Anti-uav target identification method and system based on multi-source information fusion

Abstract

This invention discloses an anti-drone target recognition method based on multi-source information fusion, comprising: acquiring motion parameters of the target object through wide-area search and precise tracking; driving an optoelectronic tracking module to track the target object based on the motion parameters; acquiring keyframes based on real-time video data collected by the optoelectronic module through target object change degree and contour clarity filtering; acquiring foreground images based on keyframes through background removal processing; acquiring multi-source auxiliary data in real time; performing spatiotemporal matching of multi-source auxiliary data and foreground images based on high-precision timestamps to acquire information unit data; arranging the foreground images in chronological order based on the information unit data and mapping the multi-source auxiliary data with the foreground images to acquire fused information data; and completing target recognition based on the fused information data.

Description

Technical Field

[0001] This invention belongs to the field of anti-drone technology, and in particular relates to an anti-drone target identification method and system based on multi-source information fusion. Background Technology

[0002] With the rapid evolution of drone technology and its widespread penetration in both military and civilian fields, the airspace security threats it poses are becoming increasingly severe. Counter-drone systems, as a key means of countering this threat, rely on their ability to rapidly detect and accurately identify drone targets for core effectiveness.

[0003] Currently, mainstream target recognition technologies primarily rely on the analysis and processing of video image information acquired by optoelectronic devices. However, this type of technology still has the following shortcomings: Low fusion rate of multi-source heterogeneous information leads to insufficient data utilization. In counter-drone applications, in addition to video stream data, auxiliary information is also included, such as parameters of the photoelectric sensor itself, including focal length, field of view, and target distance, as well as the characteristics of drone remote control links or image transmission signals obtained through electronic reconnaissance equipment. Currently, a discrete architecture is typically used to independently collect and store this multi-source heterogeneous data, lacking an effective automatic correlation mechanism. In the target identification stage, manual data correlation analysis by operators is often required. This approach is not only inefficient but also prone to misjudgment or missed judgment due to information misalignment or omission of key information, making it difficult to leverage the synergistic enhancement effect of multi-source information.

[0004] Second, the feature representation is limited in scope and has poor environmental adaptability. Traditional recognition algorithms often focus on extracting single visual features such as shape, texture, and color, or perform only simple data layer fusion. When the target is in complex conditions such as long-distance imaging, partial occlusion, sudden changes in lighting, or severe weather, the distinguishability of visual features will be greatly reduced, resulting in a significant decrease in the accuracy and robustness of the recognition model, making it difficult to meet the application requirements in complex real-world environments.

[0005] Third, it has poor adaptability to dynamic scale changes. As the UAV target approaches or moves away from the detection platform, its size in the imaging field of view undergoes drastic nonlinear changes. Traditional neural network models are quite sensitive to such large-scale scale changes, which can easily lead to unstable feature extraction and thus limit the final recognition accuracy. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention proposes a multi-source information fusion-based anti-drone target recognition method and system. It achieves spatiotemporal alignment of heterogeneous multi-source data through high-precision timestamps, constructs a structured fusion information carrier containing foreground images and auxiliary data, calculates the actual physical dimensions of the target to constrain the recognition results, and combines scale normalization processing to improve recognition stability in dynamic scenes. This invention breaks down the information silos of single sensors, achieving deep fusion of the target's essential physical features and visual features, effectively solving the problem of target recognition under dynamic scale changes and complex environments. Simultaneously, the fully automated execution significantly improves recognition efficiency and response speed, meeting the real-time requirements of anti-drone systems.

[0007] The first aspect of the present invention provides a method for anti-drone target identification based on multi-source information fusion, comprising: S1: Obtain the motion parameters of the target object through wide-area search and precise tracking, and drive the photoelectric tracking module to track the target object based on the motion parameters. The motion parameters include at least the target distance, target speed, and target orientation. S2: Based on the real-time video data collected by the photoelectric module, key frames are obtained by filtering the target object's degree of change and outline sharpness. Based on the key frames, the foreground image of the target object is obtained by background removal processing. S3: Real-time acquisition of multi-source auxiliary data, spatiotemporal matching of the multi-source auxiliary data and the foreground image based on high-precision timestamps, to obtain information unit data in spatiotemporal dimension. The information unit data is a collection of target foreground images and multi-source auxiliary data bound based on unified timestamps and spatial correlation. S4: Based on the information unit data, the foreground images are arranged in chronological order and the multi-source auxiliary data is mapped to the foreground images to obtain fused information data. The fused information data includes a time-series foreground image sequence, the photoelectric module parameters associated with the foreground images, radar signals, electronic reconnaissance signals, and structured data of environmental information. S5: Target recognition is completed based on the fused information data, specifically including parsing the labeled focal length parameters and target distance parameters from the fused information data, calculating the actual physical size of the target object through optical imaging principles, extracting visual feature information based on the foreground image through scale normalization processing, and performing feature matching and type confirmation based on the actual physical size and the visual feature information through a multimodal recognition model to obtain the target recognition result.

[0008] Preferably, step S2, which involves filtering by the degree of change of the target object, further includes: Edge images are obtained through edge detection processing based on continuous data in the real-time video data. Calculate the structural similarity index between adjacent edge images; The difference degree is calculated based on the structural similarity index: if the difference degree exceeds a preset first threshold, the corresponding video data is retained as a candidate keyframe.

[0009] Preferably, step S2, which involves obtaining keyframes through contour sharpness filtering, further includes: The ratio of the actual number of edge pixels to the theoretical total number of edge pixels in the candidate keyframe is calculated, and the ratio is used as the contour sharpness. The theoretical total number of edge pixels is obtained by estimating the perimeter of the target bounding box, using a contour fitting template, or by deriving from the segmentation results. If the outline sharpness is greater than or equal to a preset second threshold, the candidate keyframe is confirmed as a keyframe.

[0010] Preferably, obtaining the foreground image of the target object through background removal processing in step S2 further includes: Foreground extraction can be achieved through one or more of the following methods: object detection bounding box cropping, semantic segmentation, motion foreground detection, or instance segmentation.

[0011] Preferably, the step of constituting the fused information data in step S4 further includes: The fused information data is obtained by verifying and mapping the timestamp error based on the parameters of the optoelectronic module and the electronic reconnaissance signal. The parameters of the optoelectronic module include at least focal length and field of view; the electronic reconnaissance signal includes at least signal frequency and signal strength.

[0012] Preferably, in the anti-drone target identification method based on multi-source information fusion according to claim 1, the expression for calculating the actual physical size of the target object is: In the formula, s is the pixel width of the target object in the video data, s is the pixel size, D is the target distance, and f is the lens focal length.

[0013] Preferably, step S5, which involves extracting visual feature information through scale normalization, further includes: Obtain a sequence of foreground images of the target object in consecutive frames, and extract the bounding box information of the foreground images; Based on the target distance parameter or actual physical size, calculate the scale variation factor of the current keyframe relative to the reference frame; The foreground image is processed by a spatial transformation network using the scale variation factor to generate a normalized image with consistent scale. The normalized image is input into the visual feature encoder of the multimodal recognition model for feature extraction to obtain visual feature information.

[0014] Preferably, the multimodal recognition model in step S5 includes at least a visual feature encoder, an auxiliary data encoder, and a fusion inference module; the visual feature encoder is used to extract visual features from the normalized image, the auxiliary data encoder is used to parse parametric features in the fused information data, and the fusion inference module is used to fuse visual features and parametric features to complete target feature matching and type confirmation.

[0015] Preferably, the wide-area search in step S1 is completed by radar or wide-angle optoelectronic equipment, and the precise tracking is completed by closed-loop control of a narrow-field optoelectronic turntable based on motion parameters.

[0016] A second aspect of the present invention provides an anti-drone target recognition system based on multi-source information fusion, employing the anti-drone target recognition method based on multi-source information fusion described in any one of the preceding claims, comprising: The hardware interface unit includes a radar interface module, an optoelectronic acquisition module, an electronic reconnaissance interface module, and an environmental information acquisition module, which are used to acquire target motion parameters, optoelectronic video data, electronic reconnaissance signals, and environmental information, respectively. The time synchronization unit is used to add a high-precision timestamp to the data collected by the hardware interface unit to achieve time synchronization of multi-source auxiliary data. The data processing unit includes a target tracking module, an extraction and processing module, a spatiotemporal alignment module, and a fusion processing module. The target tracking module is used to obtain the motion parameters of the target object through wide-area search and precise tracking, and drive the photoelectric tracking module to track the target object based on the motion parameters. The extraction and processing module is used to complete keyframe filtering and background removal processing based on the real-time video data collected by the photoelectric module to obtain the target foreground image. The spatiotemporal alignment module is used to perform spatiotemporal matching of multi-source auxiliary data and foreground image to obtain information unit data. The fusion processing module is used to construct structured fused information data based on the information unit data. The identification and evaluation unit includes a size calculation module, a scale normalization module, a multimodal recognition module, and a threat level determination module. The size calculation module is used to calculate the actual physical size of the target based on the principle of optical imaging. The scale normalization module is used to perform scale normalization processing on the foreground image to generate a normalized image. The multimodal recognition module is used to extract visual feature information based on the normalized image and fuse it with the actual physical size to complete the target identification.

[0017] Because the present invention adopts the above technical solution, it has the following advantages and positive effects compared with the prior art: 1. Based on high-precision timestamp-driven spatiotemporal alignment of multi-source heterogeneous data, it breaks the information silos of traditional single sensors, effectively correlates video streams with multi-dimensional data such as radar, electronic reconnaissance, and environmental information, improves the information density and utilization value of the data, and provides complete contextual information for target recognition.

[0018] 2. Constructing structured fusion information data binds the foreground image and multi-source auxiliary data in a structured form, avoiding data disorder, facilitating feature extraction and fusion in subsequent models, and improving the efficiency and accuracy of data processing.

[0019] 3. The actual physical size is calculated based on focal length, distance, and pixel size and used as a recognition constraint. The problem of recognizing small targets at long distances, similar-looking targets, and targets with dynamic scale changes is solved by utilizing the essential physical characteristics of the target. This effectively improves the accuracy and environmental adaptability of target recognition. Experimental verification shows that the false recognition rate of targets can be reduced by more than 30% after adding physical size constraints.

[0020] 4. Improve recognition stability in dynamic scenes by combining scale normalization processing. The spatial transformation network normalizes target images of different scales, eliminating the influence of target distance changes on feature extraction. Experiments have verified that after scale normalization processing, the recognition accuracy in complex dynamic scenes can be improved by more than 25%.

[0021] 5. The entire process is automated, from target tracking, key frame screening, multi-source data fusion to target identification and threat level determination, without the need for manual intervention. This reduces labor costs and the risk of misjudgment, and significantly shortens the response time from target detection to output results, meeting the real-time requirements of anti-drone systems. Compared with traditional manual-assisted identification methods, the identification response time of this invention can be shortened by more than 60%.

[0022] 6. The multimodal recognition model clarifies the module composition and processing method of the model, enhancing the feasibility of the technology. At the same time, it integrates visual and physical features, improving the robustness of the model in complex environments such as partial occlusion, sudden changes in lighting, and haze. Experimental verification shows that the overall accuracy of multi-source fusion recognition is more than 35% higher than that of pure visual recognition. Attached Figure Description

[0023] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is the main process of the anti-drone target recognition method based on multi-source information fusion in this invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise ratios, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0025] It should be noted that all directional indicators (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0026] First Embodiment See Figure 1 The first aspect of the present invention provides a method for anti-drone target identification based on multi-source information fusion, comprising: S1: Obtain the motion parameters of the target object through wide-area search and precise tracking, and drive the photoelectric tracking module to track the target object based on the motion parameters. The motion parameters include at least the target distance, target speed, and target orientation. S2: Based on the real-time video data collected by the photoelectric module, key frames are obtained by filtering the target object's degree of change and outline sharpness. Based on the key frames, the foreground image of the target object is obtained through background removal processing. S3: Real-time acquisition of multi-source auxiliary data, spatiotemporal matching of multi-source auxiliary data and foreground images based on high-precision timestamps, to obtain information unit data in spatiotemporal dimensions. The information unit data is a collection of target foreground images and multi-source auxiliary data bound based on unified timestamps and spatial relationships at a single moment or time period. S4: Based on information unit data, the foreground images are arranged in time sequence and the multi-source auxiliary data is mapped to the foreground images to obtain fused information data. The fused information data includes the time-series foreground image sequence, the optoelectronic module parameters associated with the foreground images, radar signals, electronic reconnaissance signals, and structured data of environmental information. S5: Target recognition is completed based on fused information data. Specifically, this includes parsing the labeled focal length parameters and target distance parameters from the fused information data, calculating the actual physical size of the target object through optical imaging principles, extracting visual feature information based on the foreground image through scale normalization, and performing feature matching and type confirmation through a multimodal recognition model based on the actual physical size and visual feature information to obtain the target recognition result.

[0027] A dual filtering mechanism, employing both variability and contour sharpness screening, eliminates blurry, redundant, and insignificantly changing video frames, accurately identifying keyframes containing valid information. Combined with background removal, it eliminates interference from complex backgrounds, significantly reducing the amount of data and computational complexity in subsequent processing, thus improving the real-time performance and accuracy of the recognition system. High-precision timestamps are used to perform spatiotemporal matching between multi-source auxiliary data and visual foreground images, resolving the issues of information silos and spatiotemporal misalignment. By arranging foreground images chronologically and labeling auxiliary data, structured fusion information data is constructed, providing the recognition model with contextual information encompassing temporal, spatial, and physical dimensions, enhancing the comprehensiveness of feature representation. Feature extraction is performed using a multimodal large model, fusing multidimensional information in the process. This multidimensional feature fusion compensates for the shortcomings of single-vision recognition in long-distance, weak targets, or harsh environments, improving the environmental adaptability of target recognition.

[0028] Step S1: Obtain the motion parameters of the target object through wide-area search and precise tracking. The motion parameters include at least the target distance, target speed, and target orientation. Based on these motion parameters, drive the photoelectric tracking module to track the target object. Wide-area search is performed by radar or wide-angle photoelectric equipment, while precise tracking is performed by a narrow-field-of-view photoelectric turntable combined with closed-loop control of the motion parameters. The target distance parameter guides the zoom and focus control of the photoelectric module, and the target speed and orientation parameters eliminate tracking lag, achieving stable and continuous tracking of high-speed maneuvering targets.

[0029] Step S2: Keyframe Screening and Foreground Image Extraction; Based on real-time video data acquired by the photoelectric module, keyframes are obtained through target object variability screening and contour sharpness screening. Based on these keyframes, background removal processing is used to obtain the foreground image of the target object. Variation Screening: Edge detection processing is performed on continuous frame data from the real-time video data to obtain edge images. The structural similarity index of adjacent edge images is calculated, and the difference is derived. If the difference exceeds a preset first threshold (selectable range 20%~40%, determined by historical sample statistics, on-site calibration, or experience), the corresponding video frame is retained as a candidate keyframe. This method effectively resists interference from non-target motion factors such as sudden changes in illumination and environmental noise, ensuring that candidate keyframes contain effective changes in target posture or position. Contour Sharpness Screening: The ratio of the actual number of edge pixels to the theoretical total number of edge pixels in the candidate keyframe is calculated as the contour sharpness. The theoretical total number of edge pixels can be estimated based on the target bounding box perimeter, obtained from a contour fitting template, or derived from segmentation results. If the contour sharpness is greater than or equal to a preset second threshold (selectable range 70%~90%, determined by historical sample statistics, on-site calibration, or experience), it is confirmed as a valid keyframe. This method quantifies image quality issues, eliminating blurry or incomplete video frames to ensure keyframe quality. Background removal processing: It employs one or more combinations of foreground extraction methods based on object detection bounding box cropping, semantic segmentation, moving foreground detection, or instance segmentation to eliminate interference from complex backgrounds on the target and reduce the amount of data and computational complexity in subsequent processing.

[0030] Step S3: Multi-source data acquisition and spatiotemporal alignment; real-time acquisition of multi-source auxiliary data, using high-precision timestamps as a benchmark to perform spatiotemporal matching between the multi-source auxiliary data and foreground images, obtaining information unit data in spatiotemporal dimensions. Multi-source auxiliary data includes photoelectric module parameters (focal length, field of view, etc.), radar signals, electronic reconnaissance signals (signal frequency, signal strength, etc.), and environmental information, achieving complementary fusion of heterogeneous multi-source data to form a comprehensive, three-dimensional target perception capability. High-precision spatiotemporal matching: a time synchronization unit adds high-precision timestamps to all acquired data, with a time synchronization error not exceeding 100ms. Data binding is completed after verifying the timestamp error of each data point. The information unit data is a collection of target foreground images and corresponding auxiliary data bound based on a unified timestamp and spatial correlation, solving the problems of information silos and spatiotemporal misalignment.

[0031] Step S4: Constructing Structured Fusion Information Data; Based on the information unit data, the foreground images are arranged chronologically, and multi-source auxiliary data are bound one-to-one with the corresponding foreground images as structured fields to obtain fusion information data. The fusion information data is structured input data containing a sequence of foreground images arranged in chronological order, as well as photoelectric module parameters, radar signals, electronic reconnaissance signals, and environmental information associated with each foreground image. This provides the subsequent recognition model with complete contextual information including time, space, and physical parameters, enhancing the comprehensiveness of feature expression.

[0032] Step S5: Target Recognition and Feature Fusion; Target recognition is completed based on the fused information data. Specifically, this includes parsing the labeled focal length and target distance parameters from the fused information data, calculating the actual physical size of the target object using optical imaging principles, performing scale normalization on the foreground image to extract visual feature information, fusing the actual physical size and visual feature information, and performing feature matching and type confirmation through a multimodal recognition model to obtain the target recognition result. Actual physical size calculation: The calculation expression is as follows: In the formula, where For the actual physical size of the target, Let s be the pixel width of the target object in the video data, D be the target distance, and f be the lens focal length. This physical size is an essential attribute of the target and does not change with imaging distance, angle, or lighting. It can serve as a hard constraint for recognition, solving the problem of recognizing small targets at a distance or targets with similar appearances. Scale normalization and visual feature extraction: Obtain the bounding box information of the foreground image sequence, calculate the scale change factor based on the target distance or actual physical size, and generate a scale-consistent normalized image through a spatial transformation network. This normalized image is then input into the visual feature encoder of the multimodal recognition model to extract visual feature information. This method eliminates the influence of dynamic scale changes of the target on feature extraction, improves the stability of visual features, and avoids feature blurring or loss.

[0033] The multimodal recognition model includes at least a visual feature encoder, an auxiliary data encoder, and a fusion inference module. The visual feature encoder extracts visual features from the normalized image, the auxiliary data encoder parses the parametric features in the fusion information data, and the fusion inference module fuses the two types of features to complete target feature matching and type confirmation, thus making up for the shortcomings of single visual recognition in complex environments.

[0034] Taking a typical airport airspace defense scenario as an example, the deployment of the anti-drone target recognition system based on multi-source information fusion described in this invention is as follows: S1: Target Location Acquisition: A radar deployed around the airport perimeter detects a suspicious target. In precise tracking mode, it measures the target's azimuth, distance, speed, and other motion parameters. This information is then transmitted to the optoelectronic system, guiding the optoelectronic turntable to align with the target airspace. The optoelectronic equipment continuously scans the airspace. The lens focal length of the optoelectronic equipment can continuously vary between 10mm and 200mm, with a pixel size of 2.9μm and an effective ranging range of 100m to 5000m. The output video stream is sent to the optoelectronic processing module.

[0035] S2: Keyframe Extraction and Processing: The photoelectric processing module decodes the video stream. When a drone enters the field of view, the module performs edge detection processing on consecutive frames and calculates the structural similarity index difference between adjacent edge images. At a certain moment, the difference reaches 35% (higher than the preset first threshold of 30%), triggering candidate determination. Subsequently, the module extracts the foreground of the candidate frame and calculates its contour sharpness to be 88% (higher than the preset second threshold of 75%), thus confirming it as a valid keyframe. At this time, the target occupies a pixel width of 150 pixels on the sensor.

[0036] S3: Multi-source Information Acquisition and Alignment: At the timestamp of capturing this keyframe (2025-10-17 03:43:20.345), the spatiotemporal alignment module synchronously reads multi-source auxiliary data, including: photoelectric parameters (current focal length f=150mm, laser rangefinder showing target distance D=450m); electronic reconnaissance signals (detected frequency 2.45GHz, signal strength -75dBm); and environmental data (weather: clear). The spatiotemporal alignment module verifies that the timestamp error of all information is within 10ms (less than 100ms allowable error), completes data binding, and acquires information unit data.

[0037] S4: Fusion Information Map Generation: The fusion processing module uses a 1920×1080 resolution image as a base, arranges the 20 most recently captured keyframes in chronological order, and annotates the keyframes with the text: "T:20.345|f:150mm|D:450m|Freq:2.45GHz|Clarity:88%", generating fusion information data.

[0038] S5: Multimodal Target Recognition: The generated fused information data is sent to the result acquisition module. The module uses a calculation expression based on the actual physical dimensions: The estimated actual physical dimensions (wingspan) of the drone are approximately 1.3 meters.

[0039] The multimodal large model comprehensively analyzed the following information: visual features indicated the target's shape was fixed-wing; the estimated physical size was 1.3 meters, classifying it as a medium-sized UAV; signal characteristics showed 2.45 GHz as a common UAV image transmission / remote control frequency band; situational information indicated a distance of 450 meters, approaching the airport head-on. The model's final output identification result was a fixed-wing UAV, with a confidence level of 99.2%. Subsequently, the threat level determination module, considering its relatively large size of 1.3 meters, relatively close distance of 450 meters, and the identified aircraft type, comprehensively determined the threat level to be high-risk.

[0040] In another specific implementation scenario, namely in a complex environment where visibility is approximately 1.5 km and the target is partially obscured by trees, the system described in this invention is applied for target identification. The method includes the following steps: S1: Keyframe Extraction and Processing: Addressing the issue of blurred, low-contrast, and incomplete outlines in target images within the video stream, the extraction and processing module filters consecutive frames. Although the Structural Similarity (SSIM) change rate may be low, the module performs a secondary judgment using outline sharpness criteria, successfully selecting a small number of frames with an outline sharpness of 76% (higher than the preset second threshold of 75%) as valid keyframes.

[0041] S2~S4 includes multi-source information acquisition and fusion information map generation. The specific process is as follows: the spatiotemporal alignment module collects environmental information in real time, and the fusion processing module marks the environmental status as "weather: fog", "visibility: 1.5km" and "occlusion: yes" when generating fusion information data, and integrates it with the foreground image and time series data.

[0042] S5: Multimodal target recognition, the specific process is as follows: The multimodal large model, through its internal correlation analysis mechanism, identifies environmental labels such as "weather: fog" or "occlusion: yes," and dynamically reduces the dependence weight on visual features (such as shape and texture), while increasing the dependence weight on physical and signal dimension information. The model focuses on the actual physical size of the target estimated by the focal length-distance correlation analysis module (calculated result: 0.8 meters, this calculation process is not affected by foggy or hazy environments) and stable electronic reconnaissance signals (frequency: 5.8 GHz) for comprehensive judgment. Finally, the system bypasses low-quality visual interference and accurately outputs the target recognition result as "multi-rotor drone" with a confidence level of 96%, and determines the threat level as "medium risk" based on the combination of environmental and target attributes. This embodiment verifies the robustness of the present invention in complex environments and low-quality visual conditions by fusing environmental information and physical estimation features. Second Embodiment A second aspect of the present invention provides an anti-drone target recognition system based on multi-source information fusion, employing any of the aforementioned anti-drone target recognition methods based on multi-source information fusion, comprising: The hardware interface unit includes a radar interface module, an optoelectronic acquisition module, an electronic reconnaissance interface module, and an environmental information acquisition module, which are used to acquire target motion parameters, optoelectronic video data, electronic reconnaissance signals, and environmental information, respectively. The time synchronization unit is used to add high-precision timestamps to the data collected by the hardware interface unit to achieve time synchronization of multi-source auxiliary data. The data processing unit includes a target tracking module, an extraction and processing module, a spatiotemporal alignment module, and a fusion processing module. The target tracking module is used to obtain the motion parameters of the target object through wide-area search and precise tracking, and drives the photoelectric tracking module to track the target object based on the motion parameters. The extraction and processing module is used to complete key frame filtering and background removal processing based on the real-time video data collected by the photoelectric module to obtain the target foreground image. The spatiotemporal alignment module is used to perform spatiotemporal matching of multi-source auxiliary data and foreground image to obtain information unit data. The fusion processing module is used to construct structured fused information data based on the information unit data. The identification and evaluation unit includes a size calculation module, a scale normalization module, a multimodal recognition module, and a threat level determination module. The size calculation module is used to calculate the actual physical size of the target based on the principle of optical imaging. The scale normalization module is used to perform scale normalization processing on the foreground image to generate a normalized image. The multimodal recognition module is used to extract visual feature information based on the normalized image and fuse it with the actual physical size to complete the target identification.

[0043] By employing dual filtering based on variability and clarity, along with background removal, redundant and interfering data is effectively eliminated, significantly improving input data quality and system processing efficiency. Based on high-precision spatiotemporal matching and fusion of multi-source heterogeneous data, the system breaks down information silos from single sensors, providing rich multi-dimensional contextual information for target recognition. Combining physical size calculation with multimodal large-scale model feature extraction overcomes the scale variation challenge of traditional visual recognition, greatly improving the accuracy and robustness of target identification. By integrating multi-dimensional parameters to achieve intelligent threat level assessment, the system leaps from simple target recognition to battlefield situational awareness, providing scientific and precise support for countermeasure decision-making.

[0044] Through the collaborative work of the hardware interface unit and the time synchronization unit, heterogeneous data from radar, optoelectronics, and electronic reconnaissance are precisely aligned in the spatiotemporal dimension, solving the problem of fragmented and isolated information from a single sensor. The identification and evaluation unit innovatively integrates a size calculation module and a scale normalization module. By introducing actual physical dimensions based on optical principles as objective constraints and performing scale normalization processing on the visual input, it effectively overcomes the problem of abrupt changes in visual features caused by changes in distance and attitude of UAV targets, improving the identification accuracy and stability in dynamic scenes. A workflow from wide-area search, precise tracking, data processing to identification and evaluation is constructed, realizing a workflow of discovery-based identification and identification-based evaluation. Target type and threat level can be output without manual intervention, shortening response time and meeting the high timeliness requirements of counter-UAV operations. The extraction and processing modules in the data processing unit use a dual-screening mechanism to remove fuzzy and redundant frames and eliminate background interference, reducing computational load, ensuring that the multimodal identification model acquires target features, and improving the system's signal-to-noise ratio in complex backgrounds.

[0045] In the description of this application, it should be noted that the terms "inner" and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0046] It should also be noted that, unless otherwise explicitly specified and limited, the terms "setup" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0047] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific identification content executed by the system and device described above can be referred to the corresponding process in the foregoing method embodiments.

[0048] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Even if various changes are made to the present invention, if these changes fall within the scope of the claims of the present invention and their equivalents, they shall still fall within the protection scope of the present invention.

Claims

1. A method for anti-drone target identification based on multi-source information fusion, characterized in that, include: S1: Obtain the motion parameters of the target object through wide-area search and precise tracking, and drive the photoelectric tracking module to track the target object based on the motion parameters. The motion parameters include at least the target distance, target speed, and target orientation. S2: Based on the real-time video data collected by the photoelectric module, key frames are obtained by filtering the target object's degree of change and outline sharpness. Based on the key frames, the foreground image of the target object is obtained by background removal processing. S3: Real-time acquisition of multi-source auxiliary data, spatiotemporal matching of the multi-source auxiliary data and the foreground image based on high-precision timestamps, to obtain information unit data in spatiotemporal dimension. The information unit data is a collection of target foreground images and multi-source auxiliary data bound based on unified timestamps and spatial correlation. S4: Based on the information unit data, the foreground images are arranged in chronological order and the multi-source auxiliary data is mapped to the foreground images to obtain fused information data. The fused information data includes a time-series foreground image sequence, the photoelectric module parameters associated with the foreground images, radar signals, electronic reconnaissance signals, and structured data of environmental information. S5: Target recognition is completed based on the fused information data, specifically including parsing the labeled focal length parameters and target distance parameters from the fused information data, calculating the actual physical size of the target object through optical imaging principles, extracting visual feature information based on the foreground image through scale normalization processing, and performing feature matching and type confirmation based on the actual physical size and the visual feature information through a multimodal recognition model to obtain the target recognition result.

2. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, Step S2, filtering by the degree of change of the target object, further includes: Edge images are obtained through edge detection processing based on continuous data in the real-time video data. Calculate the structural similarity index between adjacent edge images; The difference degree is calculated based on the structural similarity index: if the difference degree exceeds a preset first threshold, the corresponding video data is retained as a candidate keyframe.

3. The anti-drone target identification method based on multi-source information fusion according to claim 2, characterized in that, Step S2, which involves obtaining keyframes through contour sharpness filtering, further includes: The ratio of the actual number of edge pixels to the theoretical total number of edge pixels in the candidate keyframe is calculated, and the ratio is used as the contour sharpness. The theoretical total number of edge pixels is obtained by estimating the perimeter of the target bounding box, using a contour fitting template, or by deriving from the segmentation results. If the outline sharpness is greater than or equal to a preset second threshold, the candidate keyframe is confirmed as a keyframe.

4. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, Step S2, obtaining the foreground image of the target object through background removal processing, further includes: Foreground extraction can be achieved through one or more of the following methods: object detection bounding box cropping, semantic segmentation, motion foreground detection, or instance segmentation.

5. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, Step S4, the step of constructing the fused information data, further includes: The fused information data is obtained by verifying and mapping the timestamp error based on the parameters of the optoelectronic module and the electronic reconnaissance signal. The parameters of the optoelectronic module include at least focal length and field of view; the electronic reconnaissance signal includes at least signal frequency and signal strength.

6. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, The expression for calculating the actual physical dimensions of the target object is: In the formula, s is the pixel width of the target object in the video data, s is the pixel size, D is the target distance, and f is the lens focal length.

7. The anti-drone target identification method based on multi-source information fusion according to claim 6, characterized in that, Step S5, which involves extracting visual feature information through scale normalization, further includes: Obtain a sequence of foreground images of the target object in consecutive frames, and extract the bounding box information of the foreground images; Based on the target distance parameter or actual physical size, calculate the scale variation factor of the current keyframe relative to the reference frame; The foreground image is processed by a spatial transformation network using the scale variation factor to generate a normalized image with consistent scale. The normalized image is input into the visual feature encoder of the multimodal recognition model for feature extraction to obtain visual feature information.

8. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, The multimodal recognition model in step S5 includes at least a visual feature encoder, an auxiliary data encoder, and a fusion inference module; the visual feature encoder is used to extract visual features from the normalized image, the auxiliary data encoder is used to parse the parametric features in the fused information data, and the fusion inference module is used to fuse visual features and parametric features to complete target feature matching and type confirmation.

9. The anti-drone target identification method based on multi-source information fusion according to claim 1, characterized in that, The wide-area search in step S1 is accomplished by radar or wide-angle optoelectronic equipment, and the precise tracking is accomplished by closed-loop control of a narrow-field optoelectronic turntable based on motion parameters.

10. A multi-source information fusion-based anti-drone target recognition system, employing the multi-source information fusion-based anti-drone target recognition method according to any one of claims 1-9, characterized in that, include: The hardware interface unit includes a radar interface module, an optoelectronic acquisition module, an electronic reconnaissance interface module, and an environmental information acquisition module, which are used to acquire target motion parameters, optoelectronic video data, electronic reconnaissance signals, and environmental information, respectively. The time synchronization unit is used to add a high-precision timestamp to the data collected by the hardware interface unit to achieve time synchronization of multi-source auxiliary data. The data processing unit includes a target tracking module, an extraction and processing module, a spatiotemporal alignment module, and a fusion processing module. The target tracking module is used to obtain the motion parameters of the target object through wide-area search and precise tracking, and drive the photoelectric tracking module to track the target object based on the motion parameters. The extraction and processing module is used to complete keyframe filtering and background removal processing based on the real-time video data collected by the photoelectric module to obtain the target foreground image. The spatiotemporal alignment module is used to perform spatiotemporal matching of multi-source auxiliary data and foreground image to obtain information unit data. The fusion processing module is used to construct structured fused information data based on the information unit data. The identification and evaluation unit includes a size calculation module, a scale normalization module, a multimodal recognition module, and a threat level determination module. The size calculation module is used to calculate the actual physical size of the target based on the principle of optical imaging. The scale normalization module is used to perform scale normalization processing on the foreground image to generate a normalized image. The multimodal recognition module is used to extract visual feature information based on the normalized image and fuse it with the actual physical size to complete the target identification.

Images