An unmanned aerial vehicle autonomous obstacle avoidance decision confidence evaluation method and system

Abstract

The application discloses an unmanned aerial vehicle autonomous obstacle avoidance decision confidence evaluation method and system, and particularly relates to the technical field of unmanned aerial vehicle autonomous flight control and safety decision, and is used for solving the problem that the prior art can only perform isolated confidence evaluation on single obstacle avoidance decision and cannot evaluate the overall safety level of continuous decision sequence from the system level, thereby leading to the problem of hidden accumulation of systematic collision risk; the method is realized by the following steps: acquiring completed instantaneous obstacle avoidance decision and confidence thereof, when the flight path geometry complexity meets the triggering condition, sequencing the instantaneous confidence and identifying low-confidence events therein, analyzing the time aggregation characteristics of the events and the conduction characteristics between the events, quantifying the risk evolution mode of the decision state, and generating an overall confidence evaluation conclusion for guiding the adjustment of the unmanned aerial vehicle flight control strategy; the method can effectively reveal the macro-reliability of the decision chain and improve the system-level safety of the unmanned aerial vehicle in continuous obstacle avoidance tasks.

Description

Technical Field

[0001] This invention relates to the field of autonomous flight control and safety decision-making technology for unmanned aerial vehicles (UAVs), and more specifically, to a method and system for assessing the confidence level of autonomous obstacle avoidance decisions for UAVs. Background Technology

[0002] In existing autonomous obstacle avoidance technologies for unmanned aerial vehicles (UAVs), confidence assessment based on probabilistic or statistical methods is often introduced to improve decision reliability. Specifically, the system acquires environmental perception data through onboard sensors and, based on a pre-set statistical model or hypothesis testing method, calculates a confidence score for the obstacle avoidance decision (e.g., the presence of an obstacle or feasibility) determined in the current frame or instant. This confidence score, as a quantitative indicator of the reliability of the decision, is typically compared to a fixed threshold to trigger or suppress corresponding avoidance control commands. This control logic based on the confidence score of a single decision is one of the common technical approaches to achieving reliable autonomous flight.

[0003] However, the aforementioned existing technical solutions have shortcomings: In actual flight, especially when performing continuous obstacle avoidance maneuvers, the safety of UAVs depends on a decision sequence consisting of multiple instantaneous obstacle avoidance decisions in chronological order. Existing methods only perform isolated confidence assessments and controls for each decision in the sequence, failing to assess the overall safety confidence level corresponding to the entire continuous decision sequence at the system level. This results in the flight control system potentially adopting a path composed of several high instantaneous confidence decisions without realizing that the overall failure probability of this path has significantly increased during the accumulation of the decision sequence, thus creating a systemic collision risk. The existing confidence assessment paradigm does not match the temporal continuity requirements of control decisions and cannot provide the flight control system with key information characterizing the overall reliability of the decision chain. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for evaluating the confidence of autonomous obstacle avoidance decisions of unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for assessing the confidence level of autonomous obstacle avoidance decisions by unmanned aerial vehicles (UAVs) includes the following steps:

[0007] S1. Obtain the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase;

[0008] S2. Based on the completed instantaneous obstacle avoidance decisions, determine the corresponding historical flight path segment of the UAV and determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions.

[0009] S3. When the path pattern triggering condition is met, the instantaneous confidence is serialized according to the order of the completed instantaneous obstacle avoidance decisions to form an instantaneous confidence sequence.

[0010] S4. Identify low-confidence events below a preset risk threshold in the instantaneous confidence sequence, and analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase.

[0011] S5. Based on the comprehensive risk assessment results, generate an overall confidence assessment conclusion for the drone.

[0012] Furthermore, obtain the instantaneous confidence levels corresponding to each instantaneous obstacle avoidance decision already executed by the UAV in the current continuous obstacle avoidance mission phase, including:

[0013] Real-time reception of instantaneous obstacle avoidance decisions and corresponding instantaneous confidence levels generated by the UAV during continuous obstacle avoidance missions;

[0014] The instantaneous obstacle avoidance decisions and their corresponding instantaneous confidence scores are cached according to the order in which the decisions are executed.

[0015] When an instantaneous obstacle avoidance decision is marked as completed, the instantaneous confidence level corresponding to that instantaneous obstacle avoidance decision is retrieved from the cache.

[0016] Furthermore, based on the completed instantaneous obstacle avoidance decisions, the corresponding historical flight path segments of the UAV are determined, and it is determined whether the geometric complexity of the historical flight path segments meets the preset path shape triggering conditions, including:

[0017] Extract the UAV's position coordinates at each decision moment from each completed instantaneous obstacle avoidance decision;

[0018] The extracted drone location coordinates are connected according to the time sequence of decision execution to generate historical flight path segments;

[0019] Calculate the change in heading angle between adjacent UAV position coordinates in the historical flight path segment to form a heading angle change sequence;

[0020] The amplitude and frequency characteristics of the heading angle change sequence are statistically analyzed to obtain the geometric complexity metric of the historical flight path segment.

[0021] The geometric complexity metric is compared with a preset path shape trigger threshold to determine whether the path shape trigger condition is met.

[0022] Furthermore, the setting of the preset path pattern trigger threshold includes: determining the normal fluctuation range of the geometric complexity metric based on the historical distribution of the geometric complexity metric recorded by the UAV during the calibration flight phase in open airspace; setting the upper limit of the normal fluctuation range as the initial path pattern trigger threshold; and dynamically adjusting the initial path pattern trigger threshold based on the mean and variance of the geometric complexity metric of the most recent historical flight path segment during the continuous obstacle avoidance mission phase, so as to generate the path pattern trigger threshold used in the current mission phase.

[0023] Furthermore, when the path morphology triggering condition is met, the instantaneous confidence is serialized according to the order in which the instantaneous obstacle avoidance decisions have been executed, forming an instantaneous confidence sequence, including:

[0024] Extract the decision timestamp corresponding to each decision from each completed instantaneous obstacle avoidance decision;

[0025] The extracted decision timestamps are paired with the corresponding instantaneous confidence scores;

[0026] Establish a uniform time grid based on decision timestamps;

[0027] Align the paired instantaneous confidence scores to a uniform time grid according to the decision timestamp;

[0028] The aligned instantaneous confidence scores are arranged in chronological order to form an instantaneous confidence score sequence.

[0029] Furthermore, low-confidence events below a preset risk threshold are identified in the instantaneous confidence sequence, and the temporal clustering and transmission characteristics of these low-confidence events are analyzed to obtain a comprehensive risk assessment result characterizing the risk evolution pattern of the decision-making state within the current continuous obstacle avoidance task phase. This includes:

[0030] Each instantaneous confidence level in the instantaneous confidence level sequence is compared with a preset risk threshold, and instantaneous confidence levels below the preset risk threshold and their corresponding times are marked to form a set of low-confidence events;

[0031] Based on the set of low-confidence events, the number of low-confidence events within a sliding time window is calculated, and the clustering strength is quantified according to the proportion of windows whose number exceeds the threshold of the number of events within the window.

[0032] Based on a set of low-confidence events, the time intervals between consecutive low-confidence events are statistically analyzed, and the intensity of transmission is quantified according to the degree of concentration and shortening trend of the time intervals.

[0033] Based on the quantified clustering strength and transmission strength, a comprehensive risk assessment result is generated.

[0034] Furthermore, the setting of the preset risk threshold includes: statistically analyzing the distribution of confidence values ​​in the instantaneous confidence sequence acquired during calibration environment flight; determining the initial risk threshold based on the low quantile of the confidence value distribution; calculating the moving average and standard deviation of the confidence values ​​based on the latest part of the acquired instantaneous confidence sequence during the continuous obstacle avoidance mission phase; and updating the initial risk threshold in real time based on the moving average and standard deviation to generate the currently used preset risk threshold.

[0035] Furthermore, based on the comprehensive risk assessment results, an overall confidence assessment conclusion for the drone is generated, including:

[0036] Based on the clustering intensity and conduction intensity in the comprehensive risk assessment results, the clustering level and conduction level are determined respectively;

[0037] Based on the combination of clustering level and transmission level, it is mapped to a preset overall risk level;

[0038] Based on the overall risk level, an overall confidence assessment conclusion is generated, which includes recommendations to adjust the flight control strategy or maintain the current strategy.

[0039] Furthermore, it is recommended to adjust the flight control strategy as follows: when the overall risk level is high, generate a strategy adjustment command to reduce the UAV's flight speed, increase the preset safety margin between the UAV and obstacles, and switch to a higher frequency obstacle avoidance decision mode; when the overall risk level is medium, generate a strategy adjustment command to maintain the current flight speed but increase the preset safety margin between the UAV and obstacles; when the overall risk level is low, generate a strategy adjustment command to maintain the current flight control strategy.

[0040] On the other hand, the present invention provides a confidence assessment system for autonomous obstacle avoidance decisions of unmanned aerial vehicles, comprising the following modules:

[0041] The confidence acquisition module is used to acquire the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase.

[0042] The path triggering module is used to determine the historical flight path segment of the corresponding UAV based on the completed instantaneous obstacle avoidance decisions, and to determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions.

[0043] The sequence formation module is used to serialize the instantaneous confidence scores according to the order of the completed instantaneous obstacle avoidance decisions when the path pattern triggering conditions are met, so as to form an instantaneous confidence score sequence.

[0044] The risk analysis module is used to identify low-confidence events in the instantaneous confidence sequence that are below a preset risk threshold, and to analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase.

[0045] The conclusion generation module is used to generate an overall confidence assessment conclusion for the drone based on the comprehensive risk assessment results.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. By constructing a complete assessment chain from spatial path triggering to temporal risk pattern analysis, it is possible to effectively identify and quantify the overall reliability risk of the decision chain during continuous obstacle avoidance by UAVs. Compared with traditional methods that only focus on the confidence of a single decision, the in-depth analysis of the instantaneous confidence sequence reveals the possible aggregation and propagation patterns of risks in the time dimension. It can provide timely warnings of potential collision threats constituted by a series of seemingly reliable but accumulating overall risk, thus providing key information for the flight control system to characterize the overall reliability of the decision chain and making up for the shortcomings of the existing confidence assessment paradigm and the requirement for the continuity of control decision sequence.

[0048] 2. The triggering mechanism based on the geometric complexity of the flight path ensures the timeliness and relevance of the assessment, avoids unnecessary computational overhead, and the analysis of the clustering and transmission of low-confidence events can profoundly reflect the dynamic process of the deterioration of the decision-making environment or the degradation of system performance, enabling risk assessment to move from static to dynamic. The final control strategy recommendations linked to the overall risk level allow the flight control system to adaptively adjust the aggressiveness of obstacle avoidance behavior based on the macro-reliability status of the decision chain. For example, when the risk increases, it can proactively adopt defensive strategies such as deceleration and increasing safety margin. This achieves a leap from passively responding to single decisions to actively managing sequential risks, significantly enhancing the system-level safety robustness of autonomous flight in complex scenarios. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of a confidence assessment system for autonomous obstacle avoidance decision-making of unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1: Figure 1 This invention presents a method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs), comprising the following steps:

[0053] S1. Obtain the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase;

[0054] S2. Based on the completed instantaneous obstacle avoidance decisions, determine the corresponding historical flight path segment of the UAV and determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions.

[0055] S3. When the path pattern triggering condition is met, the instantaneous confidence is serialized according to the order of the completed instantaneous obstacle avoidance decisions to form an instantaneous confidence sequence.

[0056] S4. Identify low-confidence events below a preset risk threshold in the instantaneous confidence sequence, and analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase.

[0057] S5. Based on the comprehensive risk assessment results, generate an overall confidence assessment conclusion for the drone.

[0058] When implementing the confidence assessment method for UAV autonomous obstacle avoidance decisions based on statistical hypothesis testing, the step of obtaining the instantaneous confidence of each instantaneous obstacle avoidance decision already executed by the UAV in the current continuous obstacle avoidance task phase is achieved in the following way:

[0059] The continuous obstacle avoidance mission phase refers to the entire time interval from the moment the UAV begins executing a series of continuous missions that require real-time obstacle avoidance maneuvers based on perceived environmental information, until the mission is instructed to end or be interrupted. During this mission phase, the UAV's flight control system periodically or when triggered by specific events generates instantaneous obstacle avoidance decisions. An instantaneous obstacle avoidance decision is a command-line judgment, calculated at a specific moment based on environmental data collected by current sensors and using a pre-defined statistical hypothesis testing method, regarding the existence of obstacles and how to avoid them. Simultaneously generated with this instantaneous obstacle avoidance decision is an instantaneous confidence score, a numerical result output from the same statistical hypothesis testing process, used to quantify the statistical reliability of the instantaneous obstacle avoidance decision.

[0060] Real-time reception of instantaneous obstacle avoidance decisions and corresponding instantaneous confidence levels generated by the UAV during continuous obstacle avoidance missions. Specifically, after each instantaneous obstacle avoidance decision calculation, the flight control system immediately sends out the content of the decision and the calculated instantaneous confidence level as a data packet via the internal communication bus or shared memory area; it listens to the communication bus or polls the shared memory area, and immediately reads the data packet once a new data packet is detected, thus completing real-time reception; the data packet contains at least the timestamp of the decision generation, the decision type identifier, and the instantaneous confidence level.

[0061] Instantaneous obstacle avoidance decisions and their corresponding instantaneous confidence scores are cached according to the decision execution order. Each successfully received real-time data packet is parsed, and the decision timestamp, decision type identifier, and instantaneous confidence score are extracted. The decision execution order is determined by the order of the decision timestamps, with earlier timestamps being executed first. The parsed information about the instantaneous obstacle avoidance decision is bound to its corresponding instantaneous confidence score to form a cache entry, which is then inserted into the first-in-first-out (FIFO) queue cache according to its decision timestamp order. The cache capacity can be preset, for example, to store 50 to 200 recently generated decision-confidence pairs. When the cache is full, the oldest cache entry is removed according to the FIFO principle to make room for a new entry, ensuring that the cache contains the latest decision data arranged in execution order.

[0062] When a transient obstacle avoidance decision is marked as completed, the transient confidence level corresponding to that decision is retrieved from the cache. The marking of a transient obstacle avoidance decision as completed is handled by the flight control system, which tracks the actual execution status of each transient obstacle avoidance decision issued to the actuator. When the evasive maneuver corresponding to a transient obstacle avoidance decision has been completed by the UAV, or the effective window of the decision has ended and been replaced by a subsequent decision, the transient obstacle avoidance decision is determined to be completed. Upon completion, a completion notification signal containing a unique identifier for that decision is generated. The system responsible for this evaluation method continuously monitors and receives such completion notification signals. Once a completion notification for a specific decision identifier is received, the decision identifier is used as the key... The process involves searching and matching entries in a first-in-first-out (FIFO) queue buffer. This involves iterating through each entry in the buffer queue and comparing the decision identifier stored in the entry with the decision identifier in the completion notification signal. Once a matching entry is found, the previously stored instantaneous confidence value corresponding to the instantaneous obstacle avoidance decision is precisely read and extracted from that entry. After the extraction operation is complete, the cached entry can be removed from the buffer. Through this process, the instantaneous confidence value corresponding to an instantaneous obstacle avoidance decision that has been actually executed in the current continuous obstacle avoidance mission phase is accurately obtained. This process continues, accumulating data on each executed instantaneous obstacle avoidance decision and its corresponding instantaneous confidence value for subsequent steps.

[0063] Based on the completed instantaneous obstacle avoidance decisions, the step of determining whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions is implemented in the following way:

[0064] Each completed instantaneous obstacle avoidance decision originates from the data set acquired and accumulated in the preceding steps. Each data entry contains a marked instantaneous obstacle avoidance decision and its corresponding instantaneous confidence level. The UAV's position coordinates at the moment of each decision are extracted from these completed instantaneous obstacle avoidance decisions. When each instantaneous obstacle avoidance decision is generated, in addition to the decision content and confidence level, the UAV's positioning and navigation system simultaneously records the UAV's precise position coordinates in three-dimensional space at that moment. Position coordinates are represented by latitude, longitude, and altitude, or by X, Y, and Z coordinates in a local coordinate system. Position coordinates are extracted by reading the metadata fields associated with each instantaneous obstacle avoidance decision. The extracted UAV position coordinate data corresponds one-to-one with the decision, and the decision timestamp information is retained.

[0065] The extracted UAV position coordinates are connected according to the chronological order of decision execution to generate a historical flight path segment. The chronological order of decision execution is determined by the order of decision timestamps. All extracted position coordinates are sorted from earliest to latest according to their corresponding decision timestamps. After sorting, the ordered sequence of position coordinates is represented by a series of sequentially connected line segments, with each line segment connecting two adjacent position coordinate points. This series of line segments constitutes the UAV's historical flight path segment from the earliest completed decision time to the most recent completed decision time. This historical flight path segment is a discretized representation of the UAV's actual spatial trajectory.

[0066] The heading angle change sequence is formed by calculating the change in heading angle between adjacent UAV position coordinates in the historical flight path segment. The heading angle is the angle between the projection of the UAV's heading direction onto the horizontal plane and the true north reference direction. For any two adjacent position coordinates on the historical flight path segment, the heading angle of the UAV flying from one point to the next is calculated. The calculation method is to calculate the projection vector of the direction vector from the starting point to the ending point onto the horizontal plane based on the latitude and longitude coordinates or Cartesian coordinates of the two adjacent points; then, the angle between this projection vector and the true north reference direction vector is calculated using the arctangent function, resulting in a heading angle value in degrees, ranging from 0 to 360 degrees. The difference between two adjacent heading angles is calculated as the heading angle change. When calculating the difference, the cyclical continuity between 0 and 360 degrees needs to be handled; for example, a change from 350 degrees to 10 degrees is calculated as 20 degrees. All pairs of adjacent position coordinates on the historical flight path segment are traversed, and the heading angle of each segment is calculated sequentially, then the heading angle change between adjacent segments is calculated. All the calculated changes in heading angle are arranged in order of their corresponding path segments to form a heading angle change sequence.

[0067] By statistically analyzing the amplitude and frequency characteristics of heading angle changes in a heading angle change sequence, a geometric complexity metric for historical flight path segments can be obtained. Amplitude reflects the magnitude of the heading change, while frequency reflects its frequency. One way to statistically analyze amplitude is to calculate the average of the absolute values ​​of all changes in the heading angle change sequence. Specifically, sum the absolute values ​​of all changes in the heading angle change sequence and then divide by the total number of changes to obtain the average amplitude. Another way to statistically analyze frequency is to calculate the proportion of times the absolute value of a change in the heading angle change sequence exceeds a certain significance threshold. The significance threshold can be set to a small angle value, such as 5 degrees; a higher proportion indicates more frequent heading changes. A comprehensive measurement method can be used, such as multiplying the average amplitude by an amplitude weighting coefficient and adding the proportion of high-frequency changes multiplied by a frequency weighting coefficient, with the resulting weighted sum serving as the geometric complexity metric. The amplitude weighting coefficient and frequency weighting coefficient are pre-set positive numbers, and their specific values ​​can be calibrated by analyzing data from typical flight missions; for example, both coefficients can be set to 0.5. A higher geometric complexity metric indicates a more tortuous and complex path in terms of geometry.

[0068] The geometric complexity metric is compared with a preset path pattern trigger threshold to determine whether the path pattern trigger condition is met. The preset path pattern trigger threshold is a numerical threshold. The comparison operation is performed: if the calculated geometric complexity metric is greater than or equal to the preset path pattern trigger threshold, the path pattern trigger condition is met; if the geometric complexity metric is less than the preset path pattern trigger threshold, the path pattern trigger condition is not met. The result is used to determine whether to activate the subsequent confidence sequence analysis process.

[0069] The preset path pattern trigger threshold setting includes an initialization phase and a dynamic adjustment phase. The initialization phase determines the normal fluctuation range of the geometric complexity metric based on the historical distribution of flight path segments recorded by the UAV during its open airspace calibration flight phase. The open airspace calibration flight phase refers to the phase where the UAV performs standard flights such as straight-line flight and stable hovering within a safe airspace free of dense obstacles. During this phase, the system records multiple flight paths and their calculated geometric complexity metrics. After collecting a sufficient number of samples, such as 100 geometric complexity metric samples, the mean and standard deviation of these samples are calculated. The upper bound of the normal fluctuation range can be calculated by adding K times the standard deviation to the mean, where K is a constant greater than zero, for example, K=2. The calculated upper bound of the normal fluctuation range is set as the initial path pattern trigger threshold.

[0070] During the continuous obstacle avoidance phase, the initial path shape trigger threshold is dynamically adjusted based on the mean and variance of the geometric complexity metrics of the most recent historical flight path segments to generate the path shape trigger threshold used in the current phase. A sliding window is maintained to store the geometric complexity metrics of several recently calculated historical flight path segments, for example, the most recent 20 geometric complexity metrics. When the amount of data in the window is insufficient, such as less than 5, dynamic adjustment is not performed, and the initial path shape trigger threshold is used directly. When the amount of data in the window is sufficient, the mean and variance of all geometric complexity metrics within the sliding window are calculated. One strategy for dynamic adjustment is to set the path shape trigger threshold used in the current phase as a linear combination of the initial path shape trigger threshold, the window mean, and the window variance. Specifically: Path shape trigger threshold used in the current phase = Initial path shape trigger threshold × Coefficient A + Window mean × Coefficient B + Window variance × Coefficient C. Coefficients A, B, and C are non-negative preset constants, and satisfy the condition that coefficient A + coefficient B = 1 to maintain a stable threshold magnitude. Coefficient C is used to control the sensitivity to variance changes, and its value can be determined experimentally, for example, set to 0.1. In this way, the preset path pattern trigger threshold can adaptively change with the overall level and fluctuation of the UAV's recent flight path complexity, thereby more accurately identifying abnormal, highly complex path patterns to trigger subsequent analysis.

[0071] When the path pattern triggering condition is met, the step of serializing the instantaneous confidence scores to form an instantaneous confidence score sequence according to the order in which the instantaneous obstacle avoidance decisions have been executed is implemented in the following way:

[0072] The determination that the path morphology trigger condition is met is derived from the comparison between the geometric complexity metric of the historical flight path segment and the preset path morphology trigger threshold in the preceding steps. Upon receiving the determination that the path morphology trigger condition is met, the serialization process is initiated. The operation targets the completed instantaneous obstacle avoidance decisions and their corresponding instantaneous confidence scores, which have been acquired and stored in a cache or specific data structure through previous steps.

[0073] Extract the decision timestamp corresponding to each executed instantaneous obstacle avoidance decision. Each executed instantaneous obstacle avoidance decision is associated with a decision timestamp, which records the absolute time when the decision was generated. The timestamp format is milliseconds counted since a certain epoch, or structured time data containing year, month, day, hour, minute, second, and milliseconds. The extraction of decision timestamps is completed by accessing the data structure storing the decision data and reading the timestamp field associated with each decision entry. A one-to-one correspondence is maintained between the extracted decision timestamps and the specific instantaneous obstacle avoidance decisions.

[0074] The extracted decision timestamps are paired with their corresponding instantaneous confidence scores. The corresponding instantaneous confidence score refers to the instantaneous confidence score belonging to the same instantaneous obstacle avoidance decision data entry as the extracted decision timestamp. Based on the association relationships in the data storage structure, a matching instantaneous confidence score is found for each extracted decision timestamp. The pairing operation establishes a set of multiple timestamp-instantaneous confidence score pairs, where each pair completely represents the time and reliability information of a completed instantaneous obstacle avoidance decision. This pairing set is the foundation for subsequent processing.

[0075] A uniform time grid is established based on decision timestamps. A uniform time grid refers to a sequence of time points evenly distributed along a time axis. Establishing a uniform time grid requires determining three parameters: the starting grid time point, the ending grid time point, and the grid interval. The starting grid time point is set to the minimum value among all extracted decision timestamps, i.e., the earliest decision moment. The ending grid time point is set to the maximum value among all extracted decision timestamps, i.e., the latest decision moment. The grid interval is a time length value that needs to be preset or calculated based on data characteristics, such as 50 milliseconds or 100 milliseconds. The grid interval can be set according to the typical generation frequency of instantaneous obstacle avoidance decisions; for example, if decisions are generated on average every 100 milliseconds, the grid interval can be set to 100 milliseconds. Based on the selected starting grid time point, ending grid time point, and grid interval, a series of evenly spaced grid time points are generated by repeatedly adding the grid interval from the starting grid time point until the ending grid time point is exceeded. These grid time points constitute the uniform time grid.

[0076] The paired instantaneous confidence scores are aligned to the uniform time grid according to the decision timestamps. The alignment operation involves assigning an instantaneous confidence score to each grid time point on the uniform time grid. Since decision timestamps and grid time points typically do not perfectly overlap, a mapping rule is needed. One mapping rule is to assign each grid time point to the instantaneous confidence score paired with the closest temporally preceding decision timestamp. Specifically, for each grid time point on the uniform time grid, all timestamp-instantaneous confidence score pairs with decision timestamps less than or equal to that grid time point are found in the pairing set. Then, the pair with the largest decision timestamp value is selected, and its corresponding instantaneous confidence score is assigned to that grid time point. If no decision timestamps precede a certain grid time point, a special value indicating missing data, such as a null value or a specific default confidence score, can be assigned to that grid time point. Through this alignment operation, each grid time point obtains either an instantaneous confidence score or a missing data flag.

[0077] The aligned instantaneous confidence scores are arranged in chronological order to form an instantaneous confidence score sequence. Chronological order refers to the ascending order of the grid time points. Following the grid time point order of the uniform time grid, the instantaneous confidence score assigned to each grid time point is read sequentially. These chronologically read instantaneous confidence scores are then stored in a one-dimensional array or list-like data structure. This stored instantaneous confidence score sequence is the instantaneous confidence score sequence itself. If certain grid time points were assigned special values ​​indicating missing data during the alignment process, these special values ​​are also retained at the corresponding positions in the instantaneous confidence score sequence. The final instantaneous confidence score sequence is a time-uniformly sampled or interpolated confidence score data sequence, with its temporal resolution determined by the grid time interval of the uniform time grid, for example, one data point every 100 milliseconds. This instantaneous confidence score sequence provides a regularized temporal data foundation for subsequent analysis of the clustering and transmission characteristics of low-confidence events.

[0078] The steps to identify low-confidence events below a preset risk threshold in an instantaneous confidence sequence and analyze the temporal clustering and transmission characteristics of these low-confidence events to obtain a comprehensive risk assessment result characterizing the risk evolution pattern of the decision-making state within the current continuous obstacle avoidance task phase are implemented in the following manner:

[0079] The instantaneous confidence sequence originates from the output of the serialization process performed on the completed instantaneous confidence scores in the preceding steps. The instantaneous confidence sequence is a sequence of confidence scores arranged in chronological order. It serves as the core input data for this step.

[0080] Each instantaneous confidence level in the instantaneous confidence sequence is compared with a preset risk threshold. Instantaneous confidence levels below the preset risk threshold and their corresponding times are marked to form a low-confidence event set. The preset risk threshold is a critical value used to distinguish between normal and low confidence levels. The comparison operation iterates through each instantaneous confidence value in the instantaneous confidence sequence. Each instantaneous confidence value is compared with the preset risk threshold. If an instantaneous confidence value is less than the preset risk threshold, it is determined to be a low-confidence level. The marking operation records the low-confidence value and its corresponding time in the instantaneous confidence sequence. The time information comes from the grid time point mapped to the position of the instantaneous confidence level in the instantaneous confidence sequence. Each marked low-confidence level and its corresponding time constitute a low-confidence event. All marked low-confidence events are collected to form a low-confidence event set. Each element in the low-confidence event set contains two data items: the low-confidence value and the time when the event occurred.

[0081] Based on a set of low-confidence events, the number of low-confidence events within a sliding time window is calculated, and the clustering strength is quantified by the proportion of windows whose low-confidence events exceed a threshold. The sliding time window is a fixed-length interval that slides continuously along the time axis; the length of the sliding time window can be set, for example, 1000 milliseconds. The sliding time window slides along the time axis with a fixed step size; the step size can be set, for example, 100 milliseconds. For each window position of the sliding time window, the number of low-confidence events falling within that time range is calculated. The threshold for the number of events within a window is a critical number used to determine whether there are too many low-confidence events within a certain sliding time window; the threshold can be set, for example, 3. It is determined whether the number of low-confidence events calculated within each sliding time window exceeds the threshold. The number of sliding time windows that meet the condition of having more low-confidence events than the threshold is counted out of the total number of sliding time windows. The proportion of sliding time windows that meet the condition out of the total number of sliding time windows is calculated; this proportion is called the window ratio. The window ratio is used to quantify the intensity of clustering; the higher the window ratio, the more significant the clustering of low-confidence events over time, i.e., the stronger the clustering.

[0082] Based on a set of low-confidence events, the time intervals between consecutive low-confidence events are statistically analyzed, and the transmission strength is quantified based on the concentration and shortening trend of the time intervals. Consecutive low-confidence events refer to two adjacent low-confidence events in the set, ordered by their occurrence time. The time interval between each pair of consecutive low-confidence events is calculated; the time interval equals the occurrence time of the subsequent low-confidence event minus the occurrence time of the preceding low-confidence event; the unit of the time interval is milliseconds. All calculated time intervals constitute a time interval sequence. The concentration of the time intervals reflects whether the time interval values ​​are concentrated within a small range. The method to quantify the concentration is to calculate the coefficient of variation of the time interval sequence; the coefficient of variation equals the standard deviation of the time interval sequence divided by the mean of the time interval sequence. The smaller the coefficient of variation, the more concentrated the time interval distribution. The shortening trend of the time intervals reflects whether the time intervals between consecutive low-confidence events show a gradually decreasing trend. The method for quantifying the shortening trend is to calculate the slope of the linear regression of the time interval series. A linear regression analysis is performed on the time interval series according to their occurrence order, and the resulting slope value is the quantified trend value. A negative slope indicates that the time intervals are generally shortening, and the absolute value reflects the strength of the trend. The quantification of transmissibility strength combines two indicators: the degree of distribution concentration and the shortening trend. Transmissibility strength is expressed as a weighted sum of the two sub-indices; the first sub-indicator is the reciprocal of the coefficient of variation, and the second sub-indicator is the absolute value of the shortening trend slope. Transmissibility strength equals the reciprocal of the coefficient of variation multiplied by the weighting coefficient A, plus the absolute value of the shortening trend slope multiplied by the weighting coefficient B. Weighting coefficients A and B are pre-defined non-negative numbers; their specific values ​​are determined by analyzing the correlation between risk evolution patterns and subsequent outcomes in historical flight missions, for example, to give transmissibility strength higher discriminative power for decision sequences that have historically experienced collisions or near-collisions; both weighting coefficients A and B can be set to 0.5.

[0083] A comprehensive risk assessment result is generated based on the quantified clustering and conduction strengths. Clustering and conduction strengths are two quantified values ​​reflecting risk evolution patterns from different perspectives. The comprehensive risk assessment result is generated by comparing the clustering and conduction strengths with their respective warning thresholds. These warning thresholds are pre-set empirical values; their setting is based on statistical analysis of historical safe and faulty flight data, ensuring that the probability of clustering and conduction strengths exceeding their respective warning thresholds during safe flights is very low. Based on the comparison results, the comprehensive risk status is divided into three levels: low risk, medium risk, and high risk. The classification rule is as follows: if both clustering and conduction strengths are below their respective warning thresholds, the comprehensive risk assessment result is low risk; if either clustering or conduction strength is above its warning threshold, the comprehensive risk assessment result is medium risk; and if both clustering and conduction strengths are above their respective warning thresholds, the comprehensive risk assessment result is high risk. The comprehensive risk assessment result is the determined risk level identifier.

[0084] The setting of the preset risk threshold includes an initialization phase and a real-time update phase. The distribution of confidence values ​​in the instantaneous confidence sequences acquired during calibration environment flights is statistically analyzed. Calibration environment flights refer to flights conducted in a known safe environment free of complex obstacles. Multiple instantaneous confidence sequences generated during calibration environment flights are collected, and the confidence values ​​from all instantaneous confidence sequences are merged to form a calibration environment confidence value sample set. The initial risk threshold is determined based on the low quantile of the calibration environment confidence value distribution. The low quantile refers to the value corresponding to a lower position in the cumulative probability distribution, such as the 10th percentile. The 10th percentile of the calibration environment confidence value sample set is calculated, and this calculated value is set as the initial risk threshold.

[0085] During the continuous obstacle avoidance phase, the moving average and moving standard deviation of the confidence values ​​are calculated based on the latest portion of the acquired instantaneous confidence sequence. The latest portion refers to a segment formed by truncating a fixed number of data points from the end of the instantaneous confidence sequence; the number of data points can be set, for example, to 100. The moving average is obtained by calculating the average of all confidence values ​​within this segment. The moving standard deviation is also obtained by calculating the standard deviation of all confidence values ​​within this segment. Based on the moving average and moving standard deviation, the initial risk threshold is updated in real time to generate the currently used preset risk threshold. The formula for updating the currently used preset risk threshold is: the currently used preset risk threshold equals the initial risk threshold plus a correction term. The correction term equals the difference between the moving average and the mean confidence value of the calibration environment multiplied by a coefficient E, plus the difference between the moving standard deviation and the standard deviation of the calibration environment confidence value multiplied by a coefficient F. The mean confidence value of the calibration environment is the mean of the sample set of calibration environment confidence values ​​obtained during the initialization phase. The standard deviation of the calibration environment confidence value is the standard deviation of the sample set of calibration environment confidence values ​​obtained during the initialization phase. Coefficients E and F are pre-set adjustment coefficients; their values ​​are determined experimentally to ensure that the updated preset risk threshold maintains a stable ability to identify abnormally low confidence levels under different flight environments; coefficient E can be set to 0.5, and coefficient F can be set to 0.3. Through this update mechanism, the currently used preset risk threshold can adapt to changes in the overall confidence level and volatility under different mission phases or environments.

[0086] Based on the comprehensive risk assessment results, the steps to generate the overall confidence assessment conclusion for the drone are implemented in the following manner:

[0087] Based on the clustering intensity and conduction intensity from the comprehensive risk assessment results, the clustering level and conduction level are determined respectively. Clustering intensity is the window proportion value calculated in the preceding steps. Conduction intensity is the comprehensive quantitative value calculated in the preceding steps. Determining the clustering level requires comparing the clustering intensity with a preset clustering level threshold range. The preset clustering level threshold range is pre-defined by statistically analyzing the distribution characteristics of clustering intensity under different risk states in a large amount of historical flight mission data. Historical flight mission data includes flight mission data that were completed safely and flight mission data that experienced warnings or abnormal events. The analysis method involves calculating the statistical distribution of clustering intensity in the safe flight mission data, for example, its 95th percentile, and using this value as a reference boundary for dividing low and medium clustering levels; simultaneously, calculating the statistical distribution of clustering intensity in the flight mission data before abnormal events, for example, its 50th percentile, and using this value as a reference boundary for dividing medium and high clustering levels. For example, the clustering level threshold intervals can be: low clustering level corresponds to a clustering intensity less than 0.3; medium clustering level corresponds to a clustering intensity greater than or equal to 0.3 and less than 0.6; and high clustering level corresponds to a clustering intensity greater than or equal to 0.6. Determining the conductivity level requires comparing the conductivity intensity with a preset conductivity level threshold interval. The preset conductivity level threshold intervals are pre-defined based on the distribution characteristics of conductivity intensity in historical flight mission data using similar statistical analysis methods. For example, the defined conductivity level threshold intervals can be: low conductivity level corresponds to a conductivity intensity less than 0.4; medium conductivity level corresponds to a conductivity intensity greater than or equal to 0.4 and less than 0.8; and high conductivity level corresponds to a conductivity intensity greater than or equal to 0.8. The comparison operation involves comparing the calculated clustering intensity value with the preset clustering level threshold interval to determine its clustering level. Simultaneously, the calculated conductivity intensity value is compared with the preset conductivity level threshold interval to determine its conductivity level. Clustering rank and conduction rank are two independent classification results.

[0088] Based on the combination of clustering and conduction levels, a preset overall risk level is mapped. The preset overall risk levels include low risk, medium risk, and high risk. This mapping relationship is achieved through a preset level combination mapping table. The level combination mapping table defines the overall risk level corresponding to each possible combination of clustering and conduction levels. The establishment of the level combination mapping table is based on a post-hoc analysis of the risk characteristics of historical flight safety events and corresponding decision sequences. Its design principle is that when both risk characteristic levels are high, it maps to a high risk level; when at least one risk characteristic level is high, it maps to a medium risk level; and when both risk characteristic levels are low, it maps to a low risk level. An example of a specific mapping rule is: if the clustering level is low and the conduction level is low, the overall risk level maps to a low risk level; if the clustering level is high or the conduction level is high, the overall risk level maps to a high risk level; for other combinations, such as a medium clustering level and a low conduction level, the overall risk level maps to a medium risk level. Based on this mapping rule, and according to the determined clustering and conduction levels, the corresponding overall risk level can be obtained by querying the level combination mapping table. The overall risk level is a final qualitative assessment of the overall risk level of the decision-making state within the current continuous obstacle avoidance task phase.

[0089] Based on the overall risk level, an overall confidence assessment conclusion is generated, which includes suggestions for adjusting the flight control strategy or maintaining the current strategy. The overall confidence assessment conclusion is a structured output, the core of which is a flight control strategy recommendation given for the current overall risk level. The generation of strategy recommendations follows a pre-defined strategy adjustment rule base. When the overall risk level is high, strategy adjustment instructions are generated to reduce the UAV's flight speed, increase the preset safety margin between the UAV and obstacles, and switch to a higher frequency obstacle avoidance decision mode. Reducing the UAV's flight speed means multiplying the UAV's current cruising speed or maximum permissible speed command value by a deceleration coefficient less than 1; the deceleration coefficient is pre-set based on the UAV platform's maneuverability and mission requirements, for example, setting the deceleration coefficient to 0.7. Increasing the preset safety margin between the UAV and obstacles means increasing the default minimum distance threshold used in the UAV obstacle avoidance algorithm to determine collision risk by an increment; the increment is pre-set based on sensor accuracy and environmental complexity, for example, increasing the original preset safety margin of 1.5 meters by 0.5 meters to 2.0 meters. Switching to a higher-frequency obstacle avoidance decision-making mode means shortening the cycle of obstacle avoidance decision calculations in the control loop. The specific shortened cycle value is preset based on the processor's computing power, for example, switching from once every 100 milliseconds to once every 50 milliseconds. When the overall risk level is medium, a strategy adjustment command is generated to maintain the current flight speed but increase the preset safety margin between the drone and obstacles. Increasing the preset safety margin between the drone and obstacles involves adding an increment value; this increment value can be smaller than the increment value in high-risk situations, for example, an increase of 0.3 meters. The flight speed remains unchanged. The obstacle avoidance decision-making mode also remains unchanged at its current frequency. When the overall risk level is low, a strategy adjustment command is generated to maintain the current flight control strategy. This means that no adjustments are made to the UAV's flight speed command value, preset safety margin parameters, or obstacle avoidance decision calculation cycle; all control parameters remain unchanged. The generated overall confidence assessment conclusion, in addition to including the specific strategy adjustment commands mentioned above, may also include a descriptive text of the overall risk level. The overall confidence assessment conclusion, as the final output of this assessment method, will be sent to the UAV's flight control system or flight monitoring interface for automatic execution or pilot reference.

[0090] Example 2: Figure 2 A schematic diagram of the structure of an autonomous obstacle avoidance decision confidence assessment system for unmanned aerial vehicles (UAVs) according to the present invention is provided. The system includes the following modules:

[0091] The confidence acquisition module is used to acquire the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase.

[0092] The path triggering module is used to determine the historical flight path segment of the corresponding UAV based on the completed instantaneous obstacle avoidance decisions, and to determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions.

[0093] The sequence formation module is used to serialize the instantaneous confidence scores according to the order of the completed instantaneous obstacle avoidance decisions when the path pattern triggering conditions are met, so as to form an instantaneous confidence score sequence.

[0094] The risk analysis module is used to identify low-confidence events in the instantaneous confidence sequence that are below a preset risk threshold, and to analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase.

[0095] The conclusion generation module is used to generate an overall confidence assessment conclusion for the drone based on the comprehensive risk assessment results.

[0096] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0097] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0098] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0099] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0102] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the confidence level of autonomous obstacle avoidance decisions by unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1. Obtain the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase; S2. Based on the completed instantaneous obstacle avoidance decisions, determine the corresponding historical flight path segment of the UAV and determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions. S3. When the path pattern triggering condition is met, the instantaneous confidence is serialized according to the order of the completed instantaneous obstacle avoidance decisions to form an instantaneous confidence sequence. S4. Identify low-confidence events below a preset risk threshold in the instantaneous confidence sequence, and analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase. S5. Based on the comprehensive risk assessment results, generate an overall confidence assessment conclusion for the drone.

2. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Obtain the instantaneous confidence levels corresponding to each instantaneous obstacle avoidance decision already executed by the UAV in the current continuous obstacle avoidance mission phase, including: Real-time reception of instantaneous obstacle avoidance decisions and corresponding instantaneous confidence levels generated by the UAV during continuous obstacle avoidance missions; The instantaneous obstacle avoidance decisions and their corresponding instantaneous confidence scores are cached according to the order in which the decisions are executed. When an instantaneous obstacle avoidance decision is marked as completed, the instantaneous confidence level corresponding to that instantaneous obstacle avoidance decision is retrieved from the cache.

3. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Based on the completed instantaneous obstacle avoidance decisions, the historical flight path segments of the corresponding UAV are determined, and it is judged whether the geometric complexity of the historical flight path segments meets the preset path shape triggering conditions, including: Extract the UAV's position coordinates at each decision moment from each completed instantaneous obstacle avoidance decision; The extracted drone location coordinates are connected according to the time sequence of decision execution to generate historical flight path segments; Calculate the change in heading angle between adjacent UAV position coordinates in the historical flight path segment to form a heading angle change sequence; The amplitude and frequency characteristics of the heading angle change sequence are statistically analyzed to obtain the geometric complexity metric of the historical flight path segment. The geometric complexity metric is compared with a preset path shape trigger threshold to determine whether the path shape trigger condition is met.

4. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, The preset path pattern trigger threshold setting includes: determining the normal fluctuation range of the geometric complexity metric based on the historical distribution of the geometric complexity metric recorded by the UAV during the calibration flight phase in open airspace; setting the upper limit of the normal fluctuation range as the initial path pattern trigger threshold; and dynamically adjusting the initial path pattern trigger threshold based on the mean and variance of the geometric complexity metric of the most recent historical flight path segment during the continuous obstacle avoidance mission phase, so as to generate the path pattern trigger threshold used in the current mission phase.

5. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, When the path pattern triggering condition is met, the instantaneous confidence is serialized according to the order in which the instantaneous obstacle avoidance decisions have been executed, forming an instantaneous confidence sequence, including: Extract the decision timestamp corresponding to each decision from each completed instantaneous obstacle avoidance decision; The extracted decision timestamps are paired with the corresponding instantaneous confidence scores; Establish a uniform time grid based on decision timestamps; Align the paired instantaneous confidence scores to a uniform time grid according to the decision timestamp; The aligned instantaneous confidence scores are arranged in chronological order to form an instantaneous confidence score sequence.

6. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Identify low-confidence events below a preset risk threshold in the instantaneous confidence sequence, and analyze the temporal clustering and transmission characteristics of these low-confidence events to obtain a comprehensive risk assessment result characterizing the risk evolution pattern of the decision-making state within the current continuous obstacle avoidance task phase, including: Each instantaneous confidence level in the instantaneous confidence level sequence is compared with a preset risk threshold, and instantaneous confidence levels below the preset risk threshold and their corresponding times are marked to form a set of low-confidence events; Based on the set of low-confidence events, the number of low-confidence events within a sliding time window is calculated, and the clustering strength is quantified according to the proportion of windows whose number exceeds the threshold of the number of events within the window. Based on a set of low-confidence events, the time intervals between consecutive low-confidence events are statistically analyzed, and the intensity of transmission is quantified according to the degree of concentration and shortening trend of the time intervals. Based on the quantified clustering strength and transmission strength, a comprehensive risk assessment result is generated.

7. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 6, characterized in that, The setting of the preset risk threshold includes: statistically analyzing the distribution of confidence values ​​in the instantaneous confidence sequence acquired during calibration environment flight; determining the initial risk threshold based on the low quantile of the confidence value distribution; calculating the moving average and standard deviation of the confidence values ​​based on the latest part of the acquired instantaneous confidence sequence during the continuous obstacle avoidance mission phase; and updating the initial risk threshold in real time based on the moving average and standard deviation to generate the currently used preset risk threshold.

8. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Based on the comprehensive risk assessment results, an overall confidence assessment conclusion for the drone is generated, including: Based on the clustering intensity and conduction intensity in the comprehensive risk assessment results, the clustering level and conduction level are determined respectively; Based on the combination of clustering level and transmission level, it is mapped to a preset overall risk level; Based on the overall risk level, an overall confidence assessment conclusion is generated, which includes recommendations to adjust the flight control strategy or maintain the current strategy.

9. The method for evaluating the confidence level of autonomous obstacle avoidance decisions for unmanned aerial vehicles (UAVs) according to claim 8, characterized in that, Recommended adjustments to the flight control strategy include: when the overall risk level is high, generating a strategy adjustment command to reduce the drone's flight speed, increase the preset safety margin between the drone and obstacles, and switch to a higher frequency obstacle avoidance decision mode; when the overall risk level is medium, generating a strategy adjustment command to maintain the current flight speed but increase the preset safety margin between the drone and obstacles; and when the overall risk level is low, generating a strategy adjustment command to maintain the current flight control strategy.

10. A confidence assessment system for autonomous obstacle avoidance decisions of unmanned aerial vehicles (UAVs), used to implement the confidence assessment method for autonomous obstacle avoidance decisions of UAVs as described in any one of claims 1-9, characterized in that, Includes the following modules: The confidence acquisition module is used to acquire the instantaneous confidence level of each instantaneous obstacle avoidance decision that the UAV has completed in the current continuous obstacle avoidance mission phase. The path triggering module is used to determine the historical flight path segment of the corresponding UAV based on the completed instantaneous obstacle avoidance decisions, and to determine whether the geometric complexity of the historical flight path segment meets the preset path shape triggering conditions. The sequence formation module is used to serialize the instantaneous confidence scores according to the order of the completed instantaneous obstacle avoidance decisions when the path pattern triggering conditions are met, so as to form an instantaneous confidence score sequence. The risk analysis module is used to identify low-confidence events in the instantaneous confidence sequence that are below a preset risk threshold, and to analyze the temporal clustering and transmission characteristics of low-confidence events to obtain a comprehensive risk assessment result that characterizes the risk evolution pattern of the decision state within the current continuous obstacle avoidance task phase. The conclusion generation module is used to generate an overall confidence assessment conclusion for the drone based on the comprehensive risk assessment results.

Images