Unmanned aerial vehicle autonomous obstacle avoidance method and device and storage medium

Abstract

The invention relates to the technical field of unmanned aerial vehicle control, and discloses an unmanned aerial vehicle autonomous obstacle avoidance method and device and a storage medium, and the unmanned aerial vehicle autonomous obstacle avoidance method comprises the following steps: S101, generating an echo over-dispersion curvature; step S102, generating a free domain contraction field; step S103, generating a risk occupation mapping; step S104, constructing a distance field with symbols; step S105, selecting a yaw angle speed and a speed scaling factor; and S106, generating a final obstacle avoidance control instruction. According to the method, abnormal features of fine line bodies and other targets prone to missing detection are captured by extracting the echo over-dispersion curvature, false freestyle misjudgment is avoided from the source, a signed distance field fuses geometric distance and risk potential, a reliable basis containing risk information is provided for trajectory planning, a control instruction is finally superposed with a repulsive acceleration item, and the trajectory planning accuracy is improved. The track stability and the short-distance obstacle avoidance response speed are both considered, so that the collision risk is reduced, and the reliability of autonomous obstacle avoidance of the unmanned aerial vehicle is improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and more specifically, to a method, apparatus, and storage medium for autonomous obstacle avoidance of UAVs. Background Technology

[0002] Autonomous obstacle avoidance is a core technology for the safe execution of aerial photography and inspection tasks by unmanned aerial vehicles (UAVs). Currently, the mainstream method relies on lidar to collect point cloud data and construct occupancy maps and distance fields to achieve obstacle perception and trajectory planning. Lidar determines the spatial occupancy status by emitting beams and receiving echoes. Traditional occupancy mapping directly classifies ray segments that are not hit by the echoes as free space, overlays fixed free evidence increments to generate a map, and then constructs an Euclidean distance field based on this map to provide a safe distance basis for trajectory planning.

[0003] However, drone operation scenarios contain numerous thin targets such as power lines and tree branches. These targets are slender and have small laser reflective cross-sections, making them easy for lidar beams to penetrate or miss, resulting in very low echo counts in the corresponding areas. Traditional occupancy mapping relies solely on the presence or absence of echoes to determine spatial status, making it difficult to distinguish between true free space and false free space caused by missed detection of thin targets. This leads to the generation of occupancy mappings that include false free zones. The distance field constructed based on this erroneous map will incorrectly quantify safe distances, causing trajectory planning to classify false free zones as passable paths.

[0004] Existing improvement methods mostly employ fixed obstacle dilatation or multimodal sensor fusion. However, fixed dilatation cannot adapt to scenarios with varying distances and risks, while multimodal fusion increases hardware costs and computational complexity, and struggles to accurately capture the features of thin lines. The existence of false free zones leads to errors in obstacle avoidance decisions by UAVs, making them highly susceptible to collisions with undetected thin lines during flight. This severely impacts operational safety and has become a key technical bottleneck restricting the reliable operation of UAVs in complex scenarios. Summary of the Invention

[0005] This invention provides a method, apparatus, and storage medium for autonomous obstacle avoidance of unmanned aerial vehicles (UAVs), thereby solving the technical problems mentioned in the background section.

[0006] This invention provides an autonomous obstacle avoidance method for unmanned aerial vehicles (UAVs), comprising the following steps: Step S101: Project the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, count the dispersion of echo counts within the grid, calculate the second-order rate of change of the dispersion along the radial direction and normalize it to generate the echo over-dispersion curvature. Step S102: Input the over-dispersion curvature of the echo and its radial accumulation into the radial contraction network, and output the free domain contraction field; Step S103: The free evidence increment of the forward missed ray segment is negatively modulated by the free domain contraction field to generate a risk-occupied mapping that suppresses false free regions. Step S104: Construct an Euclidean distance field based on the risk-based occupation map, and integrate the free domain contraction field along the ray path to obtain the risk potential. Subtract the risk potential from the Euclidean distance field to construct a signed distance field. Step S105: Adjust the search speed of the trajectory according to the over-dispersion curvature of the echo, calculate the first contact time of the trajectory in the signed distance field, and select the yaw rate and velocity scaling factor with the longest survival time and the largest forward displacement. Step S106: The yaw rate and velocity scaling factor are parsed into heading angle command and velocity command, and the repulsive acceleration term generated by the signed distance field gradient is superimposed to generate the final obstacle avoidance control command.

[0007] This invention provides an autonomous obstacle avoidance device for unmanned aerial vehicles (UAVs), comprising: The echo overdispersion curvature generation module projects the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, counts the dispersion of echo counts within the grid, calculates the second-order rate of change of this dispersion along the radial direction and normalizes it to generate the echo overdispersion curvature. The free domain contraction field generation module inputs the echo over-dispersion curvature and its radial accumulation into the radial contraction network and outputs the free domain contraction field. The occupation mapping generation module negatively modulates the free evidence increment of the forward missed ray segment through a free domain contraction field to generate a risky occupation mapping that suppresses false free regions. The signed distance field generation module constructs an Euclidean distance field based on the risk-based occupation map, integrates the free domain contraction field along the ray path to obtain the risk potential, and subtracts the risk potential from the Euclidean distance field to construct the signed distance field. The yaw rate calculation module adjusts the trajectory search speed based on the echo over-dispersion curvature, calculates the first edge contact time of the trajectory in the signed distance field, and selects the yaw rate with the longest survival time and the largest forward displacement, along with the velocity scaling factor. The obstacle avoidance control command generation module parses the yaw rate and velocity scaling factor into heading angle command and velocity command, and superimposes the repulsive acceleration term generated by the signed distance field gradient to generate the final obstacle avoidance control command.

[0008] The present invention provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, it performs the above-described autonomous obstacle avoidance method for unmanned aerial vehicles. The readable medium may be more specifically adopted and may include, but is not limited to, portable disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0009] The beneficial effects of this invention are as follows: This invention captures anomalous features of easily missed targets such as thin lines by extracting the over-dispersion curvature of the echo. A free domain contraction field is generated through a radial contraction network, dynamically and negatively modulating the incremental free evidence, thus avoiding misjudgments of false free zones from the source. The signed distance field integrates geometric distance and risk potential, providing a reliable basis for trajectory planning with risk information. During trajectory decision-making, the speed is adaptively adjusted based on risk, selecting the parameter with the longest survival time and the largest forward displacement to balance obstacle avoidance safety and travel efficiency. Finally, the control command is superimposed with a repulsive acceleration term, taking into account both trajectory stability and near-range obstacle avoidance response speed. It can adapt to various complex scenarios without additional hardware, effectively reducing collision risk and improving the reliability and practicality of UAV autonomous obstacle avoidance. Attached Figure Description

[0010] Figure 1 This is a flowchart of the UAV autonomous obstacle avoidance method of the present invention; Figure 2 This is a schematic diagram of the autonomous obstacle avoidance device for unmanned aerial vehicles (UAVs) of the present invention; Figure 3 This is a schematic diagram of the autonomous obstacle avoidance calculation logic of the UAV of the present invention.

[0011] In the figure: Echo over-dispersion curvature generation module 201, free domain contraction field generation module 202, occupied mapping generation module 203, signed distance field generation module 204, yaw angular velocity calculation module 205, obstacle avoidance control command generation module 206. Detailed Implementation

[0012] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0013] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0014] like Figures 1-3 As shown, the autonomous obstacle avoidance method for drones includes the following steps: Step S101: Project the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, count the dispersion of echo counts within the grid, calculate the second-order rate of change of the dispersion along the radial direction and normalize it to generate the echo over-dispersion curvature. Step S102: Input the over-dispersion curvature of the echo and its radial accumulation into the radial contraction network, and output the free domain contraction field; Step S103: The free evidence increment of the forward missed ray segment is negatively modulated by the free domain contraction field to generate a risk-occupied mapping that suppresses false free regions. Step S104: Construct an Euclidean distance field based on the risk-based occupation map, and integrate the free domain contraction field along the ray path to obtain the risk potential. Subtract the risk potential from the Euclidean distance field to construct a signed distance field. Step S105: Adjust the search speed of the trajectory according to the over-dispersion curvature of the echo, calculate the first contact time of the trajectory in the signed distance field, and select the yaw rate and velocity scaling factor with the longest survival time and the largest forward displacement. Step S106: The yaw rate and velocity scaling factor are parsed into heading angle command and velocity command, and the repulsive acceleration term generated by the signed distance field gradient is superimposed to generate the final obstacle avoidance control command.

[0015] In one embodiment of the present invention, the forward sector point cloud is projected onto a two-dimensional grid divided angularly and radially. The dispersion of echo counts within the grid is statistically analyzed, and the second-order rate of change of this dispersion along the radial direction is calculated and normalized to generate the echo over-dispersion curvature. Specifically, this includes: Set horizontal angular resolution With respect to the thickness of the shell , make the corner bucket , making the distance from the shell Statistics fall into the corner bucket and distance shell The number of echo points within a corresponding two-dimensional grid is denoted as the echo count. ; For distance shell Calculate the echo count within this layer. Variance in the angular dimension with the mean Calculate the dispersion using the following formula. : ; Extracting distance from the shell Discreteness Radial adjacent next distance shell Discreteness and the radially adjacent upper distance shell layer Discreteness Calculate the second rate of change using the following formula. : ; Set the decay rate constant with distance With numerical stability constant Calculate the geometric and temporal baselines using the following formulas. : ; The second rate of change Divide by geometric and temporal baselines The echo over-dispersion curvature is generated according to the following formula. : .

[0016] It should be noted that horizontal angular resolution represents the angular interval between two adjacent detection beams in the horizontal direction of the lidar. Forward sector represents a sub-region of the field of view within a specific angular range centered on the UAV's current heading, reflecting the detection range in the UAV's forward direction. Corner barrel represents a single angular interval obtained after dividing the forward sector according to the horizontal angular resolution, reflecting the discretized unit in the angular direction. Range shell thickness represents the radial thickness of a single shell set when dividing the range shell, reflecting the step size of radial sampling. Range shell represents a single annular band region obtained after dividing the forward sector according to the range shell thickness. Two-dimensional grid represents a two-dimensional grid structure formed by the interweaving of corner barrels and range shells. Echo point count represents the number of lidar echo points falling within a single two-dimensional grid, reflecting the target detection situation within that grid. Variance represents the dispersion index of the number of echo points in each corner barrel within the same range shell, reflecting the distribution difference of the number of echo points in the angular direction. Mean represents the average level of the number of echo points in each corner barrel within the same range shell, reflecting the central tendency of the number of echo points in the angular direction. Dispersion represents the ratio of the variance to the mean of the number of echo points within the same distance shell, reflecting the degree of over-dispersion in the echo distribution.

[0017] It should be noted that the current range shell represents the target range shell for which dispersion and second-order rate of change calculations are being performed. The second-order rate of change represents the radial curvature of the dispersion, reflecting the acceleration or deceleration trend of dispersion with distance. The range attenuation constant represents the range attenuation coefficient that integrates the effects of beam divergence and atmospheric attenuation on a single frame's temporal sequence. The numerical stability constant represents a fixed small value used to avoid division by zero in the calculation. The geometric and temporal baseline represents a normalized baseline composed of the range attenuation constant and the numerical stability constant, used to eliminate the interference of natural sparsity trends on the second-order rate of change. The echo over-dispersion curvature represents the result of the second-order rate of change after normalization by the geometric and temporal baseline, reflecting the target-induced anomalous over-dispersion characteristics.

[0018] It is important to note that the horizontal angular resolution must be strictly matched to the hardware detection parameters of the LiDAR, and its value should preferably be set between 0.1 degrees and 1 degree. A value that is too small will result in an excessive number of corner barrels, increasing the computational load; a value that is too large will reduce the accuracy of angular statistics. For example, if the nominal horizontal angular resolution of the LiDAR is 0.2 degrees, the forward sector is set to ±30 degrees, and the total angular range of the forward sector is 60 degrees, the corresponding number of corner barrels is 60 degrees divided by 0.2 degrees, resulting in 300. This number ensures statistical accuracy without excessively increasing the computational load. The distance shell thickness needs to be set comprehensively based on the obstacle avoidance detection range and accuracy requirements of the UAV, and its value should preferably be set between 0.5 meters and 2 meters. A smaller thickness can be chosen when the detection range is small to improve radial resolution, while a larger thickness can be appropriately increased to balance computational efficiency and accuracy. For example, if the obstacle avoidance detection range of a drone is set to 0 to 100 meters, and a distance shell thickness of 1 meter is selected, 100 distance shells can be divided, each covering a radial range of 1 meter. This not only accurately captures the echo characteristics at different distances but also controls the total number of grids, ensuring real-time calculation. The numerical stability constant is preferably set to 0.0001.

[0019] It should be noted that the distance attenuation constant is a key parameter that integrates the lidar beam divergence angle, atmospheric attenuation coefficient, and single-frame scan time. Its value is preferably set between 0.01 m / m and 0.05 m / m. The estimation method is as follows: first, obtain the lidar beam divergence angle (in milliradians), atmospheric attenuation coefficient (in meters), and single-frame scan time (in seconds). Then, estimate the distance attenuation constant using the formula: distance attenuation constant = beam divergence angle (in milliradians) multiplied by 0.01 + atmospheric attenuation coefficient + single-frame scan time multiplied by 0.1. For example, if the lidar beam divergence angle is 0.1 milliradians, the atmospheric attenuation coefficient is 0.01 m / m, and the single-frame scan time is 0.1 seconds, then the distance attenuation constant is 0.1 multiplied by 0.01 + 0.01 + 0.1 multiplied by 0.1, resulting in 0.021 m / m. This value accurately reflects the intensity of the natural sparsity trend. The geometric and temporal baseline is a normalized baseline composed of a range attenuation rate constant and a numerical stability constant. It is defined as the larger of the square of the range attenuation rate constant and the numerical stability constant. The square of the range attenuation rate constant quantifies the strength of the natural sparsity trend caused by the combined effects of sensor geometric characteristics (beam divergence), single-frame temporal sequence, and environmental factors (atmospheric attenuation). This isolates the interference of the natural sparsity trend on the second-order rate of change, ensuring that the final echo over-dispersion curvature only reflects the abnormal over-dispersion characteristics induced by the target.

[0020] In one embodiment of the present invention, the echo over-dispersion curvature and its radial accumulation are input into a radial contraction network to output a free domain contraction field, specifically including: For radial positions Echo over-dispersion curvature at the location Utilizing the distance from the shell thickness Calculate the radial cumulative amount using the following formula. : ; Set window radius Cut off by radial position For the over-dispersion curvature segment of the echo centered on the segment, the missing terms at the segment boundary are extended using existing terms in the vicinity, and a neighborhood sequence is constructed according to the following formula. : ; Extracting radial position Neighborhood sequence and radial cumulative amount Assemble the input vector according to the following formula : ; Setting parameters Radial contraction network and its forward mapping Set the numerical compression function , input vector Input the radial shrinkage network and generate the radial position according to the following formula. Free domain contraction field : .

[0021] It should be noted that the echo overdispersion curvature sequence represents the set of echo overdispersion curvatures distributed along each distance shell, reflecting the target's abnormal overdispersion characteristics at different radial positions. The origin distance shell represents the distance shell corresponding to the radial starting position, reflecting the starting reference for calculating the radial accumulation. The radial accumulation represents the cumulative result of the product of the echo overdispersion curvature from the origin distance shell to the current distance shell and the distance shell thickness, reflecting the risk intensity of radial accumulation. The window radius represents the number of preceding and following distance shells selected when constructing the neighborhood sequence. The echo overdispersion curvature segment represents the subset of echo overdispersion curvatures truncated by the window radius with the current distance shell as the center, reflecting the local feature distribution at the current position. Missing values ​​represent the curvature values ​​corresponding to distance shells not covered at the boundary of the curvature segment. The neighborhood sequence represents the curvature segment extended by the missing boundary values. The radial position value represents the radial distance corresponding to the current distance shell. The training parameters represent the set of weights and biases, etc., determined by the radial shrinking network through offline training. The forward mapping represents the forward computation process performed by the radial contraction network on the input vector, used to transform input features into intermediate mapping results. The compression function represents a mathematical function that restricts the network mapping results to the interval between zero and one. The free domain contraction field represents the set of contraction intensities corresponding to each distance shell, reflecting the modulation amplitude of free evidence at different radial positions.

[0022] It should be noted that the window radius is preferably set between 3 and 7. A window radius that is too small will make it difficult to fully capture the correlation between radial local features, while a window radius that is too large will introduce irrelevant features and increase the computational load. The compression function is set to the Sigmoid function. The radial shrinkage network uses 3 to 5 causal convolutional layers (to avoid future information leakage), combined with 1 to 2 fully connected layers. The input dimension is twice the window radius plus 2 (neighborhood sequence length + radial position + radial accumulation), and the output dimension is 1. Training parameters are obtained through offline datasets. The dataset must contain echo over-dispersion curvature features and corresponding manually labeled shrinkage strength labels for different scenarios. A binary cross-entropy loss function combined with a smoothing regularization term is used for training. The parameters are saved after iteration until the loss converges. For example, the dataset includes scenes with thin lines such as wires and tree branches, as well as scenes without interference. The label 1 indicates strong shrinkage is needed, and 0 indicates no shrinkage is needed. After training, the network can automatically learn the shrinkage rules corresponding to different features.

[0023] In one embodiment of the present invention, the free evidence increment of the forward-missing ray segment is negatively modulated by a free domain contraction field to generate a risk-based occupancy map that suppresses false free regions, specifically including: Based on a two-dimensional grid, targeting corner buckets Retrieve echo count The minimum distance shell index greater than zero is denoted as If no distance shell index that meets the conditions is found, then let Equal to the maximum distance shell index Extract the corresponding radial distance as the first echo distance. ; Filter out angles towards the center equals corner bucket And radial distance Less than the distance of the first echo voxels Construct a set of voxels for forward-missing ray segments. ; Utilizing the distance and shell thickness Calculate voxels using the following formula Distance shell index of belonging : ; Based on the distance shell index Extracting the contraction strength from the free domain contraction field The negative modulation weights are calculated using the following formula. : ; Setting free evidence increment Obtaining voxels Current log-likelihood value In voxel sets The log-likelihood value is updated using the following formula: ; Collect updated voxel log-likelihood values This constitutes a risk-based occupation mapping.

[0024] It should be noted that angular position refers to a single angular orientation in a 2D grid divided by angular buckets. The nearest distance shell represents the distance shell at a given angular position where the echo count is greater than zero and the distance to the sensor is closest, reflecting the radial level at which a target is first detected in that angular direction. The first echo distance represents the radial distance corresponding to the nearest distance shell (or the maximum scan distance if not detected), reflecting the initial detection distance of the target on that angular ray segment. The maximum scan distance represents the upper limit of the lidar's farthest detection distance, reflecting the radial range boundary of the point cloud acquisition. The forward-missing ray segment voxel set represents the collective voxels that are angularly matched and whose radial distance is less than the first echo distance, reflecting the potential free-space voxel group where no target was detected. The contraction intensity value represents the contraction strength in the free domain contraction field corresponding to the voxel's home distance shell, reflecting the degree of suppression of free evidence at that radial position. The negative modulation weight represents the coefficient used to adjust the free evidence increment. The free evidence increment constant represents the fixed increment used to accumulate free-space evidence in the traditional occupancy mapping. The log-likelihood value represents a log-scaled index of the probability that a voxel is classified as occupied or free, reflecting the confidence level of the voxel's spatial state. The risky occupancy map represents the set of voxel log-likelihoods updated by negative modulation.

[0025] It should be noted that the maximum scanning distance should be matched to the hardware detection capability of the LiDAR, preferably ranging from 50 meters to 200 meters. The value should be less than the hardware's nominal maximum detection distance to ensure detection reliability. For example, if the LiDAR hardware's nominal maximum detection distance is 200 meters, the maximum scanning distance can be set to 180 meters. If applied to close-range indoor scenarios, if the LiDAR hardware's nominal maximum detection distance is 50 meters, the maximum scanning distance can be set to 40 meters to avoid invalid calculations caused by exceeding the hardware's effective detection range. The free evidence increment constant should preferably be set between 0.5 and 2. A value that is too large will lead to excessively rapid accumulation of free evidence, easily forming false free zones; a value that is too small will result in slow map updates, affecting obstacle avoidance real-time performance. For example, in open outdoor scenarios, a value of 1.5 can be set to accelerate free space confirmation; in complex indoor scenarios, a value of 0.8 can be set to reduce the probability of generating false free zones.

[0026] It should be noted that the negative modulation weight is calculated by subtracting the contraction intensity value from the numerical value. Its design logic is to directly convert the intensity of the free domain contraction field into the suppression ratio of free evidence. The greater the contraction intensity, the more significantly free evidence is suppressed, thus achieving dynamic adaptive modulation of the free evidence increment, rather than the traditional fixed increment update. This can weaken the accumulation of free evidence in high-risk areas, suppressing false free regions at the source. For example, when the contraction intensity value is 0.8, the negative modulation weight is 0.2, and the free evidence increment is only 20% of the original constant; when the contraction intensity value is 0.1, the negative modulation weight is 0.9, and the free evidence increment is close to the original constant, balancing risk suppression and normal free space updates.

[0027] In one embodiment of the present invention, a Euclidean distance field is constructed based on a risk-based occupancy map, and a risk potential is obtained by integrating the free domain contraction field along the ray path. The risk potential is then subtracted from the Euclidean distance field to construct a signed distance field. Specifically, this includes: Set occupancy threshold Screening for risk-based mapping log-likelihood value Greater than or equal to the occupation threshold voxels, constructing the set of occupying voxels ; For any point in space Calculate the location point To occupy the voxel set The latest voxel center Euclidean distance Construct the Euclidean distance field according to the following formula. : ; Set sensor position With respect to the thickness of the shell Calculate the location point To the sensor location Euclidean distance The cutoff level index is determined using the following formula. : ; Set radial decay rate constant Extracting the shell layers at various distances in the free domain contraction field Corresponding shrinkage strength Calculate the risk potential using the following formula. : ; Set the scaling factor Calculate the Euclidean distance field Risk and momentum The weighted differences are used to construct a signed distance field according to the following formula. : .

[0028] It should be noted that the occupancy threshold represents the critical log-likelihood value that distinguishes a voxel as occupied or free. The occupied voxel set represents the total number of voxels whose log-likelihood values ​​reach or exceed the occupancy threshold, reflecting the region in space that is determined to be an obstacle. The Euclidean distance field value represents the geometric distance from a spatial location point to the nearest occupied voxel, reflecting the actual geometric distance from that location to the obstacle. The cutoff level index represents the maximum index of the distance shell mapped from the radial distance corresponding to the location point. The weights that decay exponentially with distance represent coefficients that decrease exponentially with increasing radial distance. The risk potential represents the cumulative result of the product of the free domain contraction field strength and the exponentially decaying weights from the origin distance shell to the cutoff level index. The scaling factor represents the scaling factor that converts the risk potential into a distance deduction. The signed distance field represents the result of subtracting the risk potential and the scaling factor from the Euclidean distance field value.

[0029] It should be noted that the occupancy threshold should ideally be set between 1 and 3. A value that is too large will lead to overly strict voxel selection, missing some obstacles; a value that is too small will result in too many voxels being occupied, leading to misjudgments of free space. For example, in open outdoor scenes, a value of 1.5 can be set to reduce misjudgments of free space; in indoor scenes with dense obstacles, a value of 2.5 can be set to ensure no obstacles are missed, improving the reliability of the judgment. The exponentially decaying weight with distance is determined by a radial decay rate constant, ideally set between 0.05 per meter and 0.2 per meter. A value that is too large will cause distant risks to decay rapidly, with only near risks taking effect; a value that is too small will cause distant risks to continue to have an impact, potentially over-compressing the effective distance. For example, when the radial decay rate constant is set to 0.1 per meter, the weight at a distance of 10 meters is approximately 0.37 (the negative 1st power of the natural constant), and the weight at a distance of 20 meters is approximately 0.14, thus retaining some influence from distant risks while highlighting the dominant position of near risks.

[0030] It should be noted that the proportional coefficient should ideally be set between 0.5 meters and 2 meters. A value that is too large will result in excessive risk compression over distance, potentially leading to over-avoidance; a value that is too small will result in weak risk correction, failing to effectively mitigate the risk of false free zones. For example, in scenarios with dense fine lines, the proportional coefficient can be set to 1.8 meters to enhance the risk compression effect; in scenarios without obvious false free zones, it can be set to 0.8 meters to gently correct the geometric distance. The risk potential is calculated by multiplying and summing the free domain contraction field strength with a weight that decays exponentially with distance. This combines the radially accumulated contraction strength with the distance decay effect, reflecting both the cumulative characteristics of risk and weakening the interference of uncertain risks at a distance through exponential decay, ensuring that the risk potential accurately reflects the actual risk level at the current location. For example, if a certain location has a cutoff level index of 10, is 1 meter away from the shell thickness, has a radial attenuation constant of 0.1 per meter, and the contraction strength of each shell layer from near to far is 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, and 0.1 respectively, then the risk potential is the sum of the strengths of each shell layer multiplied by their corresponding weights. This accumulates the radial risk and naturally weakens the impact of low-intensity risks at a distance.

[0031] In one embodiment of the present invention, the search speed of the trajectory is adjusted according to the echo over-dispersion curvature, the first edge contact time of the trajectory is calculated in the signed distance field, and the yaw rate with the longest survival time and the largest forward displacement and the velocity scaling factor are selected, specifically including: Set speed sensitivity constant With minimum speed scaling constant Extracting the echo overdispersion curvature sequence The maximum value in Calculate the speed scaling factor using the following formula. : ; in, This means truncating the value to a range. ; Set the upper limit constant for yaw rate Construct a set of candidate yaw rates Based on the current location Current heading angle and nominal speed For sets Candidate yaw rate Constructing trajectory primitives : when When, construct the linear trajectory primitive according to the following formula: ; when When turning, calculate the turning radius using the following formula. And construct the primitive for the circular arc trajectory: ; ; Set time step With the upper limit of the evaluation time domain Let discrete time In the signed distance field Querying trajectory primitives The first contact time is determined according to the corresponding distance value using the following formula. : ; If the distance values ​​corresponding to all discrete time points are greater than zero, then let ; Construct the forward unit vector along the current heading angle direction Calculate the forward displacement using the following formula. : ; From the set The selection of the middle makes the first touch time The maximum yaw rate; if multiple yaw rates exist that result in the same maximum first contact time, then select the one that maximizes the forward displacement. The maximum yaw rate is taken as the optimal yaw rate. and with speed scaling factor Output them together.

[0032] It should be noted that the velocity sensitivity constant represents the coefficient of influence of the echo over-dispersion curvature on the velocity scaling factor. The minimum velocity scaling constant represents the minimum limit of the velocity scaling factor, reflecting the minimum permissible velocity ratio when the UAV avoids obstacles. The upper limit constant of yaw rate represents the maximum permissible value of the candidate yaw rate, reflecting the maximum turning capability limit of the UAV. The candidate yaw rate set represents the set of all candidate values ​​that do not exceed the upper limit of yaw rate. The trajectory primitive represents a simple analytical motion trajectory (straight line or circular arc) constructed based on the candidate yaw rate, used to quickly evaluate the safety and effectiveness of the trajectory. The turning radius represents the radius of curvature of the circular arc trajectory, reflecting the curvature of the circular arc trajectory. The time step represents the time interval for discretizing the trajectory primitive. The upper limit of the evaluation time domain represents the maximum time range for trajectory safety evaluation. The first edge contact time represents the discrete moment when the trajectory first reaches the zero value boundary of the signed distance field (or the upper limit of the evaluation time domain if none), reflecting the safe survival time of the trajectory. The forward unit vector represents the unit vector along the current heading angle direction, reflecting the reference direction of the UAV's movement. The displacement vector represents the change in position of the trajectory from the current position to the position corresponding to the first edge contact time. The forward displacement represents the projection length of the displacement vector onto the forward unit vector. The selected yaw rate represents the yaw rate selected from the candidate set that balances the longest survival time and the maximum forward displacement.

[0033] It should be noted that the speed sensitivity constant should ideally be set between 0.3 and 1.0. A value that is too large will cause a significant drop in speed even with minor risks, affecting flight efficiency; a value that is too small will result in insufficient speed adjustment during high-risk situations, impacting safety. For example, in high-risk scenarios with dense power lines, a value of 0.8 can be used to make the speed more sensitive to risks; in open, low-risk scenarios, a value of 0.4 can be used to reduce unnecessary speed fluctuations. The minimum speed scaling constant should ideally be set between 0.2 and 0.5. A value that is too small will cause the drone to be too slow, making it difficult to maintain attitude balance; a value that is too large will not sufficiently reduce the probability of collisions in high-risk scenarios. For example, the minimum speed scaling constant for multi-rotor drones can be set to 0.3, while for fixed-wing drones, which need to maintain lift, it can be set to 0.4 to ensure flight safety and attitude stability.

[0034] It should be noted that the upper limit constant for yaw rate should ideally be set between 0.5 radians per second and 2.0 radians per second. Values ​​exceeding mechanical limits will lead to steering failure, while values ​​too small will result in insufficient obstacle avoidance flexibility. For example, a value of 1.0 radians per second can be set for small consumer drones, while 0.6 radians per second can be set for industrial heavy-duty drones due to their poor maneuverability, matching their mechanical performance. The time step should ideally be set between 0.05 seconds and 0.2 seconds. Values ​​too small will increase computational load and affect real-time performance; values ​​too large may miss dangerous points on the trajectory, reducing evaluation accuracy. For example, a value of 0.05 seconds can be set in high-speed flight scenarios to improve evaluation accuracy, while a value of 0.15 seconds can be set in low-speed flight scenarios to reduce computational load. The upper limit of the evaluation time domain should ideally be set between 2 seconds and 5 seconds, and the value must cover the time required for the drone to complete one effective obstacle avoidance turn. For example, when the nominal speed of the drone is 5 meters per second, the upper limit of the evaluation time domain can be set to 3 seconds, which corresponds to an evaluation trajectory length of 15 meters. This can cover a sufficient decision space without causing computational redundancy due to excessive evaluation distance.

[0035] It should be noted that the velocity scaling factor is first dynamically calculated based on the maximum value of the echo over-dispersion curvature, allowing the velocity to adaptively adjust according to risk. Then, candidate trajectory primitives, including straight lines and circular arcs, are constructed to cover different turning scenarios. The trajectory is discretized by time step, and the first edge contact time is determined by the signed distance field, quantifying trajectory safety; forward displacement is used to measure travel efficiency. Finally, the yaw rate with the longest survival time and the largest forward displacement, along with the velocity scaling factor, are selected to achieve a balance between safety and efficiency, providing accurate trajectory parameters for subsequent control command generation.

[0036] In one embodiment of the present invention, the yaw rate and velocity scaling factor are parsed into heading angle command and velocity command, and a repulsive acceleration term generated by the signed range field gradient is superimposed to generate the final obstacle avoidance control command, specifically including: Set control cycle With current heading angle Extract the optimal yaw rate Calculate the heading angle command using the following formula. : ; in, This represents the angle normalization operation that maps angle values ​​to a standard range. Set nominal speed Extract the speed scaling factor Construct the unit heading vector Calculate the speed command using the following formula. : ; Calculate the circular arc feedforward acceleration using the following formula. : ; Set the repulsion gain constant Get current location Calculate the signed distance field At the current location gradient at Calculate the repulsive acceleration term using the following formula. : ; Calculate the acceleration command using the following formula. : ; Heading angle command Speed ​​command and acceleration commands This serves as the final obstacle avoidance control command output.

[0037] It should be noted that the control cycle represents the time interval between control command updates, reflecting the frequency of command generation. Angle normalization is a mathematical operation that maps angle values ​​to a standard range, used to avoid angle value divergence and ensure command effectiveness. The heading angle command represents the target heading angle that the UAV needs to track. The unit heading vector represents the unit vector along the direction of the heading angle command. The velocity command represents the target velocity vector that the UAV needs to track. The circular feedforward acceleration represents the preset acceleration for tracking a circular trajectory. The normal vector represents the vector perpendicular to the unit heading vector. The repulsive acceleration term represents the obstacle avoidance acceleration generated based on the signed distance field gradient, reflecting the strength of the repulsive effect of nearby obstacles. The repulsive gain constant represents a coefficient that adjusts the strength of the repulsive acceleration term, reflecting the sensitivity of the obstacle repulsive effect. The acceleration command represents the target acceleration vector that the UAV needs to track. The final obstacle avoidance control command represents a complete set of execution commands including heading angle, velocity, and acceleration.

[0038] It should be noted that the control cycle should ideally be set between 0.02 and 0.1 seconds. A value that is too small will increase the processor's computational load, while a value that is too large will result in delayed command response and inability to avoid obstacles in a timely manner. For example, a small consumer drone can be set to 0.05 seconds, which ensures rapid command updates without affecting the drone's flight stability due to excessive calculations. Industrial drones, due to their large payload and high inertia, can be set to 0.1 seconds to match their dynamic response characteristics. The repulsion gain constant should ideally be set between 0.5 and 2.0. The selection is based on adjusting the strength of obstacle repulsion. A value that is too large can cause drastic fluctuations in the drone's flight attitude, while a value that is too small will fail to effectively avoid nearby obstacles. For example, when the signed distance field between the drone and the obstacle is less than 1 meter, the repulsion gain constant can be set to 1.8 to enhance the effect of the repulsion acceleration term and quickly move away from the obstacle; when the distance is greater than 5 meters, it can be set to 0.6 to weaken the repulsion effect and ensure stable flight.

[0039] It should be noted that the circular feedforward acceleration (ensuring smooth trajectory tracking) and the repulsive acceleration term (responding to sudden obstacles at close range) are vector-superimposed to form the final acceleration command. This approach balances the continuity of trajectory tracking with the safety of near-range obstacle avoidance, avoiding the problems of a single feedforward acceleration failing to handle sudden obstacles and a single repulsive acceleration causing trajectory turbulence. For example, when a UAV is flying along a circular trajectory, if a near-range obstacle suddenly appears to the side, the circular feedforward acceleration ensures the basic trajectory direction, while the repulsive acceleration term quickly generates lateral thrust, allowing the UAV to move away from the obstacle without deviating from the main trajectory, achieving smooth obstacle avoidance.

[0040] In one embodiment of the present invention, such as Figure 2 As shown, the autonomous obstacle avoidance device for drones includes: The echo overdispersion curvature generation module 201 projects the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, counts the dispersion of echo counts within the grid, calculates the second-order rate of change of the dispersion along the radial direction and normalizes it to generate the echo overdispersion curvature. The free domain contraction field generation module 202 inputs the echo over-dispersion curvature and its radial accumulation into the radial contraction network and outputs the free domain contraction field. The occupation mapping generation module 203 negatively modulates the free evidence increment of the forward missed ray segment through the free domain contraction field to generate a risk-based occupation mapping that suppresses false free regions. The signed distance field generation module 204 constructs an Euclidean distance field based on the risk-based occupation map, integrates the free domain contraction field along the ray path to obtain the risk potential, and subtracts the risk potential from the Euclidean distance field to construct the signed distance field. The yaw rate calculation module 205 adjusts the search speed of the trajectory based on the over-dispersion curvature of the echo, calculates the first touch time of the trajectory in the signed distance field, and selects the yaw rate and velocity scaling factor with the longest survival time and the largest forward displacement. The obstacle avoidance control command generation module 206 parses the yaw rate and velocity scaling factor into heading angle command and velocity command, and superimposes the repulsive acceleration term generated by the signed distance field gradient to generate the final obstacle avoidance control command.

[0041] In one embodiment of the present invention, a computer-readable medium is provided having a program stored thereon. When the program is executed by a processor, it performs the above-described autonomous obstacle avoidance method for unmanned aerial vehicles. The readable medium may be more specifically adopted and may include, but is not limited to, portable disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0042] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0043] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An autonomous obstacle avoidance method for unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Step S101: Project the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, count the dispersion of echo counts within the grid, calculate the second-order rate of change of the dispersion along the radial direction and normalize it to generate the echo over-dispersion curvature. Step S102: Input the over-dispersion curvature of the echo and its radial accumulation into the radial contraction network, and output the free domain contraction field; Step S103: The free evidence increment of the forward missed ray segment is negatively modulated by the free domain contraction field to generate a risk-occupied mapping that suppresses false free regions. Step S104: Construct an Euclidean distance field based on the risk-based occupation map, and integrate the free domain contraction field along the ray path to obtain the risk potential. Subtract the risk potential from the Euclidean distance field to construct a signed distance field. Step S105: Adjust the search speed of the trajectory according to the over-dispersion curvature of the echo, calculate the first contact time of the trajectory in the signed distance field, and select the yaw rate and velocity scaling factor with the longest survival time and the largest forward displacement. Step S106: The yaw rate and velocity scaling factor are parsed into heading angle command and velocity command, and the repulsive acceleration term generated by the signed distance field gradient is superimposed to generate the final obstacle avoidance control command.

2. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, The forward sector is divided into corner barrels according to the set horizontal angular resolution, and the forward sector is divided into distance shells according to the set distance shell thickness. The number of echo points falling into the two-dimensional grid formed by the interweaving of corner barrels and distance shells is counted. For any distance shell, calculate the variance of the number of echo points along the corner barrel dimension within the distance shell, calculate the mean of the number of echo points along the corner barrel dimension within the distance shell, and divide the variance by the mean to obtain the dispersion of the distance shell. Extract the dispersion of the current distance shell, the dispersion of the next distance shell radially adjacent to the current distance shell, and the dispersion of the previous distance shell radially adjacent to the current distance shell. Add the dispersion of the next distance shell to the dispersion of the previous distance shell, subtract twice the dispersion of the current distance shell, and divide the calculation result by the square of the distance shell thickness to obtain the second-order rate of change. Define geometric and temporal baselines that include distance attenuation rate constants and numerical stability constants, and divide the second-order rate of change by the geometric and temporal baselines to obtain the echo over-dispersion curvature.

3. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, For the echo overdispersion curvature sequence, the cumulative sum of the product of the echo overdispersion curvature from the origin to the current distance shell and the distance shell thickness is calculated to obtain the radial cumulative amount; Based on the set window radius, the echo over-dispersion curvature segment centered on the current distance shell is extracted, and the missing values ​​at the segment boundary are extended using neighboring values ​​to generate a neighborhood sequence. The radial position value of the current distance shell, the neighborhood sequence, and the radial cumulative amount are combined to construct the input vector. The input vector is then fed into a radial shrinkage network containing training parameters for forward mapping. The mapping result is restricted to zero and one by a compression function to obtain the free domain shrinkage field corresponding to the current distance shell.

4. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, Traverse each angular position of the two-dimensional grid, search for the nearest distance shell with an echo count greater than zero, and set the radial distance corresponding to the distance shell as the first echo distance. If no shell is found, set the maximum scanning distance as the first echo distance. Voxels whose angular coordinates are located at the current angular position and whose radial distance is less than the distance of the first echo are selected to form a set of voxels for the forward-missing ray segment. For any voxel in the set, the distance shell to which the voxel belongs is determined based on the ratio of the radial distance of the voxel to the thickness of the distance shell, and the contraction intensity value corresponding to the distance shell is extracted from the free domain contraction field. The negative modulation weight of the voxel is obtained by subtracting the contraction strength value from the calculated value. Set a free evidence increment constant, calculate the product of the free evidence increment constant and the negative modulation weight, and add the product to the current log-likelihood value of the voxel to generate a risk-based occupation map.

5. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, Set an occupation determination threshold, filter voxels in the risky occupation mapping whose log-likelihood value is greater than or equal to the occupation determination threshold, and construct an occupation voxel set. Calculate the geometric distance from any point in space to the nearest voxel in the set of voxels, and use the geometric distance as the Euclidean distance field value of that point. Set the sensor position, calculate the radial distance from the sensor position to the sensor position, divide the radial distance by the distance shell thickness and round to obtain the cutoff level index; Traverse each distance shell from level zero to the cutoff level index, extract the free domain contraction field intensity corresponding to each distance shell, calculate the product of the intensity and the weight which decays exponentially with distance, and perform an accumulation operation on the product to obtain the risk potential. By setting a scaling factor, the difference between the Euclidean distance field value and the product of the risk potential and the scaling factor is calculated to obtain the signed distance field.

6. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, Extract the maximum value in the echo over-dispersion curvature sequence, calculate the difference between the value and the product of the velocity sensitivity constant and the maximum value, and truncate the difference to the interval formed by the minimum velocity scaling constant and the value to obtain the velocity scaling factor. Set an upper limit constant for yaw rate, construct a set of candidate yaw rates, and for any candidate yaw rate in the set, construct a trajectory primitive based on the current position, current heading angle, nominal speed, and speed scaling factor; where, when the candidate yaw rate is zero, the trajectory primitive is a straight trajectory. When the candidate yaw rate is not zero, the trajectory primitive is a circular arc trajectory. The turning radius of the circular arc trajectory is calculated by dividing the product of the speed scaling factor and the nominal speed by the candidate yaw rate.

7. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 6, characterized in that, Set a time step and an upper limit for the evaluation time domain. Discretize the trajectory primitives according to the time step. Query the distance values ​​of the corresponding positions at each discrete time in the signed distance field. Determine the first time when the distance value is less than or equal to zero as the first time to touch the edge. If the distance values ​​corresponding to all discrete time points are greater than zero, then determine the upper limit of the evaluation time domain as the first time to touch the edge. Construct a forward unit vector along the current heading angle direction, calculate the displacement vector between the position reached by the trajectory primitive at the first edge contact time and the current position, calculate the projection length of the displacement vector onto the forward unit vector, and obtain the forward displacement. Select the yaw rate that results in the longest first contact time from the set of candidate yaw rates; if there are multiple yaw rates that result in the longest first contact time, select the yaw rate that results in the largest forward displacement and use it as the selected yaw rate, and output it together with the speed scaling factor.

8. The autonomous obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, Set the control cycle, calculate the product of the selected yaw rate and the control cycle, add the product to the current heading angle, perform an angle normalization operation on the addition result to map the angle to a standard range, and obtain the heading angle command; Set the nominal speed, construct a unit heading vector composed of the cosine and sine values ​​of the heading angle command, calculate the product of the selected speed scaling factor and the nominal speed, calculate the product of this product and the unit heading vector, and obtain the speed command. Calculate the product of the selected speed scaling factor, nominal speed, and selected yaw rate, construct a normal vector composed of the negative sine and cosine values ​​of the heading angle command, calculate the product of the product and the normal vector, and obtain the circular feedforward acceleration. Calculate the gradient of the signed distance field at the current position, set the repulsion gain constant, calculate the product of the repulsion gain constant and the gradient, and obtain the repulsion acceleration term; Calculate the vector sum of the circular arc feedforward acceleration and the repulsion acceleration term to obtain the acceleration command; The heading angle command, speed command, and acceleration command are output as the final obstacle avoidance control command.

9. An autonomous obstacle avoidance device for unmanned aerial vehicles, characterized in that: Performing the autonomous obstacle avoidance method for unmanned aerial vehicles as described in any one of claims 1 to 8 includes: The echo overdispersion curvature generation module projects the forward sector point cloud onto a two-dimensional grid divided into angular and radial directions, counts the dispersion of echo counts within the grid, calculates the second-order rate of change of this dispersion along the radial direction and normalizes it to generate the echo overdispersion curvature. The free domain contraction field generation module inputs the echo over-dispersion curvature and its radial accumulation into the radial contraction network and outputs the free domain contraction field. The occupation mapping generation module negatively modulates the free evidence increment of the forward missed ray segment through a free domain contraction field to generate a risky occupation mapping that suppresses false free regions. The signed distance field generation module constructs an Euclidean distance field based on the risk-based occupation map, integrates the free domain contraction field along the ray path to obtain the risk potential, and subtracts the risk potential from the Euclidean distance field to construct the signed distance field. The yaw rate calculation module adjusts the trajectory search speed based on the echo over-dispersion curvature, calculates the first edge contact time of the trajectory in the signed distance field, and selects the yaw rate with the longest survival time and the largest forward displacement, along with the velocity scaling factor. The obstacle avoidance control command generation module parses the yaw rate and velocity scaling factor into heading angle command and velocity command, and superimposes the repulsive acceleration term generated by the signed distance field gradient to generate the final obstacle avoidance control command.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the autonomous obstacle avoidance method for unmanned aerial vehicles as described in any one of claims 1 to 8.

Images