An unmanned aerial vehicle obstacle avoidance method and device

Abstract

The application discloses an unmanned aerial vehicle obstacle avoidance method and device, and particularly relates to the technical field of unmanned aerial vehicle obstacle avoidance, which comprises the following steps: constructing a truss semantic topology, obtaining a working corridor set on a safety boundary grid and marking a key visual field cone; solving an initial flight path and forming a flight point observation constraint set under the constraint; completing infrared and visible light registration and projecting to the grid by using a pose prior, calculating an obstacle occupancy probability graph, extracting an obstacle boundary set and an obstacle motion trend, and then updating a local traffic corridor; on the basis, calculating a visual angle disturbance robustness index and a channel safety margin index, inputting a predefined obstacle avoidance correction model to obtain an obstacle avoidance correction coefficient, combining a basic risk score to form a judgment score, and comparing the judgment score with an obstacle avoidance threshold parameter to generate a passing or avoiding decision; solving a normal micro-displacement vector along a target obstacle normal and implementing three-vector alignment control to make the body enter the nearest safe traffic corridor.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle obstacle avoidance, and more particularly to an unmanned aerial vehicle obstacle avoidance method and device. BACKGROUND

[0002] In the tower crane arm inspection scene, the traditional inspection method mainly depends on ground telescopes, close-range manual climbing or single visible light unmanned aerial vehicle photographing, which has disadvantages in safety, efficiency and result stability, and is easily affected by personnel experience; the recognition rate of the visible light unmanned aerial vehicle method is low. Although infrared thermal imaging technology has been introduced in recent years to find potential thermal anomalies, the emissivity of the metal surface is significantly affected by the coating state, observation angle and temperature difference change, and at the same time, solar radiation, sky background radiation and wind disturbance will introduce temperature measurement errors, resulting in false temperature rise or false cooling phenomenon, causing inaccurate temperature quantification and false reporting problems.

[0003] In the flight and perception link, the narrow channel and complex rod structure of the tower crane arm are easy to cause satellite navigation signal shielding and multipath effect; arm rotation, wind field turbulence and electromagnetic interference will cause attitude jitter and motion blur, thereby affecting the subsequent image registration and target detection accuracy. The existing path planning algorithm is mainly for open scenes, based on simple grid or fixed flight line, and it is difficult to realize adaptive inspection strategy considering high coverage rate and safe distance in the truss interior or near distance contour. In the target detection level, the overall recall rate of single-mode visible light or single-frame infrared image is low when dealing with weak temperature difference, weak texture, small size and slender defects, and the general detection algorithm is easy to overfit in the high-reflective and high-texture background of metal components, resulting in limited generalization ability. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides an unmanned aerial vehicle obstacle avoidance method and device, which generates a work corridor and a key view cone through semantic topology and a safety boundary grid; obtains synchronous image data and pose labels through planning and collection integration; constructs a view angle disturbance robustness index and a channel safety margin index, and forms a judgment score through a pre-defined obstacle avoidance correction model; controls to enter the nearest safe passage corridor and generates an obstacle avoidance event vector through normal micro-displacement and three-vector alignment, so as to solve the problems proposed in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] An unmanned aerial vehicle obstacle avoidance method, comprising:

[0007] Step S1: When the unmanned aerial vehicle enters the truss work area or detects an increased obstacle density, generate a truss semantic topology by fusing the tower crane structure profile and the field surveying data, project a passable work corridor according to the minimum safety distance and label a key view cone, and output a work corridor set and a safety boundary grid;

[0008] Step S2: When receiving the job corridor set and the safety boundary grid, solve the continuous flight path according to the coverage integrity and safety distance, and curve smooth, assign the expected observation angle and ground resolution value to each waypoint, and synchronously collect multi-source perception sequence with hardware trigger to generate initial flight path and synchronous image data;

[0009] Step S3: Register infrared and visible light with pose priori and project to safety boundary grid; generate occupancy map and free space, extract obstacle boundary and motion trend, update local traffic corridor and output observation angle time function;

[0010] Step S4: When the obstacle boundary set is updated, calculate the visual angle disturbance robustness index and the channel safety margin index, input them into the pre-defined obstacle avoidance correction model to get the obstacle avoidance correction coefficient, combine it with the basic risk score to get the decision score, compare the threshold to decide to pass or avoid;

[0011] Step S5: When avoiding, adjust the pose into the safety corridor according to the obstacle normal micro-displacement; record the decision and the quantitative field to generate the obstacle avoidance event vector and issue the control; if passing, keep the current flight path.

[0012] Preferably, the safety boundary grid size is set according to the body diagonal size, attitude control margin and ranging uncertainty, each corridor segment in the grid is attached with minimum range, maximum range and passable width initial value for waypoint constraint and entry verification; obstacle inflation and connectivity screening are performed in the safety boundary grid according to the minimum safety distance to generate a job corridor set that meets the connectivity and safety distance, and the observation angle baseline and ground resolution baseline are fixed in the key view cone.

[0013] Preferably, obstacle density and occlusion rate are used as refresh triggers, when the obstacle density exceeds the entry threshold or its change rate exceeds the stability threshold, or when the occlusion rate continuously rises in a short time window, the job corridor set and the key view cone set are quickly refreshed, and a time index is attached to the refreshed result for subsequent use of the latest valid marked data.

[0014] Preferably, based on the initial flight path and the waypoint observation constraint set, the hardware trigger list is adaptively set according to the channel curvature and the expected observation angle change; when the occlusion removal ratio exceeds the threshold, the source corridor center line is slightly moved and the expected observation angle and ground resolution value are recalculated; during collection, quality priori evaluation is performed on each frame and weight reduction or additional trigger is implemented, and after collection, infrared radiation compensation is performed to generate actual temperature map sequence combined with environmental and optical conditions.

[0015] Preferably, the step S3 generates a multi-source frame set based on pose-tagged frame-level registration, removes abnormal frames according to attitude stability and position drift, and constructs a cross-modal consistency mask according to temperature gradient and edge intensity after pixel-level registration; when projecting the registration result to a safety boundary grid, the mask and ranging consistency are jointly weighted to update the obstacle occupancy probability map and the free space map.

[0016] Preferably, the step S3 extracts a connected high-occupancy area on the obstacle occupancy probability map and generates an obstacle boundary set by verifying the continuity of the visible light texture edge; the time difference is obtained by performing time difference on the layers corrected by the pose at adjacent time points to obtain the obstacle motion trend, and the passing width, the nearest obstacle distance and their change rates are calculated on the passing belt, so as to implement local weight increase and decrease and temporary closure on the operation corridor, and output the observation angle time function.

[0017] Preferably, in the step S4, the observation angle time function and the channel state are taken as core inputs, the observation angle, the observation angle change rate and the edge angle sensitivity of the obstacle boundary set neighborhood are extracted at the same time index, the angle velocity stability criterion and the edge sensitivity insensitivity criterion are unified and normalized by using the robust statistical caliber of the sliding time window to construct the visual angle disturbance robust index; and the passing width, the nearest obstacle distance, the approaching speed positive part and the corridor occupancy probability mean are jointly evaluated based on the local passing corridor, the obstacle occupancy probability map and the obstacle motion trend to form the channel safety margin index.

[0018] Preferably, after executing the avoidance decision, the local passing corridor and the safety boundary grid are taken as the feasible region, the target corridor is determined according to the channel safety margin index, the normal priority or tangential bypass strategy is selected according to the obstacle motion trend, the safety step length is calculated and the normal micro-displacement vector is generated, the connectivity verification is completed and the step is shortened gradually until the minimum safety distance and the passing width threshold are met, the observation geometry is kept and the ground resolution baseline constraint is applied, the evidence-friendly priority is triggered when the judgment score approaches the obstacle avoidance threshold parameter, and the micro-displacement vector and the effective step length are output.

[0019] Preferably, the vector alignment control is implemented in the process of entering the target corridor, the forward velocity of the track tangential vector is kept continuous, the lateral deviation is completed along the normal micro-displacement vector, and the gimbal observation vector is limited within the observation angle baseline range; after entering, the three standards of connectivity, minimum safety distance and cross-sectional passing width are verified, and when any one of them is not met, the suboptimal corridor sequence is retreated and the track tangential vector amplitude is reduced according to the channel safety margin index, the obstacle avoidance event vector is uniformly generated and delivered to the airborne controller and the ground station.

[0020] To achieve the above object, the application further provides the following technical scheme: an unmanned aerial vehicle obstacle avoidance device for realizing the unmanned aerial vehicle obstacle avoidance method, comprising:

[0021] Semantic modeling layer: used for fusing tower structure contour and field survey data to generate truss semantic topology when the unmanned aerial vehicle enters the truss operation area or detects that the obstacle density is increased, projecting the passable operation corridor on the safety boundary grid according to the minimum safety distance and labeling the key view frustum; output the operation corridor set and the safety boundary grid for calling by the lower layer;

[0022] Trajectory acquisition layer: used for receiving the operation corridor set and the safety boundary grid, solving the continuous trajectory according to the coverage integrity and the safety distance and curve smoothing, assigning the expected observation angle and the ground resolution value to each waypoint, generating the hardware trigger list and synchronously collecting the multi-source perception sequence; output the initial trajectory, the synchronous image data and the pose label for use by the lower layer;

[0023] Fusion estimation layer: used for pixel-level registration of infrared and visible light under the pose prior constraint and projecting to the safety boundary grid, generating the obstacle occupancy probability map and the free space map in combination with the ranging points, extracting the obstacle boundary set and the obstacle motion trend, updating the local passable corridor and outputting the observation angle time function for reference by the upper layer;

[0024] Risk decision layer: used for calculating the view angle disturbance robustness index and the channel safety margin index when the obstacle boundary set is updated, inputting the predefined obstacle avoidance correction model to obtain the obstacle avoidance correction coefficient, combining the basic risk score to form the decision score and comparing the decision score with the obstacle avoidance threshold parameter, outputting the pass or avoidance decision and the required quantitative basis;

[0025] Execution record layer: used for executing the micro-displacement reset along the normal line of the target obstacle and adjusting the flight height and the heading under the avoidance decision, so that the machine body enters the nearest safe passable corridor; maintaining the current initial trajectory under the pass decision; simultaneously recording the decision state and the key quantitative field to generate the obstacle avoidance event vector and issue the control instruction for control and archiving.

[0026] Technical effects and advantages of the present application:

[0027] The present application generates the operation corridor set, the safety boundary grid and the key view frustum, solves the initial trajectory and the waypoint observation constraint set based on the constraints, issues the hardware trigger list to obtain the synchronous image data and the pose label, completes the pixel-level registration based on the pose prior, and obtains the obstacle occupancy probability map, the free space map, the obstacle boundary set and the obstacle motion trend on the safety boundary grid, simultaneously outputs the observation angle time function and the local passable corridor, realizes the collaborative operation of planning, acquisition and environment modeling in the same coordinate and time scale, takes into account the minimum safety distance, the observation angle baseline and the ground resolution baseline, significantly improves the coverage, the discrimination stability and the data consistency, and effectively solves the problems of dispersed geometric constraints, disjointed imaging and occupancy estimation drift in the traditional method.

[0028] The application constructs a visual angle disturbance robustness index by constructing an edge response as a function of an observation angle over time, and constructs a channel safety margin index by constructing a local traffic corridor, an obstacle boundary set, an obstacle movement trend, and a mean corridor occupancy probability, both of which are input into a pre-defined obstacle avoidance correction model to obtain an obstacle avoidance correction coefficient, which is then combined with a basic risk score to synthesize a judgment score and compared with an obstacle avoidance threshold parameter; a normal micro-displacement vector is solved and three-vector alignment control is implemented to enter the nearest safe traffic corridor, while maintaining the baseline values of the observation angle and the GSD, and generating an obstacle avoidance event vector, realizing adaptive obstacle avoidance with safety and evidence, consistent with adjudication and execution, effectively solving the problems of frequent misjudgment, secondary detour, and evidence quality decline in high-reflective truss scenes. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A flowchart of the unmanned aerial vehicle obstacle avoidance method of the application. DETAILED DESCRIPTION

[0030] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0031] At the same time, it should be understood that the sizes of the various parts shown in the drawings are not drawn in accordance with the actual proportional relationship for the sake of description.

[0032] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way limiting of the application or its applications or uses.

[0033] Techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered as part of the specification in appropriate circumstances.

[0034] Reference Figure 1 A flowchart of the unmanned aerial vehicle obstacle avoidance method of the application, the application provides a kind of unmanned aerial vehicle obstacle avoidance method as shown in Figure 1 The method comprises the following steps:

[0035] Step S1: fuse the profile and the mapped semantic topology; project the work corridor according to the minimum safety distance and label the key view cone, output the work corridor set and the safety boundary grid;

[0036] Step S2: solve and smooth the continuous flight path based on the work corridor set and the safety boundary grid; assign the desired observation angle and ground resolution to the flight point, and simultaneously collect and generate the initial flight path and image;

[0037] Step S3: Register infrared and visible light with pose prior, project to safety boundary grid, generate occupancy map and free space, extract obstacle boundary and motion trend, update local passable corridor and output observation angle time function;

[0038] Step S4: When obstacle boundary set is updated, calculate visual angle disturbance robustness index and channel safety margin index, input them into predefined obstacle avoidance correction model to get obstacle avoidance correction coefficient, combine it with basic risk score to get decision score, compare with threshold to decide to pass or avoid;

[0039] Step S5: When avoiding, adjust pose to enter safety corridor with small displacement along normal of obstacle; record decision and generate obstacle avoidance event vector with quantified fields and send control; if passing, keep current track.

[0040] Explanatory notes, the core purpose of step S1 is:

[0041] When the unmanned aerial vehicle enters the truss working area or detects an increase in obstacle density, the space structure and working safety are integrated modeled for semanticization, passability, and observability before track solving and imaging collection, forming three types of basic data that can directly drive planning and collection: working corridor set, safety boundary grid, and key view cone set, and solidifying key constraints such as minimum safety distance, observation angle baseline, and ground resolution baseline, providing a unified spatial reference system and observation geometric constraint for the integrated execution of planning, collection, and compensation in step S2, and preparing the observation angle time sequence and visible light occlusion information needed for the visual angle disturbance robustness index and thermal texture intersection sharpness index in step S4.

[0042] Further, the specific implementation of step S1 is:

[0043] Sub-step S101: With structural profile data and field survey data as input, complete coordinate reference alignment and build truss semantic topology, inject weld toe, node plate, and end seal plate into weld toe weight, node plate weight, and end seal plate weight, respectively, and generate surface state label and visibility label based on cross-verification of structural profile data and field survey data, so that subsequent spatial constraints and observation priorities converge to high-risk components at the semantic level, and output truss semantic topology with semantic and visibility attributes;

[0044] Sub-step S102: Under the constraint of minimum safety distance, perform spatial expansion on the truss semantic topology as an obstacle set and remove the obstacle expansion body, construct a safety boundary grid based on the body diagonal size, attitude control margin, and ranging uncertainty, perform connectivity analysis to form a working corridor set, generate a key view cone set along the corridor center line and solidify the observation angle baseline, minimum view distance, maximum view distance, and ground resolution baseline, while setting the occlusion removal proportion, output the working corridor set and key view cone set carrying the source corridor identifier and high-risk category identifier.

[0045] It is explained that the source corridor identifier refers to an identifier marking the processing path of the data from the source to the fusion layer, which is formed by concatenating the source identifier, processing node sequence, caliber version signature, and time interval marker, and is used for tracing and auditing.

[0046] Further, the obstacle density and the occlusion rate are used as refresh triggers. When the obstacle density exceeds the entry threshold or its change rate exceeds the stability threshold, or when the occlusion rate continuously rises within a short time window, the set of work corridors and the set of key view cones are quickly refreshed, and the refresh results are indexed by time for subsequent use of the latest valid marked data.

[0047] Further, the size and level of the safety boundary grid are set according to the field wind disturbance and positioning accuracy, so that the grid step size is not less than the allowed displacement accuracy of attitude control, and the initial value of the passable width and the source corridor identifier are preset for each corridor segment for subsequent entry verification and rollback switching.

[0048] Further, the space band marked as high priority by the weld toe weight or the node plate weight is given a controlled relaxation in corridor width and curvature constraints to ensure that the key component's view angle is improved under the condition of meeting the minimum safety distance, providing a direct basis for the assignment of the desired observation angle and ground resolution value of the step S2.

[0049] It is explained that the core goal of step S2 is to convert the set of work corridors, the set of key view cones, and the safety boundary grid output by step S1 into an initial flight path that can be directly executed and synchronous image data, and to assign a desired observation angle and ground resolution value to each waypoint while solving the continuous flight path, and then generate a hardware trigger list and a pose label, so that subsequent pixel-level registration and obstacle occupancy analysis can be performed under the unified time scale, unified geometric assumption, and unified observation constraint, avoiding data mismatch and evidence quality fluctuations caused by the disconnection of planning and acquisition.

[0050] Further, the specific implementation of step S2 is:

[0051] Sub-step S201: Inject high-risk category identifiers in the key view cone into the path cost with the set of work corridors and the safety boundary grid as hard constraints, first eliminate the grids that will cause insufficient minimum safety distance, and then solve the continuous flight path and curve smoothing on the remaining passable bands; During the generation process, read the line-of-sight accessibility, occlusion removal ratio, and curvature limit of each channel segment synchronously, select the channel segment that can provide a stable observation angle and sufficient imaging scale, output an initial flight path containing a path sequence and a channel segment identifier, and solidify the source corridor identifier, source view cone identifier, and time index for each path segment as a determination input for subsequent waypoint constraints and trigger timing;

[0052] Sub-step S202: At each candidate waypoint of the initial flight path, read the corresponding key view cone parameters and minimum safety distance, calculate the expected observation angle according to the observation angle baseline, minimum view distance and maximum view distance, and convert the ground resolution value with the lens focal length calibration and target view distance; when the occlusion rejection ratio of the waypoint exceeds the threshold, perform a slight translation along the source corridor center line to find a new position with less occlusion and recalculate the expected observation angle and ground resolution value; constrain the expected observation angle change amplitude of adjacent waypoints for turning waypoints to avoid excessive pan-tilt angular velocity; finally form a waypoint observation constraint set, record the source corridor identifier, source view cone identifier, expected observation angle, ground resolution value and path sequence number, which are used for unified geometry target of synchronous acquisition and pose labeling;

[0053] Explanation and description, the source view cone identifier refers to the identifier defining the data range accessible in this narrative, derived from dimension limit, time window limit and caliber version signature, used for constraint retrieval and generation;

[0054] Sub-step S203: generate hardware trigger list and acquisition time table based on the initial flight path and the waypoint observation constraint set; the hardware trigger list specifies the trigger sequence, trigger interval and pre-focusing or integration time of infrared and visible light for each waypoint, and the acquisition time table records the planned time and planned pose of each trigger to bind the pose label at the time of acquisition; when the channel curvature increases or the expected observation angle changes rapidly, the trigger density is increased adaptively, and when the segment is straight or changes slowly, the trigger density is reduced, so that the acquisition resources are concentrated to cover the unstable geometry area; the two tables are sent to the on-board controller and imaging equipment and the confirmation marks are written back to ensure that the synchronous image data and pose labels correspond one by one and carry the source corridor identifier, source view cone identifier and time index;

[0055] Sub-step S204: execute acquisition according to the hardware trigger list, the device side binds the pose label and the actual trigger time to each frame of image in real time, and performs rapid quality prior evaluation according to the waypoint observation constraint set, if exposure saturation, motion blur or expected observation angle deviation is detected, the frame is down-weighted or additional trigger is added at adjacent time, to improve the available proportion of synchronous image data; after acquisition, calibrate the internal and external parameters of the infrared camera and non-uniformity correction based on the infrared raw image, and implement infrared radiation compensation combining environmental temperature, relative humidity, ranging and observation angle, lens transmittance and sky background radiation, to generate actual temperature map sequence; finally output the initial flight path, waypoint observation constraint set, synchronous image data, pose label and actual temperature map sequence, wherein the image and label correspond to the same index number, for pixel-level registration, grid projection and obstacle occupancy probability map generation in step S3.

[0056] It is explained that the core goal of step S3 is to realize the same position fusion of the synchronized image data output by step S2 and the pose label on the unified space carrier of the safety boundary grid, and to obtain the obstacle occupancy probability graph and the free space graph by combining the ranging point calculation, on the basis of which the obstacle boundary set and the obstacle motion trend are extracted and the local traffic corridor is updated accordingly; the above data will directly provide input for the judgment calculation of step S4, in which the obstacle boundary set and the obstacle motion trend will participate in the calculation of the channel safety margin index, and the observation angle time sequence derived from the pose label will participate in the calculation of the visual angle disturbance robustness index, all names and meanings are consistent with and uniquely mapped to step S1 and step S2.

[0057] Further, the specific implementation of step S3 is:

[0058] Sub-step S301: After receiving the synchronized image data, the pose label and the actual temperature graph sequence, frame-level pairing is completed according to the time sequence and index number of the pose label, each frame of the actual temperature graph sequence frame is aligned in time with the visible light image at the corresponding time, and is bound with the pose label and the ranging point of the same index, to obtain a multi-source frame set containing the actual temperature graph sequence frame, the visible light image, the pose label and the ranging point, and carrying the source corridor identifier, the source view cone identifier and the time index; the attitude stability and the position drift in the pose label are used as consistency check, and frames with low attitude stability or abnormal position drift are removed, to ensure that the data entering the subsequent processing can be reliably explained in time and geometry, and to avoid the estimation deviation caused by mismatch and distortion;

[0059] Sub-step S302: In the multi-source frame set, the actual temperature graph sequence frame and the visible light image of the same frame are implemented pixel-level registration based on the external parameter relationship provided by the pose label as a geometric prior, so that both describe the temperature information and texture information of the same space point in the same pixel coordinate; in order to suppress the cross-modal inconsistency caused by strong reflection of metal, after registration, a cross-modal consistency mask is constructed according to the spatial coincidence and direction consistency of temperature gradient strength and edge strength, which is used as the pixel reliability weight of subsequent space projection, to ensure that the thermal anomaly and the structural boundary are emphasized at the same spatial position, and to improve the positioning accuracy of obstacle boundary extraction; the above data objects inherit the source corridor identifier, the source view cone identifier and the time index;

[0060] Sub-step S303: Taking the safety boundary grid as the target space grid, projecting the registration result obtained in sub-step S302 to the grid coordinate system under the constraints of the pose label and the ranging point to form temperature grids and visible light texture grids with the same resolution, and taking the cross-modal consistency mask as the pixel credibility weight; estimating the obstacle occupancy probability of each grid cell according to the landing point density and consistency of the ranging point in the grid, marking the cells not covered by the landing point or the prominent feature as the free space candidate, and generating the obstacle occupancy probability map and the free space map through connectivity correction and boundary adhesion correction; in this process, the pose stability is used to suppress the projection error, the ranging consistency is used to weaken the false occupancy caused by single-frame texture, and the cross-modal consistency mask is used to improve the common contribution degree of the boundary area, all of which work together to ensure convergence and stability in the complex metal structure environment; all products inherit the source corridor identification, source view cone identification and time index;

[0061] Sub-step S304: extracting the connected high-occupancy area on the obstacle occupancy probability map as the obstacle candidate according to the probability threshold, and verifying its geometric continuity on the visible light texture grid to generate the obstacle boundary set; then, performing time difference on the obstacle occupancy probability map of the adjacent time, aligning the grid coordinates of the two time according to the displacement and pose change of the pose label, and calculating the obstacle motion trend (including displacement direction and speed level); at the same time, on the traffic lane of the safety boundary grid, the traffic width and the nearest obstacle distance of each work corridor section are calculated, the time variation rate of the distance is paired with the obstacle motion trend, and the traffic lane with high shrinkage risk is marked, providing input for the channel safety margin index of step S4; the products inherit the source corridor identification, source view cone identification and time index;

[0062] Sub-step S305: based on the obstacle occupancy probability map, the free space map, the obstacle boundary set and the obstacle motion trend, performing local update on the previous work corridor on the safety boundary grid, down-weighting or temporarily closing the corridor section covered by high occupancy probability for a long time or continuously approached by obstacles, and up-weighting or expanding the corridor section stably maintained by the free space map, outputting the updated local traffic corridor and attaching the nearest obstacle distance, the nearest obstacle motion direction and the probability confidence of each section; at the same time, the observation angle information of each frame is extracted from the pose label and organized as an observation angle time function according to the time index, which is directly referenced when calculating the visual angle disturbance robustness index in step S4; all outputs continue to carry the source corridor identification, the source view cone identification and the time index, keeping the consistency of the step name, the source and the use range across steps.

[0063] The core purpose of step S4 is explained as follows: on the premise that the obstacle occupancy probability map, free space map, obstacle boundary set, obstacle motion trend and local traffic corridor have been output in step S3, two types of criteria that are interpretable, calibratable and closely coupled with observation geometry, i.e., view angle disturbance robustness index and channel safety margin index, are constructed, and both are input into a predefined obstacle avoidance correction model to generate an obstacle avoidance correction coefficient, which is then combined with the basic risk score calculated based on the data of step S3 to form a decision score; finally, a one-time decision is made by using an obstacle avoidance threshold parameter, and the output is a decision of passing or avoiding and a quantitative basis that can be directly executed by step S5.

[0064] Further, the specific implementation of step S4 is as follows:

[0065] Sub-step 401: using observation angle time function and channel state as core inputs, the observation angle, observation angle change rate and edge sensitivity of the obstacle boundary set neighborhood at the same time index are extracted, the angle velocity stability criterion and edge sensitivity insensitivity criterion are unified by using the robust statistical criterion of the sliding time window, the view angle disturbance robustness index is constructed, and the passing width, the nearest obstacle distance, the positive part of the approaching velocity and the average corridor occupancy probability are jointly evaluated based on the local traffic corridor, the obstacle occupancy probability map and the obstacle motion trend to form the channel safety margin index; the two indexes carry the source corridor identifier and the time index, are input into the predefined obstacle avoidance correction model to complete the monotonic saturated mapping, and output the obstacle avoidance correction coefficient, the involved weight and shape factor are offline calibration read-only;

[0066] Sub-step 402: the basic risk score is calculated based on the obstacle occupancy probability map, the obstacle boundary set, the obstacle motion trend and the local traffic corridor, and the involved terms include the average path occupancy probability, the nearest obstacle distance, the positive part of the approaching velocity and the passing width change rate; the decision score is obtained by combining the obstacle avoidance correction coefficient and the basic risk score, and compared with the obstacle avoidance threshold parameter to make a decision of passing or avoiding; in order to suppress boundary jitter, a hysteresis band or a minimum holding time is set, and the source corridor identifier, the time index and the adopted index value and component value are written in the decision record for the execution layer and the traceability, and the only adjustable item in the running period is the obstacle avoidance threshold parameter.

[0067] Further, the observation angle time function, the passing width, the nearest obstacle distance, the positive part of the approaching velocity, the average corridor occupancy probability and the average path occupancy probability are all provided by step S3 in real time and carry the source corridor identifier and the time index, the time window and the sampling step of the two indexes and the basic risk score are consistent with step S3, and the lower limit truncation and saturation clipping are implemented for the minimum distance and the minimum velocity to ensure the numerical stability.

[0068] Further, the channel safety margin index and the base risk score component avoid repeated weighting, preferentially entering the distance-related quantities into the channel safety margin index, and the path occupancy probability mean and the passage width change rate into the base risk score, and limiting the index weight range in the predefined obstacle avoidance correction model, so that the model bears the credibility correction instead of repeated amplification, and ensures the monotonicity and interpretability of the judgment score.

[0069] Further, the view angle disturbance robustness index is calculated as follows:

[0070] ;

[0071] wherein the view angle disturbance robustness index is dimensionless, and a larger value indicates that the observation geometry and imaging content are more robust; the observation angle time function is in degrees, and is obtained by differencing adjacent time points and dividing by the time interval; the edge response function is in pixel intensity, and is defined in the neighborhood of the obstacle boundary set; the edge angular sensitivity is in pixel intensity per degree; The robust operator takes the ninth decile of the input sequence, and outputs the same dimension as the input; represents a saturation mapping, which monotonically compresses a non-negative quantity to zero to one;

[0072] Further, the channel safety margin index is obtained as follows:

[0073] ;

[0074] wherein the channel safety margin index is dimensionless, and a larger value indicates that the channel is safer; represents the passage width, in meters, which is the passable width of the current corridor cross section; represents the passage width change rate, in meters per second, and a negative value indicates contraction; represents the nearest obstacle distance, in meters, which is calculated from the obstacle boundary set and the corridor center line; represents the positive part of the approaching velocity, in meters per second, which is the velocity of the obstacle pointing to the corridor along the normal, and is zero for non-positive values; represents the corridor occupancy probability mean, which is dimensionless, and is the average value of the forward short-time path; and are scale constants, which are dimensionless, and are offline calibrated and read-only.

[0075] Further, the obstacle avoidance correction coefficient is obtained as follows:

[0076] ;

[0077] wherein the obstacle avoidance correction factor is dimensionless, used for credibility correction of the risk score; is a weight coefficient, dimensionless, taking value between zero and one, offline calibrated and read-only; is a shape factor, positive, controlling the smoothness and saturation of the aggregate surface, offline calibrated and read-only; the shape factor is obtained by fixing other read-only parameters (such as weight coefficient, scale constant, risk weight, speed upper limit, width variation upper limit, distance regularization constant, observation angle time function sampling frequency) on the historical flight data of representative scenarios, and performing grid or Bayesian search on the candidate , taking the weighted sum of miss avoidance rate, miss avoidance rate, decision boundary jitter, and passing efficiency as the objective function, and selecting the solution with the best comprehensive performance and numerical stability at extreme values; meanwhile, the constraint and falling into the stable interval determined by platform measurement, and verifying no abnormal amplification with threshold undercutoff and saturation clipping; finally, the selected is written into the model configuration, which is not updated adaptively during runtime, but only iterated with version release.

[0078] Further, the basic risk score is obtained by:

[0079] ;

[0080] wherein, basic risk score, dimensionless, with higher value indicating higher immediate risk; represents the mean path occupancy probability, dimensionless, from step S3; represents the distance to the nearest obstacle, unit: meter, defined as before; represents the distance regularization constant, unit: meter, small positive number, to avoid divergence caused by extremely small distance; represents the positive part of the approaching speed, unit: meter per second, defined as before; represents the speed upper limit, unit: meter per second, set by platform and specification; represents the passing width variation rate, unit: meter per second, defined as before; represents the width variation upper limit, unit: meter per second, set by planning and control capability; is a weight coefficient for each term, dimensionless, offline calibrated and read-only, and the sum is one or close to one.

[0081] Further, the decision score is represented by the product of the basic risk score and the obstacle avoidance correction factor Non-dimensional, as the only quantitative basis for entering the execution layer; single-point decision is implemented based on the comparison between the decision score and the obstacle avoidance threshold parameter: when the decision score is higher than the threshold, immediately generate an "avoidance" decision, lock the target obstacle and output the target corridor candidate, normal direction, nearest obstacle distance, approaching speed positive part and passing width, etc. execution required parameters; when the decision score is lower than the threshold, generate a "pass" decision, maintain the current initial flight path and the established collection plan; in order to suppress boundary jitter, introduce a hysteresis band or minimum holding time, when the decision score falls into the hysteresis band, keep the last time decision unchanged.

[0082] Explain, the core purpose of step S5: based on the explicit decision of passing or avoiding output in step S4, the decision score, obstacle avoidance correction coefficient, channel safety margin index, obstacle boundary set and local passing corridor are translated into control quantities that can be directly issued for avoidance cases, using micro displacement along the normal of the target obstacle, entering the nearest safe passing corridor, maintaining the integrated execution strategy of observation geometry, while building an obstacle avoidance event vector containing the full quantity basis field to support recording, accountability and subsequent analysis; in the passing case, the continuous execution of the initial flight path is maintained and the key state is recorded in the same data structure, so that the control link and the decision link maintain unified terminology and unique mapping, avoiding any ambiguity and conflict;

[0083] Further, the specific implementation of the step S5 is:

[0084] Sub-step S501: read the decision, decision score, channel safety margin index and obstacle motion trend, locate the threat boundary segment in the feasible region of local passing corridor and safety boundary grid, select the candidate corridor segment that meets the minimum safety distance, passing width threshold and low average forward corridor occupancy probability according to the nearest obstacle distance and normal, and determine the target corridor according to the channel safety margin index sorting; according to the obstacle motion trend, execute the joint strategy of normal priority or tangential bypass, and solidify the source corridor identifier and the target corridor identifier;

[0085] Sub-step S502: calculate the safety step length according to the nearest obstacle distance, passing width and minimum safety distance, solve the normal micro displacement vector along the boundary normal and do connectivity check, if not satisfied, shorten the step until safe; at the same time, impose observation geometry preservation and ground resolution baseline constraints, enable evidence-friendly priority when the decision score approaches the obstacle avoidance threshold parameter, output the micro displacement vector and effective step length.

[0086] Sub-step S503: align the target with the flight path tangential vector, normal micro displacement vector and gimbal observation vector, maintain the forward speed continuity, complete the lateral deviation and make the body trajectory fall into the target corridor connectivity band, at the same time limit the observation angle time function within the observation angle baseline range; output the entering state and real-time image availability flag;

[0087] Sub-step S504: Verify the entry according to the target corridor connectivity, minimum safety distance, and cross-section passing width. If any of the three standards is not met, return to the sub-optimal corridor sequence sorted by the channel safety margin index, and reuse the safety step length and observation geometry constraints. Limit the number of returns and reduce the magnitude of the flight path tangent vector. Output the entry success or failure flag and the adopted corridor identifier.

[0088] Sub-step S505: Generate an obstacle avoidance event vector based on the avoidance and evasion results, and synchronize it to the onboard controller and ground station for control and archiving.

[0089] An unmanned aerial vehicle obstacle avoidance device is also proposed to implement the above unmanned aerial vehicle obstacle avoidance method, comprising:

[0090] Semantic modeling layer: used to generate a truss semantic topology by fusing the tower structure profile and field surveying data when the unmanned aerial vehicle enters the truss working area or detects an increase in obstacle density. Project the passable working corridor on the safety boundary grid according to the minimum safety distance and label the key view cone. Output the working corridor set and safety boundary grid for lower layer calling;

[0091] Flight path collection layer: used to receive the working corridor set and safety boundary grid, solve the continuous flight path according to the coverage integrity and safety distance, and curve smooth. Assign the expected observation angle and ground resolution value to each waypoint, generate a hardware trigger list, and synchronize the multi-source perception sequence. Output the initial flight path, synchronized image data, and pose label for lower layer use;

[0092] Fusion estimation layer: used to perform pixel-level registration and projection to the safety boundary grid under the pose prior constraint for infrared and visible light. Generate the obstacle occupancy probability map and free space map in combination with the ranging points. Extract the obstacle boundary set and obstacle motion trend, update the local passing corridor, and output the observation angle time function for upper layer judgment reference;

[0093] Risk decision layer: used to calculate the view angle disturbance robustness index and channel safety margin index when the obstacle boundary set is updated. Input the predefined obstacle avoidance correction model to obtain the obstacle avoidance correction coefficient. Combine it with the basic risk score to form the judgment score, and compare it with the obstacle avoidance threshold parameter. Output the pass or avoidance decision and the required quantitative basis;

[0094] Execution record layer: used to perform micro-displacement reset along the normal line of the target obstacle and adjust the flight height and heading when the avoidance decision is made, so that the aircraft enters the nearest safe passing corridor. Maintain the current initial flight path when the pass decision is made. Record the decision state and key quantitative fields to generate an obstacle avoidance event vector and issue control instructions for control and archiving.

[0095] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the scope of protection of the present application.

Claims

1. A method for obstacle avoidance by unmanned aerial vehicles (UAVs), characterized in that, include: Step S1: When the UAV enters the truss operation area or detects an increase in obstacle density, the truss semantic topology is generated by fusing the tower crane structure outline with the on-site survey data. The passable operation corridor is projected according to the minimum safe distance and key field cones are marked. The operation corridor set and safety boundary grid are output. Step S2: When the set of work corridors and safety boundary grids are received, solve the continuous track and smooth the curve, assign the desired observation angle and ground resolution value to each waypoint, and use hardware to trigger the synchronous acquisition of multi-source sensing sequences to generate the initial track and synchronous image data. Step S3: Register infrared and visible light with pose priors and project them onto the safety boundary grid; generate occupancy map and free space, extract obstacle boundaries and motion trends, update local passageways and output the observation angle time function; Step S4: When updating the obstacle boundary set, calculate the viewpoint disturbance robustness index and the channel safety margin index, input the two into the predefined obstacle avoidance correction model to obtain the obstacle avoidance correction coefficient, combine them with the basic risk score to form a judgment score, and compare the threshold to decide whether to pass or avoid. Step S5: When avoiding an obstacle, make a small displacement in the obstacle normal and adjust your attitude to enter the safe corridor; record the decision and quantization fields to generate an obstacle avoidance event vector and issue control; if you pass through, maintain the current track.

2. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, The size of the safety boundary grid is set based on the diagonal size of the aircraft, attitude control margin, and ranging uncertainty. Each corridor segment within the grid is accompanied by initial values ​​for minimum line of sight, maximum line of sight, and passable width, which are used for waypoint constraints and entry verification. Obstacle expansion and connectivity screening are performed within the safety boundary grid according to the minimum safe distance to generate a set of operational corridors that meet connectivity and safe distance requirements. The observation angle baseline and ground resolution baseline are then fixed in the key field of view cone.

3. The obstacle avoidance method for unmanned aerial vehicles according to claim 2, characterized in that, Obstacle density and occlusion rate are used as refresh trigger values. When the obstacle density exceeds the entry threshold or its rate of change exceeds the stability threshold, or when the occlusion rate rises continuously within a short time window, the work corridor set and the key view cone set are quickly refreshed, and a time index is added to the refresh result for subsequent use of the latest valid labeled data.

4. The obstacle avoidance method for unmanned aerial vehicles according to claim 3, characterized in that, Based on the initial track and waypoint observation constraint set, the hardware trigger list is adaptively set according to the changes in channel curvature and desired observation angle; when the occlusion removal ratio exceeds the threshold, it is slightly moved along the center line of the source corridor and the desired observation angle and ground resolution value are recalculated; during acquisition, a quality prior assessment is performed on each frame and weight reduction or additional triggering is implemented; after acquisition, infrared radiation compensation is performed in combination with environmental and optical conditions to generate the actual temperature map sequence.

5. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, Step S3 generates a multi-source frame set based on frame-level pairing of pose labels, removes abnormal frames by attitude stability and position drift, and constructs a cross-modal consistency mask based on temperature gradient and edge intensity after pixel-level registration. When projecting the registration result onto the safety boundary grid, the obstacle occupancy probability map and free space map are jointly updated by weighting the mask and ranging consistency.

6. The obstacle avoidance method for unmanned aerial vehicles according to claim 5, characterized in that, Step S3 extracts connected high-occupancy regions from the obstacle occupancy probability map and generates an obstacle boundary set by verifying the continuity of visible light texture edges; performs time difference analysis on the pose-corrected layers at adjacent times to obtain the obstacle movement trend; at the same time, calculates the passage width and the distance to the nearest obstacle and their rate of change on the passageway; based on this, implements local weighting and temporary closure of the work corridor, and outputs the observation angle time function.

7. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, In step S4, the observation angle time function and channel state are used as the core inputs. Under the same time index, the observation angle, the rate of change of the observation angle, and the edge pair angle sensitivity of the neighborhood of the obstacle boundary set are extracted. The robust statistical caliber of sliding time window is used to unify the angular velocity stability criterion and the edge sensitivity insensitivity criterion to construct the viewpoint disturbance robustness index. At the same time, based on the local passage corridor, the obstacle occupancy probability map, and the obstacle movement trend, the passage width, the nearest obstacle distance, the positive part of the approach speed, and the mean of the corridor occupancy probability are jointly evaluated to form the channel safety margin index.

8. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, After executing the obstacle avoidance decision, the feasible region is the local passage corridor and the safety boundary grid. The target corridor is determined by sorting according to the channel safety margin index. The normal priority or tangential detour strategy is selected according to the obstacle movement trend. The safe step length is calculated and the normal micro displacement vector is generated. The connectivity verification and step-by-step shortening are completed until the minimum safe distance and passage width threshold are met. At the same time, the observation geometry preservation and ground resolution baseline constraints are applied. When the judgment score is close to the obstacle avoidance threshold parameter, the evidence collection friendly priority is triggered, and the micro displacement vector and effective step length are output.

9. The obstacle avoidance method for unmanned aerial vehicles according to claim 8, characterized in that, During the entry into the target corridor, vector alignment control is implemented to maintain the continuous forward velocity of the track tangential vector, complete the lateral offset along the normal micro-displacement vector, and limit the gimbal observation vector within the observation angle baseline range. After entry, the system is verified according to three criteria: connectivity, minimum safe distance, and cross-sectional passage width. If any criterion is not met, the system retreats according to the suboptimal corridor sequence sorted by the channel safety margin index and reduces the amplitude of the track tangential vector. A unified obstacle avoidance event vector is generated and sent to the airborne controller and ground station.

10. A drone obstacle avoidance device for implementing the drone obstacle avoidance method according to any one of claims 1-9, characterized in that, include: Semantic Modeling Layer: When a drone enters the truss operation area or detects an increase in obstacle density, it integrates the tower crane structure outline with on-site survey data to generate the truss semantic topology, projects passable operation corridors onto the safety boundary grid according to the minimum safe distance, and marks key viewpoint cones; it outputs a set of operation corridors and safety boundary grids for use by lower layers; Track acquisition layer: It is used to receive the set of work corridors and safety boundary grids, solve the continuous track and smooth the curve according to the coverage integrity and safety distance, assign the expected observation angle and ground resolution value to each waypoint, generate a hardware trigger list and synchronously acquire multi-source sensing sequences; output the initial track and synchronous image data and pose label for use by the lower layer. Fusion estimation layer: Under pose prior constraints, infrared and visible light are pixel-level registered and projected onto the safety boundary grid. Combined with the ranging points, obstacle occupancy probability map and free space map are generated. The obstacle boundary set and obstacle movement trend are extracted. The local passageway is updated and the observation angle time function is output for the upper layer to use in the decision. Risk adjudication layer: When updating the obstacle boundary set, it calculates the viewpoint disturbance robustness index and the channel safety margin index, inputs the predefined obstacle avoidance correction model to obtain the obstacle avoidance correction coefficient, combines it with the basic risk score to form a judgment score, compares it with the obstacle avoidance threshold parameter, and outputs the pass or avoidance decision and the required quantitative basis. Execution Record Layer: Used to perform micro-displacement reset along the normal of the target obstacle and adjust the flight altitude and heading under the avoidance decision to bring the aircraft into the nearest safe passage corridor; maintain the current initial track under the passage decision; and simultaneously record the decision status and key quantification fields to generate obstacle avoidance event vectors and issue control commands for control and archiving.

Images