Obstacle avoidance method and system for unmanned aerial vehicle

Abstract

The invention discloses an obstacle avoidance method and system for an unmanned aerial vehicle, and the method comprises the steps: collecting the three-dimensional point cloud and image information of a laser radar, binocular vision and a millimeter-wave radar, and obtaining an obstacle comprehensive parameter set; carrying out dimensionless processing on the obstacle avoidance feature vector to obtain an obstacle avoidance feature vector; inputting the obstacle avoidance feature vector into a discrimination model constructed based on a double-membership adaptive fuzzy decision tree, and outputting an obstacle avoidance action type according to an obstacle avoidance rule; and calling a corresponding trajectory generator to calculate the vertical speed, the lateral speed and the course instruction of the target in real time, driving a propulsion motor to complete maneuvering, and when the vertical clearance coefficient, the horizontal clearance coefficient and the distance safety margin all exceed preset exit thresholds again, automatically terminating the obstacle avoidance mode and restoring to the original planned route. According to the invention, the problems of low precision and insufficient stability of real-time obstacle avoidance decision making of the unmanned aerial vehicle in a complex environment are effectively solved.

Description

Technical Field

[0001] This invention relates to the technical field of unmanned aerial vehicle (UAV) flight control, and more particularly to an obstacle avoidance method and system for UAVs. Background Technology

[0002] In recent years, with the rapid growth in demand for low-altitude applications, the development of UAV technology has shown a trend towards diversification and refinement. Among them, obstacle avoidance technology, as an important guarantee for the safe flight of UAVs, has become one of the current research hotspots. Early UAV obstacle avoidance technologies mostly used single-type sensors, such as ultrasonic sensors or monocular vision solutions. Although these were simple in structure and low in cost, they had obvious limitations in their applicability in complex environments. With the rise of multi-sensor fusion and artificial intelligence algorithms, combined sensing and perception solutions such as lidar, binocular vision, and millimeter-wave radar have begun to be widely used.

[0003] CN115033019A discloses an obstacle avoidance method for UAVs based on obstacle type determination. This method focuses on the identification of obstacle movement characteristics and the determination of safe landing areas. However, it does not fully consider the real-time maneuvering obstacle avoidance requirements in complex flight environments and lacks a comprehensive obstacle avoidance strategy for real-time flight trajectories. CN113821046A discloses an autonomous obstacle avoidance method based on deep learning. It uses binocular vision and convolutional neural networks to achieve end-to-end prediction to simplify the obstacle avoidance process. However, the convolutional neural network model is easily limited by the training data scenario and has limited generalization ability in unknown or out-of-training-set environments. Its real-time processing speed and robustness are also insufficient.

[0004] Existing drone obstacle avoidance technologies are typically limited to single-dimensional feature recognition or prediction based on pre-trained models, lacking sufficient real-time perception capabilities for complex environments and devoid of obstacle avoidance decision-making schemes that are dynamically optimized for different obstacle spatial layouts. Furthermore, existing solutions lack systematic quantitative analysis of the drone's own attitude and motion characteristics, making it difficult to effectively achieve a smooth transition in drone attitude control and flight trajectory during obstacle avoidance. Therefore, how to construct accurate and flexible obstacle avoidance strategy models in real time based on multi-sensor data fusion, and effectively balance the contradiction between drone safety and flight stability, has become one of the important technical challenges that urgently need to be solved in the field of drone obstacle avoidance.

[0005] Existing UAV obstacle avoidance technologies suffer from poor real-time adaptability to complex flight environments, limited single-dimensional perception features, and poor generalization performance. This invention provides a UAV obstacle avoidance method that integrates multi-source perception data. By constructing multi-dimensional feature vectors and employing a dual-membership adaptive fuzzy decision tree discrimination model, it can accurately output the optimal obstacle avoidance action type in real time and generate dynamically smooth flight trajectory commands, effectively solving the problems of low accuracy and insufficient stability in real-time obstacle avoidance decision-making for UAVs in complex environments. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: During the flight control cycle of the UAV, three-dimensional point cloud and image information of lidar, binocular vision and millimeter-wave radar are collected simultaneously. The collected information is then unified by external parameter calibration and timestamp alignment to obtain a set of comprehensive obstacle parameters describing the relative height, horizontal gap and relative distance of the obstacle.

[0009] Based on the current attitude, height, maximum diagonal size, and preset safety distance of the UAV, the set of comprehensive obstacle parameters is processed to be dimensionless, and three characteristic coefficients reflecting vertical clearance, horizontal clearance, and distance safety margin are obtained respectively, and they are combined into an obstacle avoidance feature vector.

[0010] The obstacle avoidance feature vector is input into a discrimination model constructed based on a dual-membership adaptive fuzzy decision tree. The discrimination model uses two membership functions to perform parallel fuzzy mapping on the vertical clearance and horizontal clearance, with distance safety margin as the safety weight, and outputs the obstacle avoidance action type according to the obstacle avoidance rules.

[0011] Based on the obstacle avoidance action type, the corresponding trajectory generator is invoked to calculate the target's vertical speed, lateral speed, and heading command in real time, and the propulsion motor is driven to complete the maneuver. When the vertical clearance coefficient, horizontal clearance coefficient, and distance safety margin all exceed the preset exit threshold again, the obstacle avoidance mode is automatically terminated and the original planned route is restored.

[0012] As a preferred embodiment of the obstacle avoidance method for UAVs described in this invention, the step of outputting the obstacle avoidance action type according to the obstacle avoidance rules includes:

[0013] When the vertical clearance coefficient is higher than the first threshold and the distance safety margin is higher than the second threshold, choose the descent or crossing action.

[0014] When the vertical clearance coefficient is lower than the third threshold and the horizontal clearance coefficient is higher than the fourth threshold, choose the ascending or overtaking maneuver.

[0015] When the horizontal clearance coefficient is lower than the fifth threshold and the distance safety margin is higher than the sixth threshold, choose lateral or detour maneuvers.

[0016] When the conditions for descent, crossing or ascent, and overflight are met simultaneously, and the conditions for lateral or circling are also met, a combined maneuver is selected.

[0017] As a preferred embodiment of the obstacle avoidance method for UAVs described in this invention, based on the obstacle avoidance action type, a corresponding trajectory generator is invoked to calculate the target's vertical velocity, lateral velocity, and heading command in real time, including:

[0018] Read the obstacle avoidance action type output by the discrimination model, and establish a correspondence table between the obstacle avoidance action type and the current flight mode of the UAV;

[0019] Based on the correspondence table, a trajectory template matching the obstacle avoidance action type is retrieved, and the trajectory template records the reference displacement curves according to the vertical axis, horizontal axis and heading axis respectively;

[0020] The curve is recalibrated by combining the current attitude, terrain height, and local wind field intensity to ensure that the obstacle avoidance trajectory is compatible with environmental constraints;

[0021] The recalibrated triaxial displacement curves are first and second differentiated to obtain the velocity curves. Then, the target's vertical velocity, lateral velocity, and rate of change of heading angle are extracted based on the smoothing results of the velocity curves.

[0022] With the attitude stability of the UAV as a constraint, extreme value corrections are performed on the target's vertical velocity, lateral velocity, and heading angle change rate to generate three-channel control commands that can be issued.

[0023] As a preferred embodiment of the obstacle avoidance method for unmanned aerial vehicles (UAVs) described in this invention, driving the propulsion motor to complete the maneuver includes:

[0024] The three-channel control commands are mapped to the desired thrust distribution values ​​for each propulsion motor, and the commands are distributed to each propulsion motor according to the thrust distribution matrix.

[0025] After the allocation is completed, the field orientation control module of the motor controller is invoked to convert the desired thrust allocation value into the corresponding current reference value;

[0026] Based on the current motor speed and temperature rise, the maximum torque-to-current ratio control strategy is selected, and the current reference value is adjusted in real time so that the output thrust of the propulsion motor can keep up with the three-channel control commands.

[0027] During the maneuver, the motor current, speed and vibration amplitude are continuously monitored. If any indicator is detected to exceed the safety limit, load reduction compensation is triggered, the thrust distribution coefficient is automatically reduced, and the remaining obstacle avoidance actions are completed within the safety threshold.

[0028] Once all monitored indicators have returned to normal levels, and the judgment model confirms that the obstacle avoidance maneuver has ended, the propulsion motor will resume operation according to the thrust requirements of the original planned route.

[0029] As a preferred embodiment of the obstacle avoidance method for unmanned aerial vehicles (UAVs) described in this invention, obtaining the comprehensive obstacle parameter set includes:

[0030] Using lidar, the space in front is scanned in layers to obtain the initial point cloud outline of obstacles;

[0031] The binocular vision module is invoked to perform depth recovery of obstacle textures, and the shape error of the lidar point cloud is corrected by feature point matching.

[0032] Millimeter-wave radar is used to perform Doppler ranging in the direction of the obstacle's center to eliminate distance drift caused by obstruction and weak reflection;

[0033] The corrected data is fused according to a unified coordinate system to extract the three-dimensional coordinates of the top, bottom and side boundary points of the obstacle, and then the relative height difference and horizontal gap of the obstacle are calculated.

[0034] By combining the current position of the UAV with the center coordinates of the obstacle, the relative distance to the obstacle is calculated, and the timestamp and confidence level are recorded simultaneously to form the comprehensive parameter set of the obstacle.

[0035] As a preferred embodiment of the obstacle avoidance method for UAVs described in this invention, obtaining the obstacle avoidance feature vector includes:

[0036] Using the aircraft height as a normalized benchmark, the relative height difference between the top of the obstacle and the top of the drone is converted into a vertical clearance coefficient;

[0037] Using the maximum diagonal size of the drone as a normalization benchmark, the horizontal clearance of obstacles is converted into a horizontal clearance coefficient;

[0038] Using a preset safe distance as a normalization benchmark, the relative distance to obstacles is converted into a distance safety margin coefficient;

[0039] The vertical clearance coefficient, the horizontal clearance coefficient, and the distance safety margin coefficient are arranged in sequence and linear normalization is performed to obtain a vector representation of fixed length.

[0040] A sensor data fusion confidence label is appended to the end of the vector representation to form the obstacle avoidance feature vector.

[0041] As a preferred embodiment of the obstacle avoidance method for UAVs described in this invention, the obstacle avoidance feature vector is input into a discriminant model constructed based on a dual-membership adaptive fuzzy decision tree, including:

[0042] Parallel fuzzy mapping of the vertical clearance coefficient and the horizontal clearance coefficient was performed using trigonometric membership functions and Gaussian membership functions respectively, resulting in two sets of membership degree values.

[0043] Based on the distance safety margin coefficient, safety weights are assigned to the two sets of membership values ​​to form a weighted and fused comprehensive membership degree.

[0044] The comprehensive membership degree is compared with the obstacle avoidance rule matching table. If the comprehensive membership degree does not belong to any single action range, fuzzy interpolation is performed to obtain the closest action type.

[0045] The confirmed obstacle avoidance action type is output and passed to the trajectory generator to ensure that the UAV completes obstacle avoidance according to the optimal action path.

[0046] As a preferred embodiment of the obstacle avoidance method for unmanned aerial vehicles (UAVs) described in this invention, the discrimination model includes a safety assessment layer, a fuzzy mapping layer, and an action decision layer, wherein:

[0047] The security assessment layer divides the environment into high-security zones, warning zones, and danger zones according to the distance safety margin coefficient, and updates the thresholds of each zone in real time to provide dynamic security weights.

[0048] The fuzzy mapping layer consists of vertical dimension nodes and horizontal dimension nodes. The two nodes use triangular membership functions and Gaussian membership functions in parallel to perform dual membership mapping on the vertical and horizontal net space coefficients and output two sets of membership degrees.

[0049] After receiving the weights from the safety assessment layer and the membership set from the fuzzy mapping layer, the action decision layer first performs weighted fusion, and then calls the obstacle avoidance rule matching table to determine the final action type.

[0050] The obstacle avoidance rule matching table is initially trained offline using a genetic algorithm, and then adaptively fine-tuned during flight using a gradient descent strategy based on real-time environmental factors to ensure the long-term stability of the model.

[0051] As a preferred embodiment of the obstacle avoidance system for unmanned aerial vehicles (UAVs) described in this invention, it includes one or more processors;

[0052] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the procedures described above for an obstacle avoidance method for a drone.

[0053] As a preferred embodiment of the computer-readable medium for storing software according to the present invention, the software includes instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the process described above for an obstacle avoidance method for unmanned aerial vehicles.

[0054] The beneficial effects of this invention are as follows: This invention ensures high-precision detection of obstacle information through multi-source sensor fusion and high-precision spatiotemporal alignment; improves the generalization performance of the discrimination model through the standardized construction of feature vectors; ensures the accuracy of obstacle avoidance action decisions in complex environments through the design of a dual-membership adaptive fuzzy decision tree model; and ensures efficient execution of obstacle avoidance maneuvers and rapid recovery of the original planned route through the coordinated cooperation of the trajectory generator and the propulsion motor. The overall solution solves the problems of insufficient information fusion, low decision accuracy, and poor flight stability in existing UAV obstacle avoidance technologies in complex environments, significantly improving the safety and adaptability of UAV autonomous flight. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0056] Figure 1 This is a schematic flowchart of the obstacle avoidance method for unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0058] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0060] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates an obstacle avoidance method for drones, which specifically includes the following steps:

[0061] S1. During the UAV flight control cycle, simultaneously collect 3D point cloud and image information from lidar, binocular vision, and millimeter-wave radar. Through external parameter calibration and timestamp alignment, coordinate unify the collected information to obtain a comprehensive set of obstacle parameters describing the relative height, horizontal gap, and relative distance of the obstacle.

[0062] The following points need to be explained in this step:

[0063] At the beginning of each flight control cycle, the lidar performs a high-speed scan of the fan-shaped airspace in front of the UAV according to a layered strategy of alternating between high and low altitudes from bottom to top.

[0064] For each layer of the scan results, distance threshold culling is first performed to filter out noise, and then voxel downsampling is performed on the remaining point cloud to generate the first batch of obstacle point cloud contours.

[0065] Record the scan timestamp for each point cloud layer to provide a baseline for subsequent time alignment;

[0066] Simultaneously trigger the left and right cameras to acquire image frames that match the coverage of the LiDAR scanning area;

[0067] A dense disparity map is obtained through a binocular matching algorithm, and a depth map is calculated using the relationship between disparity, focal length, and baseline to form a visual point cloud.

[0068] By employing point-to-surface constraints based on feature point matching, the visual point cloud and the lidar point cloud are iteratively registered with the nearest point within a local window, thereby correcting the shape error of the lidar point cloud.

[0069] The millimeter-wave radar performs continuous wave frequency modulation in the direction of the center of the overlapping field of view with the binocular laser, and uses a constant false alarm rate threshold to detect and extract high-confidence echoes.

[0070] The detected echoes are subjected to Doppler frequency shift calculation to obtain the accurate distance value in the direction of the obstacle center, and the depth drift section is reinforced by one-dimensional linear interpolation.

[0071] The augmented data is assigned a timestamp synchronized with the millimeter-wave frame and labeled with the corresponding confidence level.

[0072] Using the pre-calibrated extrinsic parameter matrix, the lidar point cloud, visual point cloud, and millimeter-wave ranging results are all converted to the body coordinate system with the UAV's center of mass as the origin;

[0073] Based on a high-precision clock, out-of-range frames are discarded from all sensor data, so that the timestamp difference between each data frame is controlled within one-tenth of the flight control cycle.

[0074] A weighted probability occupancy grid model is used to fuse three-source data to jointly evaluate the point cloud density, visual texture consistency and millimeter wave confidence within the same grid, thus obtaining the obstacle consistency probability field.

[0075] Perform connected component labeling in a consistency probability field, and extract obstacle candidates for regions with probabilities higher than a set threshold.

[0076] For each candidate object, calculate the three-dimensional coordinate difference between the top, bottom, and left and right boundary points, and output the relative height difference and horizontal gap of the obstacle.

[0077] Combine the current localization solution of the UAV to calculate the Euclidean distance from the center of the obstacle to the center of mass of the UAV, and package it together with the corresponding timestamp and fusion confidence score;

[0078] All obstacles are sorted in ascending order by the five-tuple of relative height difference, horizontal gap, relative distance, confidence level, and timestamp, and stored in a cache queue to form a complete set of comprehensive obstacle parameters.

[0079] For example, the flight control cycle is set to 30ms.

[0080] The specific value of the above-mentioned threshold can be set according to the actual situation. In this embodiment, there is no unique requirement for it.

[0081] The joint evaluation method and Euclidean distance calculation method of the above-mentioned grid model can be carried out using existing technologies and methods, and will not be elaborated in this example.

[0082] It should be noted that by accurately registering obstacle information in the spatial and temporal dimensions, the accuracy and real-time performance of obstacle detection are improved, avoiding misjudgment problems caused by blind spots or time delay errors of a single sensor, and ultimately achieving the beneficial effect of accurately describing the location and spatial characteristics of obstacles.

[0083] S2. Based on the UAV's current attitude, height, maximum diagonal dimension, and preset safety distance, the obstacle comprehensive parameter set is processed to be dimensionless, obtaining three characteristic coefficients reflecting vertical clearance, horizontal clearance, and distance safety margin, and these are combined into an obstacle avoidance feature vector. Note that the following should be noted in this step:

[0084] The drone's real-time attitude (pitch angle, roll angle, yaw angle), body height, and maximum diagonal dimension are read from the inertial measurement unit, barometer, and internal status register, and the preset safe distance is obtained by calling the model configuration file.

[0085] The top height value of each target in the obstacle comprehensive parameter set is retrieved and the difference is calculated with the current top height of the aircraft. Then, based on the aircraft height, the height difference is converted into a percentage and recorded as the vertical clearance coefficient. If the height difference is negative (i.e. the top of the obstacle is higher than the top of the aircraft), it is treated as 0.

[0086] Extract the minimum gap between the two boundary points of the obstacle in the horizontal direction and compare it with the maximum diagonal size of the UAV. Convert the ratio of the obtained gap to the maximum diagonal size of the UAV into a percentage form to obtain the horizontal clearance coefficient. If the gap is less than the maximum diagonal size of the UAV, the horizontal clearance coefficient is set to 0.

[0087] The relative distance from the center of the obstacle to the center of mass of the machine is read, and the preset safety distance is subtracted. The ratio of the remaining distance to the preset safety distance is converted into a percentage to obtain the distance safety margin coefficient. If the relative distance is less than or equal to the preset safety distance, the distance safety margin coefficient is set to 0, indicating that the machine has entered the danger zone.

[0088] Following a fixed order of vertical clearance, horizontal clearance, and distance safety margin, the three coefficients are written into an initial vector of length 3, and linear normalization is performed on each coefficient: values ​​greater than 100% are truncated to 100%, and values ​​less than 0 are truncated to 0, to ensure that the vector elements are stable between 0 and 1.

[0089] Furthermore, the lidar point cloud density score, binocular vision texture consistency score, and millimeter-wave radar echo confidence score are obtained. The weighted average is calculated according to the weight of each sensor in the fusion model to obtain a fusion confidence score between 0 and 1. This confidence score is appended to the end of the normalized vector to form an obstacle avoidance feature vector with a length of 4.

[0090] The obstacle avoidance feature vector and its timestamp are written into a circular buffer and marked as a discriminable state for subsequent discrimination models to call in real time; if there are old vectors in the buffer that have not been removed, the oldest entry is overwritten according to the first-in-first-out rule.

[0091] It should be noted that by obtaining three characteristic coefficients—vertical clearance, horizontal clearance, and distance safety margin—and combining them into an obstacle avoidance feature vector, a standardized description of the spatial relationship of obstacles is achieved. This provides a unified and comparable scale for obstacle characteristics under different flight states and environmental conditions, eliminates decision-making errors caused by differences in dimensions and states, and improves the generalization and robustness of the obstacle avoidance discrimination model.

[0092] S3. Input the obstacle avoidance feature vector into the discrimination model constructed based on a dual-membership adaptive fuzzy decision tree. The discrimination model uses two membership functions to perform parallel fuzzy mapping on the vertical and horizontal clearances, using the distance safety margin as the safety weight, and outputs the obstacle avoidance action type according to the obstacle avoidance rules. Note that the following should be noted in this step:

[0093] The triangular membership function and the Gaussian membership function are respectively used to perform parallel fuzzy mapping on the vertical clearance coefficient and the horizontal clearance coefficient, obtaining two groups of membership values;

[0094] According to the magnitude of the distance safety margin coefficient, safety weights are assigned to the two groups of membership values to form a comprehensive membership degree after weighted fusion;

[0095] The comprehensive membership degree is compared with the obstacle avoidance rule matching table. If the comprehensive membership degree does not belong to any single action interval, fuzzy interpolation is performed to obtain the closest action type;

[0096] Output the confirmed obstacle avoidance action type and transmit it to the trajectory generator to ensure that the UAV completes obstacle avoidance according to the optimal action path.

[0097] As an example, the mathematical expression formula of the discrimination model in this embodiment is as follows:

[0098]

[0099] Among them, λ is the upper limit of the comprehensive membership curve integral, ω(q) is the fusion confidence weighted membership function, q is the normalized safety distance index, v j is the weight of the jth action rule, μ j is the vertical mapping center value, ξ j is the horizontal mapping center value, is the model online attenuation factor, χ is the attitude stability compensation term, ψ is the real-time wind field coefficient, is the environmental light sensitivity coefficient, is the sum of safety weight fusions, ρ is the flight speed normalization term, θ is the lateral airflow disturbance term, and F is the final comprehensive membership degree discrimination value;

[0100] The value range of F is [0, 1]. F ≤ 0.33 represents the dangerous area, 0.33 < F ≤ 0.66 represents the warning area, and F > 0.66 represents the high safety area.

[0101] Furthermore, according to the obstacle avoidance rules, the obstacle avoidance action types are output, including:

[0102] When the vertical clearance coefficient is higher than the first threshold and the distance safety margin is higher than the second threshold, select the descending and crossing actions;

[0103] When the vertical clearance coefficient is lower than the third threshold and the horizontal clearance coefficient is higher than the fourth threshold, select the ascending and flying over actions;

[0104] When the horizontal clearance coefficient is lower than the fifth threshold and the distance safety margin is higher than the sixth threshold, select the lateral and bypassing actions;

[0105] When the conditions for descent, crossing or ascent, and overflight are met simultaneously, and the conditions for lateral or circling are also met, a combined maneuver is selected.

[0106] It should be noted that the first threshold is based on 60% of the aircraft height, and is reduced by 3% for every 5 m / s increase in average wind speed. If the landing gear extension exceeds 10% of the aircraft height, it is increased by the same percentage. The second threshold is set at 1.2 times the level flight braking distance, with an additional 0.5m added for every 10 m / s increase in flight speed, and a uniform 10% increase for high altitude or dusty environments. The third threshold is set at 45% of the aircraft height, and is increased by 2% for every 10° increase in pitch angle, and decreased by 1% for every 10% increase in load. The fourth threshold is based on the maximum aircraft height. Starting at 65% of the diagonal dimension, the threshold is increased by 4% for every 3m / s of crosswind, by 2% for a heading error exceeding 5°, and by 3% for a battery level below 30%. The fifth threshold is initially set at 40% of the maximum diagonal dimension of the aircraft, decreasing by 2% for every 0.5g of lateral acceleration, and increasing by 3% for a visual obstruction rate exceeding 30%. The sixth threshold is 90% of the level flight braking distance, increasing by 0.3m for every 1000m increase in altitude, by 0.2m for a millimeter-wave signal-to-noise ratio below the threshold, and decreasing by 5% for a fusion confidence score above 0.85 and increasing by 5% for a confidence score below 0.60.

[0107] Furthermore, the discrimination model includes a security assessment layer, a fuzzy mapping layer, and an action decision layer, wherein:

[0108] The safety assessment layer divides the environment into high-safety zones, warning zones, and danger zones according to the distance safety margin coefficient, and updates the thresholds of each zone in real time to provide dynamic safety weights.

[0109] The fuzzy mapping layer consists of vertical dimension nodes and horizontal dimension nodes. The two nodes use triangular membership functions and Gaussian membership functions in parallel to perform dual membership mapping on the vertical and horizontal netness coefficients and output two sets of membership degrees.

[0110] After receiving the weights from the safety assessment layer and the membership set from the fuzzy mapping layer, the action decision layer first performs weighted fusion, and then calls the obstacle avoidance rule matching table to determine the final action type.

[0111] The obstacle avoidance rule matching table is initially trained offline using a genetic algorithm, and then adaptively fine-tuned during flight using a gradient descent strategy based on real-time environmental factors to ensure the long-term stability of the model.

[0112] In an optional implementation, the security assessment layer, fuzzy mapping layer, and action decision layer each perform the following steps:

[0113] (1) Security assessment layer initialization

[0114] The distance safety margin coefficient in the obstacle avoidance feature vector is called and dynamically amplified based on the measured wind speed level and flight speed.

[0115] Based on the amplified safety margin coefficient, the system divides the area into a high-safety zone (coefficient greater than 0.66), a warning zone (coefficient between 0.33 and 0.66), and a danger zone (coefficient less than 0.33) in real time.

[0116] The three zones are assigned weight coefficients of 0.4, 0.7, and 1.0 respectively, and are recalculated in each flight control cycle to generate the current safety weight table;

[0117] (2) Parallel processing of fuzzy mapping layer

[0118] At the vertical dimension nodes, two sets of mappings are constructed for the vertical clearance coefficient: triangular membership functions (low, medium, high) and Gaussian membership functions (narrow peak, wide peak).

[0119] At the horizontal dimension nodes, two sets of mappings are constructed for the horizontal clearance coefficient in the same manner;

[0120] The vertical and horizontal two-dimensional mapping results are organized into membership set A and membership set B respectively, and the current timestamp is added simultaneously.

[0121] (3) Weighted fusion of comprehensive membership degree

[0122] Read the security weight table output by the security assessment layer and assign corresponding weights to each membership value of set A and set B;

[0123] The weighted membership degrees are normalized and fused to obtain the comprehensive membership degree representing the overall passability of the obstacle;

[0124] If the overall membership degree does not match any of the four predefined action intervals (descent-crossing, ascent-flying, lateral-circling, combination-maneuvering), then fuzzy interpolation is performed on the adjacent intervals according to the weights on both sides to extract the closest action type candidate.

[0125] (4) Comparison of action decision-making layer rules

[0126] a. Descent-crossing is triggered based on a vertical clearance coefficient greater than 0.60 and a distance safety margin greater than 0.50;

[0127] b. Triggering ascent-flyover based on a vertical clearance coefficient below 0.45 and a horizontal clearance coefficient above 0.65;

[0128] c. Trigger a lateral detour based on a horizontal clearance coefficient of less than 0.40 and a distance safety margin of more than 0.55;

[0129] d. If the triggering conditions of a or b and c are met simultaneously, output the combined maneuver.

[0130] e. Push the finally confirmed obstacle avoidance action type to the trajectory generator, along with the comprehensive membership value as the decision confidence level;

[0131] (5) Offline and online two-stage rule adaptation

[0132] Offline phase: The genetic algorithm is used to iteratively search for the optimal threshold set and mapping parameters on the simulation dataset until convergence occurs when the average misclassification rate is below 2%.

[0133] Online phase: The success rate of the action is calculated once every 50 consecutive control cycles. If it is lower than the set threshold, the gradient descent fine-tuning mechanism is activated to perform small-scale adjustments on the membership center of the triangular membership function and the standard deviation of the Gaussian membership function. The adjustment range does not exceed 5% of the initial value to ensure that the model gradually adapts and stabilizes with the environment.

[0134] Preferably, by outputting the optimal obstacle avoidance action type through obstacle avoidance rules, dynamic and accurate classification of obstacle spatial layout is achieved. Furthermore, through an adaptive weight adjustment strategy, safety weights are promptly increased when the safety margin is small, thereby improving the safety and accuracy of the judgment. Ultimately, this achieves the beneficial effect of stably outputting reliable obstacle avoidance decisions under complex environments and uncertain conditions.

[0135] S4. Based on the obstacle avoidance maneuver type, the corresponding trajectory generator is invoked to calculate the target's vertical velocity, lateral velocity, and heading command in real time, and the propulsion motor is driven to complete the maneuver. When the vertical clearance coefficient, horizontal clearance coefficient, and distance safety margin all exceed the preset exit threshold again, the obstacle avoidance mode is automatically terminated, and the original planned route is restored. Note that the following points should be noted in this step:

[0136] Read the obstacle avoidance action type output by the discrimination model and establish a correspondence table between the obstacle avoidance action type and the current flight mode of the UAV (such as cruise, climb, hover);

[0137] Based on the correspondence table, a trajectory template matching the obstacle avoidance action type is retrieved. The trajectory template records the reference displacement curves according to the vertical axis, horizontal axis and heading axis respectively.

[0138] The curve is recalibrated by combining the current attitude, terrain height, and local wind field intensity to ensure that the obstacle avoidance trajectory is compatible with environmental constraints;

[0139] The recalibrated triaxial displacement curves are first and second differentiated to obtain the velocity curves. Then, the target's vertical velocity, lateral velocity, and rate of change of heading angle are extracted based on the smoothing results of the velocity curves.

[0140] With the attitude stability of the UAV as a constraint, extreme value corrections are performed on the target's vertical velocity, lateral velocity, and rate of change of heading angle to form three-channel control commands that can be issued.

[0141] The three-channel control commands are mapped to the desired thrust distribution values ​​for each propulsion motor, and the commands are distributed to each propulsion motor according to the thrust distribution matrix.

[0142] After the allocation is completed, the field orientation control module of the motor controller is invoked to convert the desired thrust allocation value into the corresponding current reference value;

[0143] Based on the current motor speed and temperature rise, the maximum torque-to-current ratio control strategy is selected, and the current reference value is adjusted in real time to ensure that the output thrust of the propulsion motor can keep up with the three-channel control commands.

[0144] During maneuvering, the motor current, speed, and vibration amplitude are continuously monitored. If any indicator exceeds the safety limit, load reduction compensation is triggered, automatically reducing the thrust distribution coefficient, and stopping within a safe threshold (e.g., phase current less than 0.8I). rated Mechanical speed less than 0.95N max Root mean square axial vibration acceleration ≤ 5 mm / s² 2 Complete the remaining obstacle avoidance maneuvers within ) ;

[0145] Once all monitored indicators have returned to normal levels, and the judgment model confirms that the obstacle avoidance maneuver has ended, the propulsion motor will resume operation according to the thrust requirements of the original planned route.

[0146] As an example, the method for performing extreme value correction is as follows: compare the above three target values ​​with the platform's allowable limits (pitch ±35° / s, roll ±45° / s, yaw ±60° / s). If any one exceeds the limit, scale it proportionally to the limit according to the principle of minimum reduction, and leave an additional 10% safety margin.

[0147] As an example, the maximum torque-to-current ratio control strategy is as follows: the ESC MCU converts the expected thrust of each motor into the reference value of the AC motor's dq-axis current. Below the base speed, it operates at the current angle with the highest torque / current utilization rate, and dynamically calculates the optimal d-axis current. When the speed approaches 95% of the base speed, field weakening begins, gradually reducing the d-axis positive current and limiting the q-axis current to no more than 80% of the rated current to avoid saturation and overheating. When the temperature rises to the 90℃ alarm threshold, it automatically switches to torque protection mode, reducing the d-axis current by 10% and the overall thrust by 5%.

[0148] In an optional implementation, when the vertical clearance, horizontal clearance, and distance safety margin coefficients are all greater than the exit threshold (e.g., 0.70) for two consecutive control cycles and all safety indicators return to normal, the flight controller issues a mode deactivation command:

[0149] (1) The trajectory generator stops updating the obstacle avoidance curve and switches to the original route tracking curve;

[0150] (2) The thrust allocation matrix is ​​restored to the normal cruise weights;

[0151] (3) The ESC exits the field weakening or load reduction state and returns to standard MTPA control to achieve a smooth transition.

[0152] Preferably, by judging the obstacle avoidance action type output by the model, the corresponding trajectory generator is called to calculate the target's vertical velocity, lateral velocity, and heading command in real time, and the propulsion motor is driven to perform maneuvering actions in real time. After the obstacle threat is removed, the obstacle avoidance mode is automatically exited and the original planned route is restored. This realizes the dynamic optimization of trajectory and the smooth transition of body attitude during the obstacle avoidance process, ensuring that the UAV can quickly restore the mission route while maneuvering to avoid obstacles, thus improving flight efficiency and airframe flight stability.

[0153] In addition to the above embodiments, other aspects of the present invention also disclose an obstacle avoidance system for unmanned aerial vehicles, including one or more processors and a memory.

[0154] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the obstacle avoidance method for a drone as described in the foregoing embodiments, in particular... Figure 1 The flowchart of the method is shown.

[0155] Other aspects disclosed in the embodiments of the present invention also provide a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the obstacle avoidance method for a drone described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0156] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0157] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0158] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0159] In any case, the language can be either compiled or interpreted.

[0160] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0161] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0162] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0163] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0164] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0165] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An obstacle avoidance method for unmanned aerial vehicles (UAVs), characterized in that, include: During the UAV flight control cycle, three-dimensional point cloud and image information from lidar, binocular vision and millimeter-wave radar are collected simultaneously. The collected information is then unified by external parameter calibration and timestamp alignment to obtain a set of comprehensive obstacle parameters describing the relative height, horizontal gap and relative distance of the obstacle. Based on the current attitude, height, maximum diagonal size, and preset safety distance of the UAV, the set of comprehensive obstacle parameters is processed to be dimensionless, and three characteristic coefficients reflecting vertical clearance, horizontal clearance, and distance safety margin are obtained respectively, and they are combined into an obstacle avoidance feature vector. The obstacle avoidance feature vector is input into a discrimination model constructed based on a dual-membership adaptive fuzzy decision tree. The discrimination model uses two membership functions to perform parallel fuzzy mapping on the vertical clearance and horizontal clearance, with distance safety margin as the safety weight, and outputs the obstacle avoidance action type according to the obstacle avoidance rules. Based on the obstacle avoidance action type, the corresponding trajectory generator is invoked to calculate the target's vertical speed, lateral speed, and heading command in real time, and the propulsion motor is driven to complete the maneuver. When the vertical clearance coefficient, horizontal clearance coefficient, and distance safety margin all exceed the preset exit threshold again, the obstacle avoidance mode is automatically terminated and the original planned route is restored.

2. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, The output of obstacle avoidance action types according to obstacle avoidance rules includes: When the vertical clearance coefficient is higher than the first threshold and the distance safety margin is higher than the second threshold, choose the descent or crossing action. When the vertical clearance coefficient is lower than the third threshold and the horizontal clearance coefficient is higher than the fourth threshold, choose the ascending or overtaking maneuver. When the horizontal clearance coefficient is lower than the fifth threshold and the distance safety margin is higher than the sixth threshold, choose lateral or detour maneuvers. When the conditions for descent, crossing or ascent, and overflight are met simultaneously, and the conditions for lateral or circling are also met, a combined maneuver is selected.

3. The obstacle avoidance method for unmanned aerial vehicles according to claim 1 or 2, characterized in that, Based on the obstacle avoidance action type, the corresponding trajectory generator is invoked to calculate the target's vertical velocity, lateral velocity, and heading commands in real time, including: Read the obstacle avoidance action type output by the discrimination model, and establish a correspondence table between the obstacle avoidance action type and the current flight mode of the UAV; Based on the correspondence table, a trajectory template matching the obstacle avoidance action type is retrieved, and the trajectory template records the reference displacement curves according to the vertical axis, horizontal axis and heading axis respectively; The curve is recalibrated by combining the current attitude, terrain height, and local wind field intensity to ensure that the obstacle avoidance trajectory is compatible with environmental constraints; The recalibrated triaxial displacement curves are first and second differentiated to obtain the velocity curves. Then, the target's vertical velocity, lateral velocity, and rate of change of heading angle are extracted based on the smoothing results of the velocity curves. With the attitude stability of the UAV as a constraint, extreme value corrections are performed on the target's vertical velocity, lateral velocity, and heading angle change rate to generate three-channel control commands that can be issued.

4. The obstacle avoidance method for unmanned aerial vehicles according to claim 3, characterized in that, The propulsion motor drives the movement, including: The three-channel control commands are mapped to the desired thrust distribution values ​​for each propulsion motor, and the commands are distributed to each propulsion motor according to the thrust distribution matrix. After the allocation is completed, the field orientation control module of the motor controller is invoked to convert the desired thrust allocation value into the corresponding current reference value; Based on the current motor speed and temperature rise, the maximum torque-to-current ratio control strategy is selected, and the current reference value is adjusted in real time so that the output thrust of the propulsion motor can keep up with the three-channel control commands. During the maneuver, the motor current, speed and vibration amplitude are continuously monitored. If any indicator is detected to exceed the safety limit, load reduction compensation is triggered, the thrust distribution coefficient is automatically reduced, and the remaining obstacle avoidance actions are completed within the safety threshold. Once all monitored indicators have returned to normal levels, and the discrimination model confirms that the obstacle avoidance maneuver has ended, the propulsion motor will resume operation according to the thrust requirements of the original planned route.

5. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, The obtained set of comprehensive obstacle parameters includes: Using lidar, the space in front is scanned in layers to obtain the initial point cloud outline of obstacles; The binocular vision module is invoked to perform depth recovery of obstacle textures, and the shape error of the lidar point cloud is corrected by feature point matching. Millimeter-wave radar is used to perform Doppler ranging in the direction of the obstacle's center to eliminate distance drift caused by obstruction and weak reflection; The corrected data is fused according to a unified coordinate system to extract the three-dimensional coordinates of the top, bottom and side boundary points of the obstacle, and then the relative height difference and horizontal gap of the obstacle are calculated. By combining the current position of the UAV with the center coordinates of the obstacle, the relative distance to the obstacle is calculated, and the timestamp and confidence level are recorded simultaneously to form the comprehensive parameter set of the obstacle.

6. The obstacle avoidance method for unmanned aerial vehicles according to claim 1, characterized in that, The obstacle avoidance feature vector is obtained by: Using the aircraft height as a normalized benchmark, the relative height difference between the top of the obstacle and the top of the drone is converted into a vertical clearance coefficient; Using the maximum diagonal size of the drone as a normalization benchmark, the horizontal clearance of obstacles is converted into a horizontal clearance coefficient; Using a preset safe distance as a normalization benchmark, the relative distance to obstacles is converted into a distance safety margin coefficient; The vertical clearance coefficient, the horizontal clearance coefficient, and the distance safety margin coefficient are arranged in sequence and linear normalization is performed to obtain a vector representation of fixed length. A sensor data fusion confidence label is appended to the end of the vector representation to form the obstacle avoidance feature vector.

7. The obstacle avoidance method for unmanned aerial vehicles according to claim 1 or 6, characterized in that, The obstacle avoidance feature vector is input into a discrimination model constructed based on a dual-membership adaptive fuzzy decision tree, including: Parallel fuzzy mapping of the vertical clearance coefficient and the horizontal clearance coefficient was performed using trigonometric membership functions and Gaussian membership functions respectively, resulting in two sets of membership degree values. Based on the distance safety margin coefficient, safety weights are assigned to the two sets of membership values ​​to form a weighted and fused comprehensive membership degree. The comprehensive membership degree is compared with the obstacle avoidance rule matching table. If the comprehensive membership degree does not belong to any single action range, fuzzy interpolation is performed to obtain the closest action type. The confirmed obstacle avoidance action type is output and passed to the trajectory generator to ensure that the UAV completes obstacle avoidance according to the optimal action path.

8. The obstacle avoidance method for unmanned aerial vehicles according to claim 7, characterized in that, The discrimination model includes a security assessment layer, a fuzzy mapping layer, and an action decision layer, wherein: The security assessment layer divides the environment into high-security zones, warning zones, and danger zones according to the distance safety margin coefficient, and updates the thresholds of each zone in real time to provide dynamic security weights. The fuzzy mapping layer consists of vertical dimension nodes and horizontal dimension nodes. The two nodes use triangular membership functions and Gaussian membership functions in parallel to perform dual membership mapping on the vertical and horizontal net space coefficients and output two sets of membership degrees. After receiving the weights from the safety assessment layer and the membership set from the fuzzy mapping layer, the action decision layer first performs weighted fusion, and then calls the obstacle avoidance rule matching table to determine the final action type. The obstacle avoidance rule matching table is initially trained offline using a genetic algorithm, and then adaptively fine-tuned during flight using a gradient descent strategy based on real-time environmental factors to ensure the long-term stability of the model.

9. An obstacle avoidance system for unmanned aerial vehicles (UAVs), characterized in that, include: One or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of an obstacle avoidance method for an unmanned aerial vehicle as described in any one of claims 1 to 8.

10. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of an obstacle avoidance method for an unmanned aerial vehicle as described in any one of claims 1 to 8.

Images