Unmanned aerial vehicle unknown environment target search and positioning method and system based on evidence accumulation

Abstract

This invention discloses a method and system for target search and localization in unknown environments using unmanned aerial vehicles (UAVs) based on evidence accumulation. The method includes: observing candidate regions on an occupied map and estimating target positions; updating the existence confidence level using the target observation results as evidence, updating the position confidence level based on the target position and its covariance, and adjusting the state of the candidate regions based on the existence confidence level and position confidence level; estimating the detector reliability based on historical target observation data, converting it into evidence and incorporating it into the existence confidence level; calculating the observation gain for each candidate region in each state, and calculating the total observation gain; calculating a viewpoint score based on the total observation gain and flight cost; selecting the viewpoint with the highest score as the next observation position for the UAV; generating a flight trajectory for target search; updating the boundary candidate regions; and continuing until all candidate regions have been searched. This invention achieves target search and localization in unknown environments with higher search reliability and flight efficiency.

Description

Technical Field

[0001] This invention relates to the field of target search and localization technology, specifically to a method and system for target search and localization in unknown environments for unmanned aerial vehicles (UAVs) based on evidence accumulation. Background Technology

[0002] When drones perform target search and localization tasks in unknown environments, they cannot obtain environmental maps in advance, nor do they know the location and number of targets. They need to autonomously complete exploration, inspection, and localization. Existing technologies have the following three main limitations in this task:

[0003] (1) The determination of the completion of area inspection depends on geometric conditions. Existing methods usually compare the observation distance with a preset threshold. If the distance is met, the inspection is considered to be completed. This criterion measures whether the UAV has been to a sufficiently close position, rather than whether the detector has obtained sufficient information about the target. Under conditions of poor viewing angle or heavy occlusion, even if the distance is close enough, the detector may still miss the target. Moreover, the geometric condition is a one-time determination. Once the area is marked, it will not be revisited, and the missed detection cannot be corrected.

[0004] (2) The reliability of the detector depends on the prior parameters. Existing methods usually measure the performance parameters of the detector for specific targets and environments before the task and use them as fixed values, or directly use the single-frame detection confidence to replace the reliability. The former is bound to specific targets and environments and is no longer applicable after the scene changes. The latter is the output of the detector itself. High confidence does not mean that the detection is correct, and low confidence does not mean that the target does not exist. It cannot truly reflect the actual performance of the detector.

[0005] (3) The observation planning lacks specificity for different regions; most existing methods adopt a uniform evaluation standard for all regions, or divide exploration and target confirmation into independent stages and execute them sequentially. However, in actual searches, unconfirmed areas, confirmed areas with inaccurate locations, and unexplored spaces coexist, and the required observation types are different. Neither a uniform standard nor a phased approach can allocate observation resources in a targeted manner to the areas that need them most. Summary of the Invention

[0006] To overcome the technical shortcomings of existing technologies, such as reliance on geometric distance for region inspection completion determination, reliance on prior parameters for detector reliability, and lack of specificity for observation planning for different regions, this invention provides a method and system for UAV target search and localization in unknown environments based on evidence accumulation. This invention models the determination of region inspection completion as an evidence accumulation process, maintaining existence confidence and location confidence for each candidate region. Each observation is converted into evidence based on the detector's current reliability and incorporated into the confidence level; a determination is made when sufficient evidence is available. The geometric consistency of multiple detection locations is used to distinguish between true and false detections. Based on this, the detection rate, false alarm rate, and location measurement covariance of the detector under various observation conditions are estimated online, without prior determination of detector performance. The observation benefit is calculated based on the type of evidence currently lacking in each region, and different types of needs are uniformly prioritized and competed for. Flight resources are allocated to the regions most in need of observation, eliminating the need for prior determination of detector performance. This method achieves target search and localization in unknown environments with higher search reliability and flight efficiency, and is applicable to different detectors and UAV platforms.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for searching and locating unmanned aerial vehicles (UAVs) in unknown environments based on evidence accumulation, comprising the following steps:

[0009] Extract candidate regions from the map, observe targets, and estimate their locations;

[0010] Each target observation result is used as evidence to update the existence confidence; the location confidence is updated based on the estimated target location and its covariance; and the state of the candidate region is adjusted based on the existence confidence and location confidence.

[0011] The reliability of the detector is estimated based on the historical observation data of the target. The detection rate, false alarm rate and position measurement covariance of the detector under various observation conditions are estimated. The reliability is converted into evidence and incorporated into the existence confidence.

[0012] Calculate the observation reward for each candidate region under each state, and then calculate the total observation reward.

[0013] The viewpoint score is calculated based on the total observation gains and flight costs. The viewpoint with the highest score is selected as the next observation position for the UAV. A flight trajectory is generated to search for targets, and the boundary candidate regions are updated until the target search for all candidate regions is completed.

[0014] As a preferred technical solution, the degree of certainty is expressed as a logarithmic probability:

[0015] ;

[0016] in, Indicates candidate region There is a goal in it. Indicates candidate region There is no target in it. For all valid observations in the candidate region;

[0017] Each observation result is used as evidence to update the level of confidence, and the update rule is as follows:

[0018] ;

[0019] in, This indicates a degree of certainty. This is the ratio of the probability of the observed result occurring in the two scenarios: the presence of the target and its absence.

[0020] When the target exists:

[0021] ;

[0022] When the target does not exist:

[0023] ;

[0024] in, For observation conditions The detection rate is below This represents the false alarm rate.

[0025] As a preferred technical solution, the observation conditions are expressed as follows:

[0026] ;

[0027] in, The distance from the camera to the target. Let be the angle between the camera's optical axis and the normal to the target surface. This represents the percentage of time when the view is obstructed.

[0028] As a preferred technical solution, the state of the candidate region is adjusted based on the existence confidence and the location confidence, specifically including:

[0029] When the confidence level exceeds the judgment threshold, the candidate region is confirmed or excluded. When the existence of the target is confirmed, it enters the coarse confirmation state. When the uncertainty of the position confidence level converges to within the preset accuracy, the candidate region changes from the coarse confirmation state to the fine confirmation state, and the three-dimensional coordinates of the target position are output.

[0030] The threshold for judgment includes an upper threshold and a lower threshold, specifically expressed as follows:

[0031] ;

[0032] ;

[0033] in, Tolerable false positive rate, A certain level of certainty exists to account for a tolerable false exclusion rate. Above the upper threshold When the target is confirmed to exist, it is below the lower threshold. Eliminate targets in time.

[0034] As a preferred technical solution, the location certainty follows a three-dimensional Gaussian distribution. express, For the estimated target location, Its covariance;

[0035] The initial value for location confidence is set as follows: the mean is the result of covariance-weighted fusion of the detection locations that have passed the consistency test. Covariance It is proportional to the size of the candidate region;

[0036] Each new valid observation is fused according to its information format:

[0037] ;

[0038] ;

[0039] in, The observation conditions for the i-th target observation are as follows: Indicates the target position in the i-th target observation. Let be the covariance under the observation conditions of the i-th target observation.

[0040] As a preferred technical solution, the consistency verification specifically includes:

[0041] The same target is detected multiple times from different locations. All detections are paired, and the squared Mahalanobis distance is calculated for each pair. A chi-square test is performed on the squared Mahalanobis distance. Pairs that do not exceed a critical value are considered consistent. The proportion of consistent pairs to the total number of pairs is counted. If the proportion exceeds a preset threshold, the candidate region is considered to contain a real target. All detection locations that participated in consistent pairing are weighted by covariance and fused to obtain a preliminary estimate of the target location. .

[0042] As a preferred technical solution, the reliability of the detector is estimated based on historical observation data of the target, specifically including:

[0043] The reliability of the detector is estimated based on the detection rate, false alarm rate, and position measurement covariance. The detection rate is the probability that the detector will detect the target when it exists, specifically:

[0044] After confirming the existence of the target, the detection rate is estimated based on the historical observation of the target. The frame-by-frame judgment is made to determine whether the preset observation condition range is met. The frame that meets the observation condition range is counted as one geometric observation opportunity. The detection rate estimate is calculated by using the number of geometric observation opportunities as the denominator and the actual number of detected frames as the numerator.

[0045] The false alarm rate is the probability that the detector will falsely report an error when the target does not exist. The position measurement covariance is obtained by statistically analyzing the deviation of the detected position from the target position.

[0046] As a preferred technical solution, the detection rate and location measurement covariance vary with observation conditions;

[0047] The observation conditions space is divided into multiple intervals. Observations falling within the same interval are considered to have similar observation conditions. The corresponding detection rate and location measurement covariance are obtained from the accumulated observation statistics within the interval. Each time a new observation is obtained, the statistics of the interval to which it belongs are updated. When estimating the reliability of the detector, the estimated value of the interval to which its observation conditions belong is taken. When the observation conditions are near the boundary between two intervals, linear interpolation is performed on the adjacent intervals.

[0048] As a preferred technical solution, the observation benefits are calculated for each candidate region separately, specifically including:

[0049] The candidate region is initially in an undetermined state. The payoff in the undetermined state is the decrease in confidence level after observation, expressed as:

[0050] ;

[0051] in, Indicates the candidate region. For entropy with a degree of certainty, This indicates that there is currently a degree of certainty. In viewpoint The existence confidence level is updated after observation. This represents the expectation, which is the average of the observed results according to their respective probabilities of occurrence.

[0052] For a coarsely confirmed state, the benefit is the reduction in the range of uncertainty regarding the location confidence after observation, expressed as:

[0053] ;

[0054] in, In viewpoint Observe and fuse the covariance after the new measurements;

[0055] For a candidate boundary region to be explored, the reward is the number of newly observed boundary voxels from that viewpoint.

[0056] The present invention also provides a UAV target search and localization system based on evidence accumulation, for implementing the above-mentioned UAV target search and localization method based on evidence accumulation, including: a perception and mapping module, a candidate region management module, a detection reliability estimation module, an evidence accumulation and state update module, and a viewpoint planning and flight control module.

[0057] The perception and mapping module is used to construct an occupation map, perform target observation and estimate target position and measure covariance, and output observation conditions.

[0058] The candidate region management module is used to extract candidate regions occupying the map and adjust the status of the candidate regions based on the existence confidence and location confidence.

[0059] The detection reliability estimation module is used to estimate the reliability of the detector based on the target's historical observation data, and to estimate the detector's detection rate, false alarm rate and position measurement covariance under various observation conditions, converting reliability into evidence and incorporating it into the existence confidence level.

[0060] The evidence accumulation and state update module is used to update the existence confidence level by taking each target observation result as evidence, update the position confidence level according to the estimated target position and its covariance, update the state of the candidate region according to the existence confidence level and the position confidence level, and transmit it to the candidate region management module.

[0061] The viewpoint planning and flight control module is used to perform viewpoint planning and flight control. It calculates the observation benefits for candidate areas under each state, calculates the viewpoint score based on the total observation benefits and flight costs, selects the viewpoint with the highest score as the next observation position of the UAV, generates a flight trajectory for target search, updates the boundary candidate areas, and continues until the target search of all candidate areas is completed.

[0062] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0063] (1) This invention transforms the inspection completion from a one-time judgment to a progressive inference. Each observation, regardless of the strength of the conclusion, is used as evidence to participate in the accumulation, and finally converges to a reliable judgment. The judgment threshold is derived from the tolerable error rate, ensuring that the false judgment probability does not exceed the upper limit specified by the user, which can reduce the risk of missed detection and false detection, and make the error probability controllable.

[0064] (2) Existing methods require the determination of detector performance parameters for specific targets and environments before the task, and the determination needs to be repeated after the scene changes. This invention uses the geometric consistency of multiple detection positions to determine the authenticity of the detection, and directly estimates the reliability from the observation data during the search process. When the target or environment changes, the system automatically updates from the new data. No manual intervention or prior parameters are required, and it is applicable to different target types and environmental conditions.

[0065] (3) The present invention calculates the observation benefits based on the type of evidence currently lacking in each area. Different types of needs are uniformly ranked and competed at each step. Flight resources are allocated to the areas most in need of observation, avoiding indiscriminate repeated visits and achieving high efficiency in flight resource utilization.

[0066] (4) The reliability estimation of the present invention is based on the statistical performance of the detector. When the detector is replaced, the evidence accumulation and observation plan do not need to be modified. It can be run on various UAV platforms, does not depend on a specific detector, and is convenient for engineering deployment. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the UAV target search and localization method based on evidence accumulation according to the present invention.

[0068] Figure 2 This is a schematic diagram of the state transition of the candidate region in this invention;

[0069] Figure 3 This is a schematic diagram illustrating the authenticity determination based on geometric consistency of the present invention;

[0070] Figure 4 This is a schematic diagram illustrating the viewpoint benefits of the present invention based on evidence status;

[0071] Figure 5 This is a schematic diagram of the main process of the UAV performing target search and localization tasks in an unknown environment according to the present invention;

[0072] Figure 6 This is a schematic diagram of the overall architecture of the UAV target search and localization system based on evidence accumulation according to the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] Example 1

[0075] like Figure 1 As shown, this embodiment provides a method for UAV target search and localization in unknown environments based on evidence accumulation. This method is achieved through steps including evidence accumulation-based inspection completion determination, online estimation of detector reliability, and observation planning driven by evidence status. The UAV generates detection results during flight observation; these results are converted into evidence based on reliability and accumulated in corresponding areas. Each area updates its status based on the accumulated evidence, and then selects the next observation location based on the status of each area, driving a new round of flight. Specifically, the method includes the following steps:

[0076] S1: Determination of completion of inspection;

[0077] This embodiment models the determination of the completion of area inspection as an evidence accumulation process. The judgment is made based on the sufficiency of the accumulated detection evidence for the area, independent of geometric conditions. For each candidate area, an existence confidence score and a location confidence score are maintained. The existence confidence score reflects whether the area contains a target. Each time the UAV completes an observation of the area, whether it detects or not, it is added as a piece of evidence to the existence confidence score. The weight of each piece of evidence is determined by the detector's reliability at that time. Higher reliability results in a greater increase in the existence confidence score upon detection and a greater reduction upon non-detection. When the accumulated existence confidence score exceeds a judgment threshold, the area is confirmed to contain a target or is excluded.

[0078] The judgment threshold is set as follows: The user specifies the tolerable false confirmation rate and false rejection rate according to the task requirements. Based on the sequential test theory, the judgment threshold is derived from these two error rates. The derivation theoretically ensures that the actual false confirmation probability and false rejection probability do not exceed the upper limit specified by the user.

[0079] Specifically, the decision threshold is calculated from the tolerable error rate, let... Tolerable false positive rate, To achieve a tolerable false exclusion rate, the upper and lower thresholds are as follows:

[0080] ;

[0081] ;

[0082] in, Higher than Confirm the target exists in time, below The target is excluded if the error occurs, and observation continues if the error falls between these two thresholds. Within this threshold, the false positive probability does not exceed [a certain threshold]. The probability of false exclusion does not exceed .

[0083] In this embodiment, the drone continuously extracts candidate regions from the occupied map during flight, and records one of the candidate regions as... The degree of certainty is expressed as a logarithmic probability:

[0084] ;

[0085] in, express There is a goal in it. express There is no target in it. This refers to all valid observations of the area to date;

[0086] For each valid observation obtained, the existence confidence level is updated using the observation result, which can be either detected or not detected. The update rule is as follows:

[0087] ;

[0088] in, This is the ratio of the probability of the result occurring in the two scenarios of the target's presence and absence. When detecting:

[0089] ;

[0090] When not detected:

[0091] ;

[0092] in, For observation conditions The detection rate is below The false alarm rate is given by the detection reliability estimate. The detection rate varies with the observation conditions (distance, viewing angle, occlusion), while the false alarm rate is taken as a constant.

[0093] Before updating, it is necessary to determine whether the current observation detected or did not detect the region. This step is called detection association, and the rules are as follows: If the position measurement output by the detector in a certain frame falls within the region... Inside the enclosure, or to The distance to the geometric center does not exceed the associated threshold. Then associate the detection with When multiple detections in a frame meet the conditions, the one with the highest confidence is selected. When the same detection meets the conditions of multiple regions, it is assigned to the region with the closest geometric center. Detections that cannot be associated with any region are temporarily stored. After accumulating a certain number of frames in the vicinity of the same location, a decision is made on whether to create a new region. If there is no detection associated with it within the current frame's field of view, it is recorded as undetected. Undetected is also considered as evidence and included in the confidence level.

[0094] For regions confirmed to contain targets, the target's 3D coordinates are further estimated using location confidence. Each detection generates a location measurement, with varying accuracy under different observation conditions. The measurements are then weighted and fused according to accuracy, with higher-accuracy measurements receiving greater weight. As the number of effective observations increases, the uncertainty range of the fused result monotonically decreases, converging to the required accuracy for the mission, at which point the target location is output.

[0095] In this embodiment, the location certainty follows a three-dimensional Gaussian distribution. express, For the estimated target location, Its covariance;

[0096] Each detection will yield a measurement of the target location. Its covariance is , These are the observation conditions for this study. Location measurements obtained during the pending phase are cached and used only after the region's existence is confirmed.

[0097] After the region is transferred to coarse confirmation, the initial value for location confidence is set as follows: the mean is the result of the covariance-weighted fusion of detection locations that passed the consistency test in the geometric consistency discrimination. Covariance It is proportional to the size of the region.

[0098] Each subsequent new valid test result is then fused according to the information format:

[0099] ;

[0100] ;

[0101] When the trace of the covariance matrix Reduced to the accuracy threshold When the location accuracy meets the standard, the area enters the precise confirmation state.

[0102] like Figure 2 As shown, based on the two levels of certainty mentioned above, each candidate region is categorized into one of four states: pending, excluded, coarsely confirmed, and finely confirmed. A region initially enters the pending state. When the certainty is sufficient to confirm the target's existence, it transitions to the coarsely confirmed state; when it is sufficient to exclude the target, it transitions to the excluded state. In the coarsely confirmed state, observation continues to refine the location. When the uncertainty range of the location certainty converges below the accuracy threshold, it transitions to the finely confirmed state. The pending and coarsely confirmed states are active states, and the UAV continues observation; the excluded and finely confirmed states are terminated states, and observation ceases. The state transition is unidirectional; once a region's existence is confirmed, it will not revert to the pending state.

[0103] Specifically, in the pending state, the determination is based on the level of certainty of existence: if the level of certainty of existence is higher than the upper threshold... When the threshold is lower than the minimum threshold, the process switches to coarse confirmation. If the situation falls into either of the two categories, then exclusion is considered; otherwise, observation continues.

[0104] With the initial assessment complete, continue observations of the area to refine the location. When the trace of the covariance matrix... Reduced to the accuracy threshold At this point, the system transitions to a precision measurement state and outputs the target's three-dimensional coordinates.

[0105] In the process of accumulating evidence, the weight of each piece of evidence depends on the reliability of the detector at that time. This reliability varies with the target type and environmental conditions, cannot be predetermined before the mission, and needs to be estimated online during the search process.

[0106] S2: Online estimation of detection reliability;

[0107] In this embodiment, the reliability of the detector is estimated online using observation data during the search process, without relying on any prior parameters. The reliability of the detector is measured by three quantities: detection rate. False alarm rate Covariance of position measurement The detection rate is the probability that the detector will detect the target when it exists, the false alarm rate is the probability that the detector will falsely report the target when it does not exist, and the position measurement covariance reflects the magnitude of the deviation between the detected position and the true position. The detection rate and the position measurement covariance vary with the observation conditions, which are described by distance, viewing angle and degree of occlusion. The false alarm rate is independent of the observation conditions and is taken as a constant.

[0108] Specifically, with Indicates the observation conditions, where, The distance from the camera to the target. Let be the angle between the camera's optical axis and the normal to the target surface. The proportion of obstructed view, the detection rate, and the covariance of location measurement vary with observation conditions, denoted as . and ;

[0109] The observation conditions are continuously measured, making it impossible to estimate each value individually. This embodiment follows... , , The three dimensions divide the observation condition space into a finite number of intervals. Observations falling within the same interval are considered to have similar conditions. The corresponding detection rate and location measurement covariance are obtained from the accumulated observation statistics within the interval. Each time a new observation is obtained, the statistics of its corresponding interval are updated. When the reliability of a detection is needed, the estimated value of the interval to which the observation conditions belong is taken. When the conditions are near the boundary between two intervals, linear interpolation is performed on the adjacent intervals. When the sample size of an interval is insufficient, the estimated value of the adjacent interval is borrowed.

[0110] Online reliability estimation requires obtaining detection samples from real targets, such as... Figure 3 As shown, this embodiment utilizes the geometric consistency of multiple detection positions of a real target to determine the authenticity of detections, thereby obtaining reliable samples for estimation. The real target is stationary; when the UAV detects the same target multiple times from different positions, the resulting position measurements converge in three-dimensional space. False detections do not correspond to any real object, and the positions of multiple detections are scattered in space. Therefore, when multiple detection positions in a region pass the statistical consistency test, the region is determined to contain a real target. Since the measurement accuracy of each detection is different, Mahalanobis distance is used to measure whether the position difference between two detections is within the allowable range of their respective measurement accuracy.

[0111] Specifically, after a candidate region accumulates at least 4 detections, a true / false detection is triggered. All detections in that region are paired up, and for each pair of detections... and Calculate the squared Mahalanobis distance:

[0112] ;

[0113] in, For the combined covariance of the errors of the two measurements, A chi-square test was performed with a significance level of 0.05. Pairs that did not exceed the critical value were considered to be consistent.

[0114] In this embodiment, the proportion of consistent pairings to the total number of pairings is counted. When the proportion exceeds 50%, the region is determined to contain a real target. All detection locations participating in consistent pairings are weighted by covariance and fused to obtain a preliminary estimate of the target location. .

[0115] After confirming the real target, the detection rate is estimated using the target's historical observations. After the target is confirmed, the previous flight trajectory is traced back, and the target is judged frame by frame to determine whether it falls within the camera's field of view, whether the line of sight is blocked, and whether the distance is within the effective range. Frames that meet all the conditions are counted as one geometric observation opportunity. Regardless of whether the detector has an output at that time, frames that meet the conditions are counted.

[0116] Within each interval, the unbiased detection rate estimate is obtained by dividing the geometrical chance of observation by the actual number of detected frames. Location measurement covariance The detected position within each interval is relative to the target position. The deviation statistics are obtained; after the target is confirmed, new observations are incrementally merged into the corresponding interval in the same way;

[0117] False alarm rate The false alarm rate is estimated by using the number of false detection frames as the numerator and the number of geometric observation opportunities in the corresponding region as the denominator, obtained from the scattered detection statistics that do not constitute spatial clusters. The false alarm rate is then combined across intervals into a constant.

[0118] S3: Evidence-driven observation planning;

[0119] The UAV uniformly samples viewpoint positions within a pre-defined 3D grid within a known, mapped space. The grid resolution is denoted as [value missing]. Positions falling within the voxel or safety margin are discarded. At each retained position, viewpoints are generated by dividing the 360-degree angle into several discrete yaw angles. Pitch and roll are consistent with the aircraft by default.

[0120] viewpoint With candidate regions The visibility between them is determined in two steps: First, the candidate regions are... Geometric center transformation to viewpoint The corresponding camera coordinate system is used to determine whether the view falls within the view cone based on the camera's horizontal and vertical field of view angles and near and far clipping planes; the second step is to determine whether the view falls within the view cone along the viewpoint. To candidate area Calculate the occlusion ratio of the line of sight at the geometric center That is, the proportion of the length of the line of sight through which the obstacle passes to the total length, and the geometric center falling into the visual cone and When the occlusion ratio is less than the maximum permissible occlusion ratio, the viewpoint For candidate regions visible.

[0121] For each candidate viewpoint All undetermined regions visible to it are grouped into a set. The area was roughly identified and classified. Boundary candidate regions are included .

[0122] This embodiment calculates the observation gains based on the types of evidence currently lacking in each area, and selects the viewpoint with the highest gains to execute the flight.

[0123] like Figure 4 As shown, the observation benefits are calculated for three types of objects: for undetermined areas, the amount of uncertainty in the confidence level can be reduced after observation is calculated; for coarsely confirmed areas, the amount of uncertainty in the confidence level of location can be reduced after observation is calculated; for unknown spaces to be explored, the number of new unknown boundaries that can be observed from this viewpoint is calculated.

[0124] Specifically, for the undetermined area The return is defined as the decrease in confidence level after observation:

[0125] ;

[0126] in, The entropy is the current degree of certainty. In viewpoint The existence confidence level is updated after observation, and it is expected that the average of the detection and non-detection probabilities is taken.

[0127] For the coarsely confirmed area The benefit is defined as the reduction in the range of uncertainty regarding the location confidence after observation:

[0128] ;

[0129] in, In viewpoint Observe and fuse the covariance after the new measurements;

[0130] For the candidate boundary regions to be explored ,income The number of new boundary voxels that can be observed from this viewpoint can be weighted according to the distance from the viewpoint to the voxels.

[0131] A single viewpoint can typically observe multiple areas simultaneously. The total information gain is the sum of three types of gains. Viewpoint evaluation is calculated as the total information gain minus the flight cost. The UAV selects the viewpoint with the highest score as the next observation location, employing a two-layer planning approach: global and local. The global layer organizes high-value viewpoints into a visit sequence, while the local layer proceeds to each viewpoint in that sequence, planning obstacle avoidance paths and generating smooth trajectories. New observations continuously arrive during flight, updating evidence, status, and reliability estimates accordingly. The two-layer planning is dynamically adjusted. The search mission ends when all areas reach a terminated state and there are no more unknown spaces to explore.

[0132] Specifically, the total information gain from a viewpoint is the sum of three types of gains:

[0133] ;

[0134] The three types of returns can be adjusted according to preset weights;

[0135] Viewpoint rating is defined as:

[0136] ;

[0137] in, The distance from the drone's current position to the viewpoint Path length, By adjusting the relative weights of benefits and costs, the drone selects the viewpoint with the highest score as the next observation position.

[0138] In this embodiment, the global layer uses the viewpoints with the highest scores as nodes, models them as a traveling salesman problem, and obtains a visiting order. The local layer then proceeds to the next viewpoint according to this order. First, an obstacle avoidance path is planned in the known map, and then a smooth trajectory that meets the dynamic constraints is fitted using B-splines and handed over to the flight controller for execution.

[0139] New observations continue to arrive during the flight, and the confidence level, regional status, and reliability estimates are updated accordingly. The two-layer plan is dynamically adjusted, and the search mission ends when all regions enter a terminated state and there is no unknown space to be explored.

[0140] In this embodiment, when the UAV performs target search and localization tasks in an unknown environment, it explores the unknown space to build an environmental map, checks the explored areas to determine if targets exist, and locates confirmed targets to estimate their three-dimensional coordinates. The exploration generates new areas to be checked, and confirmed targets require further localization. The detection results generated by the UAV's flight observations are converted into evidence based on reliability and accumulated in corresponding areas. Each area updates its status based on the accumulated evidence, and then selects the next observation position based on the status of each area, driving a new round of flight. This continues until the flight budget is exhausted or termination conditions are triggered. Figure 5 As shown, the specific execution process is as follows:

[0141] (1) Initialization: For each candidate region, the existence confidence is set to the initial value, the location confidence is set to the weak prior, and the region state is set to undetermined;

[0142] (2) Determine whether the task has been terminated: When all areas are in a terminated state and there is no unknown space to be explored, the task is terminated and the search and location results are output; otherwise, continue to the next round of observation.

[0143] (3) Planning viewpoints: Generate candidate viewpoints for this round, calculate the benefit of each viewpoint based on the type of evidence currently lacking in each region, deduct the flight cost to obtain a score, and select the viewpoint with the highest score;

[0144] (4) Flight execution: The flight proceeds to the selected viewpoint according to a two-tiered plan, encompassing both global and local aspects, and receives detection frames along the way. Each detection frame is accompanied by its observation conditions;

[0145] (5) Update confidence: For each frame of observation results, determine the evidence components according to the reliability of the detector at that time, and update the existence confidence and location confidence respectively;

[0146] (6) Determine the status of the region: If there is a region with a confidence level higher than the upper threshold, it will be transferred to the coarse confirmation state. Among them, the historical observations of the region whose detection location passes the statistical consistency test will be used to update the reliability estimate. If there is a region with a confidence level lower than the lower threshold, it will be transferred to the exclusion state. When the location uncertainty of the coarse confirmation region drops below the accuracy threshold, it will be transferred to the fine confirmation state and the target location will be output. Regions in the termination state will be removed from the candidate pool, and regions in the active state will continue to participate in subsequent observations.

[0147] (7) Update candidate regions: Update the candidate regions as the occupied map increases, extract new regions for newly added spaces, remove invalid regions for disappeared spaces, update the boundary candidate regions, and return to step (2).

[0148] Example 2

[0149] like Figure 6As shown, this embodiment provides a UAV target search and localization system based on evidence accumulation, which is used to implement the UAV target search and localization method based on evidence accumulation in Embodiment 1. The system includes: a perception and mapping module, a candidate region management module, a detection reliability estimation module, an evidence accumulation and state update module, and a viewpoint planning and flight control module.

[0150] The perception and mapping module is used to acquire environmental information, including a mapping unit and a detection unit. The mapping unit uses LiDAR to record which locations in the environment are occupied by obstacles and which locations are free space, and estimates the real-time pose of the UAV. The detection unit uses a camera to perform target detection on each frame of RGB image. When a target is detected, it outputs the target's category and confidence level, and estimates the target's three-dimensional position in the world coordinate system by combining depth information. It also provides the measurement covariance of the position. Each frame of detection also records the conditions of the observation, including distance, viewing angle, and degree of occlusion.

[0151] The candidate region management module is used to extract and maintain two types of candidate regions from the occupied map. The first type is located inside the explored space: adjacent occupied voxels are merged into the same group, and each group is a candidate region, representing a space that needs to be checked to see if it contains a target. If the region is too long or too large, it is split along the main extension direction. The second type is located at the boundary between the explored and unexplored spaces: adjacent voxels on the boundary are grouped into a group to form a boundary candidate region, representing an unknown direction that can be explored by the drone. Each candidate region maintains a current state, with four types: pending, excluded, coarsely confirmed, and finely confirmed.

[0152] The detection reliability estimation module is used to estimate the reliability of the detector online. The reliability of the detector is measured by three quantities: detection rate, false alarm rate, and position measurement covariance. It obtains the signal of a confirmed area from the candidate area management module, and then obtains the historical observations of the area before it was confirmed from the perception and mapping module. Based on this, it calculates the reliability of the detector under various observation conditions. The estimation results are output to the evidence accumulation and state update module and the viewpoint planning and flight control module.

[0153] The evidence accumulation and status update module is used to update the existence confidence and location confidence of candidate regions. Each time a new observation is obtained, the evidence component for this detection is determined based on the reliability given by the detection reliability estimation module, and accumulated onto the two confidence levels respectively. When the existence confidence exceeds the judgment threshold, the region is confirmed or excluded. When the uncertainty of the location confidence converges to within a preset precision, the region transitions to a precise confirmation state. The updated status is written back to the candidate region management module.

[0154] The viewpoint planning and flight control module is used to select the next observation position based on the status of each candidate region. It calculates the observation benefits for the undetermined region, the coarsely confirmed region, and the boundary candidate region respectively. Taking into account the benefits and flight costs, it selects the viewpoint with the highest score and executes the flight. New observations obtained during the flight are sent to the perception and mapping module to form a closed loop.

[0155] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. An evidence accumulation based method for unknown environment target search and localization of unmanned aerial vehicles, characterized in that, Includes the following steps: Extract candidate regions from the map, observe targets, and estimate their locations; Each target observation result is used as evidence to update the existence confidence; the location confidence is updated based on the estimated target location and its covariance; and the state of the candidate region is adjusted based on the existence confidence and location confidence. The reliability of the detector is estimated based on the historical observation data of the target. The detection rate, false alarm rate and position measurement covariance of the detector under various observation conditions are estimated, and the reliability is converted into evidence and incorporated into the existence confidence. Calculate the observation reward for each candidate region under each state, and then calculate the total observation reward. The viewpoint score is calculated based on the total observation gains and flight costs. The viewpoint with the highest score is selected as the next observation position for the UAV. A flight trajectory is generated to search for targets, and the boundary candidate regions are updated until the target search for all candidate regions is completed.

2. The cumulative evidence-based UAV unknown environment target search and localization method according to claim 1, wherein, The degree of certainty of existence is expressed as a logarithmic probability: ； in, Indicates candidate region There is a goal in it. Indicates candidate region There is no target in it. For all valid observations in the candidate region; Each observation result is used as evidence to update the level of confidence, and the update rule is as follows: ； in, This indicates a degree of certainty. This is the ratio of the probability of the observed result occurring in the two scenarios: the presence of the target and its absence. When the target exists: ； When the target does not exist: ； in, For observation conditions The detection rate is below This represents the false alarm rate.

3. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 2, characterized in that, The observation conditions are expressed as follows: ； in, The distance from the camera to the target. Let be the angle between the camera's optical axis and the normal to the target surface. This represents the percentage of time when the view is obstructed.

4. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 1, characterized in that, The state of candidate regions is adjusted based on the confidence level of existence and the confidence level of location, specifically including: When the confidence level exceeds the judgment threshold, the candidate region is confirmed or excluded. When the existence of the target is confirmed, it enters the coarse confirmation state. When the uncertainty of the position confidence level converges to within the preset accuracy, the candidate region changes from the coarse confirmation state to the fine confirmation state, and the three-dimensional coordinates of the target position are output. The threshold for judgment includes an upper threshold and a lower threshold, specifically expressed as follows: ； ； in, Tolerable false positive rate, A certain level of certainty exists to account for a tolerable false exclusion rate. Above the upper threshold The target is confirmed to exist when it falls below the lower threshold. Eliminate targets in time.

5. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 1, characterized in that, The location confidence is distributed in a three-dimensional Gaussian pattern. express, For the estimated target location, Its covariance; The initial values ​​for location confidence are set as follows: the mean is the result of covariance-weighted fusion of the detected locations that have passed the consistency test. Covariance It is proportional to the size of the candidate region; Each new valid observation is fused according to its information format: ； ； in, The observation conditions for the i-th target observation are as follows: Indicates the target position in the i-th target observation. Let be the covariance under the observation conditions of the i-th target observation.

6. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 5, characterized in that, The consistency check specifically includes: The same target is detected multiple times from different locations. All detections are paired, and the squared Mahalanobis distance is calculated for each pair. A chi-square test is performed on the squared Mahalanobis distance. Pairs that do not exceed a critical value are considered consistent. The proportion of consistent pairs to the total number of pairs is counted. If the proportion exceeds a preset threshold, the candidate region is considered to contain a real target. All detection locations that participated in consistent pairing are weighted by covariance and fused to obtain a preliminary estimate of the target location. .

7. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 1, characterized in that, The reliability of the detector is estimated based on historical observation data of the target, specifically including: The reliability of the detector is estimated based on the detection rate, false alarm rate, and position measurement covariance. The detection rate is the probability that the detector will detect the target when it exists, specifically: After confirming the existence of the target, the detection rate is estimated based on the historical observation of the target. The frame-by-frame judgment is made to determine whether the preset observation condition range is met. The frame that meets the observation condition range is counted as one geometric observation opportunity. The detection rate estimate is calculated by using the number of geometric observation opportunities as the denominator and the actual number of detected frames as the numerator. The false alarm rate is the probability that the detector will falsely report an error when the target does not exist. The position measurement covariance is obtained by statistically analyzing the deviation of the detected position from the target position.

8. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 1, characterized in that, Detection rate and location measurement covariance vary with observation conditions; The observation conditions space is divided into multiple intervals. Observations falling within the same interval are considered to have similar observation conditions. The corresponding detection rate and location measurement covariance are obtained from the accumulated observation statistics within the interval. Each time a new observation is obtained, the statistics of the interval to which it belongs are updated. When estimating the reliability of the detector, the estimated value of the interval to which its observation conditions belong is taken. When the observation conditions are near the boundary between two intervals, linear interpolation is performed on the adjacent intervals.

9. The method for searching and locating unmanned aerial vehicle (UAV) targets in unknown environments based on evidence accumulation according to claim 4, characterized in that, The observation benefits are calculated separately for each candidate region, including: The candidate region is initially in an undetermined state. The payoff in the undetermined state is the decrease in confidence level after observation, expressed as: ； in, Indicates the candidate region. For entropy with a degree of certainty, This indicates that there is currently a degree of certainty. In viewpoint The existence confidence level is updated after observation. This represents the expectation, which is the average of the observed results according to their respective probabilities of occurrence. For a coarsely confirmed state, the benefit is the reduction in the range of uncertainty regarding the location confidence after observation, expressed as: ； in, In viewpoint Observe and fuse the covariance after the new measurements; For a candidate boundary region to be explored, the reward is the number of newly observed boundary voxels from that viewpoint.

10. A UAV target search and localization system for unknown environments based on evidence accumulation, characterized in that, The method for searching and locating unmanned aerial vehicle targets in unknown environments based on evidence accumulation, as described in any one of claims 1-9, includes: a perception and mapping module, a candidate region management module, a detection reliability estimation module, an evidence accumulation and state update module, and a viewpoint planning and flight control module. The perception and mapping module is used to construct an occupation map, perform target observation and estimate target position and measure covariance, and output observation conditions. The candidate region management module is used to extract candidate regions occupying the map and adjust the status of the candidate regions based on the existence confidence and location confidence. The detection reliability estimation module is used to estimate the reliability of the detector based on the target's historical observation data, and to estimate the detector's detection rate, false alarm rate and position measurement covariance under various observation conditions, converting reliability into evidence and incorporating it into the existence confidence level. The evidence accumulation and state update module is used to update the existence confidence level by taking each target observation result as evidence, update the position confidence level according to the estimated target position and its covariance, update the state of the candidate region according to the existence confidence level and the position confidence level, and transmit it to the candidate region management module. The viewpoint planning and flight control module is used to perform viewpoint planning and flight control. It calculates the observation benefits for candidate areas under each state, calculates the viewpoint score based on the total observation benefits and flight costs, selects the viewpoint with the highest score as the next observation position of the UAV, generates a flight trajectory for target search, updates the boundary candidate areas, and continues until the target search of all candidate areas is completed.

Images