A monocular vision distance estimation method, system and product for an aerial target

Abstract

This application provides a monocular vision distance estimation method, system, and product for aerial targets. The method is applied to a monocular camera and includes: acquiring a monocular image containing an aerial target, prior geometric parameters of the aerial target, and camera intrinsic parameters; performing target detection on the monocular image to obtain a two-dimensional bounding box of the aerial target in the monocular image; calculating a first candidate depth and a second candidate depth based on the prior geometric parameters and the camera intrinsic parameters, according to the width and height of the two-dimensional bounding box; selecting the smaller value between the first candidate depth and the second candidate depth as the estimated target depth; and determining the distance of the aerial target relative to the camera based on the estimated target depth and the center point of the two-dimensional bounding box. This invention, by employing a bounding box model combined with a minimum value selection strategy, can perform stable and reliable distance measurement of aerial targets based solely on monocular vision. Furthermore, this method is computationally efficient and easily integrated and deployed.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a method, system and product for estimating the monocular visual distance of aerial targets. Background Technology

[0002] With the rapid development of the low-altitude economy and the drone industry, real-time and reliable ranging of aerial targets within the airspace is a key foundational technology for core scenarios such as autonomous obstacle avoidance, multi-drone formation, and low-altitude security for drones. Monocular vision systems, with their advantages of low hardware cost, simple structure, easy integration, and low power consumption, have become a core research and development direction for ranging solutions on small drone platforms.

[0003] However, mainstream ranging solutions for aerial targets have many shortcomings: end-to-end monocular ranging solutions based on deep learning heavily rely on expensive labeled data with accurate ground truth distance values, have poor model generalization, and their black-box nature leads to weak interpretability, making them unable to pass compliance certification for safety-critical scenarios; in addition, traditional geometric rule monocular ranging solutions only use the single dimension of the target bounding box to calculate distance, which is prone to fatal distance underestimation when facing the variable attitudes of aerial targets such as pitch, yaw, and roll, resulting in extremely poor robustness and failing to meet the safety requirements of practical applications; furthermore, although multi-sensor fusion ranging solutions have high ranging accuracy, they have high hardware costs and complex system structures, making them unsuitable for the stringent cost, size, and power consumption constraints of small drones.

[0004] Therefore, the relevant solutions are difficult to achieve robust ranging of aerial targets with varying attitudes under the premise of low cost and high interpretability, and are also difficult to fundamentally avoid the safety risks of underestimating distance, thus failing to meet the actual application needs of the industry. Summary of the Invention

[0005] The purpose of this application is to provide a monocular visual distance estimation method, system, and product for aerial targets, which can at least solve the problems of poor accuracy and high deployment cost of monocular visual ranging methods when ranging aerial targets with varying attitudes in related technologies.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a monocular visual distance estimation method for aerial targets, applied to a monocular camera, comprising: Acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters; Target detection is performed on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image; Based on the prior geometric parameters and the camera intrinsic parameters, the first candidate depth and the second candidate depth are calculated according to the width and height of the two-dimensional bounding box, respectively. The smaller value between the first candidate depth and the second candidate depth is selected as the target estimated depth; Based on the estimated depth of the target and the center point of the two-dimensional bounding box, the distance of the aerial target relative to the camera is determined.

[0007] A second aspect of this application provides a monocular visual range estimation system for aerial targets, comprising: The acquisition module is used to acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters; The detection module is used to perform target detection on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image; The calculation module is used to calculate the first candidate depth and the second candidate depth based on the prior geometric parameters and the camera intrinsic parameters, according to the width and height of the two-dimensional bounding box, respectively. The selection module is used to select the smaller value between the first candidate depth and the second candidate depth as the target estimated depth; The determination module is used to determine the distance of the aerial target relative to the camera based on the target's estimated depth and the center point of the two-dimensional bounding box.

[0008] A third aspect of this application provides an electronic device, including a memory and a processor, wherein the processor is configured to execute a computer program stored in the memory, and when the processor executes the computer program, it implements the steps in the monocular visual distance estimation method for aerial targets described in the first aspect of the embodiments of this application.

[0009] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the monocular visual distance estimation method for aerial targets described in the first aspect of the embodiments of this application.

[0010] As can be seen from the above, the embodiments of this application first acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters. Then, target detection is performed on the monocular image to obtain a two-dimensional bounding box of the aerial target in the monocular image. Next, based on the prior geometric parameters and the camera intrinsic parameters, a first candidate depth and a second candidate depth are calculated according to the width and height of the two-dimensional bounding box, respectively. The smaller value between the first candidate depth and the second candidate depth is selected as the estimated depth of the target. Finally, based on the estimated depth of the target and the center point of the two-dimensional bounding box, the distance of the aerial target relative to the camera is determined. Compared to traditional methods, this invention addresses the core technical problems of monocular vision ranging techniques for aerial targets with varying attitudes. These problems include low distance estimation due to projection distortion caused by single-dimensional ranging, poor ranging reliability, and an inability to balance low cost and high interpretability. This invention achieves stable and reliable ranging of aerial targets based solely on monocular vision without requiring expensive ground truth distance data, thus eliminating the safety risk of distance estimation in principle. Furthermore, this method is computationally efficient and easy to integrate and deploy, effectively reducing the deployment cost of the solution.

[0011] It should be understood that the description in this section is not intended to identify key or important features of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the related technologies or the technical solutions in the embodiments of this application, the drawings used in the description of the related technologies or the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application, and not all embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the monocular visual distance estimation method for aerial targets provided in this application embodiment; Figure 2 A schematic diagram of the bounding box model provided in the embodiments of this application under different target poses; Figure 3 A flowchart illustrating the overall process of the monocular visual distance estimation method for aerial targets provided in this application embodiment; Figure 4 A detailed flowchart illustrating the monocular visual distance estimation method for aerial targets provided in this application embodiment; Figure 5 A schematic diagram of the program modules of the monocular visual distance estimation system for aerial targets provided in the embodiments of this application; Figure 6 A module block diagram of the electronic device provided in the embodiments of this application; Figure 7 A block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application more apparent and understandable, this application will be clearly and completely described below in conjunction with its embodiments and accompanying drawings. Throughout, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. It should be understood that the various embodiments of this application described below are merely illustrative of this application and are not intended to limit this application. That is, all other embodiments obtained by those skilled in the art based on the various embodiments of this application without creative effort are within the scope of protection of this application. Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0015] With the rapid development of the low-altitude economy, drones are widely used in aerial photography and mapping, security inspection, logistics and delivery, and formation flying. Meanwhile, real-time and accurate distance measurement of drone targets within airspace has become a key foundational technology for core scenarios such as autonomous obstacle avoidance, multi-drone formation flight, dynamic target tracking, and low-altitude security in key areas. Among various ranging technologies, monocular vision systems, with their core advantages of low hardware cost, simple structure, easy integration, and low power consumption and size, have become the preferred research and development direction for ranging solutions on small consumer and industrial drone platforms.

[0016] Currently, the mainstream technical implementation paths for monocular ranging and similar ranging technologies targeting aerial targets can be mainly divided into the following three categories: The first category is monocular ranging solutions based on deep learning regression. This type of solution employs an end-to-end convolutional neural network architecture to directly establish the mapping relationship between monocular image pixels and target distance, outputting the target distance value through network regression. However, this type of solution has fundamental technical flaws: First, model training heavily relies on massive amounts of labeled image data with accurate ground truth distance values, which typically require expensive equipment such as LiDAR for collection and labeling, resulting in extremely high data acquisition and labeling costs and a high deployment threshold. Second, the model has poor generalization ability and is highly sensitive to target types, flight attitudes, and scene environments outside the training set, easily leading to ranging failures in complex airspace environments. Third, the end-to-end network is a "black box" model; the calculation process lacks clear physical meaning, has weak interpretability, and cannot be debugged and verified step-by-step, making it difficult to pass compliance certifications for safety-critical scenarios such as drone obstacle avoidance and low-altitude security.

[0017] The second category is monocular ranging methods based on traditional geometric rules. This type of method is based on the classic pinhole projection principle, calculating the distance between the target and the camera by the ratio of the target's pixel size in the image to its known physical size. This approach is clear in principle and simple in calculation, and has been widely used in ranging scenarios for regular targets such as ground vehicles. However, it has a fatal technical shortcoming for aerial drone targets: aerial drones exhibit highly variable pitch, yaw, and roll attitudes during flight, causing their two-dimensional bounding box in the image plane to change drastically and irregularly with their attitude. Existing methods typically only use the width or height of the bounding box for distance calculation, failing to determine which dimension reflects the target's true physical size under the current attitude. This results in extremely large ranging errors, easily leading to dangerous distance underestimation. Distance underestimation can directly cause serious safety incidents such as drone collisions and security countermeasure failures, making it completely unsuitable for the practical application requirements of aerial target ranging.

[0018] The third category is ranging solutions based on multi-sensor fusion. These solutions typically employ a combination of multiple sensors, such as binocular stereo vision, LiDAR, and vision-IMU fusion, to calculate target depth information through multi-source data fusion. While this type of solution can achieve high ranging accuracy, its complex system architecture, high hardware cost, and large power consumption and size make it difficult to integrate into small UAV platforms with strict limitations on cost, weight, and power consumption, hindering its widespread application.

[0019] In summary, the relevant technologies cannot achieve robust and reliable monocular distance estimation for aerial targets with varying attitudes under the premise of low cost, low complexity, high interpretability, and no dependence on ground truth distance data. In particular, they cannot fundamentally avoid the safety risks of distance underestimation from a geometrical perspective, making it difficult to meet the core application requirements of safety-critical scenarios such as autonomous obstacle avoidance of UAVs and low-altitude security.

[0020] Therefore, there is a need for a detection method that can identify and locate aerial targets at low cost, low complexity, and high interpretability, in order to solve the problems of poor accuracy and high deployment cost of monocular vision ranging methods when ranging aerial targets with varying attitudes in related technologies.

[0021] This application provides a monocular visual range estimation method for aerial targets. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating the monocular visual distance estimation method for aerial targets provided in this application embodiment. The monocular visual distance estimation method for aerial targets includes the following steps 101 to 104.

[0022] Step 101: Obtain a monocular image containing the aerial target, the prior geometric parameters of the aerial target, and the camera intrinsic parameters.

[0023] First, three types of basic data are needed for ranging: the first type is the monocular image of the target aerial object captured by the monocular camera, which contains the complete imaging information of the target aerial object in the airspace scene; the second type is the prior geometric parameters corresponding to the target aerial object, which are known basic parameters related to the physical size of the target; the third type is the camera intrinsic parameters of the monocular camera used for this imaging, which are parameters that are pre-calibrated and reflect the inherent imaging characteristics of the monocular camera.

[0024] Step 102: Perform target detection on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image.

[0025] After acquiring a monocular image, target detection processing needs to be performed on the acquired monocular image to identify the aerial target to be tested contained in the monocular image, and output a two-dimensional bounding box that can completely define the imaging area of ​​the target. This two-dimensional bounding box can clearly represent the specific location of the aerial target to be tested in the monocular image plane, as well as the width and height of the target imaging.

[0026] In this embodiment of the application, the prior geometric parameters include the radius of the equivalent bounding sphere of the air target; wherein, the equivalent bounding sphere is the smallest sphere that completely contains the air target, and the radius of the equivalent bounding sphere is half of the maximum physical size of the air target.

[0027] Specifically, the prior geometric parameters include the radius of the equivalent bounding sphere corresponding to the aerial target being measured. For example, Figure 2 As shown, the equivalent bounding sphere is the smallest sphere capable of completely encompassing the entire physical structure of the aerial target under test. Regardless of the spatial distribution of the target's fuselage, rotor, and auxiliary payloads, this sphere can completely contain all of the target's physical structures within it, with no structural components exceeding the sphere's boundaries. The radius of the equivalent bounding sphere is taken as half of the maximum physical size of the aerial target under test. The maximum physical size of the aerial target under test is the maximum straight-line distance between any two points on the target in three-dimensional space, including but not limited to the maximum span of the diagonal rotor tip of a rotary-wing UAV, the wingspan of a fixed-wing UAV, and the rotor diameter of an unmanned helicopter. The radius of this equivalent bounding sphere provides a unified target physical size benchmark for subsequent depth calculations and is unaffected by attitude changes such as pitch, yaw, and roll during the flight of the aerial target under test.

[0028] It is understandable that the core function of the object detection step is to obtain the axial two-dimensional bounding box of the aerial target in the monocular image. Therefore, this solution is not limited to a specific object detection model; any modern 2D object detector capable of outputting an accurate axial two-dimensional bounding box is applicable. Optional detection models include, but are not limited to, the two-stage detection model Faster R-CNN, the single-stage detection model RetinaNet, the Transformer-based end-to-end detection model DETR, and the keypoint-based detection model CenterNet. In practical engineering applications, a suitable detection model can be flexibly selected based on the computing power, real-time requirements, and detection accuracy needs of the deployment platform. The replacement of different detection models does not deviate from the core technical principles of this invention.

[0029] Step 103: Based on prior geometric parameters and camera intrinsic parameters, calculate the first candidate depth and the second candidate depth according to the width and height of the two-dimensional bounding box, respectively.

[0030] Based on pre-acquired prior geometric parameters and camera intrinsic parameters, and combined with the basic geometric principles of camera imaging, the width and height of the two-dimensional bounding box are used as two independent calculation dimensions to calculate two independent depth estimation results, namely the first candidate depth and the second candidate depth. By calculating the depth through two independent dimensions, the imaging characteristics of the aerial target under different flight attitudes can be covered, adapting to scenarios with fluctuations in imaging size caused by changes in target attitude.

[0031] In this embodiment, based on prior geometric parameters and camera intrinsic parameters, the calculation of a first candidate depth and a second candidate depth according to the width and height of the 2D bounding box includes: calculating the normalized width and normalized height of the 2D bounding box according to the pixel coordinates of the 2D bounding box, the pixel width of the monocular image, and the pixel height; wherein, the expressions for the normalized width and normalized height are: , , in, The pixel width of the monocular image. The pixel height of the monocular image, ( , ) is the first coordinate of the 2D bounding box along the axis in the monocular image. , () represents the second coordinate of the 2D bounding box along the axis in the monocular image. For normalized width, The normalized height is used; based on prior geometric parameters and camera intrinsic parameters, the first candidate depth and the second candidate depth are calculated according to the normalized width and normalized height of the 2D bounding box, respectively.

[0032] Specifically, the 2D bounding box is an axial bounding box aligned with the pixel coordinate system of the monocular image. The first coordinate of the axis is the top-left pixel coordinate of the 2D bounding box in the monocular image, and the second coordinate is the bottom-right pixel coordinate of the 2D bounding box in the monocular image. Using these two sets of coordinates, the pixel width and height of the 2D bounding box in the monocular image can be directly determined. The normalized width is the ratio of the pixel width of the 2D bounding box to the total pixel width of the monocular image, and the normalized height is the ratio of the pixel height of the 2D bounding box to the total pixel height of the monocular image; both values ​​range from 0 to 1. Normalization eliminates the influence of monocular image resolution differences on bounding box size calculation, establishing a unified size benchmark for depth calculation of images with different resolutions. Subsequently, based on prior geometric parameters and camera intrinsic parameters, the first and second candidate depths are calculated according to the normalized width and normalized height of the 2D bounding box, respectively.

[0033] Further, in some embodiments of this application, the camera intrinsic parameters include a first focal length corresponding to the image width direction and a second focal length corresponding to the image height direction; based on prior geometric parameters and camera intrinsic parameters, calculating a first candidate depth and a second candidate depth according to the normalized width and normalized height of the two-dimensional bounding box respectively includes: based on a preset imaging model, the first focal length and the second focal length, calculating the first candidate depth and the second candidate depth according to the normalized width and normalized height of the two-dimensional bounding box respectively; wherein, the calculation expressions for the first candidate depth and the second candidate depth are respectively: , , in, As the first candidate depth, As the second candidate depth, Let be the radius of the equivalent enclosing sphere. This corresponds to the first focal length in the image width direction. This is the second focal length corresponding to the height direction of the image.

[0034] Specifically, the intrinsic parameters of a monocular camera include at least a first focal length corresponding to the width of the image and a second focal length corresponding to the height of the image. The first and second focal lengths are inherent parameters of the monocular camera obtained through pre-calibration, and correspond to the equivalent focal lengths along the x-axis (width direction) and y-axis (height direction) in the pixel coordinate system, respectively. They are the core reference parameters for the camera's imaging geometry.

[0035] Based on the preset imaging model, the first focal length, and the second focal length, and combined with the normalized width and normalized height of the previously calculated two-dimensional bounding box, the first and second candidate depths are calculated respectively. The preset imaging model is a pinhole perspective imaging model, which is a general geometric model for pinhole imaging in monocular camera imaging. It establishes a linear mapping relationship between the target's three-dimensional physical space dimensions, the pixel size of the two-dimensional imaging plane, the camera focal length, and the target depth.

[0036] Step 104: Select the smaller value between the first candidate depth and the second candidate depth as the target estimated depth.

[0037] The calculated first and second candidate depths are filtered by minimum value, and the smaller of the two values ​​is selected as the estimated depth of the aerial target. This process ensures that regardless of the flight attitude of the aerial target, the final depth estimate will not be less than the actual distance between the target and the camera, thus fundamentally avoiding the safety risks caused by underestimation of distance.

[0038] In this embodiment of the application, after selecting the smaller value between the first candidate depth and the second candidate depth as the target estimated depth, the method further includes: obtaining a weighted confidence score based on the aspect ratio of the normalized width and normalized height of the two-dimensional bounding box; and optimizing the target estimated depth based on the weighted confidence score to obtain the final target estimated depth.

[0039] Specifically, the aspect ratio, the ratio of the normalized width to the normalized height of the two-dimensional bounding box, directly reflects the completeness of the projection of the target in the current flight attitude onto the monocular image plane in both width and height dimensions. When the aspect ratio is closer to 1, it indicates that the target's projection in both dimensions is not significantly compressed, and the reliability of the candidate depth calculated for that dimension is higher. Conversely, the larger the difference between the aspect ratio and 1, the more significant the compression of the projection in one dimension due to the target's attitude, and the lower the reliability of the candidate depth corresponding to that dimension. Based on the numerical distribution of the aspect ratio, weighted confidence levels are matched for the first and second candidate depths, respectively. These weighted confidence levels characterize the reliability of the candidate depth in the corresponding dimension, and the sum of the weighted confidence levels for the two dimensions is 1. Furthermore, this scheme uses the initial target estimation depth obtained through a minimum value strategy as a safety fallback benchmark, and combines the weighted confidence levels corresponding to the candidate depths in the two dimensions to perform weighted correction and optimization of the initial target estimation depth. The entire optimization process is constrained by the condition that "the final result is not less than the initial target estimated depth". Under the premise of ensuring that the distance measurement result will not underestimate the distance and without destroying the original safety fallback characteristics, the depth calculation results of the two dimensions are balanced by confidence weighting to reduce the deviation between the distance measurement result and the actual distance, improve the distance measurement accuracy, and finally obtain the final target estimated depth that takes into account both safety and accuracy.

[0040] Step 105: Based on the target's estimated depth and the center point of the 2D bounding box, determine the distance of the aerial target relative to the camera.

[0041] Finally, based on the obtained target estimated depth, combined with the position information of the center point of the two-dimensional bounding box in the monocular image, the spatial position of the target in the air relative to the monocular camera is calculated. The depth value along the optical axis of the camera is the straight-line distance of the target in the air relative to the camera, thus completing the complete monocular visual distance estimation for the target in the air.

[0042] In this embodiment, the camera intrinsic parameters also include the camera principal point coordinates; determining the distance of the aerial target relative to the camera based on the target estimated depth and the center point of the two-dimensional bounding box includes: calculating the center normalized coordinates of the two-dimensional bounding box according to the pixel coordinates of the two-dimensional bounding box, the pixel width and pixel height of the monocular image; wherein, the expression for the center normalized coordinates of the two-dimensional bounding box is: , , in,( , Let be the center normalized coordinates of the 2D bounding box; based on the center normalized coordinates of the 2D bounding box, the camera principal point coordinates, the pixel width and pixel height of the monocular image, and the estimated depth of the target, calculate the 3D coordinates of the aerial target in the camera coordinate system; where the 3D coordinates include the distance of the aerial target relative to the camera, and the expression for the 3D coordinates is: , , in, Let the three-dimensional coordinates of the aerial target be in the camera coordinate system. Let these be the coordinates of the camera's principal point. Estimate the depth for the target.

[0043] Specifically, the intrinsic parameters of a monocular camera also include the camera principal point coordinates. The camera principal point coordinates are the coordinates of the intersection of the monocular camera's optical axis and the imaging plane in the pixel coordinate system. They are inherent parameters obtained through pre-calibration of the monocular camera and are one of the core reference parameters for realizing the transformation from two-dimensional image pixel coordinates to three-dimensional spatial coordinates.

[0044] First, the center pixel coordinates of the 2D bounding box are calculated using the first and second axial coordinates of the 2D bounding box. Then, the center pixel coordinates are normalized using the total pixel width and total pixel height of the monocular image to eliminate the influence of different image resolutions on the calculation of the center position, resulting in normalized center coordinates with values ​​between 0 and 1.

[0045] Then, based on the center-normalized coordinates of the 2D bounding box, the camera principal point coordinates, the pixel width and pixel height of the monocular image, and the estimated target depth, the 3D coordinates of the aerial target in the camera coordinate system are calculated. Specifically, the camera coordinate system used in this scheme has the optical center of the monocular camera as the origin, the camera optical axis as the Z-axis, the horizontal width direction of the imaging plane as the X-axis, and the vertical height direction of the imaging plane as the Y-axis. The estimated target depth obtained in the preceding steps is the Z-axis coordinate value of the aerial target under test in the camera coordinate system, corresponding to the straight-line distance of the aerial target under test relative to the camera along the camera optical axis. Based on the inverse projection principle of the pinhole perspective imaging model, combined with the center-normalized coordinates, camera principal point coordinates, camera focal length, and Z-axis coordinate values, the X-axis and Y-axis coordinate values ​​of the aerial target under test in the camera coordinate system can be calculated, ultimately yielding the complete 3D spatial coordinates of the target.

[0046] Furthermore, in some embodiments of this application, the prior geometric parameters also include size information, which includes multiple size levels, each size level corresponding to an equivalent bounding sphere radius; target detection of a monocular image further includes: evaluating the size level of an aerial target to obtain the size information corresponding to the aerial target; and obtaining the equivalent bounding sphere radius corresponding to the aerial target according to the preset correspondence between the size level and the equivalent bounding sphere radius.

[0047] Specifically, the prior geometric parameters also include preset size information, which includes multiple pre-defined size levels, each corresponding to a unique equivalent bounding sphere radius. The size levels are divided according to the maximum physical size range of common aerial targets in civil airspace. For example, they can be set to four levels: micro, small, medium, and large. Each size level corresponds to a conservative maximum physical size, which can cover the actual maximum physical size of all common targets within the corresponding level. The equivalent bounding sphere radius corresponding to each size level is half of the conservative maximum physical size of that level, which is completely consistent with the definition of the equivalent bounding sphere radius mentioned above.

[0048] During target detection in monocular images, the size classification of aerial targets is simultaneously evaluated to obtain their size information. This size classification is achieved through a pre-trained target classification model, which operates in conjunction with the target detection model. After the target detection model identifies and outlines aerial targets in the monocular image, the classification model simultaneously analyzes the target's imaging features and outputs the corresponding size classification, thus completing the size classification. This eliminates the need for additional independent processing steps and is well-suited for real-time ranging applications.

[0049] Based on a pre-defined one-to-one correspondence between size levels and equivalent bounding sphere radii, the equivalent bounding sphere radius corresponding to the currently measured aerial target is obtained. This radius is the prior geometric parameter used in this ranging calculation. This method eliminates the need to obtain the precise model and physical dimensions of the aerial target beforehand, allowing for the adaptation of an equivalent bounding sphere radius that meets safety requirements. This solves the problem of obtaining prior geometric parameters for unknown aerial targets and significantly expands the applicable scenarios of this method. Furthermore, since each size level uses a conservative maximum physical dimension for radius calculation, it ensures that even if there are deviations in the size level assessment, the final ranging result will not underestimate the distance, fully preserving the safety fallback characteristic of this method.

[0050] In addition, the equivalent bounding sphere model adopted in this application embodiment can be adapted to aerial targets of any attitude and unknown type, and achieve a safe range measurement effect. In application scenarios where more prior information about the target is known, the geometric equivalent model of the target can be refined and optimized. While retaining the minimum safe range strategy, the range measurement accuracy can be further improved, such as the ellipsoidal equivalent model and the multi-part combination equivalent model.

[0051] If the three-dimensional physical dimensions of the aerial target to be measured are known in advance (such as the wingspan, fuselage length, and fuselage height of a fixed-wing UAV, and the diagonal rotor span and fuselage height of a rotary-wing UAV), the target can be equivalent to a three-axis ellipsoid model, replacing the sphere model in the core embodiment. The three axes of the ellipsoid correspond to the physical dimensions of the target's length, width, and height, respectively. During depth calculation, the target's attitude information can be combined to select the physical dimension corresponding to the principal axis most parallel to the camera's imaging plane as the calculation benchmark. The candidate depths corresponding to the width and height dimensions are calculated separately. While retaining the minimum value selection strategy, this reduces the ranging error caused by attitude distortion and improves ranging accuracy.

[0052] For large aerial targets with complex structures, the target can be decomposed into multiple independent functional sub-components. Each sub-component is modeled as an independent sphere or cuboid, and corresponding prior geometric parameters are configured for each sub-component. Two-dimensional candidate depth calculation and minimum value filtering are performed on each sub-component to obtain the estimated depth for each sub-component. Then, the estimated depths of multiple sub-components are fused (e.g., by taking the minimum value, using a confidence-weighted average, etc.) to obtain the final estimated depth of the target. This method can adapt to targets with complex structures and a wide range of attitude variations, avoiding ranging errors caused by local occlusion and attitude distortion in a single overall model, while fully adhering to the core calculation logic of this invention.

[0053] In some embodiments, the monocular camera includes optical lenses such as wide-angle lenses and fisheye lenses that have severe radial and tangential distortion. Before the target detection step, the acquired monocular image can be distorted to a linear image that conforms to the pinhole imaging model, and then the subsequent target detection and depth calculation process can be performed. Alternatively, a dedicated camera imaging model corresponding to the wide-angle / fisheye lens can be directly used to replace the pinhole model equation in the core embodiment. Depth calculation can be performed directly based on the pixel coordinates of the distorted image and the pre-calibrated intrinsic parameters of the corresponding camera model, without the need for additional image distortion correction steps. This is suitable for edge deployment scenarios with limited computing power.

[0054] It is understood that the core ranging logic of this invention does not rely on labeled data with accurate ground truth distance values; it only needs to complete the training of the 2D object detection module. Therefore, by optimizing the data strategy, the training cost can be further reduced and the model's performance in real-world scenarios can be improved. Specifically: By combining a 3D model of the target aerial object with background images from different airspace scenarios, a massive training dataset with bounding box annotations along the target axis can be generated through rendering synthesis. Using this synthetic dataset, only the 2D target detection module needs to be trained; there is no need to collect image data from real-world scenes or manually annotate ground truth distances, significantly reducing the cost and barrier to model training. Furthermore, by rendering to cover all possible flight attitudes, lighting conditions, and background scenes of the target, the generalization ability of the detection model is greatly improved, completely eliminating the dependence on real-world labeled data.

[0055] When training a 2D object detection model, domain adaptation technology can be introduced. By using a domain adaptation algorithm, the feature distribution difference between synthetic training data and real scene data can be reduced, the domain offset between synthetic data and real data can be decreased, the detection accuracy and robustness of the detection model in real spatial scenarios can be improved, and the implementation performance of this method in practical engineering applications can be further optimized.

[0056] In summary, the monocular visual distance estimation method for aerial targets disclosed in this invention is based on images acquired by a monocular camera. It obtains the prior geometric parameters of the aerial target and the camera's intrinsic parameters, and after target detection to obtain the corresponding two-dimensional bounding box, it calculates two candidate depths based on the width and height of the bounding box. The smaller of the two candidate depths is selected as the estimated target depth. Finally, the distance between the target and the camera is calculated by combining the information of the bounding box's center point. The invention also provides detailed and extended implementation methods in multiple dimensions, including target geometric models, detection models, camera models, fusion strategies, and training data strategies, fully covering the engineering implementation needs of different application scenarios. This invention is based on a clear geometric imaging principle, eliminating the need for expensive labeled data with precise distance values. It features low hardware cost, strong interpretability, and easy integration and deployment. It can adapt to the varied flight attitudes of aerial targets and completely avoids the safety risk of distance underestimation from the root of geometric principles. It achieves robust and reliable monocular visual ranging for aerial targets, fully meeting the core application needs of safety-critical scenarios such as autonomous obstacle avoidance for UAVs, multi-aircraft formation flight, and low-altitude security.

[0057] like Figure 3 The diagram shows the overall flowchart of the monocular visual distance estimation method for aerial targets provided in this embodiment of the invention. The complete implementation process of this method consists of two core stages: detection model training and target distance estimation inference. In the detection model training stage, a training dataset can be constructed through two paths: real-scene data collection and synthesis of target 3D models and airspace background image data. The training dataset includes target images labeled with target location information and camera intrinsic parameters corresponding to each image. Based on the constructed training dataset, iterative training of the detection model is completed to obtain a 2D target detection model that can stably output the axial two-dimensional bounding box of the aerial target to be tested. In the target distance estimation inference stage, the images to be detected, containing the aerial target to be tested, acquired in real time by a monocular camera, are input into the trained detection model. After the detection model processes and identifies the aerial target to be tested in the image, the true three-dimensional coordinates Tx, Ty, and Tz of the aerial target to be tested relative to the camera are calculated through geometric calculation or direct model regression. The coordinate values ​​along the camera's optical axis are the straight-line distance between the aerial target to be tested and the camera, thus completing the entire process of monocular visual distance estimation for aerial targets. The core implementation of this invention uses a geometric solution path to calculate distances. It does not rely on labeled data with accurate distance truth values. Only the training of a 2D target detection model is required to achieve complete ranging function, which greatly reduces the training cost and engineering deployment threshold of the solution.

[0058] It should be understood that the sequence number of each step in this embodiment does not imply the order in which the steps are executed. The execution order of each step should be determined by its function and internal logic, and should not constitute a unique limitation on the implementation process of this application embodiment.

[0059] In summary, the detailed process of the monocular visual distance estimation method for aerial targets involved in the embodiments of this application can be found in [reference needed]. Figure 4 Specifically: Step 401: Obtain a monocular image containing the aerial target, the prior geometric parameters of the aerial target, and the camera intrinsic parameters; Step 402: Perform target detection on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image; Step 403: Calculate the normalized width and normalized height of the 2D bounding box based on the pixel coordinates of the 2D bounding box, the pixel width and pixel height of the monocular image; Step 404: Based on prior geometric parameters and camera intrinsic parameters, calculate the first candidate depth and the second candidate depth according to the normalized width and normalized height of the two-dimensional bounding box, respectively. Step 405: Select the smaller value between the first candidate depth and the second candidate depth as the target estimated depth; Step 406: Calculate the center normalized coordinates of the 2D bounding box based on the pixel coordinates of the 2D bounding box, the pixel width of the monocular image, and the pixel height.

[0060] For a more detailed process of each step in steps 401 to 406, please refer to the relevant sections shown above. The embodiments of this application will not be repeated here.

[0061] Please see Figure 5 , Figure 5 A monocular visual range estimation system for aerial targets is provided for embodiments of this application. This system can be used to implement the monocular visual range estimation method for aerial targets involved in embodiments of this application. The monocular visual range estimation system for aerial targets mainly includes: The acquisition module 501 is used to acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters. Detection module 502 is used to perform target detection on monocular images and obtain the two-dimensional bounding box of aerial targets in the monocular images; The calculation module 503 is used to calculate the first candidate depth and the second candidate depth based on the width and height of the two-dimensional bounding box, respectively, based on prior geometric parameters and camera intrinsic parameters. The selection module 504 is used to select the smaller value between the first candidate depth and the second candidate depth as the target estimated depth. The determination module 505 is used to determine the distance of an aerial target relative to the camera based on the target's estimated depth and the center point of the two-dimensional bounding box.

[0062] In some embodiments of this example, the prior geometric parameters in the system include the radius of the equivalent bounding sphere of the air target; wherein the equivalent bounding sphere is the smallest sphere that completely contains the air target, and the radius of the equivalent bounding sphere is half the maximum physical size of the air target.

[0063] In some embodiments of this example, when the calculation module 503 performs the function of calculating the first candidate depth and the second candidate depth based on the width and height of the two-dimensional bounding box according to prior geometric parameters and camera intrinsic parameters, it is used to: calculate the normalized width and normalized height of the two-dimensional bounding box according to the pixel coordinates of the two-dimensional bounding box, the pixel width and pixel height of the monocular image; wherein, the expressions for the normalized width and normalized height are: , , in, The pixel width of the monocular image. The pixel height of the monocular image, ( , ) is the first coordinate of the 2D bounding box along the axis in the monocular image. , () represents the second coordinate of the 2D bounding box along the axis in the monocular image. For normalized width, The normalized height is used; based on prior geometric parameters and camera intrinsic parameters, the first candidate depth and the second candidate depth are calculated according to the normalized width and normalized height of the 2D bounding box, respectively.

[0064] Furthermore, in some embodiments of this example, the camera intrinsic parameters include a first focal length corresponding to the image width direction and a second focal length corresponding to the image height direction; when the calculation module 503 performs the function of calculating the first candidate depth and the second candidate depth based on the normalized width and normalized height of the two-dimensional bounding box according to prior geometric parameters and camera intrinsic parameters, it is used to: calculate the first candidate depth and the second candidate depth based on the preset imaging model, the first focal length and the second focal length, and the normalized width and normalized height of the two-dimensional bounding box; wherein, the calculation expressions for the first candidate depth and the second candidate depth are respectively: , , in, As the first candidate depth, As the second candidate depth, Let be the radius of the equivalent enclosing sphere. This corresponds to the first focal length in the image width direction. This is the second focal length corresponding to the height direction of the image.

[0065] In some embodiments of this example, the camera intrinsic parameters also include the camera principal point coordinates; when the determining module 505 performs the function of determining the distance of the aerial target relative to the camera based on the target estimated depth and the center point of the two-dimensional bounding box, it is used to: calculate the center normalized coordinates of the two-dimensional bounding box according to the pixel coordinates of the two-dimensional bounding box, the pixel width and pixel height of the monocular image; wherein, the expression for the center normalized coordinates of the two-dimensional bounding box is: , , in,( , () represents the normalized coordinates of the center of the two-dimensional bounding box; Based on the normalized center coordinates of the 2D bounding box, the camera principal point coordinates, the pixel width and pixel height of the monocular image, and the estimated target depth, the 3D coordinates of the aerial target in the camera coordinate system are calculated. The 3D coordinates include the distance of the aerial target relative to the camera, and the expression for the 3D coordinates is as follows: , , in, Let the three-dimensional coordinates of the aerial target be in the camera coordinate system. Let these be the coordinates of the camera's principal point. Estimate the depth for the target.

[0066] In some embodiments of this example, after performing the function of selecting the smaller value between the first candidate depth and the second candidate depth as the target estimation depth, the selection module 504 is further configured to: collect UAV radio signals, visual images and their corresponding ground truth annotation data from different environmental scenarios according to a preset period to form an incremental dataset; perform local incremental training on the monocular visual distance estimation model of the aerial target based on the incremental dataset to generate model update parameters; upload the model update parameters to the federated learning server for aggregation, and distribute the aggregated global update parameters to the local machine to complete the dynamic update of the local monocular visual distance estimation model of the aerial target.

[0067] In some embodiments of this example, the prior geometric parameters also include size information, which includes multiple size levels, each size level corresponding to an equivalent bounding sphere radius; when the detection module 502 performs the function of target detection on a monocular image, it is used to: evaluate the size level of the aerial target to obtain the size information corresponding to the aerial target; and obtain the equivalent bounding sphere radius corresponding to the aerial target according to the preset correspondence between the size level and the equivalent bounding sphere radius.

[0068] In detail, each module in the monocular visual distance estimation system for aerial targets provided in this embodiment of the invention adopts the same approach as described above. Figure 1 It employs the same technical means as the monocular visual distance estimation method for aerial targets and can produce the same technical effect.

[0069] Please see Figure 6 , Figure 6 A block diagram of an electronic device provided in an embodiment of this application.

[0070] like Figure 6 As shown, this application embodiment also provides an electronic device that can be used to implement the monocular visual distance estimation method for aerial targets in the foregoing embodiments. The electronic device includes a memory 601, at least one processor 602, a signal generator 603, and a signal receiver 606. The memory 601 is used to store at least one program, and when the at least one program is executed by the at least one processor 602, the at least one processor 602 executes the monocular visual distance estimation method for aerial targets provided in this application embodiment.

[0071] Please see Figure 7 , Figure 7 A block diagram of a computer-readable storage medium provided in an embodiment of this application.

[0072] like Figure 7 As shown, this application embodiment also provides a computer-readable storage medium 700, on which executable instructions 710 are stored. When the executable instructions 710 are executed, they perform the monocular visual distance estimation method for aerial targets provided in this application embodiment.

[0073] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, compact disc read-only memory (CD-ROM), or any other form of storage medium known in the art.

[0074] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., Digital Video Disk, DVD), or a semiconductor medium (e.g., Solid State Disk).

[0075] It should be noted that the various embodiments in this application are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For product-related embodiments, since they are similar to method-related embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions of the method-related embodiments.

[0076] It should also be noted that, in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] The above description of the disclosed embodiments enables those skilled in the art to implement or use the content of this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A monocular visual distance estimation method for aerial targets, characterized in that, Applied to monocular cameras, including: Acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters; Target detection is performed on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image; Based on the prior geometric parameters and the camera intrinsic parameters, the first candidate depth and the second candidate depth are calculated according to the width and height of the two-dimensional bounding box, respectively. The smaller value between the first candidate depth and the second candidate depth is selected as the target estimated depth; Based on the estimated depth of the target and the center point of the two-dimensional bounding box, the distance of the aerial target relative to the camera is determined.

2. The monocular visual distance estimation method for aerial targets according to claim 1, characterized in that, The prior geometric parameters include the radius of the equivalent bounding sphere of the aerial target; wherein the equivalent bounding sphere is the smallest sphere that completely contains the aerial target, and the radius of the equivalent bounding sphere is half the maximum physical size of the aerial target.

3. The monocular visual distance estimation method for aerial targets according to claim 2, characterized in that, The step of calculating the first candidate depth and the second candidate depth based on the prior geometric parameters and the camera intrinsic parameters, according to the width and height of the two-dimensional bounding box, includes: Based on the pixel coordinates of the 2D bounding box, the pixel width of the monocular image, and the pixel height, calculate the normalized width and normalized height of the 2D bounding box; wherein, the expressions for the normalized width and normalized height are: ， ， in, The pixel width of the monocular image. The pixel height of the monocular image, ( , ) is the first coordinate of the two-dimensional bounding box along the axis in the monocular image, ( , () represents the second coordinate of the two-dimensional bounding box along the axis in the monocular image. The normalized width, The normalized height; Based on the prior geometric parameters and the camera intrinsic parameters, the first candidate depth and the second candidate depth are calculated according to the normalized width and normalized height of the two-dimensional bounding box, respectively.

4. The monocular visual distance estimation method for aerial targets according to claim 3, characterized in that, The camera intrinsic parameters include a first focal length corresponding to the image width direction and a second focal length corresponding to the image height direction; the calculation of the first candidate depth and the second candidate depth based on the prior geometric parameters and the camera intrinsic parameters, according to the normalized width and normalized height of the two-dimensional bounding box, includes: Based on a preset imaging model, the first focal length, and the second focal length, a first candidate depth and a second candidate depth are calculated according to the normalized width and normalized height of the two-dimensional bounding box, respectively; wherein, the calculation expressions for the first candidate depth and the second candidate depth are as follows: ， ， in, The first candidate depth, The second candidate depth, Let be the radius of the equivalent enclosing sphere. The first focal length corresponding to the image width direction, This is the second focal length corresponding to the height direction of the image.

5. The monocular visual distance estimation method for aerial targets according to claim 4, characterized in that, The camera intrinsic parameters also include the camera principal point coordinates; determining the distance of the aerial target relative to the camera based on the target's estimated depth and the center point of the two-dimensional bounding box includes: Based on the pixel coordinates of the 2D bounding box, the pixel width and pixel height of the monocular image, the normalized center coordinates of the 2D bounding box are calculated; wherein, the expression for the normalized center coordinates of the 2D bounding box is: ， ， in,( , ) represents the center normalized coordinates of the two-dimensional bounding box; Based on the normalized center coordinates of the 2D bounding box, the camera principal point coordinates, the pixel width and pixel height of the monocular image, and the estimated depth of the target, the 3D coordinates of the aerial target in the camera coordinate system are calculated; wherein, the 3D coordinates include the distance of the aerial target relative to the camera, and the expression for the 3D coordinates is: ， ， in, Let the three-dimensional coordinates of the aerial target be in the camera coordinate system. The coordinates of the principal point of the camera are... Estimate the depth of the target.

6. The monocular visual distance estimation method for aerial targets according to claim 3, characterized in that, After selecting the smaller value between the first candidate depth and the second candidate depth as the target estimated depth, the method further includes: The weighted confidence level is obtained based on the aspect ratio of the normalized width and normalized height of the two-dimensional bounding box; The target estimation depth is optimized based on the weighted confidence level to obtain the final target estimation depth.

7. The monocular visual distance estimation method for aerial targets according to claim 2, characterized in that, The prior geometric parameters also include size information, which includes multiple size levels, each size level corresponding to an equivalent bounding sphere radius; the target detection of the monocular image further includes: The size of the aerial target is evaluated to obtain the size information corresponding to the aerial target; Based on the preset correspondence between size levels and equivalent bounding sphere radii, the equivalent bounding sphere radius corresponding to the aerial target is obtained.

8. A monocular visual range estimation system for aerial targets, characterized in that, include: The acquisition module is used to acquire a monocular image containing an aerial target, the prior geometric parameters of the aerial target, and camera intrinsic parameters; The detection module is used to perform target detection on the monocular image to obtain the two-dimensional bounding box of the aerial target in the monocular image; The calculation module is used to calculate the first candidate depth and the second candidate depth based on the prior geometric parameters and the camera intrinsic parameters, according to the width and height of the two-dimensional bounding box, respectively. The selection module is used to select the smaller value between the first candidate depth and the second candidate depth as the target estimated depth; The determination module is used to determine the distance of the aerial target relative to the camera based on the target's estimated depth and the center point of the two-dimensional bounding box.

9. An electronic device, characterized in that, Includes memory and processor, of which: The processor is used to execute computer programs stored in the memory; When the processor executes the computer program, it implements the steps in the monocular visual distance estimation method for aerial targets according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the monocular visual distance estimation method for aerial targets according to any one of claims 1 to 7.

Images