Unmanned aerial vehicle safety control method

Abstract

The application provides a kind of unmanned plane safety control method, belongs to unmanned plane autonomous control technical field, it aims at solving the security risk caused by the simplification of risk assessment model in prior art. The method comprises: obtaining semantic information and confidence of a plurality of targets in a scene;Semantic clustering is performed on the targets to obtain semantic category clusters;Calculate the individual layer uncertainty of each target, the category layer uncertainty of each semantic category cluster, and the global layer uncertainty representing the current scene;Calculate the target level control uncertainty of each target according to the above three layer uncertainties;Finally, control the unmanned plane based on the target level control uncertainty of each target. The present application establishes a three-layer uncertainty structure of "global layer-category layer-individual layer", which quantifies the uncertainty from different sources in detail, realizes more accurate risk identification and more sensitive control response.

Description

Technical Field

[0001] This invention relates to the field of autonomous control technology for unmanned aerial vehicles (UAVs), and more particularly to a method for safe control of UAVs. Background Technology

[0002] In recent years, with the development of large-scale language models and multimodal perception technologies, drones have begun to utilize advanced semantic information for scene understanding and decision-making. In a typical "perception-decision-control" architecture, the perception layer uses sensors such as vision and depth to acquire environmental information, the decision-making layer identifies targets and understands tasks through semantic models, and the control layer is responsible for executing safe flight maneuvers.

[0003] However, a current technical bottleneck lies in how to reliably and stably transmit the rich but uncertain semantic information (such as target category and recognition confidence) output from the decision-making layer to the lower-level controller. Existing technologies typically handle this uncertainty in a relatively simple way. One approach is to perform simple threshold filtering or weighted averaging of the confidence scores of multiple detected targets to obtain a comprehensive risk index. The drawback of this approach is that when there is a harmless target with high confidence at a distance and a dangerous target with low confidence at close range in the scene, the risk of the latter can easily be averaged or diluted by the "high score" of the former, causing the system to ignore the real threat.

[0004] Another type of technology, while attempting to incorporate semantic information into advanced controllers (such as control barrier functions (CBF) or model predictive control (MPC), often only utilizes partial information and fails to fully consider the inherent, multi-layered uncertainties of the semantic information itself. Specifically, these solutions typically mix all sources of uncertainty together to form a single risk scalar, failing to distinguish whether the uncertainty stems from poor overall scene quality, inconsistencies in class identification, or the unreliability of a particular target, ultimately making it difficult to generate accurate control commands.

[0005] Therefore, existing technologies generally suffer from poor risk assessment and an inability to distinguish sources of uncertainty, resulting in difficulties in generating accurate control commands. Summary of the Invention

[0006] This invention provides a method for safe control of unmanned aerial vehicles (UAVs) to address the common problem in existing technologies that results in crude risk assessments, an inability to distinguish sources of uncertainty, and ultimately, difficulty in generating accurate control commands.

[0007] This invention provides a method for safe control of unmanned aerial vehicles (UAVs), comprising: The semantic information and semantic confidence of multiple targets in the current scene are obtained by drones; Semantic clustering of multiple targets yields multiple semantic category clusters; The individual-level uncertainty is calculated based on the semantic information of each target. The category-level uncertainty is calculated based on the semantic confidence consistency among multiple targets within each semantic category cluster. Global layer uncertainty, used to characterize the uncertainty of the current scene, is calculated based on the semantic confidence of each of the aforementioned targets. The target-level control uncertainty is calculated based on the individual-level uncertainty of each target, the category-level uncertainty of its semantic category cluster, and the global-level uncertainty. The UAV is controlled based on the target-level control uncertainty of each of the aforementioned targets.

[0008] In some embodiments, semantic clustering of multiple targets yields multiple semantic category clusters, specifically including: A first semantic graph is constructed based on the semantic information of the multiple targets, with each target and the UAV as nodes; in the first semantic graph, the edge weight between two nodes is determined based on the semantic similarity and spatial proximity between the two nodes; Multiple semantic category clusters are obtained by clustering multiple targets based on the first semantic graph.

[0009] In some embodiments, the semantic information includes at least semantic feature embedding and spatial location; The semantic similarity between two nodes is characterized by the cosine similarity or distance between the feature embeddings of the two nodes; The spatial proximity between two nodes is characterized by the difference between the spatial positions of the two nodes.

[0010] In some embodiments, clustering multiple targets based on the first semantic graph to obtain multiple semantic category clusters specifically includes: The first semantic graph is subjected to sparsity processing to obtain a second semantic graph; the sparsity processing includes deleting edges with weights less than a preset threshold. The targets in the second semantic graph are clustered using a graph clustering method to obtain multiple semantic category clusters; the graph clustering method is one of connected component partitioning, spectral clustering, community detection, or embedded similarity clustering.

[0011] In some embodiments, the semantic information includes at least spatial location and velocity; The individual-level uncertainty is calculated based on the semantic information of each target, specifically including: The relative position and relative velocity between each target and the UAV are determined based on the spatial position and velocity of each target, and the individual layer uncertainty is calculated by combining the semantic confidence of each target.

[0012] In some embodiments, a global layer uncertainty characterizing the uncertainty of the current scene is calculated based on the semantic confidence of each of the targets, specifically including: Determine the semantic confidence score with the highest value among the semantic confidence scores of each of the stated targets; The global layer uncertainty, which characterizes the uncertainty of the current scenario, is calculated based on the semantic confidence score with the highest value and using a negative correlation mapping.

[0013] In some embodiments, the UAV is controlled based on the target-level control uncertainty of each of the targets, specifically including: Calculate scenario-level control uncertainty based on the target-level control uncertainty of each of the aforementioned objectives; Dynamic safety distances are generated based on the scenario-level control uncertainty and the target-level control uncertainty of each target, respectively. The UAV is controlled based on the scenario-level control uncertainty and the dynamic safe distance of each target.

[0014] In some embodiments, the scenario-level control uncertainty is calculated based on the target-level control uncertainty of each of the objectives, specifically including: The target-level control uncertainty of each of the aforementioned targets is aggregated by maximum value aggregation or softmax weight aggregation to obtain the scenario-level control uncertainty.

[0015] In some embodiments, the dynamic safety distance of each target is generated based on the scenario-level control uncertainty and the target-level control uncertainty, respectively, specifically including: Obtain a basic safe distance; A first safety distance is generated for each target based on the target-level control uncertainty, and the first safety distance increases exponentially with the target-level control uncertainty. The second safety distance for each target is obtained by summing the first safety distance and the basic safety distance for each target. The dynamic safety distance of each target is generated based on the second safety distance of each target and the scenario-level control uncertainty.

[0016] In some embodiments, controlling the UAV based on the scenario-level control uncertainty and the dynamic safety distance of each target specifically includes: The dynamic safety distance is used as a constraint and input into the model prediction controller; The control input for the UAV is obtained by solving the model predictive controller. The UAV is controlled based on the control input and the scenario-level control uncertainty.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the unmanned aerial vehicle (UAV) safety control method as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the unmanned aerial vehicle (UAV) safety control method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the unmanned aerial vehicle (UAV) safety control method as described above.

[0020] The UAV safety control method provided by this invention establishes a three-layer uncertainty structure of "global layer - category layer - individual layer" to determine the individual layer uncertainty, category layer uncertainty and global layer uncertainty of each target respectively, realizing multi-level calculation of uncertainty. This allows for more accurate calculation of the target-level control uncertainty of each target. Based on the target-level control uncertainty of each target, the UAV is controlled, thereby achieving a more precise control effect.

[0021] Furthermore, spatial risk modulation of individual-level uncertainty can accurately identify and amplify the real high risk of close-range, low-confidence dynamic targets; at the same time, by means of maximum value aggregation and other methods, it ensures that the risk of the most dangerous targets is not diluted by average, thereby triggering an earlier and stronger safety obstacle avoidance response, which significantly improves safety in complex dynamic environments.

[0022] Furthermore, it employs a dynamic safety distance generation mechanism that exhibits an accelerating relationship with the uncertainty of target-level control, enabling the safety boundary to expand rapidly under high-risk conditions. Compared to traditional linear adjustment methods, this invention maintains flexibility when the risk is low and provides more comprehensive and reasonable protection when the risk increases, achieving a better balance between safety and task efficiency.

[0023] Meanwhile, the entire method is mainly based on analytical formula calculation and sparse graph operations, without relying on a complex online learning process. It has low computational complexity and is suitable for real-time operation on airborne edge devices with limited computing resources. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the unmanned aerial vehicle (UAV) safety control method provided by the present invention; Figure 2 This is a schematic diagram illustrating the construction of the semantic graph in the UAV safety control method provided by the present invention; Figure 3 This is a schematic diagram illustrating the calculation of three layers of uncertainty in the UAV safety control method provided by the present invention; Figure 4 This is a schematic diagram of the spatial risk modulation function in the UAV safety control method provided by the present invention; Figure 5 This is a schematic diagram of the dynamic safety distance space expansion in the UAV safety control method provided by the present invention; Figure 6 This is a graph showing the relationship between dynamic safety distance and risk in the UAV safety control method provided by this invention; Figure 7 This is a schematic diagram of a typical scenario where a long-range high-confidence target and a short-range low-confidence dynamic target coexist. Figure 8 This is a comparison chart of the risk assessment results of the method of this invention and the traditional average risk method; Figure 9 This is a comparison chart of the changes in safe distance and dynamic target distance between the method of this invention and the traditional average risk method; Figure 10 This is a comparison chart of the control response intensity of the method of this invention and the traditional average risk method; Figure 11 A comparison chart of changes in safe distance under different risk modeling methods; Figure 12 Comparison chart of minimum obstacle distances under different risk modeling methods; Figure 13 A comparison chart of response time and control strength under different risk modeling methods; Figure 14 A comparison chart of obstacle avoidance trajectory results under different risk modeling methods. Detailed Implementation

[0026] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0027] In the description of this invention, it should be noted that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0028] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0029] Reference Figure 1 In one embodiment of this application, the drone safety control method specifically includes: Step S110: The UAV acquires semantic information and semantic confidence of multiple targets in the current scene.

[0030] Step S120: Semantic clustering is performed on multiple targets to obtain multiple semantic category clusters.

[0031] Step S130: Calculate the individual-level uncertainty of each target based on its semantic information.

[0032] Step S140: Calculate the category layer uncertainty based on the semantic confidence consistency among multiple targets within each semantic category cluster.

[0033] Step S150: Calculate the global layer uncertainty used to characterize the uncertainty of the current scene based on the semantic confidence of each target.

[0034] Step S160: Calculate the target-level control uncertainty for each target based on the individual-level uncertainty, the category-level uncertainty of its semantic category cluster, and the global-level uncertainty.

[0035] Step S170: Control the UAV based on the target-level control uncertainty of each target.

[0036] In the above technical solution, by establishing a three-layer uncertainty structure of "global layer - category layer - individual layer", the individual layer uncertainty, category layer uncertainty and global layer uncertainty of each target are determined respectively, realizing multi-level calculation of uncertainty. This allows for more accurate calculation of the target-level control uncertainty of each target. Based on the target-level control uncertainty of each target, the UAV can be controlled, thereby achieving a more precise control effect.

[0037] Specifically, drones first acquire richer semantic information (environmental representation) about the target than traditional detection boxes, providing a data foundation for subsequent refined risk assessment.

[0038] For example, a drone can be equipped with image acquisition devices, depth sensors, or multimodal perception modules, and perform inference on the currently acquired scene images, depth information, point cloud information, and task commands on an onboard computing platform to obtain a set of targets and semantic information such as the semantic category, spatial location, semantic confidence, feature embedding, task relevance weight, and velocity of each target. Spatial location can be obtained from depth information, 3D detection results, multi-view reconstruction results, temporal target tracking results, or external positioning modules; relative velocity can be provided by target state difference between consecutive frames, filtered estimation, or external perception modules.

[0039] For example, the set of targets identified at the current time t can be represented as: ; Where M is the total number of targets identified in the current scene. This represents the j-th target in the current scene. The drone's state at time t. It can be represented as: ; in, Let be the spatial position of the UAV at time t. Let be the velocity of the drone at time t. Let be the attitude or other flight state variables of the UAV at time t. For any target Its semantic information can be represented as: ; in, This represents the semantic category of target j at time t. This represents the spatial position of target j at time t. This represents the semantic confidence of target j at time t. This represents the semantic feature embedding of target j at time t. This represents the relevance weight between target j and the current task at time t. This represents the velocity of target j at time t.

[0040] Since targets in the real world do not exist in isolation but are semantically and spatially related, clustering can organize semantically similar (e.g., all are vehicles) or spatially adjacent targets. This operation solves the problem of treating all targets as independent individuals in the previous technique, thus ignoring group risks or inconsistencies in identification within categories, and lays the structural foundation for subsequent calculation of category-level uncertainty. Then, the three-layer uncertainty of "global layer - category layer - individual layer" is calculated.

[0041] Step S130 aims to quantify the risk posed by a single target, modeling the risk of each target independently to avoid being averaged by other targets. By calculating an independent individual-level uncertainty for each target, the shortcomings of the background technology, which uses averaging or weighted averaging methods, that mask high-risk individuals are addressed, achieving accurate characterization of individual danger. Step S140 aims to quantify the overall recognition quality of a certain type of target. For example, if the semantic confidence of multiple targets within a "pedestrian" cluster fluctuates, it indicates that the system's recognition of the "pedestrian" category is unstable in that scenario. By introducing category-level uncertainty, the shortcomings of the background technology, which cannot distinguish whether a single target or a certain type of target is problematic, are addressed, achieving assessment of category-level risk. Step S150 aims to assess the overall perception quality of the entire scenario from a macroscopic perspective. Global-level uncertainty is determined based on the semantic confidence of each target. For example, if the overall semantic confidence of each target is relatively high, it indicates that the system's overall perception quality of the scenario is high, meaning that the global-level uncertainty is relatively low.

[0042] After obtaining the three levels of uncertainty, they are fused to obtain the target-level control uncertainty for each objective. The principle is that the ultimate risk of an objective is determined by its individual characteristics, the characteristics of its category, and the global environment in which it exists. By weighted and fused these three levels of uncertainty, the problem of a single dimension of risk assessment in the background technology is solved, and a comprehensive and multi-level consideration of the sources of risk is achieved.

[0043] Finally, the previously calculated quantitative risks (uncertainty in controlling each target level) are transformed into specific, actionable control measures.

[0044] In one specific implementation, semantic clustering of multiple targets yields multiple semantic category clusters. Specifically, this includes: constructing a first semantic graph with each target and the drone as nodes based on the semantic information of the multiple targets; and clustering the multiple targets based on the first semantic graph to obtain multiple semantic category clusters. In the first semantic graph, the edge weight between two nodes is determined based on the semantic similarity and spatial proximity between the two nodes.

[0045] This implementation explicitly models the complex relationships between targets using a graph structure. By constructing a semantic graph, not only semantic similarity but also spatial proximity can be utilized, making the clustering results more consistent with physical and semantic logic. This design addresses the problem that simple clustering based on feature vectors may ignore spatial layout, achieving more robust category clustering. Specifically, semantic similarity ensures that targets that are similar in category are associated, while spatial proximity ensures that targets that are physically close are associated. The edge weights defined in this way can more accurately reflect the true strength of the association between targets, providing a more reliable basis for subsequent graph clustering and category-level uncertainty calculations.

[0046] Furthermore, semantic information includes semantic feature embeddings and spatial location; the semantic similarity between two nodes is characterized by the cosine similarity or distance between the feature embeddings of the two nodes; the spatial proximity between two nodes is characterized by the difference between the spatial locations of the two nodes.

[0047] Cosine similarity using feature embeddings is a well-established method for measuring high-dimensional semantic similarity, while spatial location differences (such as Euclidean distance) are a direct method for measuring physical proximity. By clarifying these computational methods, clear guidance is provided for the engineering implementation of technical solutions.

[0048] In this preferred embodiment, clustering multiple targets based on the first semantic graph to obtain multiple semantic category clusters specifically includes: performing sparse processing on the first semantic graph to obtain a second semantic graph; the sparse processing includes deleting edge weights less than a preset threshold; and using graph clustering to cluster each target in the second semantic graph to obtain multiple semantic category clusters; the graph clustering method is one of connected component partitioning, spectral clustering, community detection, or embedded similarity clustering.

[0049] Specifically, the purpose of sparsity processing here is to remove weak associations and retain strong associations, thereby reducing the computational complexity of the graph and highlighting the core structure. Using mature algorithms such as graph clustering, communities with tight internal connections, i.e., semantic category clusters, can be efficiently partitioned from the graph. This solves the problems of high computational cost and susceptibility to noise interference in clustering on fully connected graphs, achieving efficient and accurate category cluster partitioning.

[0050] Reference Figure 2 In one example, the steps for constructing a semantic graph and clustering it are as follows: 1. Construct the first semantic graph based on the semantic feature embedding and spatial location of each target.

[0051] The first semantic graph comprises nodes representing all targets and the UAV itself. Its edge weights consider both semantic similarity and spatial proximity between targets. Semantic similarity is represented by the cosine similarity or distance of the feature embeddings, while spatial proximity is represented by the positional difference between targets. The first semantic graph can be represented as: ; ; in, , , and Let be the first semantic graph at time t, and its set of nodes, set of edges, and set of edge weights. and Let be the UAV's own node at time t and the Mth target node, respectively. For any two target nodes... and Its border rights The preferred definition is: ; ; ; in, To balance the weighting coefficients of semantic similarity and spatial proximity, and These are the semantic distance scale parameter and the spatial distance scale parameter, respectively. To prevent constants with a denominator of zero, For the target node and Semantic similarity between them For the target node and Spatial proximity between them and Target nodes and semantic feature embedding, and Target nodes and Spatial location.

[0052] 2. Perform sparse processing on the first semantic graph to obtain the second semantic graph.

[0053] Specifically, when the edge power Less than the preset threshold At that time, the target node is considered and The relationship between nodes is weak, so the target node should be deleted. and Edges between them are used to obtain a sparse second semantic graph.

[0054] 3. Based on sparse semantic graphs, perform connected component partitioning, spectral clustering, community detection, or embedding similarity clustering to obtain several semantic category clusters: ; in, and The semantic category cluster set at time t and the set of the first time are respectively the set of semantic category clusters at time t and the set of the second time t. l There are L semantic category clusters, where L is the total number of semantic category clusters.

[0055] It's important to note that the "category" here includes not only the category labels directly output by the semantic model, but also dynamic aggregations based on embedding similarity in the semantic graph. For example, when the model outputs different labels such as "sedan," "SUV," and "pickup," the semantic graph can aggregate them into a "vehicle class" based on embedding similarity for category-level uncertainty calculation. Through this semantic graph, subsequent risk estimation no longer relies solely on a single target, but can leverage the semantic and spatial relationships between targets to support the calculation of category-level uncertainty.

[0056] The above is an example illustrating the construction of a semantic graph and its clustering. It should be noted that the semantic confidence of each objective will be needed in the subsequent process of constructing and calculating the three levels of uncertainty (risk modeling).

[0057] In one specific implementation, the semantic information includes spatial position and velocity; the individual-level uncertainty is calculated based on the semantic information of each target, specifically including: determining the relative position and relative velocity between each target and the UAV based on the spatial position and velocity of each target, and calculating the individual-level uncertainty by combining the semantic confidence.

[0058] Specifically, the risk of a target depends not only on the reliability of its identification but also on the degree of its spatial threat to the UAV. A rapidly approaching close-range target, even with a slightly lower confidence level, poses a far greater danger than a stationary distant target. By incorporating relative position and relative velocity into the calculation of individual-level uncertainties, the problem of neglecting the dynamic characteristics of targets in the prior art is solved, enabling dynamic assessment of spatial threats.

[0059] In one specific implementation, global layer uncertainty is calculated based on the semantic confidence of each target to characterize the uncertainty of the current scenario. Specifically, this includes: determining the semantic confidence with the largest value among the semantic confidence of each target; and calculating global layer uncertainty to characterize the uncertainty of the current scenario based on the semantic confidence with the largest value and using negative correlation mapping.

[0060] Specifically, since global layer uncertainty needs to characterize the uncertainty of the current scenario, it is necessary to comprehensively consider the semantic confidence of all targets for determination. For example, a higher comprehensive semantic confidence of all targets indicates that all targets are credible overall, meaning that global layer uncertainty is relatively low; conversely, a lower comprehensive semantic confidence of all targets indicates that all targets are untrustworthy overall, meaning that global layer uncertainty is relatively high. In this specific implementation, a clear and easy-to-implement method is provided to calculate global layer uncertainty. When the maximum value of the semantic confidence of each target is small, it usually means that the semantic confidence of all targets is relatively small, and a larger global layer uncertainty is obtained through negative correlation mapping; when the maximum value of the semantic confidence of each target is large, it usually means that the upper limit of the semantic confidence of all targets is relatively high, and a smaller global layer uncertainty is obtained through negative correlation mapping.

[0061] For example, the global layer uncertainty can be obtained by using a negative correlation mapping form of "1 minus the maximum semantic confidence among all targets".

[0062] The following is an example of calculating individual-level uncertainty, category-level uncertainty, global-level uncertainty, and target-level control uncertainty.

[0063] Reference Figure 3 In one example: 1. The first layer is the global uncertainty layer, used to characterize the credibility of the overall semantic parsing of the current scene. Its calculation method is as follows: ; in, and represents the global layer uncertainty at time t and the maximum semantic confidence of all targets, respectively.

[0064] When the semantic confidence of the most reliable target in the scene remains low, it indicates poor overall semantic parsing quality. This could be due to factors such as deteriorating lighting, severe occlusion, overall semantic understanding distortion, or out-of-model input. In this case, global layer uncertainty increases, triggering more conservative control behavior. If no target exists, then... Set as a preset conservative constant Alternatively, it can directly enter the maximum conservative mode; if the system has reliable non-semantic obstacle avoidance sensors, it can also maintain a basic safe distance and continue to perform regular obstacle avoidance when there is no semantic target.

[0065] 2. The second layer is the category-level uncertainty, used to characterize the consistency of semantic confidence among targets within the same semantic category cluster. Preferably, the dispersion of semantic confidence and intra-graph consistency are used together to represent the category-level uncertainty. For the k-th semantic category cluster... Its category layer uncertainty : ; in, and These are non-negative weighting coefficients. Let be the semantic confidence of target j at time t. Let be the edge weight between targets i and j at time t. The first term... Used to measure the semantic confidence fluctuation of targets within the same category cluster; the second item This is used to measure the consistency of semantic graph connections within the same category cluster. When the semantic confidence of multiple targets within a category cluster fluctuates significantly, or the edge weights of the semantic graph are weak, it indicates that the category recognition is unstable and higher uncertainty should be passed to the control layer.

[0066] It should be noted that if the only goal is to reduce computational complexity, only intra-class variance, range, interquartile range, or coefficient of variation can be used as the uncertainty of the class layer.

[0067] For the category cluster boundary case, when At that time, since the within-class variance and the average edge weight within the graph are not statistically significant, it can be... The value can be set to zero, a preset prior value, or estimated from the semantic confidence fluctuation of the category within the historical window. When only some valid edges exist within a category cluster, the graph consistency term is only averaged over the valid edges. When there are no valid edges, the graph consistency term can be ignored, and only the semantic confidence dispersion or prior term is retained. The above processing is used to ensure that a complete and executable risk mapping can still be formed in the cases of single-objective, few-objective, and edgeless graphs.

[0068] 3. The third layer represents individual-level uncertainty, characterizing the unreliability of a single target and amplifying or suppressing it through a spatial risk modulation mechanism. First, the relative distance between the target and the UAV is calculated: ; in, and Let be the spatial positions of target j and the UAV at time t, respectively. Let be the relative distance between target j and the UAV at time t.

[0069] Further calculate the target's approach speed relative to the UAV, preferably defined as a positive value in the direction of decreasing distance: ; in, Let be the approach velocity of target j relative to the UAV at time t. and The velocities of target j and the UAV at time t are respectively. The dimensions are the same as distance, used to avoid The denominator is zero when the value is close to zero; the result in parentheses is negative when the target is far away from the drone. After processing .

[0070] in, The larger the value, the faster the target approaches the drone.

[0071] Reference Figure 4 Spatial risk modulation function From the distance term and speed term Common components: ; ; ; in, The saturation clipping function is defined as follows: Used to restrict the variable y to an interval Inside. For modulation parameters, To prevent excessive risk amplification, an upper limit is set. The distance term rapidly increases the risk when the target enters the danger zone; the speed term provides additional amplification for high-speed approaching targets. This results in individual-level uncertainty. .

[0072] Furthermore, the target-level control uncertainty can be calculated based on the individual-level uncertainty of each target, the category-level uncertainty of its semantic category cluster, and the global-level uncertainty.

[0073] set up Let j represent the semantic category cluster to which target j belongs. Then, the target-level control uncertainty of target j can be expressed as: ; in, Let be the target-level control uncertainty of target j at time t. , and The weights are for three levels, The global layer uncertainty at time t, Let t be the semantic category cluster to which target j belongs. Uncertainty in the category layer, Let be the individual-level uncertainty of target j at time t. Let t be the correlation weight between target j and the current task at time t.

[0074] The above is a calculation example of three levels of uncertainty and target-level control uncertainty.

[0075] In one specific implementation, the UAV is controlled based on the target-level control uncertainty of each target, specifically including: calculating the scenario-level control uncertainty based on the target-level control uncertainty of each target; generating the dynamic safety distance of each target based on the target-level control uncertainty of each target; and controlling the UAV based on the scenario-level control uncertainty and the dynamic safety distance of each target.

[0076] Calculating scenario-level control uncertainties is used for macro-level regulation, while generating dynamic safety distances for each target is used for micro-level avoidance. This hierarchical control strategy solves the problem that a single control command cannot simultaneously address global and local safety, achieving more refined safety control.

[0077] Specifically, the scenario-level control uncertainty is calculated based on the target-level control uncertainty of each objective. This includes: performing maximum value aggregation or softmax weight aggregation on the target-level control uncertainty of each objective to obtain the scenario-level control uncertainty.

[0078] The principle behind maximum value aggregation is the "weakest link" effect, meaning the overall risk of a scenario is determined by the most dangerous target. This approach addresses the fundamental problem of risk being averaged and sparse in background techniques. Softmax weight aggregation, on the other hand, is a smooth approximation that highlights the maximum value while also considering other secondary risks, and it possesses the desirable mathematical property of differentiability. Providing these two aggregation methods enhances the algorithm's applicability.

[0079] For example, if maximum aggregation is used, the formula for generating scenario-level control uncertainty is: .

[0080] If Softmax weight aggregation is used, the formula for generating scene-level control uncertainty is: ; in, This represents the maximum target-level control uncertainty at time t. Used to Limited to the range Inside, Let be the scenario-level control uncertainty at time t. This is the soft maximum temperature parameter. To control the upper limit of uncertainty, Let t represent the target-level control uncertainty of target j.

[0081] The purpose of using maximum aggregation is to explicitly retain the most dangerous target and avoid the average dilution of risk between distant high-confidence targets and nearby low-confidence targets. Soft maximum aggregation can be used as a subordinate scheme for optimizers that require a continuously differentiable risk function.

[0082] In one specific implementation, the dynamic safety distance exhibits an accelerating relationship with target-level control uncertainty. The principle behind this is that risk growth is often non-linear; a small increase in uncertainty in low-risk areas may be harmless, but the same increase in high-risk areas could lead to catastrophic consequences. By designing an accelerating functional relationship (such as one containing quadratic terms), the safety distance increases moderately when risk is low and expands sharply when risk is high. This design addresses the problem of insufficient protection in high-risk areas by traditional linear adjustment strategies, achieving a more reasonable allocation of safety margins.

[0083] For example, for target j, its dynamic safety distance can be expressed as: ; in, Let be the dynamic safe distance of target j at time t. Let be the target-level control uncertainty of target j at time t. Let be the approach velocity of target j relative to the UAV at time t; and For length scale adjustment parameters, To approximate the speed margin adjustment coefficient, The preset minimum safe distance, To control the cycle or predict the step size.

[0084] To prevent the dynamic safety distance from expanding excessively under extreme risk inputs, a maximum safety distance limit can be further set. And limit the dynamic safe distance to Within the interval: .

[0085] When the uncertainty of target-level control is small, the increase in safety distance is relatively slow, ensuring control flexibility; when the uncertainty of target-level control is large, the second-order term... It plays a significant role in accelerating the increase of safe distances, thereby forming a more conservative safety boundary in high-risk scenarios.

[0086] It should be noted that the dynamic safety distance can also be a piecewise linear function, an exponential function, a saturation function, or other monotonic nonlinear functions that satisfy the characteristic of "accelerated expansion of high-risk areas". Regardless of the specific function form, the core is to transform the hierarchical uncertainty and spatial risk modulation results guided by the semantic graph into a safety boundary that can be directly used by the control layer.

[0087] Reference Figure 5 and Figure 6 In another preferred embodiment, a dynamic safety distance is generated based on the scenario-level control uncertainty and the target-level control uncertainty of each target. Figure 5 and Figure 6 In this context, the distance between the drone and the dynamic safety boundary is the dynamic safety distance. Specifically, this includes: obtaining a basic safety distance; generating a first safety distance for each target based on the target-level control uncertainty, wherein the first safety distance and the target-level control uncertainty increase at an accelerating rate; calculating the sum of the first safety distance and the basic safety distance for each target to obtain a second safety distance for each target; and generating a dynamic safety distance for each target based on the second safety distance and the scene-level control uncertainty.

[0088] Dynamic safety distance is related to both target-level control uncertainty and scenario-level control uncertainty. Target-level control uncertainty reflects the local threat level of a single target to the UAV, while scenario-level control uncertainty reflects the overall risk level or the degree of control conservatism in the current scenario. By simultaneously introducing target-level and scenario-level control uncertainties, the UAV can increase its safety distance towards a high-risk target when such a target is present locally, and uniformly tighten the safety boundary for all targets when the overall scenario semantic quality is poor or the overall scenario risk is high.

[0089] In one example, a third safety distance that is linearly positively correlated with the scenario-level control uncertainty can be generated, and the sum of the third safety distance and the second safety distance can be used as the basis for the dynamic safety distance.

[0090] For example, for target j, its dynamic safety distance can be expressed as: ; in, Let be the dynamic safe distance of target j at time t. Let t represent the scenario-level control uncertainty. Let the target-level control uncertainty of target j at time t be . Let be the approach velocity of target j relative to the UAV at time t; This is a scenario-level conservative margin adjustment coefficient. and The length scale adjustment parameter corresponds to the target-level risk. To approximate the speed margin adjustment coefficient, The preset minimum safe distance, The preset maximum safe distance, To control the cycle or predict the step size.

[0091] In the above formula, This is used to uniformly adjust the safety boundaries of all targets based on the overall risk level of the current scenario. When the uncertainty of scenario-level control is high, it indicates that the overall semantic parsing quality of the current scenario is poor or there are significantly high-risk targets. In this case, the dynamic safety distance increases overall, causing the controller to adopt a more conservative obstacle avoidance strategy. It is used to adjust the local safety boundary of target j according to its own risk level, where the second-order term makes the safety distance expand at an accelerated rate when the target-level risk increases. It is used to increase the predictive safety margin based on the target's approach speed, so that a larger safety distance can be triggered in advance when approaching a target rapidly.

[0092] When the uncertainty of target-level control is small, the increase in the safety distance corresponding to the target is relatively slow, ensuring control flexibility; when the uncertainty of target-level control is large, the second-order term... It plays a significant role in accelerating the increase of the safety distance corresponding to the target; when the uncertainty of scenario-level control is large, the safety boundaries corresponding to each target rise as a whole, thereby forming a more conservative control response when the overall perception quality is poor or the risk of multiple targets is high.

[0093] It should be noted that in other implementations, the scenario-level control uncertainty may not enter the dynamic safety distance as an additive term, but as a multiplicative adjustment factor. That is, the dynamic safety distance generated based on the second safety distance is adjusted as a whole to generate the final dynamic safety distance.

[0094] For example: ; Where η is the scene-level amplification factor. Both the additive and multiplicative forms described above are used to illustrate that: target-level control uncertainty determines the local safety boundary of a single target, while scene-level control uncertainty determines the overall control conservatism in the current scene.

[0095] In one specific implementation, the UAV is controlled based on scenario-level control uncertainty and the dynamic safety distance of each target, specifically including: inputting the dynamic safety distance as a constraint into the model predictive controller; obtaining the control input for the UAV through the model predictive controller; and controlling the UAV according to the control input and scenario-level control uncertainty.

[0096] In one example, controlling a drone based on control input and scenario-level control uncertainty can be achieved by directly adjusting the control input through scenario-level control uncertainty, or by using the adjusted control input to control the drone.

[0097] In another example, controlling the UAV based on the control input and scenario-level control uncertainty can be achieved by adjusting the constraints on the control input in the model predictive controller through scenario-level control uncertainty, thereby indirectly adjusting the control input, and then using the adjusted control input to control the UAV.

[0098] The principle of Model Predictive Controller (MPC) is to predict the future state of the system within a finite time domain and solve an optimal control problem online. By using the dynamic safety distance as a hard constraint input to MPC, it ensures that the planned trajectory meets safety requirements at all times, thereby achieving a larger minimum obstacle distance and sufficient safety margin. This design combines the risk assessment model of this invention with advanced control theory, forming a complete and high-performance "perception-decision-control" closed loop.

[0099] In an example of control solution, the discrete state of the UAV can be represented as: , , and Represent the spatial position, velocity, and attitude (or other flight state variables) of the UAV at discrete time k, respectively, and the control input. It can be a velocity command, acceleration command, attitude / thrust command, or a virtual control quantity tracked by the flight control inner loop; the corresponding discrete dynamics model is represented as follows: This example does not require the safety function to be first-order relative to all possible physical inputs, but rather selects discrete predictive constraints, first-order continuous CBF, or higher-order CBF based on the actual control input hierarchy.

[0100] The aforementioned dynamic safety distance is incorporated into the safety function to construct an obstacle avoidance boundary. This is based on the drone's position... (t), Position of target j and dynamic safety distance Based on this, a security function can be constructed: ; Let be the difference between the relative distance between target j and the UAV at time t and the dynamic safe distance. At this time, it indicates that the distance between the UAV and target j is not less than the dynamic safety distance. In the preferred discrete-time implementation, the controller directly applies the following safety constraint within the prediction domain: ; in, Let be the difference between the relative distance between target j and the UAV and the dynamic safe distance at discrete time k+1. Let be the difference between the relative distance between target j and the UAV and the dynamic safety distance at discrete time k. The discrete barrier convergence coefficients are... Values ​​≥0 represent slack variables, used to avoid unsolvable optimization problems in extreme scenarios. Since this constraint directly affects the predicted next state or the state in the prediction domain, it is adaptable to discrete dynamics models of UAVs composed of position, velocity, and attitude, and does not require... The physical control input appears explicitly in the first derivative.

[0101] Furthermore, this safety constraint, along with the UAV discrete dynamics model, trajectory tracking target, and control cost, is incorporated into the model predictive control optimization problem to obtain candidate control inputs that satisfy both safety and tracking performance requirements. The preferred model predictive control problem can be written as: ; ; in, and Let Q and k be the reference trajectory and actual position of the UAV at discrete time k, respectively. The state error and control input weight matrix is ​​given. Let k be the control input at discrete time k. The penalty coefficient for slack variables, , Represents a discrete dynamics model. and These represent the minimum and maximum values ​​of the control input, respectively.

[0102] Furthermore, candidate control inputs are obtained by solving the model predictive controller. Furthermore, uncertainty can be controlled at the scenario level. Global conservative adjustments can be made to the control input or control constraints. For example, when the control input is a speed command or a speed loop virtual control variable, adjustments can be made according to... Scaling the speed limit or output speed: ; in, For example, a conservative adjustment function that decreases as scenario-level control uncertainty increases. , The speed conservative adjustment coefficient, This is the minimum scaling factor. Therefore, when scene-level control uncertainty is low, the drone basically executes according to the model predictive controller output; when scene-level control uncertainty is high, the drone reduces its overall speed or acceleration amplitude, thereby enhancing safety.

[0103] In another implementation, it can also be based on Adjusting constraint or cost parameters in model predictive control problems. For example, it can adjust the penalty coefficient of slack variables. Follow Increases accordingly, or causes the maximum velocity constraint and maximum acceleration constraint to increase with... By increasing and tightening the control, the optimizer is more inclined to maintain a safe distance in high-uncertainty scenarios, rather than sacrificing safety constraints through larger slack variables. This process ensures that scenario-level control uncertainty participates both in the generation of dynamic safe distances and in the global conservative adjustment of control outputs or the control optimization problem.

[0104] If continuous-time control is adopted, the corresponding CBF expression needs to be selected based on the physical meaning of the control input. The continuous-time UAV dynamics model can be expressed as: ; Where x is the continuous state variable of the UAV. x is the first derivative of x, which may include the position, velocity, attitude or other flight state variables of the UAV; u is the continuous-time control input, which may be a velocity command, acceleration command, attitude / thrust command or a virtual control quantity tracked by the flight control inner loop; This is the system drift dynamics term, used to represent the natural change in the system state when there is no control input; The control input matrix is ​​used to represent the effect of control input u on changes in the system state.

[0105] When u represents the virtual control quantity of the speed loop, and the safety function When u can be explicitly expressed in the first derivative, a first-order continuous-time CBF constraint can be used: ; in, For safety functions The partial derivative with respect to the state variable x; For safety functions Terms that change explicitly over time are used to characterize the impact of factors such as target motion and dynamic safety distance changes on the safety function; These are non-negative slack variables, used to avoid the optimization problem becoming unsolvable in extreme cases; To extend class K functions, the preferred choice in engineering implementation is... , It is a positive gain.

[0106] When u represents acceleration, attitude / thrust, or other control inputs that make the relative degree of the position safety function second-order or higher, the first-order form described above is not used directly; instead, a higher-order control barrier function is employed. Taking the second-order case as an example, it can be defined as follows: ; in, Let the first-order higher-order barrier function corresponding to target j be the security function. The first-order time derivative and the extended class K function terms together constitute the equation; For safety functions The total derivative with respect to time includes the effects of UAV state changes, target motion, and dynamic safety distance changes on the safety function; This is the first extended class function K.

[0107] And apply the following higher-order CBF constraints: ; in, for Total derivative with respect to time; For the second extended class K function; It is a non-negative slack variable. In engineering implementation, it can be preferably selected as... , , , It is a positive gain.

[0108] Equivalently, in a second-order system, it can be written as containing , and The constraint form, where and These are the first and second total derivatives of the safety function, respectively, allowing acceleration and attitude / thrust inputs to be explicitly incorporated into the optimization problem through dynamic relationships. If the dynamic safety distance varies with time, its time derivative or predicted change can be given by the risk mapping and target tracking results in the previous section, or directly updated in discrete MPC through the predicted step size.

[0109] Therefore, the preferred embodiment of the present invention is discrete-time MPC safety constraints; continuous-time first-order CBF is only applicable to the velocity loop or virtual control layer; for physical control inputs such as acceleration, attitude / thrust, higher-order CBF or discrete MPC predictive constraints are used. Through the above limitations, the mapping from dynamic safety distance to control constraints is clearly feasible.

[0110] Finally, the obtained control input is sent to the UAV flight controller for execution. After the UAV executes the command, the system continues to collect image, depth, point cloud, and state information for the next moment, repeating the above steps to achieve an online closed loop of "semantic understanding—semantic graph construction—hierarchical risk mapping—control constraint generation—execution feedback." Because this invention uses analytical mapping rules and a sparse graph structure, it does not rely on complex online training, thus exhibiting high real-time performance in airborne scenarios with a limited number of targets.

[0111] In an edge deployment implementation, the semantic model can be run on an on-board GPU, NPU or edge computing module, and hierarchical uncertainty mapping and control solution can be run in a real-time control thread. The output frequency of the semantic model can be lower than the control frequency; when the semantic frame rate is low, the target position and velocity can be predicted and compensated by the tracker between two semantic update cycles.

[0112] It should be noted that this embodiment requires the application of semantic confidence of each target. To reduce the impact of inconsistent original semantic confidence distribution in different scenarios on subsequent risk modeling, a temperature scaling method is preferably used to calibrate the original semantic confidence. Correspondingly, in this embodiment, the original semantic confidence can be used to calculate relevant variables, and calibrated semantic confidence is more preferably used to calculate relevant variables.

[0113] Assuming that the semantic category of target j is the l -th semantic category, if the logit output by the semantic perception model for the l -th semantic category is , then the calibrated semantic confidence can be expressed as: ; wherein, represents the unnormalized classification score output by the semantic perception model for the predicted semantic category of target j , T is a temperature parameter, l is a semantic category index. When T>1, the overconfidence phenomenon of the model can be reduced; when 0<T<1, the category discrimination can be enhanced. The calibrated semantic confidence can replace the original semantic confidence , and is used for subsequent calculation of global-level uncertainty, category-level uncertainty, individual-level uncertainty and target-level control uncertainty.

[0114] It should be noted that the semantic perception model refers to the semantic perception model adopted by the UAV perception end, including but not limited to the semantic perception module in target detection models, semantic segmentation models, open vocabulary detection models, visual language models or vision-language-action models. For the j-th target in the current scene, the model outputs unnormalized classification scores on L semantic categories, that is, logits. The semantic category set can be expressed as , represents the j-th semantic category, where L represents the number of semantic categories, such as pedestrians, vehicles, trees, utility poles, buildings, etc.

[0115] If the semantic model used in practice cannot directly output logits, but can only output semantic confidence, similarity score, or text matching score, then an offline calibration table, confidence remapping function, historical error calibration model, or empirical calibration method based on the validation set can be used to map the original output to the calibrated semantic confidence. The above alternative methods do not affect the subsequent technical process of hierarchical uncertainty mapping based on calibrated confidence in this invention.

[0116] The above is a complete embodiment of the drone safety control method provided in this application. The effects of the above embodiment will be verified and explained below with reference to some accompanying drawings.

[0117] Figure 7 This is a schematic diagram of a typical scenario where a long-range, high-confidence target and a short-range, low-confidence dynamic target coexist. Figure 7 The image displays the current location of the UAV, a distant high-confidence static target, a close-range low-confidence dynamic target, the target approach direction, the basic safety boundary, and the dynamic safety boundary generated by the method of this invention. Figure 7 It is known that in complex scenarios with both distant high-confidence targets and close-range low-confidence dynamic targets, relying solely on average confidence can easily underestimate the risk of close-range dynamic targets. However, this invention can expand the safety boundary corresponding to close-range dynamic targets based on target-level control uncertainty.

[0118] Figure 8 This is a comparison chart of the risk assessment results of the method of this invention and the traditional average risk method. Figure 8 This paper compares the normalized risk changes of traditional average risk methods with those of the present invention during the dynamic target approach process. Figure 8 It is known that traditional average risk methods are easily affected by distant high-confidence targets, which dilutes the risk of near-distance low-confidence dynamic targets. This invention, through a three-layer uncertainty structure, spatial risk modulation, and target-level risk aggregation, can retain and amplify the risk of near-distance low-confidence dynamic targets, enabling the controller to perceive potential dangers earlier.

[0119] Figure 9 This is a comparison chart showing the changes in safe distance and dynamic target distance between the method of this invention and the traditional average risk method. Figure 9 The paper demonstrates the safe distance generated by the traditional averaging method, the dynamic safe distance generated by the method of this invention, and the relationship between the actual distance between the dynamic target and the UAV over time. Figure 9 It is understood that the present invention can increase the safety distance in advance before the dynamic target approaches the drone, so that the safety boundary tightens before the danger occurs, instead of waiting for the target to enter the basic safety radius before triggering the obstacle avoidance response.

[0120] Figure 10This is a comparison chart of the control response intensity of the method of the present invention and the traditional average risk method. Figure 10 The changes in control response intensity during the approach process of a dynamic target are compared between the traditional average risk method and the method of this invention. Figure 10 It is understood that the present invention can generate an earlier and more obvious control response when the risk gradually increases, thereby reserving more control time for the UAV to decelerate, detour or adjust its trajectory; at the same time, the control response gradually increases with the change of risk, and does not output a sudden control quantity unconditionally.

[0121] Figure 11 A comparison chart showing the changes in safe distance under different risk modeling methods. Figure 11 The paper compares the changes in safe distance generated by the average confidence method, the fixed class weight method, and the method of this invention, and also presents the changes in distance to dynamic obstacles. Figure 11 It can be seen that, compared with the average confidence method and the fixed class weight method, the method of the present invention can generate a larger dynamic safety distance in the high-risk stage, thereby providing a more sufficient safety margin for drones.

[0122] Figure 12 A comparison chart of minimum obstacle distances under different risk modeling methods. Figure 12 This study compares the minimum obstacle distance maintained during obstacle avoidance using the average confidence level method, the fixed class weight method, and the method of this invention. Figure 12 It can be seen that the method of the present invention can maintain a larger minimum obstacle distance in dynamic obstacle approach scenarios, indicating that its early identification of dangerous targets and expansion of safety boundaries can effectively improve obstacle avoidance safety.

[0123] Figure 13 A comparison chart of response time and control strength under different risk modeling methods. Figure 13 The response time and peak control input of different methods are compared. Figure 13 As can be seen, the method of the present invention can trigger a safety control response earlier and provide more sufficient control strength, thereby having higher safety redundancy when a dynamic obstacle approaches rapidly. This result demonstrates that the present invention establishes a more sensitive mapping relationship between risk perception and control response.

[0124] Figure 14 A comparison chart of obstacle avoidance trajectory results under different risk modeling methods. Figure 14 The figure illustrates the reference straight path, the trajectory generated by the average confidence method, the trajectory generated by the fixed class weight method, the trajectory generated by the method of this invention, the obstacle location, and the basic safety radius. This figure serves to demonstrate that, under the same obstacle scenario, the trajectory generated by the method of this invention deviates from the danger zone earlier and avoids entering the basic safety radius; in contrast, the trajectories generated by the traditional average confidence method and the fixed class weight method are closer to the obstacle, with a smaller safety margin. Figure 8 This further verifies that the dynamic safety distance and control constraints of the present invention can be effectively translated into actual obstacle avoidance behavior.

[0125] As can be seen from the above description, this control method has the following technical effects: 1. More accurate risk identification and more sensitive control response: By establishing a three-layer uncertainty structure of "global layer - category layer - individual layer" and combining it with spatial risk modulation of uncertainty at the individual layer, it can accurately identify and amplify the real high risk of close-range, low-confidence dynamic targets; at the same time, by using maximum value aggregation and other methods, it ensures that the risk of the most dangerous targets is not averaged out, thereby triggering an earlier and stronger safety obstacle avoidance response, which significantly improves safety in complex dynamic environments.

[0126] 2. More Reasonable Safety Margin: Employing a dynamic safety distance generation mechanism that increases exponentially with target-level control uncertainty, the safety boundary expands rapidly under high-risk conditions. Compared to traditional linear adjustment methods, this invention maintains flexibility at lower risks and provides more comprehensive and reasonable protection when risks increase, achieving a better balance between safety and task efficiency.

[0127] 3. Strong structural interpretability and controllable computational complexity: The three-layer uncertainty structure has a clear source and explicit physical meaning, facilitating the analysis, debugging, and deployment of the system's risk assessment process. Furthermore, the entire method is primarily based on analytical formula calculations and sparse graph operations, without relying on complex online learning processes, resulting in low computational complexity and suitability for real-time operation on computationally limited airborne edge devices.

[0128] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the UAV safety control methods provided by the above methods.

[0129] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the unmanned aerial vehicle safety control methods provided by the above methods.

[0130] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0131] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for safe control of unmanned aerial vehicles (UAVs), characterized in that, include: The semantic information and semantic confidence of multiple targets in the current scene are obtained by the drone, and the semantic information includes spatial location and velocity; Multiple semantic clusters are obtained by semantic clustering of multiple targets, and the semantic clusters are obtained based on semantic graphs; The relative position and relative velocity between each target and the UAV are determined based on the spatial position and velocity of each target, and the individual layer uncertainty is calculated by combining the semantic confidence of each target. The category-level uncertainty is calculated based on the semantic confidence consistency among multiple targets within each semantic category cluster and the graph consistency of the semantic category cluster. Among the semantic confidence scores of each target, the semantic confidence score with the largest value is determined, and a global layer uncertainty for characterizing the uncertainty of the current scene is calculated based on the semantic confidence score with the largest value and using negative correlation mapping. The target-level control uncertainty is calculated based on the individual-level uncertainty of each target, the category-level uncertainty of its semantic category cluster, and the global-level uncertainty. The target-level control uncertainty of each target is aggregated by maximum value or softmax weight to obtain the scene-level control uncertainty. The dynamic safety distance of each target is generated based on the scene-level control uncertainty and the target-level control uncertainty of each target. The UAV is controlled based on the scene-level control uncertainty and the dynamic safety distance of each target.

2. The unmanned aerial vehicle (UAV) safety control method according to claim 1, characterized in that, Semantic clustering of multiple targets yields multiple semantic category clusters, specifically including: A first semantic graph is constructed based on the semantic information of the multiple targets, with each target and the UAV as nodes; in the first semantic graph, the edge weight between two nodes is determined based on the semantic similarity and spatial proximity between the two nodes; Multiple semantic category clusters are obtained by clustering multiple targets based on the first semantic graph.

3. The unmanned aerial vehicle (UAV) safety control method according to claim 2, characterized in that, The semantic information also includes semantic feature embedding; The semantic similarity between two nodes is characterized by the cosine similarity or distance between the feature embeddings of the two nodes; The spatial proximity between two nodes is characterized by the difference between the spatial positions of the two nodes.

4. The unmanned aerial vehicle (UAV) safety control method according to claim 2, characterized in that, Multiple semantic category clusters are obtained by clustering the multiple targets based on the first semantic graph, specifically including: The first semantic graph is subjected to sparsity processing to obtain a second semantic graph; the sparsity processing includes deleting edges with weights less than a preset threshold. The targets in the second semantic graph are clustered using a graph clustering method to obtain multiple semantic category clusters; the graph clustering method is one of connected component partitioning, spectral clustering, community detection, or embedded similarity clustering.

5. The unmanned aerial vehicle (UAV) safety control method according to claim 1, characterized in that, The dynamic safety distance is generated based on the scenario-level control uncertainty and the target-level control uncertainty of each target, respectively, including: Obtain a basic safe distance; A first safety distance is generated for each target based on the target-level control uncertainty, and the first safety distance increases exponentially with the target-level control uncertainty. The second safety distance for each target is obtained by summing the first safety distance and the basic safety distance for each target. The dynamic safety distance of each target is generated based on the second safety distance of each target and the scenario-level control uncertainty.

6. The unmanned aerial vehicle (UAV) safety control method according to claim 1, characterized in that, The UAV is controlled based on the scenario-level control uncertainty and the dynamic safe distance of each target, specifically including: The dynamic safety distance is used as a constraint and input into the model prediction controller; The control input for the UAV is obtained by solving the model predictive controller. The UAV is controlled based on the control input and the scenario-level control uncertainty.

Images