An intelligent decision-making method based on unmanned aerial vehicle abnormal target inspection evidence collection

Abstract

The present application relates to the technical field of unmanned aerial vehicle inspection and evidence collection, and discloses an intelligent decision-making method for unmanned aerial vehicle abnormal target inspection and evidence collection, comprising: S1, obtaining current observation information collected by unmanned aerial vehicle inspection, determining suspected abnormal targets and supervision judgment boundaries; S2, determining a spatial estimation center and a spatial covariance matrix according to the current observation information, unmanned aerial vehicle pose information and measurement error parameters, and constructing a three-dimensional probability confidence ellipsoid; S3, judging whether the three-dimensional probability confidence ellipsoid crosses the supervision judgment boundary, if yes, determining a judgment ambiguity state and a judgment ambiguity degree; S4, generating a candidate observation pose and predicting and updating the three-dimensional probability confidence ellipsoid; S5, determining an ambiguity convergence amount and a next observation pose according to a cross-border volume proportion; S6, collecting new observation information according to the next observation pose and updating the ellipsoid, and completing the evidence collection when a stable condition is met, otherwise returning to S4. The stability of the inspection and evidence collection result is improved.

Description

Technical Field

[0001] This invention relates to the field of drone inspection and evidence collection technology, specifically to an intelligent decision-making method for abnormal target inspection and evidence collection based on drones. Background Technology

[0002] Drone inspection and evidence collection refers to using drones equipped with image acquisition, positioning, and attitude measurement devices to conduct aerial observation of the inspection area or object, and to collect observational information related to the suspected abnormal target when it is discovered, for subsequent regulatory judgment or evidence recording. With the application of drone inspections in scenarios such as natural resource supervision, river patrol, road management, and facility inspection, the inspection process not only needs to discover abnormal targets, but also needs to form a relatively stable judgment result based on the spatial relationship between the target and the regulatory area, boundary, or facility. Current drone inspection and evidence collection typically involves identifying suspected abnormal targets during the inspection process and supplementing the images or videos based on the target's location or preset shooting rules to obtain image or video data of the abnormal target.

[0003] However, in the current technology, when a suspected abnormal target is located near the regulatory judgment boundary, the target’s spatial position is uncertain due to the influence of UAV positioning error, attitude error and target measurement error. It is difficult to determine whether the current observation results can stably determine which side of the regulatory judgment boundary the target is located on based on this uncertainty, and it is also difficult to choose the next observation attitude that is conducive to reducing the ambiguity of the boundary judgment. This can easily lead to repeated observations or unstable boundary side judgment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs), which solves the problems of unstable judgment and difficulty in determining the direction of supplementary observation when suspected abnormal targets approach the regulatory judgment boundary.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs), comprising: S1. Obtain the current observation information collected by the drone inspection and identify suspected abnormal targets and their corresponding regulatory judgment boundaries; S2. Based on the current observation information, UAV pose information and measurement error parameters, determine the spatial estimation center and spatial covariance matrix of the suspected abnormal target, and construct a three-dimensional probability confidence ellipsoid as a set of spatial uncertainties. S3. Determine whether the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary. If it does, determine that the suspected abnormal target is in a judgment ambiguity state, and determine the degree of judgment ambiguity based on the projection uncertainty of the three-dimensional probability confidence ellipsoid in the normal direction of the regulatory judgment boundary. S4. In the state of ambiguity, generate multiple candidate observation poses and predict the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose. S5. Based on the cross-boundary volume ratio of each updated three-dimensional probability confidence ellipsoid relative to the regulatory decision boundary, determine the ambiguity convergence amount of each candidate observation posture and determine the next observation posture. The ambiguity convergence amount is used to characterize the reduction in the degree of ambiguity in the judgment of the suspected abnormal target relative to the regulatory decision boundary by the corresponding candidate observation posture. S6. Control the UAV to collect new observation information and update the three-dimensional probability confidence ellipsoid according to the next observation attitude. If the updated three-dimensional probability confidence ellipsoid does not cross the regulatory judgment boundary and meets the preset confidence conditions, so that the regulatory judgment result of the suspected abnormal target located on one side of the regulatory judgment boundary reaches the preset stability condition, then the evidence collection is completed; otherwise, S4 is executed again.

[0006] Preferably, step S1 specifically includes the following steps: Extract the inspection object or the target within the inspection area from the current observation information to obtain the target observation results that characterize the observation position and observation range of the target in the current observation coordinate system; The target observation results are matched with preset anomaly judgment conditions. When there is a matching preset anomaly judgment condition, the target is identified as a suspected anomaly target, and the anomaly type corresponding to the matching preset anomaly judgment condition is identified as the anomaly judgment type of the suspected anomaly target. According to the anomaly determination type, the regulatory determination rule is matched, and the regulatory determination boundary corresponding to the suspected anomaly target is determined in the preset regulatory space coordinate system in accordance with the regulatory determination rule.

[0007] Preferably, step S2 specifically includes the following steps: Based on the pose information of the UAV when collecting the current observation information, a spatial mapping is established between the current observation coordinate system and the preset monitoring space coordinate system. The target observation results are then transformed to the preset monitoring space coordinate system using the spatial mapping to determine the spatial estimation center of the suspected abnormal target. The measurement error parameters are converted into an observation covariance matrix, and the observation covariance matrix is ​​propagated to the preset monitoring space coordinate system according to the Jacobian matrix of the spatial mapping to obtain the spatial covariance matrix of the suspected abnormal target. The spatial covariance matrix is ​​subjected to eigenvalue decomposition. The direction of the principal axis of the ellipsoid is determined by the obtained eigenvectors. The length of the principal axis of the ellipsoid is determined by the obtained eigenvalues ​​and the preset confidence level. The three-dimensional probability confidence ellipsoid is constructed with the spatial estimation center as the center of the ellipsoid.

[0008] Preferably, obtaining the spatial covariance matrix of the suspected anomalous target specifically includes: Based on the observation position error and observation range error of the target observation results in the current observation coordinate system, the observation covariance matrix is ​​generated. The Jacobian matrix of the spatial mapping is obtained by taking the partial derivative of the observation position variable in the current observation coordinate system with respect to the spatial mapping, and the observation covariance matrix is ​​propagated to the preset monitoring spatial coordinate system through the Jacobian matrix. The spatial covariance matrix is ​​obtained by fusing the propagated covariance matrix and the pose error covariance corresponding to the UAV pose information.

[0009] Preferably, step S3 specifically includes the following steps: In the preset regulatory space coordinate system, the regulatory judgment boundary is represented as a boundary judgment function, wherein the function value sign of the boundary judgment function is used to characterize the spatial position located on different sides of the regulatory judgment boundary; Based on the function value of the spatial estimation center in the boundary determination function and the boundary normal vector of the regulatory determination boundary, the normal distance from the spatial estimation center to the regulatory determination boundary is determined, and the boundary direction projection radius is determined based on the projection of the three-dimensional probability confidence ellipsoid onto the boundary normal vector. When the normal distance is less than or equal to the projection radius of the boundary direction, it is determined that the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary, and the suspected abnormal target is determined to be in a judgment ambiguity state.

[0010] Preferably, step S4 specifically includes the following steps: Based on the length of each ellipsoidal principal axis of the three-dimensional probability confidence ellipsoid and the projection of each ellipsoidal principal axis direction onto the boundary normal vector, the contribution of each ellipsoidal principal axis to the projection radius of the boundary direction is determined, and the ellipsoidal principal axis with the largest contribution is determined as the decision ambiguity principal axis. Using the spatial estimation center as the observation pointing point, multiple candidate observation positions are determined, such that the projection magnitude of the unit direction vector of the determination ambiguity principal axis on the observation plane perpendicular to the candidate observation direction is not less than a preset projection threshold, and multiple candidate observation poses are generated according to each candidate observation position and its candidate observation direction pointing to the spatial estimation center. A predictive observation relationship for the suspected anomalous target is established based on each candidate observation pose, and an updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose is predicted based on the predictive observation relationship.

[0011] Preferably, the step of predicting the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose based on the predicted observation relationship specifically includes: For each candidate observation posture, a prediction observation function is established from the preset monitoring space coordinate system to the candidate observation coordinate system, and the partial derivative of the prediction observation function with respect to the target spatial position is obtained at the spatial estimation center to obtain the candidate observation Jacobian matrix; The predicted observation covariance matrix is ​​determined based on the candidate observation attitude and the preset measurement error model. The inverse of the spatial covariance matrix is ​​added to the observation information matrix formed by the candidate observation Jacobian matrix and the predicted observation covariance matrix, and the inverse of the addition result is obtained to obtain the predicted update spatial covariance matrix. Using the spatial estimation center as the center of the prediction ellipsoid, and based on the prediction update spatial covariance matrix and the preset confidence level, an updated three-dimensional probability confidence ellipsoid corresponding to the candidate observation attitude is constructed.

[0012] Preferably, step S5 specifically includes the following steps: The first volume portion located on the normal side of the regulatory judgment boundary and the second volume portion located on the abnormal side of the regulatory judgment boundary in the current three-dimensional probability confidence ellipsoid are determined respectively, and the current cross-boundary volume ratio is determined based on the first volume portion and the second volume portion, wherein the current cross-boundary volume ratio is used to characterize the degree of ambiguity in the judgment of the suspected abnormal target between the normal side and the abnormal side; For each candidate observation posture, the updated three-dimensional probability confidence ellipsoid is determined, and the volume portion of the updated three-dimensional probability confidence ellipsoid located on the normal side and abnormal side of the regulatory judgment boundary is determined accordingly. The predicted cross-boundary volume ratio corresponding to the candidate observation posture is then determined. Based on the reduction between the current cross-boundary volume ratio and the predicted cross-boundary volume ratio, the ambiguity convergence amount of the corresponding candidate observation posture is determined, and the ambiguity convergence amount is used to characterize the ability of the corresponding candidate observation posture to eliminate the uncertainty of the regulatory judgment result.

[0013] Preferably, determining the next observation attitude specifically includes the following steps: Candidate observation poses with ambiguity convergence greater than zero are selected to form a set of effective candidate observation poses; From the set of valid candidate observation poses, select the candidate observation pose with the largest ambiguity convergence. When there are multiple candidate observation poses with the same ambiguity convergence, the projection radius of the corresponding updated three-dimensional probability confidence ellipsoid on the boundary normal vector is determined, and the candidate observation pose with the smallest projection radius is determined as the next observation pose.

[0014] Preferably, step S6 specifically includes the following steps: Based on the new observation information collected by the UAV in the next observation attitude and the actual pose information when collecting the new observation information, the actual observation relationship is established, and the observation residual between the new observation information and the predicted observation information obtained based on the current spatial estimation center is determined. Based on the observation Jacobian matrix, observation covariance matrix, and observation residuals corresponding to the actual observation relationship, the current spatial estimation center and spatial covariance matrix are updated, and the three-dimensional probability confidence ellipsoid is reconstructed based on the updated spatial estimation center and spatial covariance matrix. Determine the probability that the suspected abnormal target is located on each side of the regulatory decision boundary. When the reconstructed three-dimensional probability confidence ellipsoid does not cross the regulatory decision boundary and the probability on one side reaches the preset confidence condition, the evidence collection of the suspected abnormal target is determined to be complete, and an evidence collection record containing the new observation information, actual pose information, updated spatial estimation center, spatial covariance matrix, regulatory decision boundary, and the probability of being located on each side of the regulatory decision boundary is generated. Otherwise, return to execute S4 based on the updated spatial estimation center and spatial covariance matrix.

[0015] This invention provides an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs). It has the following beneficial effects: 1. This invention represents suspected abnormal targets as three-dimensional probability confidence ellipsoids and determines whether they cross the regulatory judgment boundary. When there is cross-boundary ambiguity, the next observation posture is determined based on the convergence effect of the candidate observation posture on the cross-boundary volume ratio. This makes the supplementary observation process converge around the uncertainty near the regulatory judgment boundary, thereby reducing the risk of boundary judgment errors caused by target spatial position errors and improving the stability of inspection and evidence collection results.

[0016] 2. This invention determines the spatial estimation center and spatial covariance matrix based on current observation information, UAV pose information and measurement error parameters, and expresses the spatial uncertainty range of suspected abnormal targets with a three-dimensional probability confidence ellipsoid, which can avoid judging whether the target is located on one side of the regulatory judgment boundary based on a single estimation point.

[0017] 3. In this invention, evidence collection is completed when the updated three-dimensional probability confidence ellipsoid does not cross the regulatory judgment boundary and meets the preset confidence conditions; otherwise, candidate observation poses are generated and supplementary observations are performed, so that the evidence collection termination condition corresponds to the stability of the regulatory judgment result. Attached Figure Description

[0018] Figure 1 This is a flowchart of an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to the present invention. Figure 2 This is a schematic diagram of the structure of the three-dimensional probability confidence ellipsoid of the present invention; Figure 3 This is a schematic diagram of the planar structure of an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 -Appendix Figure 3 This invention provides an intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs), comprising: S1. Obtain the current observation information collected by the drone inspection and identify suspected abnormal targets and their corresponding regulatory judgment boundaries; Furthermore, S1 specifically includes the following steps: Extract the inspection object or the target within the inspection area from the current observation information to obtain the target observation results that characterize the observation position and observation range of the target in the current observation coordinate system; The target observation results are matched with the preset anomaly judgment conditions. When there are matching preset anomaly judgment conditions, the target is identified as a suspected anomaly target, and the anomaly type corresponding to the matching preset anomaly judgment conditions is identified as the anomaly judgment type of the suspected anomaly target. Based on the anomaly determination type, match the regulatory determination rules, and in accordance with the regulatory determination rules, determine the regulatory determination boundary corresponding to the suspected anomaly target in the preset regulatory space coordinate system.

[0021] Specifically, the UAV observes the inspection object or inspection area according to the preset inspection task, and obtains the current observation information. The current observation information may include visible light images, video frames, infrared images, point cloud data, UAV positioning information, flight altitude, gimbal angle and shooting time, etc. After parsing the current observation information, the inspection object or target within the inspection area is extracted to obtain the target observation results. The target observation results are used to characterize the observation position and observation range of the target in the current observation coordinate system. The observation position can be the target center point, the target contour center or the representative point of the target area. The observation range can be the target detection box, the target segmentation region, the target point cloud bounding box or the target spatial contour. Through this processing, the raw observation data collected by the UAV can be converted into target-level observation data, providing input for subsequent identification of suspected abnormal targets, calculation of spatial estimation center and construction of spatial uncertainty set. After obtaining the target observation results, the target observation results are matched with preset anomaly judgment conditions. The preset anomaly judgment conditions can be pre-configured according to the inspection business, such as target category, target size, target color, target temperature, target location, relative relationship between the target and the specified object, whether the target is located in a prohibited area or a regulatory area, etc. If the target observation results meet at least one preset anomaly judgment condition, the target is identified as a suspected anomaly target, and the anomaly type corresponding to the matching preset anomaly judgment condition is identified as the anomaly judgment type of the suspected anomaly target. This step can screen out the objects that need further judgment and evidence collection from multiple targets found during the UAV inspection process, so that the subsequent spatial uncertainty analysis is performed on the suspected anomaly targets. As an example, in the natural resource inspection scenario, if the UAV collects an image of a suspected pile of objects at the edge of a plot of land, and the target observation results show that the pile of objects is located near a preset regulatory area, and its area or shape characteristics meet the preset anomaly judgment conditions for illegal piles of objects, then the pile of objects can be identified as a suspected anomaly target, and its anomaly judgment type can be identified as suspected illegal piles of objects. After determining the anomaly type, the corresponding regulatory judgment rule is matched according to the anomaly type. Then, the regulatory judgment boundary corresponding to the suspected anomaly target is determined in the preset regulatory spatial coordinate system according to the regulatory judgment rule. The preset regulatory spatial coordinate system can be a three-dimensional spatial coordinate system uniformly used in the inspection area, or it can be a spatial coordinate system established based on geographic coordinates, elevation information, and regulatory layers. The regulatory judgment rule is used to characterize what kind of spatial boundary needs to be used to judge a certain type of anomaly target. The regulatory judgment boundary can be the land parcel red line, the boundary of the prohibited area, the river management scope line, the road berth line, the equipment safety distance boundary, the building control line, or other spatial boundaries with regulatory significance. Taking the above suspected illegal dumping as an example, if the regulatory judgment rule requires it to be judged whether the dumped material has entered the land parcel red line, then the land parcel red line is retrieved from the regulatory layer and determined as the regulatory judgment boundary corresponding to the suspected anomaly target in the preset regulatory spatial coordinate system. In this way, the regulatory judgment of the anomaly target can be converted into a spatial relationship judgment between the target spatial location and the regulatory judgment boundary, providing a boundary basis for subsequent judgment on whether the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary.

[0022] S2. Based on the current observation information, UAV pose information and measurement error parameters, determine the spatial estimation center and spatial covariance matrix of the suspected abnormal target, and construct a three-dimensional probability confidence ellipsoid as a set of spatial uncertainties. Furthermore, S2 specifically includes the following steps: Based on the pose information of the UAV when collecting current observation information, a spatial mapping between the current observation coordinate system and the preset monitoring space coordinate system is established, and the target observation results are transformed to the preset monitoring space coordinate system using the spatial mapping to determine the spatial estimation center of the suspected abnormal target. The measurement error parameters are converted into the observation covariance matrix, and the observation covariance matrix is ​​propagated to the preset monitoring space coordinate system according to the Jacobian matrix of the spatial mapping to obtain the spatial covariance matrix of the suspected abnormal target. The spatial covariance matrix is ​​decomposed into eigenvalues, and the direction of the principal axis of the ellipsoid is determined by the obtained eigenvectors. The length of the principal axis of the ellipsoid is determined by the obtained eigenvalues ​​and the preset confidence level. A three-dimensional probability confidence ellipsoid is constructed with the spatial estimation center as the center of the ellipsoid.

[0023] Specifically, based on the obtained target observation results and combined with the pose information of the UAV when collecting the current observation information, the suspected abnormal target is transformed from the current observation coordinate system to the preset monitoring space coordinate system. The UAV pose information can include the UAV's position, heading angle, pitch angle, roll angle, gimbal attitude, and camera intrinsic and extrinsic parameters. Based on the above pose information, a spatial mapping relationship between the current observation coordinate system and the preset monitoring space coordinate system is established, and the observation position in the target observation results is transformed to the preset monitoring space coordinate system. The transformed target representative point is used as the spatial estimation center of the suspected abnormal target. As an example, the UAV collects images of suspected piled objects above the edge of the plot. The target observation results give the target area center and outline range of the piled object in the image coordinates. Based on the UAV's positioning, altitude, attitude, and camera parameters when collecting the image, the target area center can be mapped to the monitoring space coordinate system, thereby obtaining the spatial estimation center of the suspected piled object relative to the plot red line. This allows subsequent monitoring judgments to be based on a unified spatial coordinate system, avoiding judgments based solely on the image planar position. After determining the spatial estimation center, the spatial uncertainty of the suspected anomalous target is determined based on the measurement error parameters. The measurement error parameters may include image target positioning error, target range extraction error, depth estimation error, camera calibration error, and UAV positioning and attitude measurement error. First, the observation error in the current observation coordinate system is converted into an observation covariance matrix. Then, combined with the Jacobian matrix spatially mapped to the target observation position, the observation covariance matrix is ​​propagated to the preset monitoring spatial coordinate system to obtain the spatial covariance matrix of the suspected anomalous target. This spatial covariance matrix is ​​used to describe the uncertainty distribution of the spatial estimation center in different directions. For example, if the UAV obliquely photographs the suspected piled-up object, the depth error of the target in the line-of-sight direction is usually greater than the image lateral positioning error. The spatial covariance matrix can reflect that the target has a large positional uncertainty in the line-of-sight direction. Through this processing, the suspected anomalous target is no longer expressed as a single spatial point, but as an uncertain spatial region with directional differences. After obtaining the spatial estimation center and the spatial covariance matrix, eigenvalue decomposition is performed on the spatial covariance matrix. The resulting eigenvectors are used to determine the principal axis directions of the three-dimensional probability confidence ellipsoid, and the resulting eigenvalues ​​are used to determine the uncertainty scale along each principal axis direction. A three-dimensional probability confidence ellipsoid is constructed with the spatial estimation center as the ellipsoid center and a pre-set confidence level. ; in, Represents a three-dimensional probability confidence ellipsoid. This represents any spatial location point in the preset regulatory spatial coordinate system. Indicates the spatial estimation center of suspected anomalous targets. This represents the spatial covariance matrix of the suspected anomalous target. Indicates the preset credit level. The threshold represents the three-dimensional chi-square distribution corresponding to the preset confidence level. That is, based on the spatial estimation center, the possible location of the suspected anomalous target is restricted to a three-dimensional ellipsoidal region whose shape and orientation are determined by the covariance matrix. In this embodiment, the spatial position estimation error of the suspected anomalous target approximately follows a three-dimensional Gaussian distribution. The spatial covariance matrix is ​​a positive definite matrix or a positive definite matrix after regularization. The threshold corresponding to the preset confidence level is a three-degree-of-freedom chi-square distribution threshold. When the preset confidence level is 0.95, the corresponding three-degree-of-freedom chi-square distribution threshold can be 7.815. For the spatial position error of the target that does not satisfy the Gaussian distribution assumption, Monte Carlo sampling can be used to obtain an equal probability confidence region, and the equal probability confidence region can be used to replace the three-dimensional probability confidence ellipsoid for subsequent cross-boundary judgment.

[0024] Furthermore, the spatial covariance matrix of the suspected anomalous target is obtained, specifically including: Based on the observation position error and observation range error of the target observation results in the current observation coordinate system, an observation covariance matrix is ​​generated; The Jacobian matrix of the spatial mapping is obtained by taking the partial derivative of the observation position variable in the current observation coordinate system with respect to the spatial mapping, and the observation covariance matrix is ​​propagated to the preset monitoring spatial coordinate system through the Jacobian matrix. The spatial covariance matrix is ​​obtained by fusing the propagated covariance matrix and the pose error covariance corresponding to the UAV pose information.

[0025] Specifically, when determining the spatial covariance matrix of suspected anomalous targets, the observation covariance matrix is ​​first generated based on the observation position error and observation range error of the target observation results in the current observation coordinate system. The observation position error can come from the target center point extraction error, image pixel positioning error, depth estimation error, or point cloud target center fitting error. The observation range error can come from the target detection box boundary deviation, target contour segmentation deviation, or point cloud enclosing range deviation. In practice, the position error variance of the target in each direction in the current observation coordinate system can be determined separately, and the variance value of the corresponding direction can be corrected according to the dispersion of the target observation range, thereby forming the observation covariance matrix. As an example, when the UAV takes oblique pictures of suspected piles of objects at the edge of the plot, the image lateral positioning error is small, while the target depth direction is greatly affected by the shooting angle. Therefore, the variance corresponding to the depth direction in the observation covariance matrix can be greater than the variance corresponding to the image lateral direction. Through this processing, the observation error of suspected anomalous targets can be expressed in matrix form, providing basic data for subsequent spatial uncertainty propagation. After obtaining the observation covariance matrix, based on the spatial mapping relationship from the current observation coordinate system to the preset monitoring spatial coordinate system, the partial derivative of the observation position variable in the current observation coordinate system is calculated to obtain the Jacobian matrix of the spatial mapping. This Jacobian matrix is ​​used to characterize the influence of small position errors in the current observation coordinate system on position errors in the preset monitoring spatial coordinate system. Subsequently, the observation covariance matrix is ​​propagated to the preset monitoring spatial coordinate system through this Jacobian matrix to obtain the propagated covariance matrix. This propagation process transforms the observation error from the observation coordinates of the camera, radar, or other sensors into the spatial coordinates used for monitoring and judgment. For example, suspected piled-up objects may only appear as pixel deviations in the target area in an image, but in the monitoring spatial coordinate system where the land parcel boundary is located, these pixel deviations will be affected by the drone altitude, shooting angle, and target distance, transforming into spatial position uncertainties of different magnitudes and directions. Through this processing, the subsequent ellipsoid construction and the monitoring and judgment boundary can be based on the same coordinate system. After obtaining the propagated covariance matrix, the spatial covariance matrix of the suspected anomalous target is further fused by combining the pose error covariance corresponding to the UAV pose information. The UAV pose error covariance can be determined based on the error parameters of the positioning module, inertial measurement unit, heading measurement module, and gimbal attitude measurement module. It is used to characterize the uncertainty in the UAV's position, attitude, and observation direction. During fusion, the propagated target observation error and the spatial positioning error caused by the UAV pose error are included in the spatial covariance matrix of the suspected anomalous target. Taking the suspected pile of objects as an example, if the UAV positioning accuracy is high but the gimbal pitch angle error is large, the spatial covariance matrix will show a large uncertainty in the direction related to the observation line of sight. Thus, the spatial covariance matrix not only reflects the error of the target observation result itself, but also reflects the influence of the pose error when the UAV collects the observation information, so that the subsequently constructed three-dimensional probability confidence ellipsoid can be closer to the spatial uncertainty range in the actual evidence collection scenario.

[0026] S3. Determine whether the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary. If it does, determine that the suspected abnormal target is in a judgment ambiguity state, and determine the degree of judgment ambiguity based on the projection uncertainty of the three-dimensional probability confidence ellipsoid in the normal direction of the regulatory judgment boundary. Furthermore, S3 specifically includes the following steps: In the preset regulatory space coordinate system, the regulatory decision boundary is represented as a boundary decision function, where the function value sign of the boundary decision function is used to characterize the spatial location located on different sides of the regulatory decision boundary; Based on the function value of the spatial estimation center in the boundary judgment function and the boundary normal vector of the regulatory judgment boundary, the normal distance from the spatial estimation center to the regulatory judgment boundary is determined, and the projection radius of the boundary direction is determined based on the projection of the three-dimensional probability confidence ellipsoid onto the boundary normal vector. When the absolute value of the normal distance is less than or equal to the projection radius of the boundary direction, it is determined that the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary, and the suspected abnormal target is determined to be in a judgment ambiguity state.

[0027] Specifically, the regulatory decision boundary is expressed functionally within the preset regulatory space coordinate system. Specifically, the regulatory decision boundary can be represented as a boundary decision function: ; in, This indicates a spatial location point in the preset regulatory spatial coordinate system. This represents the judgment function value of the spatial location point relative to the regulatory judgment boundary, thus transforming the regulatory judgment boundary into a computable spatial interface. When, it indicates that the spatial location point is located on one side of the regulatory judgment boundary, when When, it indicates that the spatial location point is located on the other side of the regulatory judgment boundary, when When the boundary is a planar boundary, the boundary normal vector is directly determined based on that planar boundary. When the boundary is a curved boundary, the boundary tangent plane is determined at the boundary point closest to the spatial estimation center, and the normal vector of the boundary tangent plane is used as the local boundary normal vector. The boundary determination function is differentiable at the corresponding location, and its gradient magnitude is greater than zero. As an example, in the land parcel red line supervision scenario, the red line boundary or the three-dimensional boundary interface formed by the expansion of the red line can be used as the supervision determination boundary. The boundary determination function is used to determine whether the spatial location of the suspected piled-up object is inside or outside the red line. After obtaining the boundary determination function, based on the function value of the spatial estimation center of the suspected abnormal target in the boundary determination function, and the boundary normal vector of the regulatory determination boundary near that location, the normal distance from the spatial estimation center to the regulatory determination boundary is determined: ; in, This represents the signed normal distance from the spatial estimation center to the regulatory decision boundary. Indicates the spatial estimation center of suspected anomalous targets. This represents the function value after substituting the spatial estimation center into the boundary judgment function. This represents the gradient of the boundary decision function at the spatial estimation center. This represents the magnitude of the gradient. Through the above, the functional offset of the spatial estimation center relative to the regulatory decision boundary can be converted into a distance along the boundary normal direction, allowing the proximity between the target center and the regulatory boundary to be quantified. It is calculated from the spatial estimation center and regulatory judgment boundary determined in real time, without setting fixed values; Subsequently, the projection radius of the boundary direction is determined based on the projection of the three-dimensional probability confidence ellipsoid onto the boundary normal vector. The unit boundary normal vector can be expressed as: ; The projection radius along the boundary direction can be expressed as: ; in, The unit normal vector representing the boundary of the regulatory determination. This represents the projection radius of the three-dimensional probability confidence ellipsoid onto the boundary normal direction. This represents the spatial covariance matrix of the suspected anomalous target. This represents the three-dimensional chi-square distribution threshold corresponding to a preset confidence level. It involves projecting the three-dimensional spatial uncertainty of a suspected abnormal target along the normal direction of the regulatory judgment boundary to obtain the uncertainty range of the target's position in the boundary judgment direction. As a feasible implementation method, a preset confidence level is used. A value of 0.95 can be used, corresponding to the threshold of the three-dimensional chi-square distribution. We can take 7.815, for the spatial covariance matrix. Determined by the aforementioned observation error, spatial mapping, and pose error; When the normal distance from the spatial estimation center to the regulatory decision boundary is less than or equal to the projection radius of the boundary direction, the following condition is met: ; If the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary, and the suspected anomalous target is in a state of judgment ambiguity, then although the spatial estimation center may be located on one side of the regulatory judgment boundary, the spatial uncertainty range of the target has extended to the other side of the boundary. Therefore, based solely on the current observation information, it is not possible to reliably determine whether the suspected anomalous target belongs to the normal side or the anomalous side. Taking the suspected stockpile as an example, if the spatial estimation center of the stockpile is 0.6m away from the land parcel red line, and the projection radius of the three-dimensional probability confidence ellipsoid in the red line normal direction is 1.1m, then the ellipsoid crosses the land parcel red line, indicating that the current observation results are insufficient to reliably determine whether the stockpile has entered the red line range. After determining the state of ambiguity, the degree of ambiguity can be further determined based on the relative relationship between the normal distance and the projection radius of the boundary direction. ; Where A represents the degree of ambiguity in the judgment, and d represents the signed normal distance from the spatial estimation center to the regulatory judgment boundary. A represents the projection radius along the boundary direction. Based on the uncertainty range along the boundary direction, it measures the degree to which the spatial estimation center is close to the regulatory judgment boundary. The closer A is to 1, the closer the spatial estimation center is to the regulatory judgment boundary and the more obvious the uncertainty along the boundary direction. When A is 0, it means that no cross-boundary ambiguity has been formed or the degree of ambiguity can be ignored. This degree of ambiguity can be used as the basis for generating candidate observation poses and evaluating the ambiguity convergence effect.

[0028] S4. In the case of ambiguity, generate multiple candidate observation poses and predict the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose. Furthermore, S4 specifically includes the following steps: Based on the lengths of the principal axes of the three-dimensional probability confidence ellipsoid and the projection of the directions of the principal axes onto the boundary normal vector, the contribution of each principal axis to the projection radius of the boundary direction is determined, and the principal axis with the largest contribution is determined as the decision ambiguity principal axis. Using the spatial estimation center as the observation pointing point, multiple candidate observation positions are determined, ensuring that the projection magnitude of the unit direction vector of the decision ambiguity axis on the observation plane perpendicular to the candidate observation direction is not less than a preset projection threshold, and generating multiple candidate observation poses based on each candidate observation position and its candidate observation direction pointing to the spatial estimation center. Based on each candidate observation pose, a predictive observation relationship for suspected anomalous targets is established, and the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose is predicted based on the predictive observation relationship.

[0029] Specifically, when a suspected abnormal target is in a state of ambiguous judgment, the main directions causing uncertainty in the boundary direction are determined based on the principal axis distribution of the three-dimensional probability confidence ellipsoid. That is, the principal axis lengths and directions are determined for the three principal axes of the three-dimensional probability confidence ellipsoid, and the contribution of each principal axis to the uncertainty of the boundary direction projection is calculated in combination with the unit normal vector of the regulatory judgment boundary. ; in, Indicates the first The contribution of each principal axis of the ellipsoid to the uncertainty of the projection along the boundary direction. Indicates the first The length of the principal axis of the ellipsoid Indicates the first The unit direction vectors corresponding to the principal axes of the ellipsoid The unit normal vector representing the boundary of the regulatory determination. This indicates the principal axis number of the ellipsoid, with values ​​of 1, 2, or 3. The physical meaning of this formula is to measure the uncertainty component generated by each principal axis in the direction normal to the regulatory decision boundary. The longer the principal axis and the closer its direction is to the boundary normal direction, the greater its contribution to boundary ambiguity. The principal axis with the largest contribution is determined as the decision ambiguity principal axis, i.e.: ; in, This indicates the decision ambiguity principal axis number, which is used to indicate the spatial uncertainty direction that should be the focus of subsequent observations for convergence; After determining the principal axis of ambiguity, multiple candidate observation positions are generated around the suspected anomalous target, using the spatial estimation center as the observation pointing point. These candidate observation positions can be generated according to preset radius, altitude range, and azimuth intervals, or they can be filtered based on UAV flight safety distance, terrain obstacles, and flight path accessibility. As a feasible implementation, 8 or 12 candidate observation positions can be generated within a range of 20m to 50m from the spatial estimation center. The candidate observation altitude can be set to 15m to 40m. For each candidate observation position, its candidate observation direction pointing to the spatial estimation center is determined, and the horizontal projection length of the principal axis of ambiguity on the observation plane perpendicular to that candidate observation direction is calculated. ; in, Indicates the first The candidate observation pose is an observability index for determining the ambiguous principal axis. This represents the unit direction vector that determines the principal axis of ambiguity. Indicates the first The unit observation direction vector corresponding to each candidate observation attitude. This represents the candidate observation pose number. The physical meaning of this formula is that it measures the degree to which the principal axis of ambiguity can be resolved in the candidate observation image plane or observation cross-section. If the candidate observation direction is too close to being parallel to the principal axis of ambiguity, the projection of that principal axis onto the observation plane is small, which is not conducive to reducing the uncertainty in that direction. The candidate observation pose satisfies... At that time, it is retained as a usable candidate observation pose, where, As a preset projection threshold, a feasible implementation method is to set it to 0.5. This threshold is used to exclude candidate poses with low observability for determining the ambiguous principal axis. After obtaining multiple candidate observation postures, predictive observation relationships for suspected abnormal targets are established for each candidate observation posture. These predictive observation relationships characterize the observability of the target position near the spatial estimation center by the UAV at the corresponding candidate observation position and candidate observation direction. Combining the candidate observation posture, sensor measurement error parameters, and the current spatial covariance matrix, the updated three-dimensional probability confidence ellipsoid corresponding to the execution of the candidate observation posture can be predicted. Taking the scenario where the suspected pile of objects is close to the red line of the plot as an example, if the current three-dimensional probability confidence ellipsoid is longer along the normal direction of the red line, the candidate observation posture close to the normal side of the red line can usually make the direction form a more obvious projection in the observation plane, thus providing observation conditions for reducing cross-boundary judgment ambiguity in the future. This allows the generation of the next observation posture to be associated with the spatial uncertainty direction near the regulatory judgment boundary.

[0030] Furthermore, based on the predicted observation relationship, the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose is predicted, specifically including: For each candidate observation attitude, a prediction observation function is established from the preset monitoring space coordinate system to the candidate observation coordinate system. The partial derivative of the prediction observation function with respect to the target spatial position is obtained at the spatial estimation center to obtain the candidate observation Jacobian matrix. The predicted observation covariance matrix is ​​determined based on the candidate observation attitude and the preset measurement error model. The inverse of the spatial covariance matrix is ​​added to the observation information matrix formed by the candidate observation Jacobian matrix and the predicted observation covariance matrix, and the inverse of the addition result is obtained to obtain the predicted update spatial covariance matrix. Using the spatial estimation center as the center of the prediction ellipsoid, and based on the prediction update spatial covariance matrix and the preset confidence level, an updated three-dimensional probability confidence ellipsoid corresponding to the candidate observation attitude is constructed.

[0031] Specifically, for each candidate observation posture, a predictive observation relationship is first established from the preset monitoring space coordinate system to the candidate observation coordinate system. This predictive observation relationship is used to represent the spatial position of the suspected abnormal target in the monitoring space coordinate system. If the UAV observes according to a certain candidate observation posture, the predictive observation result will be formed in the corresponding candidate observation coordinate system. The candidate observation coordinate system can be determined by the candidate observation position, candidate observation direction, gimbal attitude, and sensor installation relationship. For the j-th candidate observation posture, near the spatial estimation center, the change relationship of the target spatial position in the predictive observation relationship is linearized to obtain the candidate observation Jacobian matrix Hj. This candidate observation Jacobian matrix is ​​used to characterize the degree of influence of the small positional change of the target in the monitoring space coordinate system on the predictive observation result in the candidate observation coordinate system. For example, for suspected piles near the red line of the plot, the small offset of the piles along the red line normal direction under different candidate observation postures will not be reflected in the image or point cloud observation. The candidate observation Jacobian matrix can reflect this difference in observation sensitivity. After obtaining the candidate observation Jacobian matrix, the predicted observation covariance matrix Rj is determined based on the candidate observation attitude and the preset measurement error model. The predicted observation covariance matrix represents the measurement error that the UAV may generate when observing in the j-th candidate observation attitude. This error can be related to the candidate observation distance, observation angle, sensor type, image resolution, depth estimation accuracy, and gimbal stability. As a feasible implementation, when using a visible light camera for target position observation, the image positioning error can be set to 1 to 3 pixels. When combined with depth estimation or point cloud measurement, the standard deviation of depth direction measurement can be set to 0.2m to 1m based on the observation distance. The above data are only used to form the preset measurement error model, and the actual values ​​can be determined based on the UAV payload and inspection environment. Subsequently, the inverse matrix of the current spatial covariance matrix is ​​added to the observation information matrix formed by the candidate observation Jacobian matrix and the predicted observation covariance matrix, and the inverse of the addition result is obtained to obtain the predicted update spatial covariance matrix. ; in, Let represent the prediction update space covariance matrix corresponding to the j-th candidate observation pose. Represents the current spatial covariance matrix. Let represent the candidate observation Jacobian matrix corresponding to the j-th candidate observation attitude. Let represent the predicted observation covariance matrix corresponding to the j-th candidate observation attitude. This allows for the fusion of the current spatial uncertainty with the observation information provided by the candidate observation attitude, thus estimating the degree to which the spatial position uncertainty of suspected abnormal targets may converge if the UAV executes the candidate observation attitude. This can be understood as the amount of observation information provided by the j-th candidate observation attitude to the target's spatial position. The smaller the observation error and the more sensitive the observation direction is to the uncertainty of the target, the more obvious the effect of this term on the convergence of uncertainty. After obtaining the predicted updated spatial covariance matrix, the current spatial estimation center is used as the center of the predicted ellipsoid. Based on the predicted updated spatial covariance matrix and a preset confidence level, an updated 3D probability confidence ellipsoid corresponding to the j-th candidate observation posture is constructed. This updated 3D probability confidence ellipsoid is not the final ellipsoid after actual observation, but rather a prediction of the spatial uncertainty convergence result that the posture may produce before the UAV executes the candidate observation posture. Continuing with the scenario where the suspected stockpile is close to the land parcel boundary line as an example, if a certain candidate observation posture can significantly reduce the spatial uncertainty of the stockpile along the normal direction of the boundary line, the scale of the updated 3D probability confidence ellipsoid corresponding to the candidate observation posture in the normal direction of the boundary line will be reduced. Through this processing, the expected evidence collection effects of multiple candidate observation postures can be compared before the UAV actually moves, providing a basis for selecting the next observation posture based on the cross-boundary volume ratio and the amount of ambiguity convergence.

[0032] S5. Based on the cross-boundary volume ratio of each updated three-dimensional probability confidence ellipsoid relative to the regulatory decision boundary, determine the ambiguity convergence of each candidate observation posture and determine the next observation posture. The ambiguity convergence is used to characterize the reduction in the degree of ambiguity in the judgment of the suspected abnormal target relative to the regulatory decision boundary by the corresponding candidate observation posture. Furthermore, S5 specifically includes the following steps: The first volume portion located on the normal side of the regulatory judgment boundary and the second volume portion located on the abnormal side of the regulatory judgment boundary in the current three-dimensional probability confidence ellipsoid are determined respectively. The current cross-boundary volume ratio is determined based on the first volume portion and the second volume portion. The current cross-boundary volume ratio is used to characterize the degree of ambiguity in the judgment of suspected abnormal targets between the normal side and the abnormal side. For each candidate observation posture, the updated three-dimensional probability confidence ellipsoid is determined, and the volume portion of the updated three-dimensional probability confidence ellipsoid located on the normal side and abnormal side of the regulatory judgment boundary is determined accordingly. Based on this, the predicted cross-boundary volume ratio corresponding to the candidate observation posture is determined. Based on the reduction between the current cross-boundary volume ratio and the predicted cross-boundary volume ratio, the ambiguity convergence amount of the corresponding candidate observation posture is determined, and the ambiguity convergence amount is used to characterize the ability of the corresponding candidate observation posture to eliminate the uncertainty of the regulatory judgment result.

[0033] Specifically, first, determine the cross-boundary volume distribution of the current three-dimensional probability confidence ellipsoid relative to the regulatory decision boundary. That is, take the regulatory decision boundary as the spatial interface and divide the current three-dimensional probability confidence ellipsoid to obtain a first volume portion located on the normal side of the regulatory decision boundary and a second volume portion located on the abnormal side of the regulatory decision boundary. The first volume portion can be denoted as... The second volume part can be denoted as ,in, This represents the volume within the current three-dimensional probability confidence ellipsoid located on the normal side of the regulatory decision boundary. This represents the volume located on the abnormal side of the regulatory judgment boundary within the current three-dimensional probability confidence ellipsoid. In practice, two volume parts can be calculated using spatial mesh partitioning, Monte Carlo sampling, or boundary cutting. Taking a scenario where suspected dumped objects are close to the land parcel boundary line as an example, if the current three-dimensional probability confidence ellipsoid covers both the inside and outside of the boundary line, the volume inside and outside the boundary line are respectively used as the abnormal side volume and the normal side volume in subsequent calculations. Through this processing, the ambiguous state of the target relative to the regulatory judgment boundary can be further quantified from whether it crosses the boundary to the degree of crossing the boundary.

[0034] Based on the first volume portion and the second volume portion, determine the current cross-boundary volume percentage: ; in, Indicates the cross-boundary volume ratio. This represents the volume portion of the three-dimensional probability confidence ellipsoid located on the normal side of the regulatory decision boundary. This represents the volume portion of the three-dimensional probability confidence ellipsoid located on the abnormal side of the regulatory judgment boundary. The total volume of the three-dimensional probability confidence ellipsoid is represented by the ratio between the volume of the smaller side and the total volume of the ellipsoid. This ratio measures the degree of distribution balance of the three-dimensional probability confidence ellipsoid on both sides of the regulatory decision boundary. If the ellipsoid is almost entirely located on the same side, the volume of the smaller side is close to zero, and the proportion of the cross-boundary volume is close to zero, indicating low decision ambiguity. If the ellipsoid is relatively close in distribution on the normal side and the abnormal side, the proportion of the cross-boundary volume is relatively large, indicating that the current observation results are unstable in the decision on the side where the target is located. This proportion of the cross-boundary volume is calculated in real time from the current ellipsoid and the regulatory decision boundary. Among them, the cross-boundary volume ratio is a geometric distribution index of the three-dimensional probability confidence ellipsoid relative to the regulatory judgment boundary. It is used to evaluate the degree of distribution balance of the confidence ellipsoid on both sides of the regulatory judgment boundary. It is not directly equivalent to the probability that the suspected abnormal target is located on either side of the regulatory judgment boundary. The probability that the suspected abnormal target is located on each side of the regulatory judgment boundary is calculated separately based on the spatial estimation center, spatial covariance matrix and boundary judgment function. For each candidate observation attitude, the updated 3D probability confidence ellipsoid is partitioned using the same volume division method as the current 3D probability confidence ellipsoid. The volume portions of the updated 3D probability confidence ellipsoid located on the normal and abnormal sides of the regulatory judgment boundary are determined, and the predicted cross-boundary volume ratio is determined accordingly. For the ... The predicted cross-boundary volume percentage corresponding to each candidate observation posture can be denoted as: ,in, This indicates that if the drone follows the first... The system observes each candidate observation attitude to predict the extent to which the updated 3D probability confidence ellipsoid remains within the regulatory boundary. Continuing with the example of suspected debris near the land parcel boundary line, if a candidate observation attitude causes the predicted updated ellipsoid to significantly shrink to the outside of the boundary line, then the predicted cross-boundary volume corresponding to that candidate observation attitude is relatively small; conversely, if a large portion of the predicted updated ellipsoid is still distributed on both sides of the boundary line, then the predicted cross-boundary volume is relatively large. This process allows for comparison of the expected reduction effect of each candidate observation attitude on cross-boundary ambiguity before actual flight. After obtaining the current cross-boundary volume percentage and the predicted cross-boundary volume percentage for each candidate observation posture, determine the ambiguity convergence factor for each candidate observation posture: ; in, Indicates the first Ambiguity convergence quantity corresponding to each candidate observation pose This represents the current cross-boundary volume percentage corresponding to the current three-dimensional probability confidence ellipsoid. Indicates the first The predicted cross-boundary volume ratio of the predicted updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation posture is used to quantify the ability of the candidate observation posture to eliminate the uncertainty of the regulatory judgment result by comparing the change of the cross-boundary volume ratio before and after the observation. when When, it indicates that the candidate observation pose can reduce the ambiguity in the judgment of suspected abnormal targets relative to the regulatory judgment boundary, when If the candidate observation pose does not effectively contribute to reducing decision ambiguity, then the selection of the next observation pose changes from a fixed shooting angle to the expected convergence effect on regulatory decision ambiguity.

[0035] Further, determining the next observation attitude includes the following steps: Candidate observation poses with ambiguity convergence greater than zero are selected to form a set of effective candidate observation poses; From the set of valid candidate observation poses, select the candidate observation pose with the largest ambiguity convergence. When there are multiple candidate observation poses with the same ambiguity convergence, the projection radius of the corresponding updated 3D probability confidence ellipsoid on the boundary normal vector is determined, and the candidate observation pose with the smallest projection radius is determined as the next observation pose.

[0036] Specifically, after determining the ambiguity convergence of each candidate observation pose, the effectiveness of the candidate observation poses is first screened. Specifically, candidate observation poses with ambiguity convergence greater than zero are retained, forming a set of effective candidate observation poses. This set can be represented as: ; in, Represents the set of valid candidate observation poses. Indicates the candidate observation attitude number, Indicates the first The ambiguity convergence value corresponding to each candidate observation pose will filter out candidate observation poses that can reduce the cross-boundary volume ratio, ensuring that the subsequent selection process is only performed on candidate observation poses with ambiguity convergence effect. If a certain candidate observation pose... This indicates that the predicted cross-boundary volume ratio after the update is not less than the current cross-boundary volume ratio. This candidate observation attitude is not the preferred choice for the next observation attitude. If the set of effective candidate observation attitudes is empty, the candidate observation position can be regenerated or the candidate observation height, distance and azimuth range can be adjusted and the ambiguity convergence amount can be recalculated. After forming a set of effective candidate observation postures, the candidate observation posture with the largest ambiguity convergence is selected from the set. The larger the ambiguity convergence, the more the candidate observation posture prediction can reduce the cross-boundary volume ratio of the suspected abnormal target relative to the regulatory judgment boundary, and thus the stronger its ability to eliminate the uncertainty of the regulatory judgment result. As an example, if four candidate observation postures are generated around the suspected pile of objects, and their ambiguity convergence values ​​are 0.03, 0.12, 0.12 and -0.02 respectively, then the first, second and third candidate observation postures are effective candidate observation postures, and the fourth candidate observation posture is not included in the subsequent priority selection. At this time, the second and third candidate observation postures have large ambiguity convergence values ​​and can be used as candidates for the next observation posture. When multiple candidate observation poses with the same ambiguity convergence value exist, the projection radii of the updated 3D probability confidence ellipsoids corresponding to each candidate observation pose on the normal direction of the regulatory decision boundary are further compared, and the candidate observation pose with the smallest projection radius is determined as the next observation pose. This selection process can be expressed as: ; in, This indicates the candidate observation attitude number that has been determined as the next observation attitude. Represents the set of valid candidate observation poses. Indicates the first Ambiguity convergence of candidate observation poses Indicates the first The projection radius of the updated 3D probability confidence ellipsoid corresponding to each candidate observation pose on the boundary normal direction, in this formula This indicates that candidate observation poses are selected according to a sorting rule, first by comparing... And then Comparison when they are the same ,because The larger the corresponding The smaller the value, the better. Therefore, this formula can first select the attitude with a larger ambiguity convergence amount, and then select the attitude with a smaller boundary normal projection radius when the ambiguity convergence amount is the same. Thus, when multiple candidate observation attitudes have the same cross-boundary volume convergence effect, the observation attitude that can further compress the uncertainty of the boundary normal direction is given priority. This makes the next observation more conducive to the stable determination of the suspected abnormal target relative to the regulatory judgment boundary.

[0037] S6. Control the UAV to collect new observation information and update the three-dimensional probability confidence ellipsoid according to the next observation attitude. If the updated three-dimensional probability confidence ellipsoid does not cross the regulatory judgment boundary and meets the preset confidence conditions, so that the regulatory judgment result of the suspected abnormal target located on one side of the regulatory judgment boundary reaches the preset stability condition, then the evidence collection is completed; otherwise, S4 is executed again.

[0038] Furthermore, S6 specifically includes the following steps: Based on the new observation information collected by the UAV in the next observation attitude and the actual pose information when collecting the new observation information, the actual observation relationship is established, and the observation residual between the new observation information and the predicted observation information obtained based on the current spatial estimation center is determined. Based on the observation Jacobian matrix, observation covariance matrix and observation residuals corresponding to the actual observation relationship, the current spatial estimation center and spatial covariance matrix are updated, and the three-dimensional probability confidence ellipsoid is reconstructed based on the updated spatial estimation center and spatial covariance matrix. Determine the probability of a suspected anomalous target being located on each side of the regulatory decision boundary. When the reconstructed 3D probability confidence ellipsoid does not cross the regulatory decision boundary and the probability on one side reaches the preset confidence condition, the evidence collection of the suspected anomalous target is completed, and an evidence collection record containing new observation information, actual pose information, updated spatial estimation center, spatial covariance matrix, regulatory decision boundary, and the probability of being located on each side of the regulatory decision boundary is generated. Otherwise, return to execute S4 based on the updated spatial estimation center and spatial covariance matrix.

[0039] Specifically, after determining the next observation attitude, the UAV is controlled to move to the corresponding observation position, and the sensors are oriented towards the spatial estimation center of the suspected abnormal target to collect new observation information. The new observation information may include new images, video frames, point clouds, or other sensor data used to locate the target. At the same time, the actual pose information when collecting the new observation information is recorded. The actual pose information includes the UAV position, attitude, gimbal angle, and acquisition time. Based on the new observation information and the actual pose information, an actual observation relationship is established, and the current spatial estimation center is substituted into the actual observation relationship to obtain the predicted observation information. Subsequently, the new observation information is compared with the predicted observation information to determine the observation residual. The observation residual is used to characterize the deviation between the actual acquisition result and the prediction result based on the current spatial estimation center. Taking the suspected pile of objects close to the red line of the plot as an example, after the UAV performs lateral observation according to the next observation attitude, the outline of the pile of objects in the new image is offset relative to the predicted position. This offset can be used as the input to update the spatial estimation center and the spatial covariance matrix. After obtaining the observation residuals, the current spatial estimation center and spatial covariance matrix are updated based on the observation Jacobian matrix, observation covariance matrix, and observation residuals corresponding to the actual observation relationship. The observation Jacobian matrix represents the impact of the positional changes of the suspected anomalous target in the monitoring spatial coordinate system on the actual observation results, and the observation covariance matrix represents the measurement error of this actual observation. Through this update process, the spatial estimation center of the suspected anomalous target can be corrected using the new observation information, and its spatial covariance matrix can be adjusted. The updated spatial estimation center represents the estimated position of the target after this supplementary observation, and the updated spatial covariance matrix represents the spatial uncertainty after the supplementary observation. Subsequently, following the aforementioned method for constructing a three-dimensional probability confidence ellipsoid, a new three-dimensional probability confidence ellipsoid is reconstructed based on the updated spatial estimation center and spatial covariance matrix. This process enables the UAV to obtain a new set of spatial uncertainties after each active observation, providing a basis for determining whether to continue observation. After reconstructing the three-dimensional probability confidence ellipsoid, the probability that a suspected anomalous target is located on each side of the regulatory decision boundary is determined. In this embodiment, a boundary decision function is defined. The side that is abnormal is the side that is not. The side with the abnormal side is the normal side. The probability of the abnormal side can be expressed as: ; ; in, This indicates the probability that a suspected abnormal target is located on the abnormal side of the regulatory judgment boundary. This indicates the probability that a suspected abnormal target is located on the normal side of the regulatory judgment boundary. Represents the standard normal distribution function. This represents the boundary determination function corresponding to the regulatory determination boundary. Indicates the updated spatial estimation center. This represents the updated spatial covariance matrix. The superscript indicates the gradient of the boundary decision function at the center of the updated spatial estimate. This indicates transposition, which maps the updated target space uncertainty along the regulatory decision boundary in a first-order manner, thereby estimating the probability that a suspected abnormal target lies on the abnormal or normal side of the regulatory decision boundary. If in actual business... One side is defined as the abnormal side, and adjustments can be made accordingly. and The correspondence; When the reconstructed three-dimensional probability confidence ellipsoid does not cross the regulatory judgment boundary, and the probability that the suspected abnormal target is located on one side reaches the preset confidence condition, the evidence collection is considered complete. This stability condition can be expressed as: ; and: ; in, This represents the probability threshold corresponding to the preset confidence condition. As a feasible implementation method, it can be set to 0.95. This represents the signed normal distance from the updated spatial estimation center to the regulatory decision boundary. This represents the projection radius of the updated 3D probability confidence ellipsoid onto the normal direction of the regulatory decision boundary. Simultaneously, it constrains the target space uncertainty set to no longer cross the regulatory decision boundary, and the probability of the target being located on one side reaches a preset confidence level. If the above conditions are met, an evidence record is generated. The evidence record includes new observation information, actual pose information, the updated spatial estimation center, the spatial covariance matrix, the regulatory decision boundary, and the probability of being located on each side of the regulatory decision boundary. If the above conditions are not met, S4 is re-executed based on the updated spatial estimation center and spatial covariance matrix to continue generating candidate observation poses and conducting the next round of active observation.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent decision-making method based on UAV anomaly target inspection and evidence collection, characterized in that, include: S1. Obtain the current observation information collected by the drone inspection and identify suspected abnormal targets and their corresponding regulatory judgment boundaries; S2. Based on the current observation information, UAV pose information and measurement error parameters, determine the spatial estimation center and spatial covariance matrix of the suspected abnormal target, and construct a three-dimensional probability confidence ellipsoid as a set of spatial uncertainties. S3. Determine whether the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary. If it does, determine that the suspected abnormal target is in a judgment ambiguity state, and determine the degree of judgment ambiguity based on the projection uncertainty of the three-dimensional probability confidence ellipsoid in the normal direction of the regulatory judgment boundary. S4. In the state of ambiguity, generate multiple candidate observation poses and predict the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose. S5. Based on the cross-boundary volume ratio of each updated three-dimensional probability confidence ellipsoid relative to the regulatory decision boundary, determine the ambiguity convergence amount of each candidate observation posture and determine the next observation posture. The ambiguity convergence amount is used to characterize the reduction in the degree of ambiguity in the judgment of the suspected abnormal target relative to the regulatory decision boundary by the corresponding candidate observation posture. S6. Control the UAV to collect new observation information and update the three-dimensional probability confidence ellipsoid according to the next observation attitude. If the updated three-dimensional probability confidence ellipsoid does not cross the regulatory judgment boundary and meets the preset confidence conditions, so that the regulatory judgment result of the suspected abnormal target located on one side of the regulatory judgment boundary reaches the preset stability condition, then the evidence collection is completed; otherwise, S4 is executed again.

2. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S1 specifically includes the following steps: Extract the inspection object or the target within the inspection area from the current observation information to obtain the target observation results that characterize the observation position and observation range of the target in the current observation coordinate system; The target observation results are matched with preset anomaly judgment conditions. When there is a matching preset anomaly judgment condition, the target is identified as a suspected anomaly target, and the anomaly type corresponding to the matching preset anomaly judgment condition is identified as the anomaly judgment type of the suspected anomaly target. According to the anomaly determination type, the regulatory determination rule is matched, and the regulatory determination boundary corresponding to the suspected anomaly target is determined in the preset regulatory space coordinate system in accordance with the regulatory determination rule.

3. The intelligent decision-making method based on UAV abnormal target inspection and evidence collection according to claim 2, characterized in that, S2 specifically includes the following steps: Based on the pose information of the UAV when collecting the current observation information, a spatial mapping is established between the current observation coordinate system and the preset monitoring space coordinate system. The target observation results are then transformed to the preset monitoring space coordinate system using the spatial mapping to determine the spatial estimation center of the suspected abnormal target. The measurement error parameters are converted into an observation covariance matrix, and the observation covariance matrix is ​​propagated to the preset monitoring space coordinate system according to the Jacobian matrix of the spatial mapping to obtain the spatial covariance matrix of the suspected abnormal target. The spatial covariance matrix is ​​subjected to eigenvalue decomposition. The direction of the principal axis of the ellipsoid is determined by the obtained eigenvectors. The length of the principal axis of the ellipsoid is determined by the obtained eigenvalues ​​and the preset confidence level. The three-dimensional probability confidence ellipsoid is constructed with the spatial estimation center as the center of the ellipsoid.

4. The intelligent decision-making method based on UAV abnormal target inspection and evidence collection according to claim 3, characterized in that, The process of obtaining the spatial covariance matrix of the suspected anomalous target specifically includes: Based on the observation position error and observation range error of the target observation results in the current observation coordinate system, the observation covariance matrix is ​​generated. The Jacobian matrix of the spatial mapping is obtained by taking the partial derivative of the observation position variable in the current observation coordinate system with respect to the spatial mapping, and the observation covariance matrix is ​​propagated to the preset monitoring spatial coordinate system through the Jacobian matrix. The spatial covariance matrix is ​​obtained by fusing the propagated covariance matrix and the pose error covariance corresponding to the UAV pose information.

5. The intelligent decision-making method based on UAV abnormal target inspection and evidence collection according to claim 4, characterized in that, S3 specifically includes the following steps: In the preset regulatory space coordinate system, the regulatory judgment boundary is represented as a boundary judgment function, wherein the function value sign of the boundary judgment function is used to characterize the spatial position located on different sides of the regulatory judgment boundary; Based on the function value of the spatial estimation center in the boundary determination function and the boundary normal vector of the regulatory determination boundary, the normal distance from the spatial estimation center to the regulatory determination boundary is determined, and the boundary direction projection radius is determined based on the projection of the three-dimensional probability confidence ellipsoid onto the boundary normal vector. When the normal distance is less than or equal to the projection radius of the boundary direction, it is determined that the three-dimensional probability confidence ellipsoid crosses the regulatory judgment boundary, and the suspected abnormal target is determined to be in a judgment ambiguity state.

6. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 5, characterized in that, S4 specifically includes the following steps: Based on the length of each ellipsoidal principal axis of the three-dimensional probability confidence ellipsoid and the projection of each ellipsoidal principal axis direction onto the boundary normal vector, the contribution of each ellipsoidal principal axis to the projection radius of the boundary direction is determined, and the ellipsoidal principal axis with the largest contribution is determined as the decision ambiguity principal axis. Using the spatial estimation center as the observation pointing point, multiple candidate observation positions are determined, such that the projection magnitude of the unit direction vector of the determination ambiguity principal axis on the observation plane perpendicular to the candidate observation direction is not less than a preset projection threshold, and multiple candidate observation poses are generated according to each candidate observation position and its candidate observation direction pointing to the spatial estimation center. A predictive observation relationship for the suspected anomalous target is established based on each candidate observation pose, and an updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose is predicted based on the predictive observation relationship.

7. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 6, characterized in that, The step of predicting the updated three-dimensional probability confidence ellipsoid corresponding to each candidate observation pose based on the predicted observation relationship specifically includes: For each candidate observation posture, a prediction observation function is established from the preset monitoring space coordinate system to the candidate observation coordinate system, and the partial derivative of the prediction observation function with respect to the target spatial position is obtained at the spatial estimation center to obtain the candidate observation Jacobian matrix; The predicted observation covariance matrix is ​​determined based on the candidate observation attitude and the preset measurement error model. The inverse of the spatial covariance matrix is ​​added to the observation information matrix formed by the candidate observation Jacobian matrix and the predicted observation covariance matrix, and the inverse of the addition result is obtained to obtain the predicted update spatial covariance matrix. Using the spatial estimation center as the center of the prediction ellipsoid, and based on the prediction update spatial covariance matrix and the preset confidence level, an updated three-dimensional probability confidence ellipsoid corresponding to the candidate observation attitude is constructed.

8. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 7, characterized in that, S5 specifically includes the following steps: The first volume portion located on the normal side of the regulatory judgment boundary and the second volume portion located on the abnormal side of the regulatory judgment boundary in the current three-dimensional probability confidence ellipsoid are determined respectively, and the current cross-boundary volume ratio is determined based on the first volume portion and the second volume portion, wherein the current cross-boundary volume ratio is used to characterize the degree of ambiguity in the judgment of the suspected abnormal target between the normal side and the abnormal side; For each candidate observation posture, the updated three-dimensional probability confidence ellipsoid is determined, and the volume portion of the updated three-dimensional probability confidence ellipsoid located on the normal side and abnormal side of the regulatory judgment boundary is determined accordingly. The predicted cross-boundary volume ratio corresponding to the candidate observation posture is then determined. Based on the reduction between the current cross-boundary volume ratio and the predicted cross-boundary volume ratio, the ambiguity convergence amount of the corresponding candidate observation posture is determined, and the ambiguity convergence amount is used to characterize the ability of the corresponding candidate observation posture to eliminate the uncertainty of the regulatory judgment result.

9. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 8, characterized in that, Determining the next observation attitude specifically includes the following steps: Candidate observation poses with ambiguity convergence greater than zero are selected to form a set of effective candidate observation poses; From the set of valid candidate observation poses, select the candidate observation pose with the largest ambiguity convergence. When there are multiple candidate observation poses with the same ambiguity convergence, the projection radius of the corresponding updated three-dimensional probability confidence ellipsoid on the boundary normal vector is determined, and the candidate observation pose with the smallest projection radius is determined as the next observation pose.

10. The intelligent decision-making method for abnormal target inspection and evidence collection based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S6 specifically includes the following steps: Based on the new observation information collected by the UAV in the next observation attitude and the actual pose information when collecting the new observation information, the actual observation relationship is established, and the observation residual between the new observation information and the predicted observation information obtained based on the current spatial estimation center is determined. Based on the observation Jacobian matrix, observation covariance matrix, and observation residuals corresponding to the actual observation relationship, the current spatial estimation center and spatial covariance matrix are updated, and the three-dimensional probability confidence ellipsoid is reconstructed based on the updated spatial estimation center and spatial covariance matrix. Determine the probability that the suspected abnormal target is located on each side of the regulatory decision boundary. When the reconstructed three-dimensional probability confidence ellipsoid does not cross the regulatory decision boundary and the probability on one side reaches the preset confidence condition, the evidence collection of the suspected abnormal target is determined to be complete, and an evidence collection record containing the new observation information, actual pose information, updated spatial estimation center, spatial covariance matrix, regulatory decision boundary, and the probability of being located on each side of the regulatory decision boundary is generated. Otherwise, return to execute S4 based on the updated spatial estimation center and spatial covariance matrix.

Images