Low-altitude unmanned aerial vehicle global management and control method and system based on image enhancement

Abstract

The application relates to the technical field of image detection, and discloses a low-altitude unmanned aerial vehicle global management and control method and system based on image enhancement, which comprises the following steps: acquiring continuous multiple frames of images of each node sensor, extracting a current frame for semantic decoupling, performing background suppression based on a semantic mask matrix to obtain a background suppression image; extracting an entity feature tensor, and calculating a three-dimensional space-time gradient tensor to extract an aerodynamic wake feature tensor; after the two tensors are spliced, the spliced tensors are input into a feature-level enhancement network which is updated end to end by a detection loss function, an enhanced feature tensor is generated, and a detection result is obtained by inputting the enhanced feature tensor into a target detection network; based on the detection result, a target feature signature is extracted to construct a dynamic topological graph, and cross-node matching is performed on a management and control platform to realize collaborative management and control. Through multi-domain feature object decoupling and space topological graph matching, the application effectively strips complex environment clutter and supplements air flow disturbance variables, and eliminates feature truncation, so that the cross-view matching breakage problem is solved.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, specifically to a method and system for the comprehensive control of low-altitude unmanned aerial vehicles (UAVs) based on image enhancement. Background Technology

[0002] With the deepening development and utilization of low-altitude airspace, micro and small unmanned aerial vehicle (UAV) systems are widely deployed in various fields, leading to a surge in demand for security and defense in low-altitude airspace. Visual sensors, due to their passive detection and high-resolution characteristics, occupy a core position in low-altitude target management networks. Low-altitude operating environments are generally filled with complex background textures such as urban buildings and vegetation canopies, and UAV targets account for a very low percentage of pixels in the image. Conventional visual detection methods rely on static rigid physical feature extraction, which is prone to visual confusion in dense environmental clutter. Furthermore, they fail to utilize non-rigid aerodynamic perturbation cues associated with rotor wind shear, making front-end signal extraction difficult.

[0003] To improve visual signal quality, conventional techniques typically introduce an image enhancement preprocessing module before the detection stage. In existing processing architectures, enhancement processing and detection networks are completely separate. Most pre-enhancement algorithm parameters are fixed and fail to establish a closed-loop feedback channel with the back-end target detection task. This deviates from the guidance and constraints of terminal classification and localization error loss. The pre-feature mapping process is prone to accidentally deleting weak high-frequency edges of the target while smoothing background noise, resulting in latent space information truncation and limiting the overall target detection accuracy of the system. When facing a swarm of multiple UAVs, the field of view of a single sensor is prone to occlusion blind spots. When existing systems introduce distributed sensor networks to perform global collaborative tasks, they generally rely on the absolute two-dimensional pixel coordinates of individual targets for cross-view matching. Affected by the installation tilt angle of sensors in different locations, lens distortion, and changes in target attitude, it is difficult to unify the isolated location point feature reference system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for the comprehensive control of low-altitude unmanned aerial vehicles (UAVs) based on image enhancement. This solves the problems of image detection difficulties caused by weak visual features of low-altitude micro-UAVs in complex environments, as well as the problem of fragmented target image matching in multi-source sensor cross-view collaborative detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for global control of low-altitude unmanned aerial vehicles based on image enhancement, comprising: Acquire multiple consecutive frames of raw image data collected by sensors at each node in a distributed sensor network, and take the latest frame among the multiple consecutive frames of raw image data as the current frame; Semantic decoupling is performed on the current frame to obtain the semantic mask matrix of the current frame, and background suppression is performed on the current frame based on the semantic mask matrix of the current frame to obtain a background suppressed image; The background suppression image is mapped to the latent space to extract entity feature tensors. At the same time, the three-dimensional spatiotemporal gradient tensor of the original image data of the consecutive multiple frames is calculated to extract aerodynamic wake feature tensors. The entity feature tensor and the aerodynamic wake feature tensor are concatenated in the channel dimension and then input into the feature-level enhancement network for latent space feature enhancement to generate enhanced feature tensors. The enhanced feature tensor is input into the target detection network to generate detection results; wherein, the model parameters of the feature-level enhancement network are obtained by joint updating of network parameters end-to-end through a task-driven detection loss function; Based on the detection results, the feature signature of the target is extracted, a dynamic topology map of each node sensor is constructed, and cross-node matching of the dynamic topology map is performed on the global control platform to achieve collaborative control of the target.

[0006] Specifically, in the method provided by the present invention, the step of acquiring raw image data of multiple consecutive frames adopts a spatiotemporal non-uniform sampling strategy: the system defines the background region and the target region according to the semantic mask matrix of the preceding frame; a lower first sampling frequency is applied to the background region and a higher second sampling frequency is applied to the target region, thereby reducing the overall data throughput burden of the system while ensuring the high-frequency dynamic capture capability of the target region.

[0007] Furthermore, this invention introduces a pixel-level physical isolation mechanism in the background suppression stage. The system uses a semantic mask matrix to perform bitwise masking operations to separate the foreground pixel region containing high-frequency details and the background pixel region containing environmental clutter. Low-pass filtering and smoothing are performed separately on the background pixel region, and then the smoothed background pixel region is re-overlaid and fused with the foreground pixel region. This physical isolation principle preserves the edge contour features of the UAV entity without loss, while greatly reducing the interference of local environmental noise.

[0008] Furthermore, in the feature extraction and enhancement stage, this invention innovatively employs a dual-domain feature fusion architecture. In addition to extracting entity feature tensors in the spatial domain... In addition, the system extracts aerodynamic wake feature tensors that characterize the physical disturbances of airflow by calculating the three-dimensional spatiotemporal gradient tensor. The feature-level enhancement network performs nonlinear mapping and channel dimensionality upscaling on the concatenated features, following the spatial mapping logic as follows: ; In the formula, The high-dimensional enhanced feature tensor generated is used to represent the characterization. The concatenated tensor representing the entity feature tensor and the aerodynamic wake feature tensor in the channel dimension; Characterizing the dimensionality-upgrading weight control matrix; Characterizes the bias vector; Characterizing nonlinear activation function operators; Characterizes tensor convolution multiplication. The... and The parameters are not preset by human experience, but are obtained by backpropagation gradient of the loss function of the detection task and are jointly updated end-to-end during the network training stage, ensuring that the latent space feature enhancement fully serves to improve the downstream target detection accuracy.

[0009] Furthermore, in the multi-view collaboration stage, this invention transforms isolated target location information into graph structure data with global topology awareness. The system performs spatial feature pooling on the enhanced feature tensor, extracting high-dimensional abstract semantic feature vectors as feature signatures (topology nodes) representing the target's identity; simultaneously, based on consecutive multi-frame prediction sequences, it calculates the target's center coordinates and inter-frame motion velocity, and uses the following kernel function to calculate the edge connection weights between topology nodes. : ; In the formula, A parameter representing the spatial distance between any two targets; A parameter characterizing the difference in motion velocity between the two targets; and These represent the distance scaling factor and the velocity scaling factor, respectively. Their values ​​are preset constants calibrated based on historical empirical data from multiple field flight tests. This dynamic topology graph construction mechanism performs high-dimensional mathematical quantification of the relative spatial configuration and motion consistency of multiple targets within the cluster.

[0010] Furthermore, the global control platform utilizes graph neural network operators to uniformly map the dynamic topology graphs uploaded by sensors at each node into fixed-dimensional graph embedding vectors and calculates graph structure similarity. When the similarity value is greater than or equal to a preset matching threshold, it is determined that the targets tracked by the remote sensors have macroscopic structural consistency, confirming them as the same cluster. Subsequently, cross-node location information is merged, and control and coordination instructions are output to surrounding devices.

[0011] A second aspect of the present invention provides an image-enhanced low-altitude unmanned aerial vehicle (UAV) global control system, comprising: Memory, used to store computer programs; The processor, coupled to the memory, is used to implement the image enhancement-based low-altitude UAV global control method described in the first aspect above when executing the computer program.

[0012] Specifically, the global control system supports the efficient operation of the aforementioned algorithms through a physical hardware architecture. The non-volatile region of the memory permanently stores pre-trained feature network parameters and preset thresholds, while the volatile region is used to dynamically allocate time-series cache queues to carry multi-frame tensor data. The processor has an embedded tensor parallel operation acceleration engine and a scalar computation core, which interact with the memory and communication network interface through the underlying system data bus to achieve end-to-end hardware-level closed-loop execution from the underlying image sequence input and latent space tensor operation to the top-level global collaborative control command output.

[0013] This invention provides a method and system for comprehensive control of low-altitude unmanned aerial vehicles (UAVs) based on image enhancement. It offers the following advantages: 1. This invention utilizes semantic masking to suppress background and combines it with three-dimensional spatiotemporal gradient extraction of aerodynamic wakes and entity features for splicing. Spatial isolation and dynamic wake capture are linked, environmental clutter is removed and airflow perturbation variables are supplemented, solving the detection obstacles caused by the small visual area and environmental confusion of low-altitude targets.

[0014] 2. This invention is based on end-to-end joint updating of feature enhancement and target detection network parameters using the detection loss function. The dimensionality enhancement of the preceding features is controlled by the feedback of the deviation of the subsequent tasks, eliminating the information truncation caused by the combination of independent modules, and forcibly aligning the latent space feature mapping direction with the terminal classification and localization requirements.

[0015] 3. This invention extracts target feature signatures to construct a dynamic topology graph determined by distance and velocity difference parameters, performs cross-node matching, transforms isolated two-dimensional location points into a spatially related network, solves the cluster matching disruption caused by changes in viewing angle of remote sensors, and completes the unification of multi-source coordinates. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the architecture and application scenarios of the low-altitude unmanned aerial vehicle (UAV) all-domain control system of the present invention; Figure 2 This is a flowchart of the overall control method for low-altitude unmanned aerial vehicles (UAVs) of the present invention. Figure 3 This is a logic block diagram of the spatiotemporal non-uniform sampling and background suppression module of the present invention; Figure 4 This is a schematic diagram of the feature extraction and feature-level enhancement network structure of the present invention; Figure 5 This is a schematic diagram illustrating the dynamic topology graph construction and cross-node matching principle of the present invention; Figure 6 This is a schematic diagram of the hardware physical structure of the global control system of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see the appendix Figure 1 To be continued Figure 6 This invention provides a method for the comprehensive control of low-altitude unmanned aerial vehicles (UAVs) based on image enhancement. This method is applied to a distributed sensor network containing multiple heterogeneous computing nodes and optical acquisition devices to achieve collaborative perception and cross-regional tracking of multiple targets in the low-altitude airspace.

[0019] See attached document Figure 1 The image enhancement-based method for comprehensive control of low-altitude unmanned aerial vehicles may include the following steps: Acquire multiple consecutive frames of raw image data collected by sensors at each node in a distributed sensor network, and take the latest frame as the current frame; Semantic decoupling is performed on the current frame to obtain a semantic mask matrix, and background suppression is performed on the current frame based on the semantic mask matrix to obtain a background-suppressed image; The background suppression image is mapped to the latent space to extract the entity feature tensor, and the three-dimensional spatiotemporal gradient tensor of the original image data of the consecutive frames is calculated to extract the aerodynamic wake feature tensor. After concatenating the entity feature tensor with the aerodynamic wake feature tensor, the concatenation is input into a feature-level enhancement network to perform latent space feature enhancement and generate an enhanced feature tensor. The enhanced feature tensor is input into the target detection network to generate detection results; wherein, the model parameters of the feature-level enhancement network and the target detection network are obtained in advance through joint updating of the task-driven detection loss function; Based on the detection results, the feature signature of the target is extracted, a dynamic topology map of each node sensor is constructed, and cross-node matching of the dynamic topology map is performed on the global control platform to achieve collaborative control of the target.

[0020] A distributed sensor network comprises multiple sensor nodes deployed in different geographical locations. Each sensor node connects to the underlying communication network and executes the IEEE 1588 precise time protocol, achieving millisecond-level clock alignment of the network's hardware timestamps. Driven by a unified clock pulse, all sensor nodes synchronously execute image acquisition tasks. The set of sensor nodes in a distributed sensor network is defined as... : in, Represents the first in the network Individual node sensors; This indicates the total number of node sensors. This parameter is calculated based on historical experience data from previous field surveys and the effective field of view of a single optical device, and is used to ensure blind-spot-free coverage of the target area.

[0021] Any node sensor The controlled airspace is optically scanned according to a preset initial frame rate, and multiple consecutive frames of raw image data are acquired to form an image set corresponding to this time segment. : in, This indicates the time index of the current image stream sample. Indicates node sensor At any moment The system will collect the latest frame of image data. Set as the current frame. Each frame's raw image data is represented in the underlying memory as a three-dimensional matrix data structure, whose dimensions are defined as follows: ;in, Represents the pixel height of the image. These two parameters, representing the pixel width of the image, are determined based on the physical pixel array specifications of the photosensitive chip inside the node sensor. This indicates the number of color channels in the image. When using conventional optical acquisition equipment, it is set to 3, corresponding to the three physical color channels: red, green, and blue.

[0022] The system reads the current frame and performs semantic decoupling, outputting a semantic mask matrix with spatial correspondence to distinguish target pixels from non-target pixels in the image. Based on the mask separation result, the system performs background suppression processing to eliminate useless high-frequency environmental noise interference in non-target areas of the image, generating a background-suppressed image and completing the initial image cleaning in the spatial domain.

[0023] The system transforms the processed background-suppressed image into a non-Euclidean geometric latent space using a feature extraction operator, generating an entity feature tensor containing geometric distribution information. The system processes multiple consecutive frames of raw image data in parallel. The inter-frame difference and gradient changes between multiple frames of data are extracted to generate a three-dimensional spatiotemporal gradient tensor, and based on this, the aerodynamic wake feature tensor reflecting the aerodynamic physical disturbance characteristics is extracted.

[0024] The system performs tensor concatenation operations along the channel dimension to establish a joint feature matrix of static geometric entities and dynamic airflow wakes, which is then input into a feature-level enhancement network. The feature-level enhancement network performs nonlinear mapping and dimensional transformation, outputting an enhanced feature tensor containing higher-level abstract semantics, which serves as direct input to downstream detection tasks. The target detection network receives this enhanced feature tensor as the root node input data of the feature pyramid structure. The target detection network uses a YOLO single-stage detection architecture to perform forward inference operations, outputting dense candidate bounding boxes through parallel distributed regression and classification detection heads. The system calls a non-maximum suppression algorithm to remove redundant detection boxes with excessive spatial overlap. The system sets the intersection-union ratio (IUGR) removal threshold to a range of 0.35 to 0.55; in this embodiment, it is preferably set to 0.45. After cleaning, the system outputs the two-dimensional spatial coordinates of each independent target and the confidence probability of its corresponding category.

[0025] Based on the output coordinates of the target detection network, the system extracts spatial local features from the enhanced feature tensor to generate feature signatures representing the independent identity attributes of each target. The system abstracts multiple targets within the field of view as nodes in a graph structure, transforming the physical spatial distance and relative motion differences between targets into edge connection weights in the graph structure, generating a structured dynamic topology graph. The global control platform receives the dynamic topology graph uploaded by sensors from all nodes within the network, executes the similarity calculation logic of high-dimensional graph embedding vectors, and achieves cross-physical node cluster matching and identification of the same target, issuing collaborative control commands.

[0026] This embodiment details the physical process and algorithm execution logic for acquiring multiple consecutive frames of raw image data and performing background suppression. The system introduces a spatially non-uniform exposure mechanism from the front end of the sensor control link, breaking the conventional pattern of globally uniform frame rate sampling. The system reads the semantic mask matrix of the preceding frame cached in the hardware register and uses it as a priori index matrix for dividing the spatial physical attributes of the field of view.

[0027] The semantic mask matrix of the preceding frame is set as follows: The system iterates through the matrix to extract the pixel-level spatial distribution state, using the formula... Extract the set of coordinates of the target region, and use the formula Extract the set of background region coordinates. The system generates adaptive sampling control commands and sends them to the sensor pixel addressing array to identify regions belonging to the background. The pixel photosensitive unit applies the first sampling frequency Perform long-cycle exposure sampling, while simultaneously targeting the area. The pixel photosensitive unit uses a second sampling frequency Activate event-driven high-frequency discrete sampling.

[0028] The system sets the numerical relationship between the first sampling frequency and the second sampling frequency as follows: The specific values ​​of the above two frequency parameters are calculated based on the maximum bandwidth throughput limit of the sensor hardware data bus. The system merges the above-mentioned physical photosensitive signals with asynchronous time-dimensional characteristics, splices them in the memory buffer layer to generate a spatiotemporally non-uniform sampled image sequence, and uses it as the raw image data of the continuous multi-frame input to the algorithm backend. In this step, the system realizes the physical separation of the high-frequency dynamic features of the moving target from the low-frequency steady-state features of the static background, effectively avoiding overload of the front-end data bus.

[0029] The system extracts the latest frame from the spatiotemporally non-uniformly sampled image sequence as the current frame. The result is then fed into a pre-trained semantic segmentation network based on a lightweight UNet architecture. This semantic segmentation network outputs a pixel-level forward probability distribution matrix with the same resolution as the input through an encoder-decoder structure. The system calls a threshold determination function, setting the activation threshold to a range of 0.4 to 0.6, preferably 0.5 in this embodiment. When the probability of a pixel belonging to a target is greater than or equal to 0.5, the system assigns a value of 1 to the corresponding element in the mask matrix for that pixel; otherwise, it assigns a value of 0. This generates a semantic mask matrix for the current frame containing clearly defined binary boundaries. The system uses this semantic mask matrix to perform a bitwise Hadamard product operation on the current frame, decoupling the pixel components of the image in the spatial domain. Based on the high-level and low-level indices of the mask matrix, the system separates the foreground pixel region of the current frame. and background pixel area : In the formula, Characterizes pixel coordinates within a two-dimensional plane of an image. Characterizes the bitwise multiplication operation of matrix elements. A constant matrix of all 1s with dimensions exactly the same as the mask matrix.

[0030] For the separated background pixel region, the system calls a two-dimensional convolution operator to perform low-pass filtering and smoothing processing to remove high-frequency disturbance noise in the background environment. The system uses a local neighborhood weighted average rule to calculate the smoothing features: In the formula, Characterizes the smoothed background pixel region in the output after processing. The sampling radius parameter characterizing the filter window The sampling radius parameter characterizes the filter window. To address the high-frequency clutter and illumination interference commonly found in low-altitude environments, the system defines the sampling radius parameter. The value of is an odd integer between 3 and 7; in this embodiment, the system preferably... To construct a size of Two-dimensional Gaussian filter kernel, The normalized low-pass filter kernel weight coefficients, generated by a two-dimensional Gaussian distribution function, are used to suppress local sharpness jumps in background pixel regions.

[0031] The system constructs a matrix adder to re-perform spatial fusion of the smoothed background pixel region (after filtering out high-frequency noise) and the foreground pixel region (preserving the original high-frequency details). The system then executes feature addition logic to generate a background-suppressed image. : Through the aforementioned mask separation and differential frequency domain filtering operations, the system constructs a spatially asymmetric signal-to-noise ratio distribution structure within the same frame of image, providing high-contrast underlying benchmark data for the extraction of entity and aerodynamic wake features in the latent space.

[0032] In this embodiment, the system performs latent space dual-domain feature extraction and network-level channel dimension upscaling to achieve low-level data fusion of the spatial geometric distribution matrix and the temporal airflow perturbation matrix. The system reads the background suppression image and establishes a forward data transmission channel, then calls the network feature extraction layer to map it into a latent space with a non-Euclidean manifold structure. The system uses a two-dimensional convolutional kernel matrix to perform a spatial domain sliding inner product operation along the image plane, outputting an entity feature tensor. In the formula symbols, The height spatial resolution parameter characterizing the feature tensor. The spatial resolution parameter characterizing the feature tensor width. The parameter representing the channel dimension occupied by the feature tensor of this entity.

[0033] The system is configured with parallel computing branches to extract spatiotemporal perturbation information within multiple consecutive frames of raw image data. The system applies first-order partial differential operators along the three-dimensional coordinate axes to calculate the inter-frame rate of change of image pixel intensity, generating a three-dimensional spatiotemporal gradient tensor characterizing the evolution of the spatial environment's flow field. This is for applications using two-dimensional planar coordinates... With time index Given a defined three-dimensional point, the system executes discrete difference equations to calculate the horizontal spatial gradient at the corresponding location. Vertical spatial gradient and time gradient : In the formula, The system represents the underlying pixel brightness values ​​of the original image data. It concatenates the aforementioned three sets of partial derivative matrices along the tensor channel dimension to construct a structure defined as follows: The three-dimensional spatiotemporal gradient tensor.

[0034] The system invokes a 3D spatiotemporal feature extraction operator to perform latent space feature mapping on the 3D spatiotemporal gradient tensor. This 3D spatiotemporal feature extraction operator comprises a continuous three-layer 3D convolutional network, and the system sets the built-in cross-spatiotemporal receptive field convolutional kernel size parameter to be... While preserving local spatial texture, the system aggregates temporal perturbations between frames to extract non-rigid target motion contour cues present in temporal gradient changes, generating a dynamic texture feature matrix characterizing airflow physical disturbances. In conventional optical images, airflow physical disturbances manifest as high-frequency physical displacements of the surrounding lightweight background (such as vegetation leaves, grass clippings, and dust) caused by the high-speed wind shear of the UAV rotor, and optical refraction distortions (thermal scintillation effect) caused by changes in local airflow density and temperature gradients. The 3D spatiotemporal feature extraction operator is used to capture these weak non-rigid dynamic pixel displacement cues. The system outputs this matrix as an aerodynamic wake feature tensor. The system configuration output will include channel dimension parameters specifically designed to characterize dynamic airflow features. It is explicitly set to 64.

[0035] The system issues a tensor merging command, arranging entity feature tensors along the channel dimension. With aerodynamic wake feature tensor Generate an overall dimension size of The system inputs the combined composite data tensor into the input interface of the feature-level augmentation network. Utilizing the pointwise transformation matrix set and nonlinear operator set within the feature-level augmentation network, the system performs space polynomial mapping and channel dimensionality upscaling on the composite data tensor. The system control unit executes the following activation and mapping equations: In the formula, The enhanced feature tensor generated after characterization processing The up-dimensional weight control matrix representing the feature-level augmentation network is instantiated in the underlying code as follows. Pointwise convolution operators are used to achieve cross-channel information interaction. The bias vector matrix representing the offset used to deflect the feature centroid. Activation function operators that characterize nonlinear mapping properties. This represents the multidimensional tensor convolution multiplication operation. In this embodiment, the system sets the entity feature channel dimension. And the aerodynamic wake channel dimension ,symbol The system output channel quantity parameter represents the number of output channels after dimensional expansion; the system forces allocation. satisfy The inequality constraints are used to provide the capacity of the tensor space to accommodate richer semantic features.

[0036] The system completes the network channel dimensionality upscaling and feature space reconstruction tasks, and extracts the enhanced feature tensor. The data is then directly transmitted to the input node of the target detection network via a data bus. The system uses this enhanced feature tensor as the underlying data source for the bounding box coordinate prediction logic and the target category probability determination logic, establishing a data flow link from the preceding image feature extraction node to the target attribute determination node.

[0037] In this embodiment, the system performs a joint parameter training process for the feature-level enhancement network and the object detection network in an offline environment to solidify the network weight topology with optimal recognition performance. The system reads training set data containing more than 50,000 images covering various low-altitude meteorological environments, and obtains the original image sequence including manual prior verification and the corresponding real annotation data. According to the internal data structure format specifications, the system splits the real annotation data into a set of real category labels containing target classification and attribution information. and the set of true bounding box coordinates containing the physical space coordinates of the target. The system feeds the training set image sequences into the established data flow link between the feature-level enhancement network and the object detection network, performs cascaded forward inference operations, and outputs the forward inference detection results corresponding to the current training batch. The forward inference detection results encompass the system's predicted bounding boxes for multiple targets generated by target spatial localization. And the set of category confidence scores output by the system for the target category attribute. .

[0038] The system extracts the aforementioned forward inference detection results and ground truth labeled data to perform error quantification and evaluation. For the multi-object classification and recognition task, the system constructs a classification cross-entropy loss function to quantify the relative information entropy difference in the category prediction space. The system iterates through the sample data features within a batch and applies the following equation to calculate the classification cross-entropy loss. : In the formula, The total number of target categories pre-defined by the characterization system; The representation exists in the set of real category labels The category indicator variable parameter within the index The variable takes the value of 1 when it points to a true category sequence, and takes the value of 0 when it points to other non-true category sequences; The representation exists in the class confidence set The inner, attributed to the classifier output by the last layer of the object detection network. The predicted probability value of the class.

[0039] For the target spatial geometric dimension localization task, the system establishes a bounding box regression loss function to measure the spatial coordinate deviation between the network's predicted boundaries and the actual physical boundaries. The system uses the bounding box prediction set... Extract the prediction box coordinate vector and from the set of true bounding box coordinates Extract the corresponding ground truth bounding box coordinate vector The system executes two sets of spatial vector spacing metrics and region overlap calculation metrics to generate bounding box regression loss. : In the formula, A parameter representing the total number of target entities contained within the current batch of image frames; The norm operator characterizes the first-order Manhattan distance for each component dimension of the coordinate vector; The spatial operation of the intersection-union function represents the intersection-union function of the predicted bounding box and the true bounding box in a two-dimensional plane coordinate system.

[0040] The system uses a linear superposition and fusion mechanism to combine the two evaluation error terms mentioned above, constructing a complete task-driven detection loss function. As a benchmark evaluation indicator for the overall optimization direction of the network: In the formula, The weight parameters characterize the contribution of the gradient to the backpropagation of the control classification error term. The weight parameters characterize the contribution of the gradient backpropagation to the bounding box regression error term; the system reads historical initialization data of each independent loss value and calculates that the magnitude of the bounding box regression loss is greater than that of the classification cross-entropy loss, and sets the parameters accordingly. The value range is from 0.8 to 1.2, set The value range is from 0.3 to 0.5, and... The value is 1.0 and A value of 0.4 is set as the fixed system configuration to balance the magnitude difference in loss values ​​for different tasks and maintain a balanced update amplitude of the optimization vector.

[0041] The system initiates a backpropagation optimization path based on partial differential operations, and uses the constructed task-driven detection loss function. A continuous backpropagation gradient flow is generated. This flow sequentially passes through the feedforward computation graph structure of the object detection network and seamlessly extends to the latent space data input of the feature-level augmentation network. The system calls the adaptive moment estimation network optimization operator to capture this backpropagation gradient flow signal, synchronously updating the up-dimensional feature mapping matrix parameters of the feature-level augmentation network and the node connection weight parameters of the object detection network end-to-end. The system configures the network parameter training environment and calls the AdamW optimizer as the gradient descent driving operator. The system sets the base learning rate parameter to a range of 5 × 10⁻⁶. -5 Up to 5×10 -4 Preferably 1×10 -4 The training batch size parameter is set to a value range of 8 to 32, preferably 16. The system performs forward inference data delivery and backward gradient optimization operations cyclically based on the preset maximum number of iterations (100 training rounds) or the error convergence threshold parameter (loss value decay of less than 0.01% for ten consecutive rounds), completes the joint update iteration action, and outputs the model combination with fixed parameters.

[0042] In this embodiment, the system extracts the enhanced feature tensor and target detection results from the preceding process, and performs a spatial feature pruning mechanism based on the target location information contained in the detection results. The system calls the region of interest aligned spatial feature pooling operator, uses bilinear interpolation to avoid quantization errors, and consistently maps the local feature regions corresponding to the physical bounding boxes within the enhanced feature tensor to a feature matrix with a spatial resolution of 7×7. The system performs feature flattening operation using the tensor flattening operator, and then processes the feature matrix using linear dimensionality reduction logic composed of fully connected layers, compressing it into a high-dimensional abstract semantic feature vector with a fixed dimension of 256, and defining it as the feature signature representing the unique identity of the target.

[0043] The system reads the feature signatures of each target generated within the field of view of the same node's sensor and uses them as topological nodes to construct the graph structure data space; the system retrieves multi-frame consecutive bounding box prediction sequences stored in the temporal buffer queue and extracts the first... The goal is currently The system continuously detects the 2D pixel scale parameters of the bounding box output, and calculates the center coordinates of each target's spatial contour accordingly. In the formula, Characterizing the first The goal is currently The two-dimensional center coordinate vector of the time frame, and The variable representing the coordinates of the top-left corner pixel of the bounding box extracted from the bounding box prediction sequence. and The bounding box pixel width and height variables extracted from the bounding box prediction sequence are characterized. The system derives the inter-frame motion velocity vector of the target in physical space by combining the temporal sampling interval parameter. In the formula, Characterizing the first Inter-frame motion velocity vectors of each target The sampling time interval parameter characterizes the underlying sensor hardware. This parameter is calculated based on the data read from the clock cycle register of the system's image acquisition interface.

[0044] The system iterates through the target set within the current frame, using norm operations to calculate the spatial distance and velocity difference parameters between any two targets. The system then performs Euclidean norm calculations to obtain the spatial distance parameters. The system uses the same computational logic to obtain the motion speed difference parameters. The system calls the exponentially decaying kernel function to perform numerical normalization on the two physical space difference parameters mentioned above, and calculates the edge connection weights between topological nodes: In the formula, Characterizing the first The target node and the first The edge connection weight values ​​between each target node. and The distance and velocity scaling factors characterize the convergence interval of the exponential function. The system sets the distance scaling factor. The value ranges from 40.0 to 60.0, preferably 50.0; and a speed scale scaling factor is set. The value range is 5.0 to 15.0, preferably 10.0; the system generates the dynamic topology graph representing the relative spatial relationship of multiple targets in the region based on the feature signature node containing the identity attributes of each target and the generated edge connection weight parameter; the system generates the dynamic topology graph representing the relative spatial relationship of multiple targets in the region based on the feature signature node containing the identity attributes of each target and the generated edge connection weight parameter.

[0045] The system establishes an underlying communication link to synchronously upload the dynamic topology map generated by the sensors at each node to the overall control platform. The platform then uses graph convolution feature aggregation operators to perform node feature transfer and aggregation operations on the graph data. The platform extracts node feature matrices containing the signatures of their respective targets. and the adjacency matrix containing the edge connection weights Using equations Perform forward aggregation propagation, where The degree matrix representing the adjacency matrix. Characterizing the first The platform employs a multi-weighted self-weighted matrix for each layer. After multi-layer feature aggregation and propagation, it performs average pooling on the output global features, mapping the graph data containing relative position information of multiple targets into graph embedding vectors of fixed dimension length. The platform then uses the dot product similarity equation to calculate the graph embedding vectors originating from sensors at different nodes. and Graph structure similarity between them: In the formula, The macroscopic structural matching value characterizes two sets of local dynamic topology graphs, reflecting the consistency of observation topology of UAV swarms by sensors from different locations within the same airspace. The platform uses the dot product similarity equation to calculate graph embedding vectors originating from sensors at different nodes. and The system reads the pre-programmed matching threshold parameter from the internal storage area. This matching threshold parameter ranges from 0.75 to 0.90. In this embodiment, the system explicitly sets this matching threshold parameter to 0.85. When the calculated graph structure similarity is greater than or equal to the matching threshold of 0.85, the platform determines that the target cluster tracked in the field of view of different node sensors is the same UAV cluster in physical space. The platform reads the intrinsic parameter matrix of each node sensor camera and the extrinsic parameter matrix containing position and attitude information, which are pre-generated using the Zhang Zhengyou calibration method. The platform calls the homography perspective transformation algorithm, executes the multi-source coordinate system-to-conversion logic, and uniformly maps the two-dimensional pixel coordinates of each independent sensor to the WGS-84 geographic coordinate system. It merges cross-node position information, generates continuous track coordinates with absolute latitude, longitude, and altitude under the global perspective, and directly outputs control and coordination commands to surrounding defense equipment to implement interception and countermeasure actions.

[0046] In this embodiment, the present invention provides a low-altitude UAV global control system based on image enhancement, and provides a physical hardware entity to carry the underlying algorithm logic of the aforementioned global control method. The system includes a computing architecture motherboard and a memory and processor deployed on the motherboard. The memory and processor establish electrical coupling and data communication links using a system data bus. The system divides the memory into a non-volatile storage area and a volatile cache area. The system persistently writes computer program instruction code, model parameter matrices of feature-level enhancement networks and target detection networks, and preset matching thresholds and kernel function weight configurations to the non-volatile storage area. During the operating cycle, the system dynamically allocates physical memory addresses for the volatile cache area to temporarily store multiple consecutive frames of raw image data, feature tensors of various dimensions, and graph structure node sequence data.

[0047] The processor integrates a scalar computation core and a tensor parallel computation acceleration engine. The processor reads computer program instructions from memory via the system data bus, performs instruction decoding, and issues micro-instruction control flow to each hardware processing unit. The processor controls the tensor parallel computation acceleration engine to retrieve raw image data and model parameter matrices, performing semantic decoupling, spatiotemporal gradient difference calculation, latent space feature fusion and dimensionality upscaling, and bounding box regression coordinate calculation. For large-scale matrix multiplication and addition operations within the network model, the processor decomposes the tensor operation matrix and maps it to parallel computation thread flows supported by the underlying hardware.

[0048] The processor allocates underlying computing core resources according to its internal instruction set scheduling logic. The system sets the total floating-point operations required by the feature-level augmentation network during a single forward inference process to be [amount missing]. Set the number of parallel computing cores activated by the processor as follows: The physical operating frequency parameters maintained by each computing core are set as follows: The physical computational delay time of the processor performing latent space feature extraction and target detection on a single frame image. Follow the following time mapping equation: In the formula, The system calculates a loss factor characterizing the efficiency of parallel hardware computing by reading historical performance test data of the underlying hardware instruction set architecture under full pipeline load. This loss factor is used to verify the processing time of a single frame of image, ensuring that the physical computation latency is within acceptable limits. The value is strictly smaller than the sampling time interval parameter of the aforementioned sensor underlying hardware. This is to maintain the real-time throughput response capability of the overall management platform in handling multiple concurrent image streams and to prevent data overflow anomalies in the volatile buffer area.

[0049] In this embodiment, the present invention provides a computer-readable storage medium, which serves as a non-transient physical information carrier independent of the control system and stores the aforementioned computer program instruction set. An external computing device reads the machine code file stored within the computer-readable storage medium, allocates system memory and computing thread resources under the kernel scheduling of the device's operating system, and executes the entire logical process of acquiring non-uniformly sampled images, stitching entity and aerodynamic wake tensors, and constructing and matching a multi-target dynamic topology map. This computer-readable storage medium can be a magnetic storage medium, an optical storage medium, or a solid-state flash memory chip. System development and maintenance personnel utilize these media to implement the installation, deployment, and code version iteration distribution of the underlying control algorithm program of the present invention across physical hardware platforms.

Claims

1. A method for comprehensive control of low-altitude unmanned aerial vehicles (UAVs) based on image enhancement, characterized in that: Includes the following steps: Acquire multiple consecutive frames of raw image data collected by sensors at each node in a distributed sensor network, and take the latest frame among the multiple consecutive frames of raw image data as the current frame; Semantic decoupling is performed on the current frame to obtain the semantic mask matrix of the current frame, and background suppression is performed on the current frame based on the semantic mask matrix of the current frame to obtain a background suppressed image; The background suppression image is mapped to the latent space to extract entity feature tensors. At the same time, the three-dimensional spatiotemporal gradient tensor of the original image data of the consecutive multiple frames is calculated to extract aerodynamic wake feature tensors. The entity feature tensor and the aerodynamic wake feature tensor are concatenated in the channel dimension and then input into the feature-level enhancement network for latent space feature enhancement to generate enhanced feature tensors. The enhanced feature tensor is input into the target detection network to generate detection results; wherein, the model parameters of the feature-level enhancement network are obtained by joint updating of network parameters end-to-end through a task-driven detection loss function; Based on the detection results, the feature signature of the target is extracted, a dynamic topology map of each node sensor is constructed, and cross-node matching of the dynamic topology map is performed on the global control platform to achieve collaborative control of the target.

2. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The acquisition of raw image data across multiple consecutive frames includes: The background region and the target region are divided according to the semantic mask matrix of the preceding frame; The background region and the target region are sampled using a first sampling frequency and a second sampling frequency, respectively, to generate a spatiotemporally non-uniformly sampled image sequence as the original image data; wherein the second sampling frequency is higher than the first sampling frequency.

3. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The background suppression of the current frame based on the semantic mask matrix includes: The semantic mask matrix is ​​used to perform bitwise masking operations to separate the foreground pixel region and the background pixel region of the current frame; The background pixel region is smoothed by low-pass filtering to obtain a smoothed background pixel region, which is then fused with the foreground pixel region to generate the background suppressed image.

4. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The extraction of the aerodynamic wake feature tensor includes: Calculate the spatial and temporal gradients of the original image data, and construct the three-dimensional spatiotemporal gradient tensor; The three-dimensional spatiotemporal gradient tensor is feature-mapped by a three-dimensional spatiotemporal feature extraction operator to extract dynamic texture features that characterize the physical disturbance of airflow, which are then used as the aerodynamic wake feature tensor.

5. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The generation of the enhanced feature tensor includes: The feature-level enhancement network performs nonlinear mapping and channel dimensionality upscaling on the spliced ​​features to generate the enhanced feature tensor, which is then directly input into the target detection network.

6. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The model parameters are pre-updated through the following training process: Based on the forward inference detection results of the training set data and the corresponding real labeled data, an objective function including classification cross-entropy loss and bounding box regression loss is calculated as the task-driven detection loss function. Based on the backpropagation gradient of the detection loss function, the parameters of the feature-level enhancement network and the target detection network are jointly updated end-to-end.

7. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The extracted feature signature of the target includes: Based on the target location information in the detection results, spatial feature pooling is performed on the enhanced feature tensor to extract a high-dimensional abstract semantic feature vector as the feature signature representing the target identity.

8. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The construction of the dynamic topology graph includes: The feature signatures of each target within the field of view of the same node sensor are used as topological nodes; The center coordinates and inter-frame motion velocity of each target are calculated based on a multi-frame continuous bounding box prediction sequence, and the spatial distance parameter and motion velocity difference parameter between any two targets are calculated respectively. The edge connection weights between the topological nodes are calculated using the spatial distance parameter and the motion speed difference parameter to generate the dynamic topology graph.

9. The image enhancement-based method for full-domain control of low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The cross-node matching includes: The dynamic topology graph uploaded by each node sensor is mapped into a graph embedding vector, and the graph structure similarity is calculated. When the similarity of the graph structure is greater than or equal to the preset matching threshold, it is determined that the target tracked by the sensors of different nodes is the same cluster, the cross-node location information is merged and the control and coordination command is output.

10. A low-altitude unmanned aerial vehicle (UAV) global control system based on image enhancement, characterized in that: include: Memory, used to store computer programs; A processor, coupled to the memory, is configured to implement, when executing the computer program, the image enhancement-based low-altitude unmanned aerial vehicle (UAV) global control method as described in any one of claims 1 to 9.

Images