification results with the target recognition results. Assuming the preset accuracy threshold is 90%, if the current classification accuracy is 85%, the model fine-tuning parameter adjustment is triggered. For example, increase the number of training rounds during fine-tuning or adjust the learning rate to 0.001 to regenerate the adaptive model.

[0093] Understandably, this iterative optimization can gradually improve the model's adaptability to complex environments. Preferably, when regenerating the adaptive model, a small amount of newly collected data, such as new aerial photography samples, can be introduced to enrich the diversity of the training set. This continuous optimization approach ensures that the model maintains high-precision classification capabilities in dynamic environments, providing reliable support for subsequent environmental analysis tasks.

[0094] Step S1010, obtain the final recognition results of multiple drone nodes, use the consistent hashing algorithm to perform distributed fusion on the target classification results, and generate collaborative perception results, where the consistent hashing optimizes the spatial coverage by sharing data between nodes to obtain the global target recognition result.

[0095] If the final target identification data is complete, a consistent hashing algorithm is used to distribute the data, generating a data mapping table between nodes. Based on this data mapping table, the classification results are transmitted through an inter-node data sharing mechanism to generate a data set with optimized coverage. If the spatial coverage of the data set meets the preset threshold, the node data is integrated using a distributed fusion algorithm to obtain a preliminary collaborative perception result. A consistency check is performed on the preliminary collaborative perception result to determine whether there are any data conflicts, resulting in a conflict detection result. If the conflict detection result is normal, the collaborative perception result is globally optimized to generate the final global target identification result.

[0096] For example, in the field of target identification, consistent hashing algorithms are used to distribute data. The key is to ensure even distribution of data across multiple nodes. In principle, consistent hashing constructs a virtual node ring, mapping data and nodes to the same hash space. Data is then distributed to the nearest node based on the hash value.

[0097] For example, suppose there are three nodes with 1,000 records. Each record generates a hash value and is evenly distributed across the nodes. Node A might receive 350 records, while nodes B and C receive 325 and 325, respectively. This distribution method minimizes data migration even when nodes are added or removed, making it suitable for dynamic expansion scenarios.

[0098] In one possible implementation, after generating a data mapping table between nodes, classification results are transmitted via an inter-node data sharing mechanism. This data sharing mechanism is typically based on message queues or point-to-point communication. For example, node A's classification results, including the target object's location information, need to be transmitted to nodes B and C for collaborative analysis. This can be implemented using a publish-subscribe model, with node A publishing the classification results and nodes B and C subscribing and receiving them. This mechanism ensures efficient and real-time data transmission. Specifically, if the spatial coverage of a dataset meets a preset threshold, such as 90%, the node data is integrated using a distributed fusion algorithm. The core of this distributed fusion algorithm is to merge the partial recognition results of each node into a unified perception result. For example, if node A identifies a target object in area X and node B identifies the same object in area Y, a fusion algorithm is used to calculate the target's precise location, generating a preliminary collaborative perception result. This fusion process typically relies on a weighted average or voting mechanism, with weights based on the node's credibility.

[0099] It's important to note that consistency checks on preliminary collaborative perception results are designed to detect data conflicts. For example, nodes A and B might disagree on the category of the same object, with node A identifying it as a vehicle and node B as a pedestrian. This verification process can resolve conflicts by comparing the confidence levels of each node and selecting the result with the higher confidence, or by triggering a secondary recognition process. If there are no conflicts, the conflict detection result is considered normal.

[0100] Preferably, time and space constraints can be introduced when globally optimizing the collaborative perception results. For example, based on the movement trajectory of the target object, the optimization algorithm eliminates abnormal points and generates the final global target recognition result. Specifically, if the target object is at position P1 at time t1 and at position P2 at time t2, the optimization algorithm determines whether the movement from P1 to P2 is reasonable based on the speed constraint and eliminates unreasonable data points. This optimization method improves the accuracy and robustness of the recognition results. The implementation of the above-mentioned technical themes jointly supports the complete process of efficient data allocation, collaborative processing and result optimization in the target recognition system.

[0101] For example, consistent hashing ensures balanced data distribution, data sharing mechanisms guarantee inter-node information flow, fusion algorithms and consistency checks enhance the reliability of results, and global optimization further refines the output. These interconnected and logically rigorous processes collectively enable efficient target recognition.

[0102] In step S1011, based on the global target recognition result, the Kalman filter algorithm is used to predict the target motion trajectory to generate a target trajectory prediction result, wherein the Kalman filter updates the target position and velocity through the state transfer matrix to obtain the target dynamic information.

[0103] Global target recognition is used to obtain the target's initial state data. A Kalman filter algorithm is used to perform a preliminary estimate of the target's position and velocity, resulting in an initial state estimate. Based on the initial state estimate, a state transition matrix is ​​constructed, and the state vector of the target's trajectory is updated to obtain an updated state vector. If the updated state vector deviates from the sensor observation data, the state estimate is adjusted using the Kalman filter's gain matrix to obtain an optimized state estimate. Based on the optimized state estimate, the target's position and velocity at the next moment are predicted to generate a target trajectory prediction. The trajectory prediction is analyzed for continuity. If the continuity meets a preset threshold, the trajectory prediction is deemed valid, resulting in valid trajectory data. Based on the valid trajectory data, the parameters of the state transition matrix are updated, generating an adaptive state update mechanism and a dynamically adjusted state model. Using this dynamically adjusted state model, the target state estimate is continuously iterated to generate a long-term target trajectory prediction.

[0104] For example, when global target recognition acquires the target's initial state data, sensors are typically used to capture the target's original position and velocity information. For example, in a drone collaborative tracking scenario, assume the drone is equipped with radar, a camera, and infrared sensors, and captures the target vehicle's coordinates as x = 100 meters, y = 50 meters, and its velocity as 10 meters per second. The Kalman filter algorithm reduces noise interference by fusing these observations with prior state estimates. In principle, the Kalman filter uses a two-step process: prediction and update. It predicts the target's state at the next moment, and the update corrects the prediction deviation based on sensor data.

[0105] Specifically, the drone may experience observation errors due to wind speed or sensor jitter. The Kalman filter uses weighted averaging to balance predictions and observations, generating an initial state estimate. For example, the corrected position x = 102 meters, y = 51 meters, and velocity 9.8 meters per second. Based on this initial state estimate, a state transition matrix is ​​constructed to describe the target's motion.

[0106] In one possible implementation, the state transition matrix contains a linear relationship between position and velocity. Assuming the target moves at a constant velocity, the matrix parameters reflect the displacement change within a time interval, such as 0.1 seconds. When updating the state vector, the matrix is ​​combined to calculate the target's new position, for example, predicting the position in 0.1 seconds to be x = 103 meters. It should be noted that if the target accelerates, the matrix needs to be dynamically adjusted to reflect nonlinear motion. If the state vector deviates significantly from the sensor observation, the Kalman filter's gain matrix plays a key role.

[0107] Optimally, the gain matrix adjusts the weight of the state estimate based on observation noise and prediction uncertainty. For example, if a sensor observes x = 104 meters but predicts x = 103 meters, the gain matrix might assign higher weight to the observed data, revising the estimate to x = 103.8 meters. This mechanism ensures that the state estimate is close to the true value. When predicting the target's position and velocity at the next moment, the optimized state estimate is combined with the state transition matrix for inference.

[0108] In one embodiment, the drone predicts the target vehicle's position 0.2 seconds from now to be x = 105.6 meters, maintaining a velocity of 9.8 meters per second. Trajectory continuity analysis ensures the prediction is reasonable. For example, the prediction point is checked to see if it is within the target's possible movement range. If the position jumps to x = 120 meters, it is considered discontinuous and requires re-estimation. Valid trajectory data is used to update the parameters of the state transition matrix, forming an adaptive mechanism. Specifically, if the target vehicle enters a curve, the matrix parameters are adjusted to reflect the change in curvature, such as reducing the weight of velocity. The dynamically adjusted state model is continuously iterated to generate a long-term trajectory prediction. For example, the predicted path of the vehicle 1 second into the future is a smooth curve, with the position gradually changing from x = 105.6 meters to x = 115 meters. This adaptive mechanism improves prediction robustness, especially in complex environments. It is understood that the above method forms a closed-loop mechanism through layered optimization, from initial estimation to long-term prediction. Each link supports each other, ensuring the drone's ability to continuously track the target. For example, the initial estimate provides a reliable starting point for subsequent predictions, while the dynamic matrix adjustment adapts to changes in the target's behavior. This progressive logic remains consistent within a single scenario, and the generated trajectory data supports drone mission planning.

[0109] On the other hand, Figure 2 As shown, the present invention provides a UAV low-altitude target recognition system based on multi-source perception, which mainly includes:

[0110] Multi-source data acquisition module, used to obtain multi-source perception data collected by optical, infrared, and radar sensors;

[0111] An optical preprocessing module, configured to use a preprocessing algorithm to perform denoising on the optical data to generate a first optical feature set;

[0112] An infrared correction module, configured to perform thermal imaging correction on infrared data to generate a first infrared feature set;

[0113] A radar filtering module is used to perform signal filtering on radar data to generate a first radar feature set and obtain a multi-source feature set;

[0114] A principal component analysis module is used to perform dimensionality reduction processing on the first optical feature set, the first infrared feature set, and the first radar feature set using a principal component analysis algorithm for the multi-source feature set to generate a second optical feature set, a second infrared feature set, and a second radar feature set to obtain a unified dimensional feature set;

[0115] A feature fusion module is used to fuse the second optical feature set, the second infrared feature set, and the second radar feature set based on the unified dimensional feature set using a deep belief network algorithm to generate a fused feature vector. The deep belief network extracts cross-modal features through a pre-trained restricted Boltzmann machine to obtain a fused feature set.

[0116] The feature selection module is used to select features from the fused feature set by using a recursive feature elimination algorithm if the dimension of the fused feature set is greater than the maximum allowed dimension, delete redundant features, and generate a streamlined feature set;

[0117] The target classification module is used to classify the target based on the reduced feature set using a convolutional neural network algorithm to generate target recognition results. The convolutional neural network extracts the spatial features of the target through multi-layer convolution and pooling to obtain preliminary recognition results.

[0118] The environment classification module is used to obtain the preliminary recognition results and the spatial span parameters of the current environment. It uses the support vector machine algorithm to classify urban occlusion environments and open area environments. If the spatial span parameters indicate an urban occlusion environment, the convolutional neural network's convolution kernel size is adjusted to S1 (S1 is a positive integer indicating a small convolution kernel);

[0119] The transfer learning module is used to reclassify the reduced feature set based on the adaptive recognition model using a transfer learning algorithm to generate the final recognition result. Transfer learning improves the model's adaptability to complex environments by fine-tuning the pre-trained model to obtain the target classification result.

[0120] The distributed fusion module is used to obtain the final recognition results of multiple drone nodes. It uses the consistent hashing algorithm to perform distributed fusion on the target classification results to generate collaborative perception results. The consistent hashing algorithm optimizes the spatial coverage by sharing data between nodes to obtain the global target recognition result.

[0121] The trajectory prediction module is used to predict the target motion trajectory based on the global target recognition results using the Kalman filter algorithm to generate the target trajectory prediction result. The Kalman filter updates the target position and velocity through the state transfer matrix to obtain the target dynamic information.

[0122] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for identifying low-altitude targets of unmanned aerial vehicles based on multi-source perception, characterized in that: include: Acquire multi-source perception data, preprocess and reduce the dimensionality of the multi-source perception data, and obtain a reduced-dimensionality feature set; The dimensionality reduction feature set is fused through the deep learning model to obtain the fused feature set; Perform dimension judgment on the fused feature set and obtain a streamlined feature set based on the judgment result; The simplified feature set is identified through a deep learning model to obtain preliminary identification results; spatial data is obtained, and regional classification is performed based on the preliminary identification results and spatial data to obtain environmental classification results; According to the environmental classification results, the simplified feature set is re-identified to obtain the target classification results. The target classification results of different UAVs are integrated, and the dynamic information of target recognition is obtained based on the integration results.

2. The method according to claim 1, characterized in that The process of obtaining the dimensionality reduction feature set includes: Preprocessing is performed on multi-source perception data, wherein the multi-source perception data includes optical data, infrared data, and radar data; the preprocessing includes denoising and calibration processing; and dimensionality reduction is performed on the multi-source perception data through principal component analysis to obtain a reduced-dimensionality feature set.

3. The method according to claim 1, characterized in that The process of obtaining the fusion feature set includes: A deep belief network is constructed, the restricted Boltzmann machine in the deep belief network is pre-trained, and the fusion feature set is fused through the trained deep belief network to obtain the fusion feature set.

4. The method according to claim 1, wherein The process of obtaining the simplified feature set includes: The dimensionality of the fused feature set is judged. When the dimension of the fused feature set is greater than a preset threshold, the dimensionality is reduced by a recursive feature elimination method. When the dimensionality reduction result of the recursive feature elimination method is greater than the preset threshold, redundant features are removed again by correlation analysis to obtain a streamlined feature set. Otherwise, the fused feature set is directly used as the streamlined feature set.

5. The method according to claim 1, characterized in that The process of obtaining environmental classification results includes: The spatial data and dimensionality reduction feature set are classified through the machine learning model, and the preliminary recognition results are input into the machine learning model as auxiliary information to obtain the environmental classification results.

6. The method according to claim 1, characterized in that The process of obtaining target classification results includes: According to the environmental classification results, an adaptive model is constructed, wherein the adaptive model is a deep learning model. During the construction of the adaptive model, the model parameters of the adaptive model are adjusted according to the corresponding relationship between the preset environmental classification results and the model parameters of the adaptive model; the simplified feature set is identified through the adaptive model to obtain the target classification result.

7. The method according to claim 6, characterized in that The adaptive model is trained by transfer learning.

8. The method according to claim 1, characterized in that The process of integrating the target classification results of different UAVs includes: The integrity of the target classification result is judged. When the target classification result is complete, the target classification result is distributed through the consistent hashing algorithm to obtain a data mapping table between different drones. According to the data mapping table, the target classification result is transmitted through the data sharing mechanism between drone nodes to obtain a coverage-optimized data set. The coverage rate of the coverage-optimized data set is judged. When the coverage rate is greater than a threshold, the target classification results identified by the drone nodes are integrated through a distributed fusion algorithm to obtain a preliminary collaborative perception result. The preliminary collaborative perception structure is subjected to consistency verification and global optimization to obtain the final global target recognition result, i.e., the integration result.

9. The method according to claim 1, characterized in that According to the integration results, the target motion trajectory in the integration results is predicted by the Kalman filtering method to obtain the dynamic information of target recognition.

10. A UAV low-altitude target recognition system based on multi-source perception, characterized in that: Used to execute the method according to any one of claims 1 to 9.

Images