A battlefield target intelligent identification and mapping system and method based on deep learning
By using deep learning technology to extract multimodal features and dynamically adjust recognition paths in battlefield target identification, the problem of recognition efficiency and accuracy in complex battlefield environments is solved, achieving high-precision target identification and adaptive optimization, which is applicable to various types of battlefield target identification tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DONGZHI QIANWEI TECHNOLOGY CO LTD
- Filing Date
- 2025-06-30
- Publication Date
- 2026-05-29
AI Technical Summary
Existing battlefield target identification methods struggle to achieve high-precision identification in complex and ever-changing tactical scenarios. They lack detailed identification of battlefield types and cannot adapt to environments with diverse terrains, diverse targets, and dynamic changes, leading to decreased identification efficiency and accuracy. Furthermore, they lack continuous modeling of target movement trajectories and effective feedback mechanisms, failing to meet the identification needs under multi-task concurrency.
By employing a deep learning-based intelligent battlefield target identification and mapping method, battlefield image data is collected for multimodal feature extraction and scene recognition. The recognition path is dynamically adjusted to perform dynamic intelligent target recognition and dense target recognition. A battlefield simulation scene is constructed and accurately labeled. The method supports the identification of unknown targets and database updates, and achieves adaptive optimization and multi-layer target nesting.
It improves the adaptability and accuracy of target identification tasks in complex battlefield environments, realizes real-time prediction of multi-target trajectories and complete identification of high-density areas, enhances identification efficiency and closed-loop control capabilities, has self-learning capabilities and equipment identification expansion capabilities, and is suitable for various types of battlefield target identification tasks.
Smart Images
Figure CN120833468B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and more specifically, to a deep learning-based intelligent battlefield target identification and mapping system and method. Background Technology
[0002] With the rapid development of intelligent battlefield perception and command and control systems, parsing multi-source battlefield data has become one of the core capabilities of informationized warfare in modern combat scenarios. Especially in complex tactical scenarios with dense dynamic targets and ever-changing environments, how to quickly complete target detection and real-time labeling has become an important direction for improving real-time combat response capabilities and battlefield situational understanding. However, achieving high-precision battlefield target recognition in the traditional target recognition process remains challenging.
[0003] In existing technologies, battlefield target recognition largely relies on fixed rules or a single model architecture, lacking detailed recognition of battlefield scene types. This makes it difficult to adapt to complex battlefield environments with diverse terrain, varied targets, and dynamic changes, such as urban warfare, jungle operations, and high-altitude maneuvers, resulting in a significant decrease in target recognition efficiency and accuracy. Most existing recognition methods employ static deep learning structures, unable to dynamically adjust the recognition direction and adapt the model structure based on battlefield type labels, leading to low recognition efficiency and uneven task load, failing to meet the recognition requirements under multi-task concurrency. Furthermore, most existing methods are limited to static calibration result output, lacking the ability to continuously model target motion trajectories and adapt to various scenarios. Coordinate mapping in spatial simulation environments can lead to a disconnect between target information and spatial layout. Furthermore, when targets in battlefield images are occluded or overlap due to excessive density, the lack of detailed target region analysis results in frequent detection errors or target loss in existing recognition methods, failing to meet the high-precision requirements of tactical operations. Additionally, traditional recognition methods cannot effectively identify target types, such as new equipment not included in the database, posing a risk of missing typical targets. Moreover, existing target recognition methods lack effective feedback mechanisms, failing to adaptively optimize the target recognition structure based on recognition results, leading to long-term performance lag and hindering the evolution and updating of the overall recognition system.
[0004] In view of this, the present invention proposes a battlefield target intelligent identification and mapping system and method based on deep learning to solve the above problems. Summary of the Invention
[0005] To overcome the aforementioned shortcomings of existing technologies and achieve the above objectives, this invention provides the following technical solution: a battlefield target intelligent identification and mapping method based on deep learning, comprising:
[0006] S1. Collect battlefield image data and perform data cleaning to obtain enhanced battlefield image data;
[0007] S2. Perform multimodal feature extraction on the battlefield augmented image data to obtain a scene feature vector set; perform scene recognition on the battlefield augmented image data based on the scene feature vector set, and output battlefield type labels;
[0008] S3. Based on the battlefield type label, the battlefield enhanced image data is scheduled to obtain the scheduled image data and sent to the intelligent recognition path built based on the deep learning network to execute the target recognition process;
[0009] S4. Perform dynamic target intelligent recognition on the scheduling image data and output dynamic target trajectory data; perform dense target recognition on the scheduling image data and output high-density battlefield target detection data; perform auxiliary discrimination on the scheduling image data and update the preset database based on the discrimination results;
[0010] S5. Obtain battlefield spatial information corresponding to the battlefield augmented image data, and construct a battlefield simulation scene based on the battlefield spatial information; accurately annotate the battlefield targets in the battlefield simulation scene to obtain battlefield target visualization plotting data;
[0011] S6. Update the intelligent identification path to the optimized processing path and send it to the preset database for storage.
[0012] Furthermore, the method for scene recognition of the battlefield augmented image data includes:
[0013] Extract discriminative feature vectors and perform battlefield environment discrimination on battlefield augmented image data, outputting battlefield type labels.
[0014] Furthermore, the method for scheduling battlefield augmented imagery data based on battlefield type labels includes:
[0015] Based on the battlefield type label, generate scene identifiers corresponding to the battlefield type and establish a mapping relationship between scene identifiers and battlefield type labels; divide the intelligent recognition path of the preset control terminal based on the battlefield type label; deploy a deep learning network model in each intelligent recognition path and collect historical battlefield enhanced image data of the battlefield type label corresponding to the intelligent recognition path to train the deep learning network model; activate the corresponding intelligent recognition path using the scene identifier, and at the same time use the battlefield enhanced image data as the input data of the intelligent recognition path.
[0016] Furthermore, the method for dynamic target intelligent recognition of the scheduled image data includes:
[0017] The process involves converting scheduling image data into a battlefield image frame sequence; acquiring basic feature information of battlefield targets; performing preliminary screening of the battlefield image frame sequence to output a target candidate set; and selecting battlefield targets from the target candidate set while outputting the target category probability.
[0018] Battlefield targets are selected from the battlefield image frame sequence and their position change information is obtained. Based on the position change information, the future trajectory of the battlefield targets is predicted to obtain the future frame prediction path of the battlefield targets. Based on the future frame prediction path, the error of the current battlefield target's identified position change is corrected, and the dynamic target trajectory data of the battlefield targets is output.
[0019] Furthermore, the method for performing dense target recognition on the scheduled image data includes:
[0020] Deep features are extracted from each frame in the battlefield image frame sequence to obtain deep features of the frame images; the target distribution density of each frame image is calculated based on the deep features of the frame images; high-density target areas are determined based on the target distribution density; initial detection points for candidate targets are set and feature perception is performed on the initial detection points of candidate targets to output feature perception data; the feature perception data is matched and judged with the basic feature information of battlefield targets and battlefield target bounding boxes are selected.
[0021] Each detection box retains battlefield targets that appear consecutively in neighboring frames, and performs continuous frame recognition on occluded targets; obtains the positional changes of occluded targets; determines the complete outline of occluded targets based on positional changes and basic feature information of battlefield targets, performs battlefield target box selection, and generates target category probabilities; and constructs high-density battlefield target detection data based on the recognized battlefield image frame sequence.
[0022] Furthermore, the method for performing auxiliary discrimination on the scheduled image data includes:
[0023] The process involves determining the location of an unknown target and extracting its pixel features to generate an unknown target feature vector; obtaining the feature vectors of known targets corresponding to known battlefield targets, and calculating the feature distance between the unknown target feature vector and the known target feature vector; calculating the confidence score and setting a confidence score matching threshold; marking the unknown target as an associated target when the confidence score is greater than or equal to the confidence score matching threshold, and otherwise marking it as a typical unknown target; extracting the feature information of typical unknown targets as new entries in a preset database, and merging the feature information of associated targets with existing entries in the preset database.
[0024] Furthermore, the method for constructing the battlefield simulation scenario includes:
[0025] Extract pixel change features and construct the change trajectory of the region based on the pixel change features; quantify the change amplitude of the change trajectory, set the change amplitude threshold, and filter the spatial reconstruction region; merge the timestamp and the spatial reconstruction region to output the battlefield environment change layer; construct the battlefield environment spatial model based on the battlefield spatial information and perform time-series matching with the battlefield environment change layer to generate a battlefield simulation scene.
[0026] Furthermore, the method for accurately labeling battlefield targets in the battlefield simulation scenario includes:
[0027] The process involves: acquiring the image frame index of each battlefield target; constructing a top-level battlefield target mapping layer; mapping the image coordinates corresponding to the latest and initial timestamps of each battlefield target to the top-level battlefield target mapping layer based on the image frame index; constructing a battlefield target motion layer and dividing it into layers to generate battlefield target motion sub-layers; mapping dynamic target trajectory data to the corresponding battlefield target motion sub-layers; constructing a battlefield dense area layer and annotating the density of the battlefield dense area layer based on high-density battlefield target detection data; generating a structured plotting layer; nesting the structured plotting layer and the battlefield simulation scene to generate battlefield target visualization plotting data.
[0028] Furthermore, the method for obtaining the optimized processing path includes:
[0029] Real-time feedback data is collected, and the parameters of the intelligent recognition path are optimized based on the feedback data to obtain an optimized processing path.
[0030] A deep learning-based intelligent battlefield target identification and mapping system, used to implement a deep learning-based intelligent battlefield target identification and mapping method, is characterized by comprising:
[0031] The data acquisition module is used to collect battlefield image data and perform data cleaning to obtain enhanced battlefield image data.
[0032] The scene recognition module is used to extract multimodal features from battlefield augmented image data to obtain a scene feature vector set; based on the scene feature vector set, it performs scene recognition on the battlefield augmented image data and outputs battlefield type labels.
[0033] The path construction module schedules battlefield augmented image data based on battlefield type labels, obtains the scheduled image data, and sends it to the intelligent recognition path built based on deep learning network to execute the target recognition process.
[0034] The fine-grained recognition module is used to perform dynamic target intelligent recognition on scheduling image data and output dynamic target trajectory data; to perform dense target recognition on scheduling image data and output high-density battlefield target detection data; and to perform auxiliary discrimination on scheduling image data and update the preset database based on the discrimination results.
[0035] The visualization plotting module is used to acquire battlefield spatial information corresponding to battlefield augmented image data, construct battlefield simulation scenarios based on battlefield spatial information, and accurately annotate battlefield targets in the battlefield simulation scenarios to obtain battlefield target visualization plotting data.
[0036] The parameter optimization module is used to update the intelligent identification path to an optimized processing path and send it to a preset database for storage; the modules are connected to each other via wired and / or wireless means.
[0037] The technical effects and advantages of the intelligent battlefield target identification and mapping system and method based on deep learning in this invention are as follows:
[0038] Accurate target labeling was achieved by performing scene recognition, path scheduling, trajectory prediction, dense target processing, unknown target assisted recognition, and scene construction on collected battlefield image data. Target recognition performance was improved by optimizing path parameters. Intelligent recognition paths based on scene labeling were constructed, and dedicated recognition models were dynamically loaded according to battlefield type, significantly improving the task adaptability and accuracy of target recognition in complex battlefield environments. Real-time prediction of multi-target trajectories was achieved by fusing inter-frame detection results with pre-trained temporal models, and target localization accuracy in continuous time was enhanced through error correction. In high-density area recognition, candidate detection point mechanisms and occlusion completion were integrated. The strategy ensures the integrity and stability of target identification in scenarios with dense targets and obscuring interference; it also supports the identification and judgment of unknown targets and dynamic expansion and updating of the database, possessing good self-learning capabilities and equipment identification expansion capabilities; in addition, in the simulation scene construction and target annotation stages, multi-layer target nesting is realized based on a three-dimensional coordinate system, improving visualization clarity and multi-layer controllability; in terms of feedback optimization path scheduling, the identification path parameters are reconstructed based on time-segmented error feedback, effectively improving identification efficiency and closed-loop control capabilities; it has extremely high practical deployment value, is suitable for target identification tasks in various types of battlefields, and helps to improve the execution efficiency of dynamic target management and intelligent command systems. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of a battlefield target intelligent identification and mapping method based on deep learning according to the present invention;
[0040] Figure 2 This is a schematic diagram of a battlefield target intelligent identification and mapping system based on deep learning according to the present invention. Detailed Implementation
[0041] The technical solutions of 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.
[0042] Example 1
[0043] Please see Figure 1 As shown in this embodiment, a battlefield target intelligent identification and mapping method based on deep learning includes:
[0044] S1. Collect battlefield image data and video data to obtain battlefield image data; perform data cleaning and sharpness enhancement on the battlefield image data to obtain enhanced battlefield image data;
[0045] S2. Perform multimodal feature extraction on the battlefield augmented image data to obtain a scene feature vector set; perform scene recognition on the battlefield augmented image data based on the scene feature vector set, and output battlefield type labels;
[0046] S3. Based on the battlefield type label, perform data scheduling on the battlefield enhanced image data, obtain the scheduled image data, and send it to the intelligent recognition path based on the deep learning network of the corresponding battlefield type to execute the target recognition process;
[0047] S4. Perform dynamic target intelligent recognition on the scheduling image data and output dynamic target trajectory data; perform dense target recognition on the scheduling image data and output high-density battlefield target detection data; perform auxiliary discrimination on the scheduling image data and update the preset database based on the discrimination results;
[0048] S5. Obtain battlefield spatial information corresponding to the battlefield augmented image data, and construct a battlefield simulation scene based on the battlefield spatial information; accurately annotate the battlefield targets in the battlefield simulation scene to obtain battlefield target visualization plotting data;
[0049] S6. Update the intelligent identification path to the optimized processing path and send it to the preset database for storage.
[0050] Battlefield imagery data was obtained by capturing battlefield images and videos from multiple angles using drones. The data was then cleaned by denoising and removing invalid frames. Clarity enhancement was achieved by improving image contrast, edge sharpness, and super-resolution reconstruction, resulting in clearer battlefield enhanced imagery data, which is convenient for subsequent processing.
[0051] Methods for scene recognition of battlefield augmented imagery data include:
[0052] Extract discriminative feature vectors and perform battlefield environment discrimination on battlefield augmented image data, outputting battlefield type labels.
[0053] The scene feature vector set is formatted and encoded to obtain a scene feature encoding dataset. In this embodiment, features are extracted from battlefield enhanced image data through a multimodal feature extraction channel. The obtained scene feature vector set includes features such as color features, texture features, and spatial features of the current scene. By normalizing the obtained scene feature vectors and encoding them using principal component analysis algorithm, these encodings are mapped to a unified vector space, eliminating inconsistencies in structural dimensions and numerical distribution, and laying a data foundation for subsequent processing.
[0054] A feature association relationship is established for all scene feature codes in the scene feature coding dataset to obtain a feature association coding matrix. The cosine similarity between scene feature codes is calculated and used as an association discrimination index. A similarity threshold is set, and scene feature codes with a cosine similarity greater than the similarity threshold are combined to form an association unit. All association units are used to form a feature association coding matrix. The feature association coding matrix is used to represent the semantic feature relationship between scene feature codes.
[0055] A feature association topology is constructed based on the feature association coding matrix, and neighborhood aggregation is performed on the feature association topology to generate a feature association coding dataset. In the feature association topology, each scene feature code in each feature association coding matrix is used as a node, and the semantic correlation between two scene feature codes is used as an edge. Neighborhood aggregation is performed on the nodes with correlation based on a graph neural network, so that each aggregated node set includes the semantic features of each type of related nodes. For example, the scene features of the same building complex in the data are aggregated into one node set. Multi-scale feature extraction is performed on the feature association coding dataset to obtain multi-scale features. In this embodiment, a CNN convolutional neural network is used to extract global and local semantic features from the feature association coding dataset, and the output results are concatenated to generate multi-scale features. The multi-scale features are used to capture the regional difference information at different scales in the image data corresponding to the feature association coding dataset.
[0056] The system combines multi-scale features to model the dynamic change patterns of the battlefield environment and outputs a discriminative feature vector. This model utilizes a pre-trained recurrent neural network (GRU) model to enhance the ability to distinguish battlefield types. It learns scene feature change patterns between frames or images at different times within the received multi-scale features. The output discriminative feature vector includes information such as spatial features, semantic relationships, dynamic change patterns, and inter-frame change trends of the image or video. Based on the discriminative feature vector, the system performs battlefield environment discrimination on the enhanced battlefield image data and outputs battlefield type labels. In this embodiment, a Softmax classifier is used to specifically classify the discriminative feature vector, resulting in battlefield type labels that include various battlefield types such as jungle battlefield, urban battlefield, maritime battlefield, and air combat battlefield.
[0057] Methods for scheduling battlefield augmented imagery data based on battlefield type labels include:
[0058] Based on battlefield type labels, scene identifiers corresponding to the battlefield types are generated, and a mapping relationship between scene identifiers and battlefield type labels is established. The label matching conditions are constructed by combining expert experience and military literature, including the characteristic patterns, terrain structure and tactical performance elements of each battlefield type corresponding to each label. For example, in air combat scenarios, the sky area accounts for a high proportion, the boundary texture is sparse, the frequency of buildings is low, and the background changes slowly. For each label matching condition and the corresponding battlefield type label, a unique scene identifier is generated, and this scene identifier is constructed into a coding structure that the system can recognize.
[0059] Based on the battlefield type label, the intelligent recognition path of the preset control terminal is divided. In this embodiment, the control terminal constructs the most suitable intelligent recognition path for each battlefield type label. For example, when the battlefield type label is urban street fighting, the intelligent recognition path adopts high-precision edge detection and geometric structure recognition, while favoring small target tracking; when the battlefield type label is air combat, the intelligent recognition path focuses more on tracking high-speed targets.
[0060] Deep learning network models are deployed in each intelligent identification path, and historical battlefield augmented image data corresponding to the battlefield type label of the intelligent identification path are collected to train the deep learning network models. The type of deep learning network model deployed in each intelligent identification path is different. For example, the YOLO model is deployed in the air combat battlefield path, which has both temporal modeling and attention mechanisms and can easily capture high-speed moving targets; the HRNet model is deployed in the naval combat battlefield path, which has strong background modeling capabilities and excellent adaptability to water surface turbulence. By using historical battlefield augmented image data corresponding to the battlefield type label to train the model, it is ensured that each intelligent identification path focuses on identifying one battlefield environment, thereby improving execution efficiency.
[0061] Scene identifiers are used to activate corresponding intelligent recognition paths, while battlefield augmented imagery data is used as input data for these paths. Activating the corresponding intelligent recognition path using scene identifiers enables adaptive switching of the intelligent recognition path based on the input data. Whenever the scene described by the input data matches a certain scene identifier, the corresponding intelligent recognition path is activated, and a deep learning network model is loaded to prepare for target recognition. However, due to various factors that may prevent the models deployed in each path from performing all recognition tasks, the battlefield targets to be identified in different battlefield environments may vary. These factors often include three common issues: displacement deviation, dense target distribution, and inability to identify new equipment. Therefore, in subsequent steps, the scheduling imagery data allocated to each intelligent recognition path undergoes general processing to address these three common issues, serving as an extension and supplement to each intelligent recognition path to improve overall recognition accuracy.
[0062] Methods for dynamic target intelligent recognition of scheduled image data include:
[0063] The scheduling image data is converted into image frames and sorted based on timestamps to obtain a battlefield image frame sequence. In order to facilitate the capture of motion information in the time dimension from continuous scenes, the scheduling image data is decomposed into frames, and the video data and image sequence are converted into a battlefield image frame sequence sorted based on timestamps.
[0064] The process involves: acquiring basic feature information of battlefield targets; performing preliminary screening of the battlefield image frame sequence to output a target candidate set, where the basic feature information of the battlefield targets includes basic information such as the size, color and texture distribution, and shape of the battlefield targets to be identified; scanning the battlefield image frame sequence frame by frame, matching the basic feature information of the battlefield targets with each frame image, setting the regions containing feature information that may be the battlefield targets to be identified as target candidate regions, and integrating all target candidate regions to form a target candidate set; for example, if the battlefield target to be identified is a certain type of tank, then in the frame image, the regions where rectangular, low-lying objects with obvious edge contours are located are selected as target candidate regions.
[0065] A target recognition window is constructed to select battlefield targets in the target candidate set and output the target category probability. In this embodiment, a deep learning network model deployed in the corresponding intelligent recognition path is used to construct a detection box. The detection box is used to select battlefield targets and output the target category probability around the detection box. The target category probability refers to the recognition probability of the target category in the detection box.
[0066] Battlefield target bounding boxes are selected for N consecutive frames in a battlefield image frame sequence, and the position change information of the battlefield targets is obtained. In this embodiment, the battlefield target to be identified is bounded in N consecutive frames by calling a model, where N can be dynamically adjusted based on the specific number of frames. The position change information includes the coordinate position of the battlefield target in each frame image, as well as the velocity and displacement direction of the battlefield target. Combined with the timestamps corresponding to the frames, the position change information of the battlefield target over time in multiple frames images is obtained.
[0067] Based on position change information, the future trajectory of battlefield targets is predicted to obtain the predicted path of the battlefield targets in future frames. In this embodiment, a pre-trained LSTM model is used to capture the trajectory change trend in the position change information of battlefield targets, so as to predict the position of the battlefield targets in several future frames of images.
[0068] Error correction is performed on the identified position changes of the current battlefield target based on the predicted path of the future frame. The dynamic target trajectory data is output, obtaining the position information of the battlefield target at a certain frame at the current moment and the position change information of several frames in the past, thus predicting the predicted path for the future frame. This predicted path is then compared with the detected motion paths of the battlefield target in the image sequence of the same number of frames following the predicted path, and the error between the two paths is calculated. The error calculation formula is: Δdis=(a1-a,b1-b,c1-c,d1-d); where Δdis represents the difference between the two paths. The error compensation value of the detection box of the battlefield target in each frame of the image; a1 represents the x-coordinate of the center point of the detection box in the predicted path; a represents the x-coordinate of the center point of the detection box in the motion path; b1 represents the y-coordinate of the center point of the detection box in the predicted path; b represents the y-coordinate of the center point of the detection box in the motion path; c1 represents the width of the detection box in the predicted path; c represents the width of the detection box in the motion path; d1 represents the height of the detection box in the predicted path; d represents the height of the detection box in the motion path; This trajectory data can be used to adjust the error of the recognition position, and can also be used to complete the motion trajectory when the battlefield target is temporarily lost.
[0069] Methods for dense target identification of scheduled image data include:
[0070] Deep feature extraction is performed on each frame image in the battlefield image frame sequence to obtain deep features of the frame image. In order to prevent situations where some frames in the battlefield image frame sequence have drastic changes in the image environment, which may prevent normal feature extraction from fully extracting features, this embodiment combines feature extraction with a spatial pyramid pooling strategy to obtain deep features of the frame image. In cases where the environment changes drastically, such as a sudden scene destruction in a certain frame image that causes the battlefield target to be identified to be partially obscured, it is necessary to obtain more microscopic features through deep feature extraction to facilitate subsequent processing.
[0071] The target distribution density of each frame image is calculated based on deep features of the frame images. In this embodiment, the response intensity of each pixel in the frame image is calculated using the superposition of Gaussian kernel functions. By constructing a sliding window of adjustable size to traverse the frame images, the average of the sum of the response intensities of all pixels in each sliding window is the target distribution density of the corresponding region of that window. The formula for calculating the response intensity is as follows: ρ(x,y) represents the response intensity of the pixel at coordinates (x,y) in the frame image; Num represents the total number of pixels; G σ1 This represents a Gaussian kernel function with parameter σ1, used to measure coordinates (x... k ,y k The degree of influence of a pixel at coordinates (x, y) on a pixel at coordinates (x, y).
[0072] Density heatmaps are drawn for each frame image based on the target distribution density. A distribution density threshold is set and the density heatmaps are divided into density levels to identify high-density target areas. The higher the target distribution density in a certain area of the density heatmap, the more numerous and denser the potential targets in that area. Areas with target distribution density greater than the preset distribution density threshold are identified as high-density target areas, which facilitates subsequent focused identification of potential targets in that area.
[0073] In a high-density target region, candidate target initial detection points are set and feature perception is performed on the candidate target initial detection points. Feature perception data is output. In each local response maximum region of the high-density target region, candidate target initial detection points are set. The Top-k algorithm is used to select pixels in the neighborhood of the candidate target initial detection points and perform feature perception. These pixels are introduced into the intermediate perception layer of the CNN model. The texture, lighting and semantic attributes of the region formed by these pixels are fused to obtain feature perception data.
[0074] The feature perception data is matched and judged with the basic feature information of the battlefield target to select the battlefield target. In this embodiment, both the feature perception data and the basic feature information of the battlefield target are converted into vector representations. The Euclidean distance between the two vectors is calculated, and an Euclidean distance threshold is set. If the Euclidean distance is less than the preset Euclidean distance threshold, it is considered that the feature perception data and the basic feature information of the battlefield target can be matched. Then, the detection box can be used to select the battlefield target. However, at this time, the selection is performed on a region, not on each individual target, which provides a basis for subsequent elimination of overlapping conflicts.
[0075] When contour overlap occurs in the detection box, battlefield targets that appear consecutively in the nearest frames are retained for each detection box, and continuous frame recognition is performed on the occluded targets. There are two situations where battlefield target contours overlap: one is that multiple battlefield targets overlap together, and the other is that the battlefield target to be identified is partially occluded by other non-target objects in the environment. In this embodiment, battlefield targets that appear completely in several consecutive frames are extracted and retained in the detection box. If multiple other moving targets appear in consecutive frames but are not fully displayed in some frame images, these moving targets are regarded as occluded targets and marked. The position change of the occluded targets is obtained by gradually expanding the continuous frame recognition range. In this embodiment, a temporal sliding window combining backward tracking and forward expansion is constructed, starting from the frame image where any occluded target first appears, and traversing forward and backward respectively. The movement trajectory of the occluded target in consecutive frames is determined by extracting the coordinate position, displacement direction, and velocity of the occluded target in other frame images, which is the position change.
[0076] Based on the positional changes and basic feature information of battlefield targets, the complete outline of the occluded target is determined, battlefield target bounding boxes are selected, and target category probabilities are generated. Since the occluded target does not completely disappear, some feature information about these occluded targets can be extracted from other frame images. At the same time, there may be images in other frame images that fully display the outline of the occluded target, but it cannot be confirmed whether the occluded target to be identified is in that image. Therefore, in this embodiment, the possible complete outline of each occluded target is restored based on the basic feature information of the battlefield target. Combined with the positional changes of each occluded target in consecutive frames, the outlines in other frame images that match the positional changes are matched with the possible complete outlines. Finally, the frame image that can reflect the complete outline of the corresponding occluded target is found, and the battlefield target bounding box is selected in that image while generating target category probabilities. The high-density battlefield target detection data is constructed based on the battlefield image frame sequence after recognition. The battlefield image frame sequence after recognition realizes the recognition of each individual target in the high-density area. In order to prevent too many detection boxes from appearing in the frame image at any time and making it difficult to see, the frame image at each time is layered, and each layer corresponds to one battlefield target.
[0077] Methods for performing auxiliary discrimination on scheduled image data include:
[0078] The location of unknown targets is determined and pixel features are extracted to generate an unknown target feature vector. In this process, any frame in the battlefield image sequence that is selected by a detection box but cannot match the basic feature information of known battlefield targets is considered an unknown target. In this embodiment, an unknown target refers to new equipment or weapons appearing on the battlefield that are not included in a preset database, such as new tanks or new fighter jets. If a target is not included in the database, it cannot be identified. The unknown target feature vector is obtained by extracting features from all pixel blocks in the area marked as an unknown target, including through techniques such as texture analysis.
[0079] Obtain the known target feature vector corresponding to the known battlefield target, and calculate the feature distance between the unknown target feature vector and the known target feature vector. In this embodiment, the feature distance is the Mahalanobis distance.
[0080] Confidence is calculated based on feature distance. A confidence matching threshold is set. When the confidence is greater than or equal to the confidence matching threshold, the unknown target is marked as an associated target; otherwise, it is marked as a typical unknown target. The associated target indicates that it has a relationship with the known battlefield target, such as the associated target being an extension of the known battlefield target. The feature information of the typical unknown target is added as a new entry in the preset database. The feature information of the associated target is merged with the associated known battlefield target entries in the preset database. By adding the typical unknown target and the associated target to the preset database, the database is self-updated, which is convenient for future use of the database for target identification.
[0081] Methods for constructing battlefield simulation scenarios include:
[0082] Pixel change features of any region in the battlefield augmented image data as a function of timestamps are extracted, and a change trajectory of the region is constructed based on these features. Pixel change features refer to information such as the brightness value, color channel, and edge gradient of pixel blocks within the selected region. The pixel change features are combined with timestamps to obtain the change trajectory of the region over a period of time. The magnitude of the change trajectory is quantified, a threshold is set, and spatial reconstruction regions are selected based on the magnitude and the threshold. The formula for calculating the magnitude is: Where ΔTR represents the magnitude of change of any trajectory; M represents the total number of pixels in the region corresponding to the trajectory; R represents the region corresponding to the trajectory; I t (X0, Y0) represents the pixel value at coordinate (X0, Y0) at time t; t-1(X0,Y0) represents the pixel value of coordinate (X0,Y0) at time t-1; the trajectory of change with a change amplitude greater than the preset change amplitude threshold is judged as a region of large change and needs to be used as a spatial reconstruction region.
[0083] The battlefield environment change layer is output by fusing timestamps and spatial reconstruction regions. The spatial reconstruction region is combined with the frame image corresponding to each timestamp to form the environmental change layer of that region over time. This layer can fully reflect the changes in the battlefield environment caused by the movement of obstacles or the destruction of terrain. A battlefield environment spatial model is constructed based on the three-dimensional coordinates of the battlefield environment, and the battlefield environment change layer is matched with the battlefield environment spatial model in time to generate a battlefield simulation scene. In this embodiment, the battlefield environment spatial model is constructed based on battlefield spatial information using modeling technology. The battlefield spatial information includes information reflecting spatial attributes such as GPS coordinates, battlefield elevation difference data, and battlefield area corresponding to the battlefield augmented image data. At the same time, the battlefield environment change layer is synchronously nested into the battlefield environment spatial model according to the timestamp, so that the model not only has a static structure, but can also dynamically reflect the changes in the battle area, the battlefield destruction area, and the cover, improving the readability of the generated battlefield simulation scene.
[0084] Methods for accurately labeling battlefield targets in battlefield simulation scenarios include:
[0085] The image frame index of each battlefield target is obtained, and a top-level battlefield target mapping layer is constructed. Based on the image frame index, the image coordinates corresponding to the latest timestamp and the earliest timestamp of each battlefield target are mapped to the top-level battlefield target mapping layer. In this embodiment, the earliest and latest coordinates of the battlefield target are extracted to obtain the first and last frame coordinates of the battlefield target's lifecycle in the data of this embodiment. The first and last frame coordinates of each battlefield target are projected onto the top-level battlefield target mapping layer to represent the start and end points of the battlefield target's movement trajectory. The first and last frame coordinates of each battlefield target are marked in the top-level battlefield target mapping layer to make the presentation more intuitive and to avoid the situation where the coordinates of the battlefield target at every moment appear in one layer.
[0086] A battlefield target motion layer is constructed and divided into layers, generating battlefield target motion sub-layers. The number of sub-layers in the battlefield target motion layer is equal to the number of battlefield targets, and they are labeled using category tags, so that each battlefield target motion sub-layer corresponds to a unique battlefield target. The dynamic target trajectory data is mapped to the corresponding battlefield target motion sub-layer. In this embodiment, the dynamic target trajectory data is converted into a coordinate representation, forming structured trajectory point data composed of several coordinate points. The path of the occluded tactical target is obtained through continuous frame recognition simulation. By projecting the dynamic target trajectory data of each battlefield target onto the corresponding battlefield target motion sub-layer, the real motion trajectory of each battlefield target is restored. Users can directly search for the category tag of the battlefield target to view the corresponding motion trajectory.
[0087] A battlefield density area layer is constructed. Based on high-density battlefield target detection data, the density of the battlefield density area layer is labeled. The density heatmap of each frame image in the high-density battlefield target detection data is superimposed on the battlefield density area layer. In this embodiment, color is used to distinguish the density level: red represents high-density areas, yellow represents medium-density areas, blue represents evacuation areas, and green represents open areas. The battlefield density area layer can be used for important decisions such as battlefield fire deployment and movement route design.
[0088] A structured plotting layer is generated, and then nested with the battlefield simulation scene to generate battlefield target visualization plotting data. This involves combining a battlefield target motion layer, a battlefield dense area layer, and a top-level battlefield target mapping layer. The structured plotting layer includes tactical target motion trajectories, first and second frame coordinates, and high-density area markers. Users can adjust the layer structure as needed. For example, if the motion trajectory of a specific battlefield target is required, the corresponding battlefield target motion sub-layer can be moved to the top layer, and other layers can be hidden. By binding the structured plotting layer with the coordinates of the battlefield simulation scene and designing functions such as layer transparency adjustment and enabling layer dynamic states, the battlefield target visualization plotting data is obtained.
[0089] Optimized methods for obtaining processing paths include:
[0090] Feedback data is collected in real time, and the parameters of the intelligent recognition path are optimized based on the feedback data to obtain the optimized processing path. In this process, the feedback data is divided into time periods for feature extraction to obtain the intelligent recognition difference vector. The intelligent recognition difference vector belongs to the vector corresponding to each intelligent recognition path within any consecutive time period. In this embodiment, the feedback data is divided into several time periods based on timestamps. The intelligent recognition difference vector is constructed by extracting indicators such as the response efficiency, misjudgment probability, and resource utilization rate of the intelligent recognition path in each time period. This vector is used to reflect the running performance of the intelligent recognition path in the corresponding time period.
[0091] All intelligent recognition difference vectors are integrated to obtain a path difference feedback set, and a correspondence between the path parameters and intelligent recognition difference vectors of each intelligent recognition path is established. The path parameters include the model parameters and discrimination thresholds of the deep learning network model deployed in the intelligent recognition path. Establishing a correspondence between the path parameters and intelligent recognition difference vectors is to enable the intelligent recognition difference vectors to be used as a basis for judgment in subsequent operations, and to optimize the relevant path parameters in a targeted manner.
[0092] An effectiveness evaluation model is constructed and used to evaluate the path recognition performance of the path difference feedback set. The model outputs a path recognition performance score for each intelligent recognition path. In this embodiment, the effectiveness evaluation model is a multilayer perceptron network model. The historical path difference feedback set is collected and labeled with time period labels. The historical path difference feedback set is used as the training set of the multilayer perceptron network model. The path recognition performance score of a certain intelligent recognition path in each time period is calculated through the loss function of the model.
[0093] A recognition performance score threshold is set, and intelligent recognition paths with a recognition performance score lower than the threshold are identified as optimization target paths. Parameter analysis is performed on the path parameters of the optimization target paths, and the path parameters are reconstructed based on the parameter analysis results. The optimized parameters are then applied to the optimization target paths. In this embodiment, the path parameters of the optimization target paths are extracted, and the parameter set is reconstructed using a Bayesian optimization algorithm to obtain the optimized parameters. The optimized parameters are then applied to the corresponding optimization target paths to obtain the optimized target paths.
[0094] The optimized target path and the intelligent recognition path with a recognition effect score greater than or equal to the recognition effect score threshold are integrated to obtain the optimized intelligent recognition path. The optimized intelligent recognition path is sent to a preset database for storage, and the optimized target path in the database is used to directly replace the original intelligent recognition path to achieve closed-loop parameter optimization of the system.
[0095] This embodiment achieves accurate battlefield target labeling by performing scene recognition, path scheduling, trajectory prediction, dense target processing, unknown target assisted recognition, and scene construction on acquired battlefield image data. Simultaneously, it improves target recognition performance by optimizing recognition path parameters. By constructing an intelligent recognition path based on scene label segmentation and dynamically loading a dedicated recognition model according to battlefield type, it significantly improves the task adaptability and recognition accuracy of target recognition in complex battlefield environments. By fusing inter-frame detection results with a pre-trained temporal model, it achieves real-time prediction of multi-target trajectories and enhances target localization accuracy in continuous time through error correction. In high-density area recognition, it integrates candidate detection point mechanisms and occlusion detection. The completion strategy ensures the integrity and stability of identification in scenarios with dense targets and obscuring interference; it also supports the identification and discrimination of unknown targets and dynamic expansion and updating of the database, possessing good self-learning capabilities and equipment identification expansion capabilities; in addition, in the simulation scene construction and target annotation stages, multi-layer target nesting is realized based on a three-dimensional coordinate system, improving visualization clarity and multi-layer controllability; in terms of feedback optimization path scheduling, identification path parameters are reconstructed based on time-segmented error feedback, effectively improving identification efficiency and closed-loop control capabilities; it has extremely high practical deployment value, is suitable for target identification tasks in various types of battlefields, and helps to improve the execution efficiency of dynamic target management and intelligent command systems.
[0096] Example 2
[0097] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A deep learning-based intelligent battlefield target identification and mapping system is provided, comprising:
[0098] The data acquisition module is used to collect battlefield image data and perform data cleaning to obtain enhanced battlefield image data.
[0099] The scene recognition module is used to extract multimodal features from battlefield augmented image data to obtain a scene feature vector set; based on the scene feature vector set, it performs scene recognition on the battlefield augmented image data and outputs battlefield type labels.
[0100] The path construction module schedules battlefield augmented image data based on battlefield type labels, obtains the scheduled image data, and sends it to the intelligent recognition path built based on deep learning network to execute the target recognition process.
[0101] The fine-grained recognition module is used to perform dynamic target intelligent recognition on scheduling image data and output dynamic target trajectory data; to perform dense target recognition on scheduling image data and output high-density battlefield target detection data; and to perform auxiliary discrimination on scheduling image data and update the preset database based on the discrimination results.
[0102] The visualization plotting module is used to acquire battlefield spatial information corresponding to battlefield augmented image data, construct battlefield simulation scenarios based on battlefield spatial information, and accurately annotate battlefield targets in the battlefield simulation scenarios to obtain battlefield target visualization plotting data.
[0103] The parameter optimization module is used to update the intelligent identification path to an optimized processing path and send it to a preset database for storage; the modules are connected to each other via wired and / or wireless means.
[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0105] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0106] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.
[0107] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0108] In the description of this invention, "several" means one or more, and "a large number" means two or more.
[0109] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0110] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.
[0111] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
Claims
1. A battlefield target intelligent identification and mapping method based on deep learning, characterized in that, include: S1. Collect battlefield image data and perform data cleaning to obtain enhanced battlefield image data; S2. Perform multimodal feature extraction on the battlefield augmented image data to obtain a scene feature vector set; Scene recognition is performed on battlefield augmented image data based on scene feature vector sets, and battlefield type labels are output. S3. Data scheduling of battlefield augmented imagery data based on battlefield type labels, including: Based on the battlefield type label, generate scene identifiers corresponding to the battlefield type and establish a mapping relationship between scene identifiers and battlefield type labels; divide the intelligent recognition path of the preset control terminal based on the battlefield type label; deploy a deep learning network model in each intelligent recognition path and collect historical battlefield enhanced image data of the battlefield type label corresponding to the intelligent recognition path to train the deep learning network model; activate the corresponding intelligent recognition path using the scene identifier, and at the same time use the battlefield enhanced image data as the input data of the intelligent recognition path. The scheduled image data is obtained and sent to the intelligent recognition path built based on a deep learning network to execute the target recognition process; S4. Perform dynamic target intelligent recognition on the scheduling image data and output dynamic target trajectory data; perform dense target recognition on the scheduling image data and output high-density battlefield target detection data; perform auxiliary discrimination on the scheduling image data and update the preset database based on the discrimination results; S5. Obtain battlefield spatial information corresponding to the battlefield augmented image data, and construct a battlefield simulation scene based on the battlefield spatial information; accurately annotate the battlefield targets in the battlefield simulation scene to obtain battlefield target visualization plotting data; S6. Update the intelligent recognition path to an optimized processing path, including: real-time collection of feedback data, optimization of parameters of the intelligent recognition path based on the feedback data, obtaining the optimized processing path, and sending it to a preset database for storage.
2. The battlefield target intelligent identification and mapping method based on deep learning according to claim 1, characterized in that, The methods for scene recognition of enhanced battlefield image data include: Extract discriminative feature vectors and perform battlefield environment discrimination on battlefield augmented image data, outputting battlefield type labels.
3. The battlefield target intelligent identification and mapping method based on deep learning according to claim 2, characterized in that, The method for dynamic target intelligent recognition of scheduled image data includes: The process involves converting scheduling image data into a battlefield image frame sequence; acquiring basic feature information of battlefield targets; performing preliminary screening of the battlefield image frame sequence to output a target candidate set; and selecting battlefield targets from the target candidate set while outputting the target category probability. Battlefield targets are selected from the battlefield image frame sequence and their position change information is obtained. Based on the position change information, the future trajectory of the battlefield targets is predicted to obtain the future frame prediction path of the battlefield targets. Based on the future frame prediction path, the error of the current battlefield target's identified position change is corrected, and the dynamic target trajectory data of the battlefield targets is output.
4. The battlefield target intelligent identification and mapping method based on deep learning according to claim 3, characterized in that, The method for performing dense target identification on scheduled image data includes: Deep features are extracted from each frame in the battlefield image frame sequence to obtain deep features of the frame images; the target distribution density of each frame image is calculated based on the deep features of the frame images; high-density target areas are determined based on the target distribution density; initial detection points for candidate targets are set and feature perception is performed on the initial detection points of candidate targets to output feature perception data; the feature perception data is matched and judged with the basic feature information of battlefield targets and battlefield target bounding boxes are selected. Each detection box retains battlefield targets that appear consecutively in neighboring frames, and performs continuous frame recognition on occluded targets; obtains the positional changes of occluded targets; determines the complete outline of occluded targets based on positional changes and basic feature information of battlefield targets, performs battlefield target box selection, and generates target category probabilities; and constructs high-density battlefield target detection data based on the recognized battlefield image frame sequence.
5. The battlefield target intelligent identification and mapping method based on deep learning according to claim 4, characterized in that, The method for performing auxiliary discrimination on the scheduled image data includes: The process involves determining the location of an unknown target and extracting its pixel features to generate an unknown target feature vector; obtaining the feature vectors of known targets corresponding to known battlefield targets, and calculating the feature distance between the unknown target feature vector and the known target feature vector; calculating the confidence score and setting a confidence score matching threshold; marking the unknown target as an associated target when the confidence score is greater than or equal to the confidence score matching threshold, and otherwise marking it as a typical unknown target; extracting the feature information of typical unknown targets as new entries in a preset database, and merging the feature information of associated targets with existing entries in the preset database.
6. The battlefield target intelligent identification and mapping method based on deep learning according to claim 5, characterized in that, The methods for constructing battlefield simulation scenarios include: Extract pixel change features and construct the change trajectory of the region based on the pixel change features; quantify the change amplitude of the change trajectory, set the change amplitude threshold, and filter the spatial reconstruction region; merge the timestamp and the spatial reconstruction region to output the battlefield environment change layer; construct the battlefield environment spatial model based on the battlefield spatial information and perform time-series matching with the battlefield environment change layer to generate a battlefield simulation scene.
7. The battlefield target intelligent identification and mapping method based on deep learning according to claim 6, characterized in that, The methods for accurately marking battlefield targets in battlefield simulation scenarios include: The process involves: acquiring the image frame index of each battlefield target; constructing a top-level battlefield target mapping layer; mapping the image coordinates corresponding to the latest and initial timestamps of each battlefield target to the top-level battlefield target mapping layer based on the image frame index; constructing a battlefield target motion layer and dividing it into layers to generate battlefield target motion sub-layers; mapping dynamic target trajectory data to the corresponding battlefield target motion sub-layers; constructing a battlefield dense area layer and annotating the density of the battlefield dense area layer based on high-density battlefield target detection data; generating a structured plotting layer; nesting the structured plotting layer and the battlefield simulation scene to generate battlefield target visualization plotting data.
8. A deep learning-based intelligent battlefield target identification and mapping system, used to implement the deep learning-based intelligent battlefield target identification and mapping method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect battlefield image data and perform data cleaning to obtain enhanced battlefield image data. The scene recognition module is used to extract multimodal features from battlefield augmented image data to obtain a scene feature vector set; Scene recognition is performed on battlefield augmented image data based on scene feature vector sets, and battlefield type labels are output. The path construction module schedules battlefield augmented image data based on battlefield type labels, obtains the scheduled image data, and sends it to the intelligent recognition path built based on deep learning network to execute the target recognition process. The fine recognition module is used to perform intelligent dynamic target recognition on the scheduling image data and output dynamic target trajectory data; Perform dense target identification on scheduling image data and output high-density battlefield target detection data; Perform auxiliary discrimination on the scheduled image data and update the preset database based on the discrimination results; The visualization plotting module is used to acquire battlefield spatial information corresponding to battlefield augmented image data and construct battlefield simulation scenarios based on the battlefield spatial information. Accurately label battlefield targets in battlefield simulation scenarios to obtain visualized battlefield target plotting data; The parameter optimization module is used to update the intelligent identification path to an optimized processing path and send it to a preset database for storage; the modules are connected to each other via wired and / or wireless means.