Unmanned aerial vehicle precision approach control method, system and device

Abstract

The application discloses a kind of unmanned aerial vehicle accurate approach control method, system and equipment.Method includes: by pure center point detection network, inference is carried out to airborne image, obtains target center point, confidence and horizontal and vertical anisotropic gaussian radius;Based on anisotropic gaussian radius, target scale representation and shape representation are constructed, and the observation reliability weight is constructed in combination with the scale and shape prediction residual of continuous frame and confidence;The bandwidth of yaw and pitch channel linear self-disturbance rejection controller is adjusted online using the weight to generate angular velocity instruction;According to the difference between scale representation and preset scale, forward speed instruction is generated to constrain safe approach.The application does not need boundary box regression or depth information, can adaptively adjust control response according to observation reliability degree, significantly improve the alignment stability, control smoothness and task safety of unmanned aerial vehicle accurate approach.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a method, system, and device for precise close-range control of UAVs. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have been widely used in tasks such as target inspection, target tracking, and logistics delivery. These tasks can essentially be modeled as a precise approach control process for UAVs. For example, target inspection and tracking involve approaching targets at a safe distance, while logistics delivery involves approaching targets at a desired distance of zero. In these tasks, the accuracy of target visual perception, the stability of center point localization, and the robustness of the control response directly determine the reliability and safety of the task execution. Currently, conventional UAV perception and control schemes typically employ a serial processing framework: a detection model outputs a target bounding box, the target center point is calculated based on the bounding box, pixel errors are calculated, and the data is input to the controller.

[0003] However, this conventional approach still has significant shortcomings in perception model design, target information representation, and perception-control fusion. First, the perception paradigm is redundant. Most mainstream target detection algorithms adopt a bounding box paradigm, relying on bounding box regression to indirectly obtain the target's center point and scale information. However, downstream precise control tasks actually require the target's accurate location and compact state variables that directly reflect the target's observation scale and shape stability. The width and height information of the bounding box is not necessary, and the jitter of the bounding box in scenarios with small targets or large rotational changes can introduce additional control errors. Second, the perception accuracy is insufficient. The optimization objective of existing bounding box detection methods is mainly focused on bounding box localization, with the center point coordinates indirectly calculated from the bounding box. There is a lack of direct optimization for center point localization, making it difficult to meet the requirements of precise proximity control in terms of point localization accuracy, precision, and recall. Furthermore, there is a disconnect between perception and control. Existing loosely coupled processes rarely consider the specific needs of back-end control during the perception model design phase. They also fail to effectively incorporate visual observation states such as detection confidence, target size changes, shape ratio changes, and line-of-sight angle jumps into the dynamic adjustment process of controller parameters. As a result, under conditions of visual detection jitter, rapid changes in target size, or external disturbances, fixed-parameter PID or conventional control methods are unable to adjust the control response in a timely manner, which can easily lead to problems such as attitude oscillation, response lag, and unsmooth approach process, significantly affecting the approach performance of UAVs in complex environments. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to overcome the loose coupling problem between visual perception output and controller parameter adjustment in the existing precise approach control of UAVs.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for precise approach control of unmanned aerial vehicles, comprising the following steps: Acquire target images captured by the UAV's onboard visual sensor; The target image is input into a pre-trained pure center point detection network, and the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius are obtained through inference. Based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius, a target scale representation and a target shape representation are constructed; and based on the pixel deviation between the target center point and the preset image reference point, the horizontal viewing angle error and the vertical viewing angle error are calculated. Based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on consecutive frames, and the shape prediction residual determined by the shape representation quantity based on consecutive frames, construct the yaw channel observation reliability weight and the pitch channel observation reliability weight. Using the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, the bandwidth of the corresponding channel's linear active disturbance rejection controller is adjusted online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively. The total disturbance is estimated and compensated by the extended state observer in the linear active disturbance rejection controller, generating yaw rate and pitch rate commands. Based on the difference between the target scale characterization quantity and the preset target scale, a forward speed command is generated. The yaw rate command, pitch rate command, and forward rate command are sent to the UAV flight controller to control the UAV to approach the target.

[0006] This invention also provides a precise approach control system for unmanned aerial vehicles (UAVs), comprising: The image acquisition module is used to acquire target images captured by the UAV's onboard vision sensor; The center point detection module is used to input the target image into a pre-trained pure center point detection network and infer the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius. The visual state construction module is used to construct target scale representation and target shape representation based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius; and to calculate horizontal viewing angle error and vertical viewing angle error based on the pixel deviation between the target center point and the preset image reference point. The observation reliability calculation module is used to construct the yaw channel observation reliability weight and the pitch channel observation reliability weight based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on the continuous frames, and the shape prediction residual determined by the shape representation quantity based on the continuous frames. The linear active disturbance rejection control module is used to take the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, adjust the bandwidth of the corresponding channel's linear active disturbance rejection controller online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively, estimate and compensate for the total disturbance through the extended state observer in the linear active disturbance rejection controller, generate yaw rate command and pitch rate command; and generate forward velocity command based on the difference between the target scale characterization quantity and the preset target scale. The flight control interface module is used to send the yaw rate command, pitch rate command and forward speed command to the UAV flight controller to control the UAV to approach the target.

[0007] This invention also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the UAV precise approach control method as described in any of the preceding embodiments.

[0008] The beneficial effects of this invention are as follows: By directly inferring the target center point and its horizontal and vertical anisotropic Gaussian radii from airborne visual images, and constructing target scale representation, shape representation, and observation reliability weights driven by continuous inter-frame prediction residuals, the reliability of target visual observation can be quantified in real time, and the bandwidth of the linear active disturbance rejection controller can be adaptively adjusted online accordingly. When the target detection confidence is high and the scale and shape change smoothly, the controller responds quickly to line-of-sight errors with a high bandwidth, achieving agile alignment. When detection jitter, occlusion, or scale abrupt changes occur, the controller bandwidth is automatically reduced, effectively suppressing attitude oscillations caused by mis-adjustment. At the same time, a forward velocity command is generated based on the difference between the scale representation and the preset target scale, automatically achieving deceleration and stopping safety constraints during target approach. Therefore, this solution significantly improves the stability of alignment, the smoothness of control, and the safety of mission execution during precise UAV approach without relying on target bounding boxes, actual physical dimensions, or depth information. Attached Figure Description

[0009] The specific structure of the present invention will now be described in detail with reference to the accompanying drawings.

[0010] Figure 1 This is a schematic diagram illustrating the rationality proof of center point detection in an embodiment of the present invention; Figure 2 This is a flowchart of the UAV precise approach control method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the PCPNet pure centroid detection network structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the neighborhood-supervised offset learning enhancement method according to an embodiment of the present invention; Figure 5 This is a visualization of the anisotropic Gaussian radius in an embodiment of the present invention; Figure 6 This is a structural diagram of a linear active disturbance rejection controller according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure for online bandwidth adjustment based on reliability weights according to an embodiment of the present invention; Figure 8 This is a visualization of the detection results of the pure centroid detection network in an embodiment of the present invention. Figure 9 This is a comparison chart of the performance metrics of the pure center point detection network in this invention embodiment; Figure 10 This is a schematic diagram illustrating the precise approach of a drone to a target area according to an embodiment of the present invention. Figure 11 This is a block diagram of a drone precise approach control device according to an embodiment of the present invention; Figure 12 This is a schematic block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0013] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0014] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0015] Before describing the specific control method, let's first explain the rationale behind choosing the pure center point detection paradigm in this invention. In downstream tasks such as UAV approach control, target bounding box representation is not necessary; only the target's category and location are required, meaning point representation is perfectly sufficient. There are many ways to locate targets using points (such as the top-left corner point, bottom-right corner point, etc.), but only the center point possesses rotational invariance from the UAV's perspective. For example... Figure 1 As shown, taking a rectangular target as an example, without loss of generality, assuming that the two types of detection boxes have no offset, the coordinates of the center point of the ordinary horizontal bounding box detection are: (1) In the rotating frame detection paradigm, the center point of the target can be represented as: (2) Subtract the x-coordinates (and the y-axis similarly). (3) From formula (3), it can be seen that the projection of the true physical center point of the target onto the image before and after rotation is unique and invariant, that is, the center point is invariant compared to the ordinary horizontal bounding box or the rotated detection box. The pure center point detection network (PCPNet) proposed in this invention directly learns and outputs the target center point, without relying on the width and height regression of the bounding box. Its predicted center point coordinates are not affected by the target rotation and aspect ratio changes, and have better rotation invariance and localization robustness. This provides a stable and high-quality state observation source for precise approach control in the backend.

[0016] The following is combined with Figure 2 The overall flowchart shown below provides a detailed explanation of the UAV precise approach control method of the present invention.

[0017] The present invention provides a method for precise approach control of an unmanned aerial vehicle (UAV), comprising the following steps: S10. Acquire the target image collected by the UAV's onboard visual sensor.

[0018] In this step, a sequence of images of the target area is acquired in real time using a monocular or binocular vision sensor (such as a CMOS camera) mounted on the drone. This image serves as the raw input for subsequent perception and control processing, and its resolution, frame rate, and other parameters can be set according to the actual task requirements.

[0019] S20. Input the target image into a pre-trained pure center point detection network and infer the target center point, center point confidence, horizontal anisotropic Gaussian radius and vertical anisotropic Gaussian radius. The pure center point detection network includes a feature extraction module, a feature pyramid module, and center point detection heads of multiple scales connected in sequence. The output of each scale center point detection head includes a heatmap branch, an anisotropic Gaussian radius branch, an offset branch, and a center point quality auxiliary branch, wherein the anisotropic Gaussian radius includes a horizontal anisotropic Gaussian radius and a vertical anisotropic Gaussian radius.

[0020] In this module, after the high-level features of the feature extraction module, a sparse hybrid expert feature enhancement module is connected in the form of learnable residuals; the sparse hybrid expert feature enhancement module includes multiple expert branches and a router module for adaptively generating fusion weights for each expert branch based on the input features.

[0021] In a specific embodiment, the reasoning to obtain the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius includes: S21. Multiply the center point heatmap branch with the center mass branch to obtain the center point mass heatmap. Then, use the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius to perform anisotropic max pooling decoding to obtain candidate center points. S22. The candidate center point is corrected by combining the output of the center point offset branch to obtain the final target center point; S23. The confidence level of the center point is the probability value that the target belongs to the center point.

[0022] In the above embodiments, reference is made to Figure 3 The diagram shows the PCPNet architecture, which consists of three main parts: a feature extraction module, a feature pyramid module, and a multi-scale center point detection head. The feature extraction module extracts multi-level features from the input image; the feature pyramid module (such as PAN or FPN) enhances the information interaction between features at different scales; and the multi-scale center point detection head outputs a center point heatmap, anisotropic Gaussian radii (including horizontal and vertical radii), and center point offsets, respectively. Additionally, a center point quality auxiliary branch can be output to learn the quality score of the center points.

[0023] Furthermore, the total loss function of the entire model is: (4) in, Loss detection at the center point For the loss of the hybrid expert model, The loss weights are used for the hybrid expert model.

[0024] (5) Among them, the loss terms in detection loss (5) are, in order: heatmap loss (using Focal Loss), offset loss (using SmoothL1 loss), anisotropic Gaussian radius loss (using SmoothL1 loss), and quality auxiliary branch loss (using BCE loss). The preceding parameters are their respective weights. These are the weights of each layer. In this embodiment, the parameters are set as follows: To improve the feature representation capability of the feature extraction module under conditions of target scale variation, background interference, or viewpoint variation, this embodiment connects a sparse hybrid expert feature enhancement module in a learnable residual form after its high-level features. This module includes multiple expert branches and a router module, where each expert branch can be implemented using a lightweight inverse residual convolutional block. The router module adaptively generates fusion weights for each expert branch based on the input features using global average pooling, convolution, and the SiLU activation function. In this embodiment, the total number of hybrid experts is four, and the actual number of experts activated after adopting sparse activation is two, thereby enhancing the nonlinear expressive capability of the model while maintaining computational efficiency.

[0025] In training the PCPNet network, in addition to using conventional data augmentation techniques such as Mosaic, image stretching, and color transformation, two additional augmentation training techniques were designed specifically for the center point detection task: neighborhood-supervised offset learning augmentation and high-precision layer-assisted localization augmentation.

[0026] The principle of neighborhood-supervised offset learning enhancement is as follows: Figure 4 As shown. Traditional center point detection typically generates a supervision signal only at the location of the target center point, resulting in a small number of positive samples. This embodiment, when generating the supervision signal for the center point offset branch, not only supervises the center point itself but also generates supervision signals within the neighborhood region centered on the center point. Specifically, for the original image target center point... The coordinates mapped to the feature map are: (6) in, This represents the downsampling factor. The grid where the center point is located is... (7) in, This indicates a floor operation. Define the neighborhood Gaussian weights. (8) in, It is the vertical index of the feature map. It is a horizontal index; , These are the anisotropic Gaussian radii, and their calculation follows the formula below. (9) in, The overlap rate is 0.7 in this embodiment. Let be the original height and width of the target. Then the neighborhood supervision range is... (10) In this embodiment, .

[0027] Figure 5 The visualization results of the anisotropic Gaussian are presented by... Figure 4 As can be seen from the explanation, the principle of its calculation is shown in formula (9). This can be derived from... Figure 5 As can be seen, anisotropic Gaussian has a natural advantage in conforming to the aspect ratio of objects, and it conforms to the shape of the target better than isotropic Gaussian, and it also learns a larger number of samples (nearly 6 times more).

[0028] S30. Based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius, construct the target scale representation quantity and the target shape representation quantity; and calculate the horizontal viewing angle error and the vertical viewing angle error according to the pixel deviation between the target center point and the preset image reference point. In a specific embodiment, the step of constructing the target scale characterization quantity and the target shape characterization quantity based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius includes: S31. The square root of the product of the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius is used as the target scale characterization quantity. S32. The maximum and minimum ratio of the horizontal anisotropic Gaussian radius to the vertical anisotropic Gaussian radius is used as the target shape characterization quantity.

[0029] This step transforms the output of PCPNet into visual state variables that are directly relevant to control.

[0030] Specifically, Figure 6 The diagram shows the structure of the linear active disturbance rejection controller (AVR) of this embodiment. For ease of explanation, only yaw angle control is shown (pitch angle control is similar; the horizontal line-of-sight angle error is used for yaw channel control, and the vertical line-of-sight angle error is used for pitch channel control; both use the same type of linear AVR controller, and the controller bandwidth is adjusted online by the observation reliability weight). It must be noted that since the ideal line-of-sight angle is 0, the line-of-sight angle error is equal to the line-of-sight angle itself; this point needs to be considered in the following explanation. That is, the reference line-of-sight angle error illustrated in the figure. The value is 0 because for the UAV to align with the target, the target center and the UAV's forward axis should be parallel, hence the line-of-sight angle is 0. The calculation formula can be obtained from the principles of optical imaging: (11) in, The target center point output by PCPNet. This is the image reference point (the center point is the default). For camera The camera focal length in the direction.

[0031] Based on the principle of a first-order linear active disturbance rejection controller, the yaw angle channel can be modeled as a first-order uncertain system: (12) in, This includes the total disturbance caused by factors such as wind disturbance and target motion. To control the input, To control the gain parameters, The rate of change of the controlled variable is the rate of change of the yaw line-of-sight angle. To achieve anti-disturbance control against disturbances, a linear extended state observer is introduced to estimate the disturbance of the entire system. The total system disturbance is first canceled out before the control input is given, i.e., the observation is introduced. Based on the first-order Luenberger-type extended state observer, the entire linear extended state observer can be obtained as follows: (13) in, The output of the observer is calculated, that is, the yaw angle and line-of-sight angle error calculated in real time according to formula (11) after PCPNet real-time inference. The observation residual is used to measure the deviation between the observer's estimate and the actual measurement. The observer bandwidth depends on the camera acquisition frame rate and the PCPNet inference speed.

[0032] For the entire control system, it is desirable that the controlled variable converges dynamically according to the first-order expectation, i.e., the state error feedback is... (14) Solving equations (13) and (14) together yields the total disturbance compensation as follows: (15) in, The controller bandwidth is used to control the drone's speed adjustment. This is a limiting function because, in engineering practice, the maximum adjustment angle of the drone's yaw angle is finite.

[0033] In this embodiment, Unlike traditional linear active disturbance rejection controllers, this invention is deeply coupled with the real-time detection results of PCPNet, meaning the controller bandwidth is adjusted online. During the continuous approach of a UAV to a target region, the relative distance between the UAV and the target gradually changes, causing the target's scale representation in the image to change continuously over time. Therefore, if the observation reliability is judged solely based on the difference in scale representation between two adjacent frames, the continuous increase in target scale caused by normal UAV approach may be misjudged as an observation anomaly. To avoid this problem, this invention constructs observation reliability weights based on prediction residuals from consecutive multiple frames. The specific adjustment process can be achieved by... Figure 7 As shown. Specifically, let the first... The PCPNet inference decodes the target center point and anisotropic radius. With confidence level Since there is a potential mapping relationship between the anisotropic Gaussian radius and the shape and size of the target, scale and shape representation quantities can be reasonably constructed based on this.

[0034] (16) in, The scaling factor is the normalization coefficient. To minimize the amount of division, prevent the denominator from being 0.

[0035] S40. Based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on the continuous frames, and the shape prediction residual determined by the shape representation quantity based on the continuous frames, construct the yaw channel observation reliability weight and the pitch channel observation reliability weight. In a specific embodiment, constructing the yaw channel observation reliability weight and the pitch channel observation reliability weight based on the center point confidence, the scale prediction residual determined based on the scale representation quantity of consecutive frames, and the shape prediction residual determined based on the shape representation quantity of consecutive frames includes: S41. Obtain the scale and shape representations of the current frame and the two previous consecutive frames; S42. Based on the scale representation of the previous two frames, use linear extrapolation to predict the scale prediction value of the current frame, and use the normalized difference between the actual scale representation of the current frame and the predicted scale value as the scale prediction residual. S43. Based on the shape representation values ​​of the previous two frames, use linear extrapolation to predict the shape prediction value of the current frame, and use the normalized difference between the actual shape representation value of the current frame and the predicted shape value as the shape prediction residual. S44. Based on the center point confidence, the scale prediction residual, and the shape prediction residual, generate observation reliability weights that are positively correlated with the center point confidence and negatively correlated with the scale prediction residual and the shape prediction residual.

[0036] Specifically, in this embodiment, the scale representation at the current time can be predicted using the scale representation at the previous two sampling times, that is: (17) The scale prediction residual is: (18) Similarly, the predicted shape proportion representation at the current moment can be obtained as: (19) Its shape prediction residual is: (20) Furthermore, construct reliability weights: (twenty one) in, These are the scale prediction residual weights and the shape prediction residual weights, respectively. This is a limiting function. Using the above method, when the target scale and shape proportions change smoothly according to a continuous motion trend, the scale prediction residuals and shape prediction residuals are small, and the observation reliability weight is mainly determined by the center point confidence. When the target experiences occlusion, false detection, target switching, or center point jumps, the residuals between the current scale representation or shape proportion representation and the predicted values ​​increase, and the observation reliability weight decreases. This avoids misjudging continuous scale changes during normal UAV approach as observation anomalies, improving the accuracy of reliability modeling.

[0037] S50. Using the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, the bandwidth of the corresponding channel's linear active disturbance rejection controller is adjusted online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively. The total disturbance is estimated and compensated through the extended state observer in the linear active disturbance rejection controller, generating yaw rate command and pitch rate command; and a forward speed command is generated based on the difference between the target scale characterization quantity and the preset target scale. In a specific embodiment, the step of adjusting the bandwidth of the linear active disturbance rejection controller for the corresponding channel online using the yaw channel observation reliability weight and the pitch channel observation reliability weight includes: S501. Set the yaw channel controller bandwidth to be positively correlated with the yaw channel observation reliability weight, and limit it between the preset minimum yaw bandwidth and the preset maximum yaw bandwidth. S502. Set the pitch channel controller bandwidth to be positively correlated with the pitch channel observation reliability weight, and limit it between the preset minimum pitch bandwidth and the preset maximum pitch bandwidth.

[0038] In this embodiment, the bandwidth of the entire controller is: (twenty two) in, These are the preset minimum and maximum bandwidths, respectively.

[0039] In a specific embodiment, generating a forward velocity command based on the difference between the target scale representation and a preset target scale includes: S511. When the target scale representation is less than the preset target scale, a forward velocity command positively correlated with the difference is generated. S512. When the target scale representation is equal to or greater than the preset target scale, the forward speed command is set to zero or a deceleration command is output.

[0040] In this embodiment, the performance of the PCPNet center point detection network was specifically evaluated and tested in actual operation (using the ResNet50 backbone, COCO2017 dataset, 150 training rounds). Figure 8 As shown, the entire center point detection model can effectively locate the center point of any target, whether it's a person or a bird with normal proportions, or an extreme proportion like a toothbrush or a fork, further demonstrating the algorithm's effectiveness. Furthermore, in terms of specific metrics, the box paradigm's metrics are naturally introduced into the point detection paradigm. A predicted center point... Compared with the true center point of the same category satisfy: (twenty three) If the point is considered to be matched at this pixel distance, the corresponding precision P, recall R, mean precision mAP, and F1 score (the higher the better) are all based on this standard.

[0041] (twenty four) in, pixels The distance from a successful one-to-one match represents true positives, false positives, and false negatives. mAP is the average point set detection precision, defined as follows: (25) in, Indicates category, This represents the pixel threshold. In addition, to further measure the performance of point detection, metrics such as MAE, MSE, rMAE, and RMSE (lower is better) are introduced. These metrics are commonly found in common counting evaluation standards.

[0042] (26) in, The total number of sample images. , Each sample image represents both the number of detected targets and the actual number of targets. For fairness, the performance of both the comparison algorithm and the method proposed in this invention patent are compared under this standard, and the hyperparameters of the comparison algorithms are all the officially recommended default values. Figure 9 As can be seen, the detection performance of the proposed PCPNet network has achieved leading results in many aspects, further demonstrating the effectiveness of the proposed method.

[0043] S60. The yaw rate command, pitch rate command and forward speed command are sent to the UAV flight controller to control the UAV to approach the target.

[0044] In this step, the control commands generated above are sent to the UAV flight controller in real time via an airborne communication protocol (such as MAVLink) or using the ROS / MAVROS interface. After responding to the commands, the flight controller drives the motors and control surfaces to complete continuous closed-loop control of the UAV's attitude and forward motion, ultimately enabling the UAV to approach the target area smoothly and accurately.

[0045] Figure 10 This diagram illustrates the coupling relationship between perception and control, as well as the geometric relationship of control, during the approach process of a UAV. After the UAV's onboard camera images the target area, the target's center point position, confidence level, and anisotropic Gaussian radius are output after PCPNet inference. The difference between the target's center point in the image and the image's own center point is calculated, and the line-of-sight angle error is obtained using the camera's intrinsic parameters. This line-of-sight angle error is then used to generate the control input. Besides participating in decoding during the inference stage, the anisotropic Gaussian radius also participates in constructing the observation reliability weights during the control stage, thereby determining whether the target's observation state is stable. This dynamically generates the controller bandwidth during the approach control process, making the entire control process smoother.

[0046] like Figure 11 As shown, this embodiment of the invention also provides a precise approach control system for unmanned aerial vehicles (UAVs), comprising: Image acquisition module 10 is used to acquire target images collected by the UAV's onboard vision sensor; The center point detection module 20 is used to input the target image into a pre-trained pure center point detection network and infer the target center point, center point confidence, horizontal anisotropic Gaussian radius and vertical anisotropic Gaussian radius. The visual state construction module 30 is used to construct target scale representation and target shape representation based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius; and to calculate horizontal viewing angle error and vertical viewing angle error based on the pixel deviation between the target center point and the preset image reference point. The observation reliability calculation module 40 is used to construct the yaw channel observation reliability weight and the pitch channel observation reliability weight based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on the continuous frames, and the shape prediction residual determined by the shape representation quantity based on the continuous frames. The linear active disturbance rejection control module 50 is used to take the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, adjust the bandwidth of the corresponding channel's linear active disturbance rejection controller online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively, estimate and compensate for the total disturbance through the extended state observer in the linear active disturbance rejection controller, generate yaw rate command and pitch rate command; and generate forward velocity command based on the difference between the target scale characterization quantity and the preset target scale. The flight control interface module 60 is used to send the yaw rate command, pitch rate command and forward speed command to the UAV flight controller to control the UAV to approach the target.

[0047] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned UAV precision approach control device and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0048] The aforementioned precise approach control device for unmanned aerial vehicles can be implemented as a computer program, which can, for example... Figure 12 It runs on the computer device shown.

[0049] Please see Figure 12 , Figure 12 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0050] See Figure 12 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0051] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a precise approach control method for an unmanned aerial vehicle.

[0052] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0053] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for precise approach control of a UAV.

[0054] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0055] The processor 502 is used to run the computer program 5032 stored in the memory to implement the UAV precise approach control method as described above.

[0056] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0057] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0058] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the UAV precise approach control method as described above.

[0059] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0060] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0061] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0062] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0063] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for precise approach control of an unmanned aerial vehicle (UAV), characterized in that, Includes the following steps: Acquire target images captured by the UAV's onboard visual sensor; The target image is input into a pre-trained pure center point detection network, and the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius are obtained through inference. Based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius, a target scale representation and a target shape representation are constructed; and based on the pixel deviation between the target center point and the preset image reference point, the horizontal viewing angle error and the vertical viewing angle error are calculated. Based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on consecutive frames, and the shape prediction residual determined by the shape representation quantity based on consecutive frames, construct the yaw channel observation reliability weight and the pitch channel observation reliability weight. Using the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, the bandwidth of the corresponding linear active disturbance rejection controller is adjusted online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively. The total disturbance is estimated and compensated by the extended state observer in the linear active disturbance rejection controller, and yaw rate command and pitch rate command are generated. And based on the difference between the target scale representation quantity and the preset target scale, a forward velocity command is generated; The yaw rate command, pitch rate command, and forward rate command are sent to the UAV flight controller to control the UAV to approach the target.

2. The precise approach control method for unmanned aerial vehicles according to claim 1, characterized in that, The pure center point detection network includes a feature extraction module, a feature pyramid module, and center point detection heads of multiple scales connected in sequence. The output of each scale center point detection head includes a heatmap branch, an anisotropic Gaussian radius branch, an offset branch, and a center point quality auxiliary branch, wherein the anisotropic Gaussian radius includes a horizontal anisotropic Gaussian radius and a vertical anisotropic Gaussian radius.

3. The precise approach control method for unmanned aerial vehicles according to claim 2, characterized in that, Following the high-level features of the feature extraction module, a sparse hybrid expert feature enhancement module is connected in the form of learnable residuals. The sparse hybrid expert feature enhancement module includes multiple expert branches and a router module for adaptively generating fusion weights for each expert branch based on the input features.

4. The precise approach control method for unmanned aerial vehicles according to claim 2, characterized in that, The reasoning yields the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius, including: The center point heatmap branch is multiplied by the center mass branch to obtain the center point mass heatmap. Then, anisotropic max pooling decoding is performed using the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius to obtain candidate center points. The candidate center point is corrected by combining the output of the center point offset branch to obtain the final target center point; The center point confidence level is the probability value that the target belongs to the center point.

5. The method according to claim 1, characterized in that, The construction of target scale and target shape characterization quantities based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius includes: The square root of the product of the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius is used as the target scale characterization quantity. The maximum and minimum ratio of the horizontal anisotropic Gaussian radius to the vertical anisotropic Gaussian radius is used as the target shape characterization quantity.

6. The method according to claim 1, characterized in that, The construction of yaw channel observation reliability weights and pitch channel observation reliability weights based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on consecutive frames, and the shape prediction residual determined by the shape representation quantity based on consecutive frames includes: Obtain the scale and shape representations of the current frame and the two previous consecutive frames; Based on the scale representation of the previous two frames, the scale prediction value of the current frame is predicted by linear extrapolation, and the normalized difference between the actual scale representation of the current frame and the predicted scale value is used as the scale prediction residual. Based on the shape representation values ​​of the previous two frames, the shape prediction value of the current frame is predicted by linear extrapolation, and the normalized difference between the actual shape representation value of the current frame and the predicted shape value is used as the shape prediction residual. Based on the center point confidence, the scale prediction residual, and the shape prediction residual, observation reliability weights are generated that are positively correlated with the center point confidence and negatively correlated with the scale prediction residual and the shape prediction residual.

7. The method according to claim 1, characterized in that, The method of adjusting the bandwidth of the linear active disturbance rejection controller for the corresponding channel online using the yaw channel observation reliability weight and the pitch channel observation reliability weight includes: The bandwidth of the yaw channel controller is set to be positively correlated with the reliability weight of the yaw channel observation, and is limited to between the preset minimum yaw bandwidth and the preset maximum yaw bandwidth. The pitch channel controller bandwidth is set to be positively correlated with the pitch channel observation reliability weight, and is limited to between a preset minimum pitch bandwidth and a preset maximum pitch bandwidth.

8. The method according to claim 1, characterized in that, The step of generating a forward velocity command based on the difference between the target scale representation quantity and the preset target scale includes: When the target scale representation is smaller than the preset target scale, a forward velocity command positively correlated with the difference is generated; When the target scale representation is equal to or greater than the preset target scale, the forward velocity command is set to zero or a deceleration command is output.

9. A precise approach control system for unmanned aerial vehicles (UAVs), characterized in that, include: The image acquisition module is used to acquire target images captured by the UAV's onboard vision sensor; The center point detection module is used to input the target image into a pre-trained pure center point detection network and infer the target center point, center point confidence, horizontal anisotropic Gaussian radius, and vertical anisotropic Gaussian radius. The visual state construction module is used to construct target scale representation and target shape representation based on the horizontal anisotropic Gaussian radius and the vertical anisotropic Gaussian radius; and to calculate horizontal viewing angle error and vertical viewing angle error based on the pixel deviation between the target center point and the preset image reference point. The observation reliability calculation module is used to construct the yaw channel observation reliability weight and the pitch channel observation reliability weight based on the center point confidence, the scale prediction residual determined by the scale representation quantity based on the continuous frames, and the shape prediction residual determined by the shape representation quantity based on the continuous frames. The linear active disturbance rejection control module is used to take the horizontal line-of-sight angle error and the vertical line-of-sight angle error as inputs, adjust the bandwidth of the corresponding channel's linear active disturbance rejection controller online using the yaw channel observation reliability weight and the pitch channel observation reliability weight, respectively, estimate and compensate for the total disturbance through the extended state observer in the linear active disturbance rejection controller, generate yaw rate command and pitch rate command; and generate forward velocity command based on the difference between the target scale characterization quantity and the preset target scale. The flight control interface module is used to send the yaw rate command, pitch rate command and forward speed command to the UAV flight controller to control the UAV to approach the target.

10. A computer device, characterized in that: The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the UAV precise approach control method as described in any one of claims 1 to 8.

Images