Small unmanned aerial vehicle identification and evaluation method and system based on onboard edge computing

Abstract

The application discloses a small unmanned aerial vehicle identification and evaluation method and system based on onboard edge computing, and the method is applied to a detection unmanned aerial vehicle carrying an edge computing device, is initialized in a detection mode, and comprises obtaining and preprocessing a target image; in the detection mode, a main path deep learning target detection network and a branch texture feature small target enhancement double-path parallel architecture are used for large-scale scanning, and a comprehensive detection result is obtained through optical flow and epipolar geometry cascade denoising; threat evaluation is performed on the unmanned aerial vehicle target based on a time sequence IOU association mechanism; after a high-threat target is confirmed, the detection mode is switched to a tracking mode, the high-threat "black flight" is locked and tracked, an alarm output is generated to the target end, and the detection mode is periodically switched back based on a verification trigger mechanism. Through the switching of the detection / tracking mode, the large-scale scanning and the continuous tracking of the high-threat target are considered under the limited onboard edge computing power, the small target detection capability is improved, and the false detection rate and the power consumption are reduced.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) identification technology, and in particular to a method and system for identifying and evaluating small UAVs based on airborne edge computing. Background Technology

[0002] With the widespread adoption of consumer and industrial drones, the phenomenon of unauthorized and indiscriminate flights by "low-altitude, slow-speed, and small" drones (drones with low altitude, slow speed, and small size) is becoming increasingly serious, posing a severe security threat to important targets such as airports, nuclear power plants, large event venues, and military restricted areas. In recent years, incidents of flight delays, privacy breaches, and even security risks caused by drone intrusions have occurred frequently, making counter-drone technology a major technological need in the field of public safety.

[0003] Current mainstream counter-drone methods mainly include radar detection, radio spectrum monitoring, electro-optical tracking, navigation deception, laser strikes, and jamming suppression. However, for the detection and countermeasures against small drones (especially "low, slow, and small" targets), existing technologies still have the following significant shortcomings: 1. Single detection methods have blind spots, making accurate identification across all scenarios difficult: Radar detection is susceptible to ground clutter interference and has weak detection capabilities for hovering or low-speed flying low-altitude small targets; radio spectrum monitoring can only identify drones with known communication protocols, failing to detect drones flying autonomously or changing frequencies; traditional optoelectronic equipment, while providing intuitive images, is usually deployed at fixed ground locations, and is severely limited by the field of view and obstructed by buildings, resulting in detection blind spots and making it difficult to achieve omnidirectional and continuous tracking and locking of low-altitude intruding drones. Simply combining multiple detection methods not only increases system complexity and cost, but also lacks the intelligent processing capabilities for multi-source data fusion and accurate target attribute identification at the edge.

[0004] 2. Existing airborne systems face limitations in computing power and power consumption, making it difficult to deploy efficient identification algorithms: Counter-drone scenarios require identification algorithms with high accuracy, low latency, and multi-category recognition capabilities. However, airborne edge computing devices suitable for small drones are generally limited by weight, size, and power consumption, making it difficult to simultaneously meet the dual requirements of computing power and battery life. Deploying lightweight identification models results in high false positive and false negative rates in complex backgrounds (such as urban building clusters or forests); deploying high-precision deep neural network models leads to large computational latency and high power consumption, severely impacting the drone's own battery life and the real-time performance of tracking response. How to achieve efficient and accurate drone target identification with limited airborne edge computing power is a pressing technical challenge that needs to be addressed.

[0005] 3. Inability to cope with changing environments: Drones and other aerial platforms possess highly flexible maneuverability, enabling them to observe targets from various perspectives during missions, including top-down, eye-level, and oblique views, as well as close-up and distant views at different distances. However, existing target detection algorithms are typically trained and optimized only for one or a subset of these perspectives, making it difficult to effectively cover all viewpoints. The fundamental reason is that the flexible observation methods of aerial platforms make it difficult to comprehensively acquire training data for each perspective, thus limiting the model's generalization ability under diverse observation conditions. Summary of the Invention

[0006] The main purpose of this application is to provide a method and system for identifying and evaluating small unmanned aerial vehicles (UAVs) based on airborne edge computing, in order to solve the technical problem of how to balance the large-scale scanning and identification of potential small UAVs in the target area with the continuous and stable tracking of high-threat "black flights" under limited airborne edge computing power, and to switch between the two as needed to save computing power and ensure real-time performance.

[0007] To achieve the above objectives, firstly, this application proposes a method for identifying and evaluating small unmanned aerial vehicles (UAVs) based on airborne edge computing, applied to UAVs equipped with edge computing devices, initially in detection mode, including the following steps: Acquire the target image of the target region and perform preprocessing; In the detection mode, a dual-path parallel detection architecture of main path detection and branch path enhancement detection is adopted to identify potential UAV targets in the preprocessed target image and obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhancement detection mechanism based on texture features. A cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints is used to confirm the moving targets in the candidate target set and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; If there are abnormally moving targets in the target area, a target threat assessment is performed on the abnormally moving targets based on a preset temporal intersection-union-comparison correlation mechanism; When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

[0008] Furthermore, the improved deep learning object detection network is an improved YOLOv8 architecture, which is configured with a piecewise linear size-adaptive confidence threshold function. The piecewise linear size-adaptive confidence threshold function is configured such that the smaller the detection box width, the lower the confidence threshold, so as to retain small targets with weak activation signals; and the larger the detection box width, the higher the confidence threshold, so as to suppress false detections of large target regions. The method for main path detection using the improved deep learning object detection network includes: The target image is preprocessed using an anchor-free detection head and a distributed focal loss (DFL) bounding box regression mechanism. Multi-scale detection box decoding is performed on feature maps of different scales to obtain the initial detection boxes and their detection confidence for each scale. For each initial detection box, based on the width of the initial detection box, the piecewise linear size adaptive confidence threshold function is queried to obtain a dynamic confidence threshold that matches the size of the initial detection box; The initial detection box is filtered using the dynamic confidence threshold. The detection confidence of the initial detection box is compared with the dynamic confidence threshold. If the detection confidence is not less than the dynamic confidence threshold, the initial detection box is determined to pass the filter and is used as the first candidate target. If the detection confidence is less than the dynamic confidence threshold, the initial detection box is determined to fail the filter and is removed. The first candidate target is used as the main road detection result for the main road, so as to be merged with the branch road detection results.

[0009] Furthermore, the method for main path detection using an improved deep learning object detection network also includes: A multi-threaded parallel architecture is used for multi-scale feature map decoding. For each step feature map, each thread reads data from the corresponding output feature map memory area and independently generates candidate boxes. In response to the completion of candidate box generation by all threads, candidate boxes of each step size are merged and then subjected to non-maximum suppression (NMS) processing.

[0010] Furthermore, a method for branch enhancement detection using a small target enhancement detection mechanism based on texture features includes: Locate the texture-salient regions in the target image, perform small target candidate box decoding only in the texture-salient regions, and combine multi-scale integral map texture verification to filter out false detections of background texture to obtain the second candidate target.

[0011] Further, the method for locating texture-salient regions in the target image, performing small target candidate box decoding only in the texture-salient regions, and simultaneously combining multi-scale integral image texture verification to filter out false detections of background texture, to obtain a second candidate target, includes: Gradient calculation is performed on the preprocessed target image to extract texture edge features and generate a texture intensity response map; Based on the texture intensity response map, candidate texture regions are obtained by dual threshold segmentation. Based on the integral image, fast response calculation is performed on each candidate texture region, and background noise texture is filtered out by sparse texture constraint to obtain the second candidate target. When the target image texture information is insufficient or the texture intensity response map cannot provide effective anchor points, a gray-level fallback supplementation strategy is adopted to generate supplementary candidate targets based on the image gray-level features. These supplementary candidate targets are then merged with the second candidate target and used to merge with the main path detection results.

[0012] Furthermore, a method for confirming the movement of candidate targets in a candidate target set and obtaining a comprehensive detection result of whether there are abnormal moving targets in the target region, based on a cascaded denoising algorithm of optical flow tracking and epipolar geometry constraints, includes: For candidate targets in the candidate target set, the following processing is performed between adjacent consecutive frames: Using the Lucas-Kanade optical flow algorithm, tracking feature points are selected by downsampling within the actual area of ​​the candidate target according to a preset grid, and the tracking feature points are tracked from the current frame to the adjacent frame to establish inter-frame point pair matching of the candidate target; Pre-screening of inter-frame point pair matching based on adaptive reprojection error: calculate the motion consistency residual of each inter-frame point pair matching, dynamically determine the adaptive pre-screening threshold based on the statistics of the motion consistency residual, and identify point pairs with residuals greater than the adaptive pre-screening threshold as outliers and remove them to obtain pre-screened matching point pairs. Based on the epipolar geometric constraints of the fundamental matrix and homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets and obtain the comprehensive detection results.

[0013] Furthermore, based on the epipolar geometric constraints of the fundamental matrix and the homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets, and a method for obtaining the comprehensive detection result includes: Simultaneously, the fundamental matrix F and homography matrix H of the pre-selected matching point pairs are estimated, wherein the fundamental matrix satisfies epipolar geometric constraints. , For the correct matching point pair; The epipolar distance residual of each pair of matching points is calculated based on the basic matrix F. An adaptive reprojection threshold is obtained based on the statistics of the epipolar distance residual and dynamically adjusted by exponential moving average. Pairs of points whose epipolar distance residual exceeds the adaptive reprojection threshold are identified as outliers. For each pair of matching points, the votes of the interior / exterior points in the two models, the basic matrix F and the homography matrix H, are counted, and the geometric consistency of the pair of matching points is confirmed based on the voting results of the dual models. When the number of interior points that pass geometric verification in the pre-screened matching point pairs of a candidate target reaches a set proportion, the candidate target is confirmed as a real moving target.

[0014] Secondly, this application also proposes a small unmanned aerial vehicle (UAV) identification and evaluation system based on airborne edge computing, applied to the detection of UAVs equipped with edge computing devices, initially in detection mode, including: The image acquisition and preprocessing module is used to acquire the target image of the target area and perform preprocessing. A dual-path parallel detection module is used to identify potential UAV targets in the preprocessed target image by adopting a dual-path parallel detection architecture of main path detection and branch path enhanced detection in detection mode, and to obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhanced detection mechanism based on texture features. The motion matching denoising module is used to confirm the motion targets in the candidate target set by a cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints, and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; The data fusion and decision-making module is used to assess the threat of abnormal moving targets based on a preset temporal cross-comparison and comparison correlation mechanism if there are abnormal moving targets in the target area. When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

[0015] Thirdly, this application also proposes an electronic device, including a processor and a memory, wherein the memory stores a computer program; when the processor executes the computer program, it implements the small unmanned aerial vehicle identification and evaluation method based on airborne edge computing as described in any one of the first aspects.

[0016] Fourthly, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the small unmanned aerial vehicle (UAV) identification and evaluation method based on airborne edge computing as described in any one of the first aspects.

[0017] The method and system for identifying and evaluating small unmanned aerial vehicles (UAVs) based on airborne edge computing provided in this application enable "local detection (large-area scanning)" to operate only in detection mode by switching between detection and tracking modes on demand. Once a high-threat "black flight" is confirmed, the system switches to tracking mode and pauses large-area detection. This frees up limited airborne edge computing power from "indiscriminate scanning of the entire target area" and concentrates it on the continuous locking and tracking of a single high-threat target. This achieves efficient allocation of computing power in the "UAV counter-UAV" scenario. Compared with the prior art, this invention has the following beneficial effects: (1) The dual-path parallel detection architecture balances accuracy and recall: the main path is based on a deep learning target detection network to ensure general detection capability, while the branch path is based on a small target enhancement detection mechanism based on texture features to enhance the characteristics of "low, slow and small" targets with weak texture and small proportion. The combination of the two effectively improves the recall rate of small targets and reduces false negatives and false positives in complex backgrounds.

[0018] (2) Cascaded denoising combines optical flow and epipolar geometry constraints to filter out false detections caused by background motion and clutter, thereby improving the accuracy and robustness of moving target identification.

[0019] (3) The computing power switching logic between detection mode and tracking mode enables the system to stop large-scale detection and focus on tracking after confirming high-threat "black flight", which significantly reduces the computing power and power consumption at the edge, extends the endurance of the detection drone itself and improves the real-time performance of tracking response. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a first flowchart illustrating an embodiment of the method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing in this application. Figure 2 This is a second flowchart illustrating an embodiment of the method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing in this application. Figure 3 This is a schematic diagram of a structure provided for an embodiment of the small unmanned aerial vehicle (UAV) identification and evaluation system based on airborne edge computing in this application; Figure 4 This is a schematic diagram of the structure of an embodiment of the electronic device in this application.

[0023] Explanation of icon numbers: 10. Memory; 20. Processor.

[0024] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Example 1

[0028] This application provides a method for identifying and evaluating small unmanned aerial vehicles (UAVs) based on airborne edge computing. It is applied to UAVs equipped with edge computing devices, initially in detection mode, and refers to... Figure 1 This includes the following steps: Acquire the target image of the target region and perform preprocessing; In the detection mode, a dual-path parallel detection architecture of main path detection and branch path enhancement detection is adopted to identify potential UAV targets in the preprocessed target image and obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhancement detection mechanism based on texture features. A cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints is used to confirm the moving targets in the candidate target set and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; If there are abnormally moving targets in the target area, a target threat assessment is performed on the abnormally moving targets based on a preset temporal intersection-union-comparison correlation mechanism; When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

[0029] Specifically, in this embodiment, target images are acquired by integrating a visible light camera and an infrared camera on the detection drone; Visible light cameras are used to acquire high-resolution color images, providing rich details of the target's appearance, while infrared cameras can capture the target's thermal radiation characteristics, enabling detection at night and in low light conditions. By working together, these two sensors can collect optical information of the target area from all directions and transmit it in real time to edge computing devices and ground data processing centers. Preprocessing includes temporal synchronization and spatial registration of visible light and infrared images, and fusion preprocessing of the two based on scene brightness. When the ambient light is sufficient, the visible light channel is used as the main channel and the infrared channel is used as the auxiliary channel. When the light is insufficient or the contrast between the target and the background is low, the infrared channel is automatically switched to be the main channel, thereby improving the robustness of subsequent local detection under complex lighting conditions. Preprocessing also includes conventional operations such as image denoising, region of interest cropping and normalization.

[0030] In this embodiment, in order to achieve efficient and accurate UAV target recognition with limited airborne edge computing power, a large-scale potential UAV target detection is first performed in detection mode to confirm the presence of abnormal moving targets in the target area. Then, a target threat assessment is performed on the abnormal moving targets. For high-threat targets with "black flight" or "random flight" phenomena, the detection mode is switched to tracking mode. After the target is stably captured, a single target tracker based on a twin network is automatically started to continuously track the high-threat target at a higher frame rate. Among them, abnormal motion targets refer to targets that have been confirmed to have abnormal motion and are suspected to be drones. Further threat assessment will be conducted based on appearance confidence, thermal characteristics, and temporal correlation threat scores. The comprehensive threat score will be used to determine whether it is a drone and whether it is a high-threat target, i.e., an illegal drone.

[0031] In this embodiment, the application adopts an automatic switching mechanism for threat assessment and detection tracking based on time-series correlation; In terms of threat assessment, the detection box of each frame is temporally cross-connected with the detection result of the previous frame. The number of consecutive hit frames of the target is accumulated. The threat score is calculated by combining the target size ratio and the detection confidence. When the number of consecutive hit frames exceeds the threshold and the threat score exceeds the threshold, it is judged as a high threat target. Meanwhile, based on the preset temporal intersection-comparison correlation mechanism, a periodic verification mechanism is also set up. Every preset interval, the detection mode is switched back to verify the tracking results, or the detection mode is automatically switched back when the tracking confidence drops sharply relative to the previous frame and exceeds a set threshold, so as to ensure tracking reliability.

[0032] Furthermore, in one embodiment, the data acquired by the optical acquisition module and the small UAV intelligent recognition algorithm module are fused and processed to comprehensively analyze the target's appearance features, thermal characteristics, motion status and other multi-dimensional information in multiple frames of images, determine the target's threat level, issue an alarm to the operator based on the threat level, and transmit the target information to the command and control system to assist in subsequent decision-making. For high-threat targets identified by the data fusion and decision-making module, the target locking procedure is immediately initiated. Based on parameters such as the target's position, size, and direction of movement, the stabilization platform of the optoelectronic pod is controlled to precisely adjust the direction of the optical acquisition module, ensuring that the target remains in the center of the field of view. Simultaneously, the field-of-view image of the optoelectronic pod, target labeling information, system status, alarm prompts, and other target information will be sent to the human-machine interface at the target end. This allows operators to manually intervene in the system as needed, such as adjusting detection parameters, switching sensor modes, and performing secondary confirmation of suspected targets, thereby achieving human-machine collaborative operation and improving the system's flexibility and reliability.

[0033] Furthermore, in one embodiment, the improved deep learning object detection network is an improved YOLOv8 architecture, which is configured with a piecewise linear size-adaptive confidence threshold function. The piecewise linear size-adaptive confidence threshold function is configured such that the smaller the detection box width, the lower the confidence threshold, so as to retain small targets with weak activation signals; and the larger the detection box width, the higher the confidence threshold, so as to suppress false detections of large target regions. The method for main path detection using the improved deep learning object detection network includes: The target image is preprocessed using an anchor-free detection head and a distributed focal loss (DFL) bounding box regression mechanism. Multi-scale detection box decoding is performed on feature maps of different scales to obtain the initial detection boxes and their detection confidence for each scale. For each initial detection box, based on the width of the initial detection box, the piecewise linear size adaptive confidence threshold function is queried to obtain a dynamic confidence threshold that matches the size of the initial detection box; The initial detection box is filtered using the dynamic confidence threshold. The detection confidence of the initial detection box is compared with the dynamic confidence threshold. If the detection confidence is not less than the dynamic confidence threshold, the initial detection box is determined to pass the filter and is used as the first candidate target. If the detection confidence is less than the dynamic confidence threshold, the initial detection box is determined to fail the filter and is removed. The first candidate target is used as the main road detection result for the main road, so as to be merged with the branch road detection results.

[0034] Specifically, in this embodiment, the main path adopts a deep learning-based multi-scale object detection network (YOLOv8 architecture). Unlike traditional anchor-based detection methods, YOLOv8 uses an anchor-free detection head design, and its core detection principle is as follows: Bounding box regression mechanism: Traditional YOLO series bounding box regression directly predicts continuous offsets (dx, dy, dw, dh), while YOLOv8 models bounding box regression as a discretized probability distribution prediction problem. Specifically, for each bounding box's four boundaries (left, top, right, and bottom), the network does not directly output continuous boundary distance values ​​at each feature map location. Instead, it outputs 16 discretized bin values, with each bin representing the probability that the boundary falls on the corresponding discrete distance. During the decoding stage, the 16 bin values ​​are normalized into a probability distribution using the softmax function, and then the expected value of this distribution is multiplied by the step size of the current feature map to obtain continuous boundary distance values. The core advantage of this DFL regression mechanism is that it transforms the continuous regression problem into a combination of discrete classification and expected regression, avoiding the dependence of anchor-based methods on the size and proportion of prior boxes. This allows the model to adapt more flexibly to the diverse scales and aspect ratios of UAV targets at different distances and from different perspectives, which is especially beneficial for targets with extremely large scale variations, such as small UAVs.

[0035] Multi-scale feature map pyramid: The YOLOv8 network decodes bounding boxes on three feature maps with different strides, each corresponding to a different receptive field size: stride=8 (shallow high-resolution feature map): small receptive field, good at detecting small targets (long-range drones); stride=16 / 32 (middle layer feature map): moderate receptive field, suitable for detecting medium-scale targets; stride=64 / 128 (deep low-resolution feature map): large receptive field, good for detecting large targets at close range.

[0036] Meanwhile, in order to further address the specific needs of target detection for small unmanned aerial vehicles, this embodiment performs piecewise linear adaptive size adjustment on the confidence threshold of the detection box at different scales. In the standard YOLOv8 detection process, detection boxes of all scales share a uniform confidence threshold. This makes the activation signal of small targets (such as long-range drones) naturally weaker than that of large targets in the feature map. Using a higher threshold will cause a large number of small targets to be incorrectly filtered out; while using a lower threshold will cause a large number of false detections in the large target area. Therefore, the implementation method sets a piecewise linear size-adaptive confidence threshold function, which dynamically adjusts the confidence threshold according to the width of the detection box:

[0037] Because small drones have a small projected area in the image, their detection signals are weak and susceptible to noise interference. Therefore, an extremely low confidence threshold is used to ensure that no potential targets are missed. As the detection box size increases, the target signal strength increases and the sources of false detection decrease. The confidence threshold is gradually increased to control the false detection rate. This piecewise linear function achieves a fine balance between high recall for small targets and low false positives for large targets. Compared with a uniform fixed threshold strategy, it significantly improves the recall rate while keeping the false positive rate controllable in the detection of small drones.

[0038] Furthermore, methods for main path detection using improved deep learning object detection networks also include: A multi-threaded parallel architecture is used for multi-scale feature map decoding. For each step feature map, each thread reads data from the corresponding output feature map memory area and independently generates candidate boxes. In response to the completion of candidate box generation by all threads, candidate boxes of each step size are merged and then subjected to non-maximum suppression (NMS) processing.

[0039] To fully utilize the multi-core NPU / CPU resources of the airborne edge computing platform, this application designs the multi-scale feature map decoding process of YOLOv8 as a multi-threaded parallel architecture. For each step feature map, each thread reads data from the corresponding output feature map memory area and independently completes the generation of candidate boxes. After all threads have completed, the candidate boxes of each step are merged and sent to the subsequent NMS processing.

[0040] Furthermore, in this embodiment, the method for branch enhancement detection using a small target enhancement detection mechanism based on texture features includes: Locate the texture-salient regions in the target image, perform small target candidate box decoding only in the texture-salient regions, and combine multi-scale integral map texture verification to filter out false detections of background texture to obtain the second candidate target.

[0041] Furthermore, the method for locating texture-salient regions in the target image, performing small target candidate box decoding only in the texture-salient regions, and simultaneously combining multi-scale integral image texture verification to filter out false detections of background texture, to obtain a second candidate target, includes: Gradient calculation is performed on the preprocessed target image to extract texture edge features and generate a texture intensity response map; Based on the texture intensity response map, candidate texture regions are obtained by dual threshold segmentation. Based on the integral image, fast response calculation is performed on each candidate texture region, and background noise texture is filtered out by sparse texture constraint to obtain the second candidate target. When the target image texture information is insufficient or the texture intensity response map cannot provide effective anchor points, a gray-level fallback supplementation strategy is adopted to generate supplementary candidate targets based on the image gray-level features. These supplementary candidate targets are then merged with the second candidate target and used to merge with the main path detection results.

[0042] Specifically, the branch uses a texture-guided small target enhancement detection mechanism to compensate for the main path's shortcomings in detecting small UAV targets. Although the main path's YOLOv8 detection network has good detection performance on abnormally moving UAV targets at normal scale, it has the following inherent limitations on small UAV targets (occupying only 8-64 pixels in the image): 1. Weak feature response: The activation region of small targets in deep feature maps is extremely small or even disappears (especially in feature maps with a stride of 32 or more), which makes it impossible for deep features to effectively represent small targets.

[0043] 2. Background confusion: Small drones are easily confused with background textures (such as urban building textures and tree tree textures) in complex backgrounds, resulting in a large number of false detections.

[0044] 3. Full-image brute-force search is not feasible: If candidate box decoding is performed on all anchor points in the feature map, the computational load is too large (a feature map with a stride of 8 has 80×80=6400 anchor points with a 640×640 input), which cannot be completed in real time on airborne edge devices.

[0045] To address the aforementioned issues, this application designs a texture-guided small target enhancement detection branch. It utilizes an independent texture analysis model to locate texture-salient regions in the image and performs small target candidate box decoding only within these regions, significantly reducing computational cost. Furthermore, it combines multi-scale integral image texture verification to further filter out false detections of background textures, as detailed below: 1. Solve for the x-axis and y-axis gradients using the Sobel operator and normalize them to obtain the texture intensity response map. Align the resolution size with the YOLOv8 feature map that is downsampled by 8 times and has a stride of 8.

[0046] 2. Texture Mask Generation Algorithm: Texture mask generation is the core preprocessing step in this branch, including: (1) Dual threshold binarization: For each pixel of the texture intensity response map, a binary map is generated based on a preset low threshold (th_low=20.0) and a high threshold (th_high=40.0).

[0047] (2) Integral image pre-computation: Integral images are calculated for high and low thresholds respectively; The introduction of integral images is a key performance optimization method: Pixel statistics within any rectangular area (such as "the number of pixels with high_bin set to 1 in the area") can be completed in O(1) time through 4 memory accesses and 3 addition and subtraction operations. Compared with the O(w×h) complexity of brute-force traversal, the computational cost of regional texture statistics is reduced from thousands of operations to single-digit operations.

[0048] (3) Low threshold texture anchor point screening: Scan each pixel of the texture response map, and add pixels that meet the low threshold to the low threshold anchor point list.

[0049] (4) High-threshold texture anchor point screening (texture-guided anchor points): High-texture anchor points must be located in positions where the texture intensity is high enough, the texture pattern is not too dense (avoiding complex texture backgrounds), and the texture pattern has local sparsity in space. This sparse texture constraint is a key design to suppress false detection of complex backgrounds such as urban building clusters and trees - complex backgrounds usually exhibit dense high-texture responses, while small drones (especially fixed-wing or quadcopter drones) present as areas with moderate local texture contrast in images due to their relatively uniform surface material.

[0050] Meanwhile, in one implementation, a method for confirming moving targets in a candidate target set based on a cascaded denoising algorithm of optical flow tracking and epipolar geometry constraints, and obtaining a comprehensive detection result of whether there are abnormal moving targets in the target region, includes: For candidate targets in the candidate target set, the following processing is performed between adjacent consecutive frames: Using the Lucas-Kanade optical flow algorithm, tracking feature points are selected by downsampling within the actual area of ​​the candidate target according to a preset grid, and the tracking feature points are tracked from the current frame to the adjacent frame to establish inter-frame point pair matching of the candidate target; Pre-screening of inter-frame point pair matching based on adaptive reprojection error: calculate the motion consistency residual of each inter-frame point pair matching, dynamically determine the adaptive pre-screening threshold based on the statistics of the motion consistency residual, and identify point pairs with residuals greater than the adaptive pre-screening threshold as outliers and remove them to obtain pre-screened matching point pairs. Based on the epipolar geometric constraints of the fundamental matrix and homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets and obtain the comprehensive detection results.

[0051] Furthermore, based on the epipolar geometric constraints of the fundamental matrix and the homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets, and a method for obtaining the comprehensive detection result includes: Simultaneously, the fundamental matrix F and homography matrix H of the pre-selected matching point pairs are estimated, wherein the fundamental matrix satisfies epipolar geometric constraints. , For the correct matching point pair; The epipolar distance residual of each pair of matching points is calculated based on the basic matrix F. An adaptive reprojection threshold is obtained based on the statistics of the epipolar distance residual and dynamically adjusted by exponential moving average. Pairs of points whose epipolar distance residual exceeds the adaptive reprojection threshold are identified as outliers. For each pair of matching points, the votes of the interior / exterior points in the two models, the basic matrix F and the homography matrix H, are counted, and the geometric consistency of the pair of matching points is confirmed based on the voting results of the dual models. When the number of interior points that pass geometric verification in the pre-screened matching point pairs of a candidate target reaches a set proportion, the candidate target is confirmed as a real moving target.

[0052] Specifically, in this embodiment, for the candidate targets output by branch detection, a cascaded denoising algorithm based on optical flow tracking and geometric constraints is used to confirm the moving targets. The matching point pairs pre-screened by LK error are input into the epipolar geometric outlier detection module. The core of this module is to simultaneously estimate the fundamental matrix (F) and the homography matrix (H), and to determine outliers through the RANSAC voting mechanism of the dual geometric model. The specific steps are as follows: 1. Fundamental Matrix (F) Estimation: The fundamental matrix F describes the epipolar geometric constraints of inter-frame matched point pairs: for correct matched point pairs... It should meet (i.e., point) Landing point (on the epipolar line in the second frame). The RANSAC estimation strategy for the fundamental matrix includes adaptive iteration number selection: Round 1: maxIters=1, quickly detect the inlier ratio; If the inlier ratio is < 70% (kMinInlierRatio=0.7): this indicates poor data quality. Increase the number of iterations to maxIters=20 to improve estimation accuracy. If the inlier rate is ≥ 70%, it indicates that the data quality is good. Use maxIters=200 for finer estimation to obtain a more accurate F matrix and inlier mask.

[0053] 2. Adaptive RANSAC reprojection threshold: Traditional RANSAC uses a fixed reprojection threshold (e.g., 3.0 pixels). However, in UAV-borne scenarios, due to the motion of the flight platform itself (vibration, attitude changes, altitude changes), the epipolar geometric residual distribution of inter-frame matching points is dynamically changing—the residual is small when the platform is stable and large when the platform is maneuvering. This implementation design proposes an adaptive threshold adjustment mechanism based on epipolar residual statistics: (1) Using the estimated fundamental matrix F, calculate the epipolar distance residual for each pair of matching points.

[0054] (2) Sort the residuals, take the data within the first 75th percentile, and then take the median.

[0055] (3) Dynamic adjustment is achieved through exponential moving average (EMA), using an asymmetric smoothing coefficient. The asymmetric EMA strategy is as follows: Rapid tightening (weight 0.3): When the residual decreases (e.g., when the platform enters a stable hovering state), the threshold should be rapidly reduced to provide stricter geometric constraints and avoid mismatches from being mixed in; Slow relaxation (weight 0.1): When the residual increases (which may be due to momentary platform vibration or airflow disturbance), the threshold should be increased slowly to avoid excessively relaxing geometric constraints due to single-frame disturbances and introducing a large number of mismatches.

[0056] (4) Dual-model voting mechanism: For each pair of matching points, calculate its inlier / outlier votes in both the F and H models; The advantage of this dual voting mechanism over the single F matrix method is that some matching point pairs whose F matrix estimation is unstable due to scene structure degradation (such as texture repetition, planar scene) can obtain correct geometric verification through the H matrix, thus avoiding the lack of robustness of the single model.

[0057] Specifically, such as Figure 2 The diagram illustrates the core workflow of this embodiment. The motion detection algorithm in the figure corresponds to a dual-path parallel detection architecture of main path detection and branch path enhancement. The main path is YOLOv8 multi-scale target detection, and the branch path is texture-guided small target enhancement detection (including texture feature extraction, anchor point grid downsampling, small target candidate box generation, grayscale bottom-up supplementation, etc.). The results of the two paths are merged and sent to the motion matching denoising module. Motion matching denoising uses LK pyramid optical flow tracking to establish inter-frame point pair matching, and performs cascaded denoising through adaptive error pre-screening and epipolar geometry outlier detection (dual RANSAC voting of the fundamental matrix and homography matrix). Finally, it confirms the final output through temporal intersection-union correlation filtering. The motion tracking algorithm in the figure corresponds to the automatic switching mechanism of detection and tracking. After the detection module stably captures the target, it starts the single target tracker (SOT) based on the Siamese network. The tracking reliability is ensured through periodic verification and confidence drop detection, and the target's motion trajectory, speed and other information are output.

[0058] In summary, this embodiment has the following beneficial effects: 1. Significantly improved small target detection capability: By enhancing the detection mechanism through texture-guided branching, combined with texture feature extraction, anchor point grid downsampling and grayscale fallback strategy, small drone targets of 8 to 64 pixels can be effectively detected with limited edge computing power, greatly reducing the false negative rate.

[0059] 2. Significantly reduced false alarm rate: By using a cascaded denoising algorithm based on LK pyramid optical flow tracking and epipolar geometry outlier detection (based on the fundamental matrix and homography matrix dual RANSAC voting), combined with adaptive reprojection threshold and temporal cross-union ratio correlation filtering, it effectively distinguishes between real moving targets and background false alarms, significantly reducing the false alarm rate in complex scenarios; the automatic switching mechanism between detection and tracking further ensures continuous and stable locking of high-threat targets.

[0060] Example 2

[0061] This application provides a small unmanned aerial vehicle (UAV) identification and evaluation system based on airborne edge computing, applied to the detection of UAVs equipped with edge computing devices. The system is initially in detection mode, including: The image acquisition and preprocessing module is used to acquire the target image of the target area and perform preprocessing. A dual-path parallel detection module is used to identify potential UAV targets in the preprocessed target image by adopting a dual-path parallel detection architecture of main path detection and branch path enhanced detection in detection mode, and to obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhanced detection mechanism based on texture features. The motion matching denoising module is used to confirm the motion targets in the candidate target set by a cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints, and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; The data fusion and decision-making module is used to assess the threat of abnormal moving targets based on a preset temporal cross-comparison and comparison correlation mechanism if there are abnormal moving targets in the target area. When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

[0062] Specifically, such as Figure 3 As shown, the video stream captured by the visible light camera is sent to the intelligent recognition algorithm module of the small UAV (corresponding to the dual-path architecture of main path detection and branch path enhancement and motion matching denoising); the infrared data captured by the infrared camera is sent to the data fusion and decision module; the data fusion and decision module integrates appearance confidence, thermal features and time-series associated threat scores, and outputs alarm signals or starts tracking commands; after receiving the command, the target locking and tracking module starts a single target tracker (SOT) based on twin network to continuously lock, and periodically switches back to the detection mode to verify the tracking results; finally, the target information (location, trajectory, threat level) is displayed in real time on the ground-based human-machine interface.

[0063] Example 3

[0064] This application also provides an electronic device, such as Figure 4 As shown, it includes a processor and a memory, the memory storing a computer program; when the processor executes the computer program, it implements the small unmanned aerial vehicle identification and evaluation method based on airborne edge computing as described in Embodiment 1.

[0065] Example 4

[0066] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the small unmanned aerial vehicle (UAV) identification and evaluation method based on airborne edge computing as described in Embodiment 1.

[0067] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying and evaluating small unmanned aerial vehicles (UAVs) based on airborne edge computing, characterized in that, For detection drones equipped with edge computing devices, the initial setup in detection mode includes the following steps: Acquire the target image of the target region and perform preprocessing; In the detection mode, a dual-path parallel detection architecture of main path detection and branch path enhancement detection is adopted to identify potential UAV targets in the preprocessed target image and obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhancement detection mechanism based on texture features. A cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints is used to confirm the moving targets in the candidate target set and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; If there are abnormally moving targets in the target area, a target threat assessment is performed on the abnormally moving targets based on a preset temporal intersection-union-comparison correlation mechanism; When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

2. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 1, characterized in that, The improved deep learning object detection network is an improved YOLOv8 architecture. The improved YOLOv8 architecture is configured with a piecewise linear size-adaptive confidence threshold function. The piecewise linear size-adaptive confidence threshold function is configured such that the smaller the detection box width, the lower the confidence threshold, so as to retain small targets with weak activation signals; and the larger the detection box width, the higher the confidence threshold, so as to suppress false detections of large target regions. The method for main path detection using the improved deep learning object detection network includes: The target image is preprocessed using an anchor-free detection head and a distributed focal loss (DFL) bounding box regression mechanism. Multi-scale detection box decoding is performed on feature maps of different scales to obtain the initial detection boxes and their detection confidence for each scale. For each initial detection box, based on the width of the initial detection box, the piecewise linear size adaptive confidence threshold function is queried to obtain a dynamic confidence threshold that matches the size of the initial detection box; The initial detection box is filtered using the dynamic confidence threshold. The detection confidence of the initial detection box is compared with the dynamic confidence threshold. If the detection confidence is not less than the dynamic confidence threshold, the initial detection box is determined to pass the filter and is used as the first candidate target. If the detection confidence is less than the dynamic confidence threshold, the initial detection box is determined to fail the filter and is removed. The first candidate target is used as the main road detection result for the main road, so as to be merged with the branch road detection results.

3. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 2, characterized in that, Methods for main path detection using improved deep learning object detection networks also include: A multi-threaded parallel architecture is used for multi-scale feature map decoding. For each step feature map, each thread reads data from the corresponding output feature map memory area and independently generates candidate boxes. In response to the completion of candidate box generation by all threads, candidate boxes of each step size are merged and then subjected to non-maximum suppression (NMS) processing.

4. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 2, characterized in that, Methods for branch enhancement detection using a small target enhancement detection mechanism based on texture features include: Locate the texture-salient regions in the target image, perform small target candidate box decoding only in the texture-salient regions, and combine multi-scale integral map texture verification to filter out false detections of background texture to obtain the second candidate target.

5. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 4, characterized in that, A method for locating texture-salient regions in a target image, performing small target candidate box decoding only within these regions, and combining multi-scale integral image texture verification to filter out false background texture detections, to obtain a second candidate target, includes: Gradient calculation is performed on the preprocessed target image to extract texture edge features and generate a texture intensity response map; Based on the texture intensity response map, candidate texture regions are obtained by dual threshold segmentation. Based on the integral image, fast response calculation is performed on each candidate texture region, and background noise texture is filtered out by sparse texture constraint to obtain the second candidate target. When the target image texture information is insufficient or the texture intensity response map cannot provide effective anchor points, a gray-level fallback supplementation strategy is adopted to generate supplementary candidate targets based on the image gray-level features. These supplementary candidate targets are then merged with the second candidate target and used to merge with the main path detection results.

6. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 1, characterized in that, A method for identifying moving targets in a target region by using a cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints, and obtaining a comprehensive detection result for the presence of abnormal moving targets in the target region, includes: For candidate targets in the candidate target set, the following processing is performed between adjacent consecutive frames: Using the Lucas-Kanade optical flow algorithm, tracking feature points are selected by downsampling within the actual area of ​​the candidate target according to a preset grid, and the tracking feature points are tracked from the current frame to the adjacent frame to establish inter-frame point pair matching of the candidate target; Pre-screening of inter-frame point pair matching based on adaptive reprojection error: calculate the motion consistency residual of each inter-frame point pair matching, dynamically determine the adaptive pre-screening threshold based on the statistics of the motion consistency residual, and identify point pairs with residuals greater than the adaptive pre-screening threshold as outliers and remove them to obtain pre-screened matching point pairs. Based on the epipolar geometric constraints of the fundamental matrix and homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets and obtain the comprehensive detection results.

7. The method for identifying and evaluating small unmanned aerial vehicles based on airborne edge computing as described in claim 6, characterized in that, Based on epipolar geometric constraints of the fundamental matrix and homography matrix, a dual RANSAC voting mechanism is used to perform geometric verification on the pre-screened matching point pairs to confirm that the candidate targets are real moving targets, and a method for obtaining the comprehensive detection result is provided, comprising: Simultaneously, the fundamental matrix F and homography matrix H of the pre-selected matching point pairs are estimated, wherein the fundamental matrix satisfies epipolar geometric constraints. , For the correct matching point pair; The epipolar distance residual of each pair of matching points is calculated based on the basic matrix F. An adaptive reprojection threshold is obtained based on the statistics of the epipolar distance residual and dynamically adjusted by exponential moving average. Pairs of points whose epipolar distance residual exceeds the adaptive reprojection threshold are identified as outliers. For each pair of matching points, the votes of the interior / exterior points in the two models, the basic matrix F and the homography matrix H, are counted, and the geometric consistency of the pair of matching points is confirmed based on the voting results of the dual models. When the number of interior points that pass geometric verification in the pre-screened matching point pairs of a candidate target reaches a set proportion, the candidate target is confirmed as a real moving target.

8. A small unmanned aerial vehicle (UAV) identification and evaluation system based on airborne edge computing, characterized in that, This is applied to detection drones equipped with edge computing devices, initially in detection mode, including: The image acquisition and preprocessing module is used to acquire the target image of the target area and perform preprocessing. A dual-path parallel detection module is used to identify potential UAV targets in the preprocessed target image by adopting a dual-path parallel detection architecture of main path detection and branch path enhanced detection in detection mode, and to obtain a candidate target set composed of the two detection results. The main path is configured as an improved deep learning target detection network, and the branch path is configured to adopt a small target enhanced detection mechanism based on texture features. The motion matching denoising module is used to confirm the motion targets in the candidate target set by a cascaded denoising algorithm based on optical flow tracking and epipolar geometry constraints, and obtain a comprehensive detection result of whether there are abnormal moving targets in the target area. If there are no abnormal moving targets in the target area, the comprehensive detection results are directly output to the target end; The data fusion and decision-making module is used to assess the threat of abnormal moving targets based on a preset temporal cross-comparison and comparison correlation mechanism if there are abnormal moving targets in the target area. When an abnormally moving target is assessed as a non-high-threat target, the comprehensive detection result is directly output to the target end; When an abnormally moving target is assessed as a high-threat target, the system switches from the detection mode to the target locking and tracking mode to continuously lock and track the high-threat target. Based on the obtained tracking results, an alarm message is generated and output to the target. Additionally, based on a preset verification trigger mechanism, the system switches from the target locking and tracking mode back to the detection mode to verify the tracking results and ensure tracking reliability.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program; when the processor executes the computer program, it implements the small unmanned aerial vehicle identification and evaluation method based on airborne edge computing as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the small unmanned aerial vehicle (UAV) identification and evaluation method based on airborne edge computing as described in any one of claims 1 to 7.

Images