Dynamic obstacle avoidance and emergency landing methods for drones

Abstract

This invention discloses a dynamic obstacle avoidance and emergency landing method for unmanned aerial vehicles (UAVs), including input and potential field map construction, dynamic estimation of the landing target point, optimized artificial potential field path planning, trap state identification and processing mechanism, and path smoothing and output. This UAV dynamic obstacle avoidance and emergency landing method intelligently identifies the landing target point, combines visual information and altitude data to dynamically assess the environment, select a safe landing point, and optimizes path planning: by using an improved artificial potential field method, it avoids local minima problems, achieving efficient path planning; it has high computational efficiency: compared to deep learning methods, it has lower computational resource requirements, is suitable for real-time use, and adapts to dynamic environments: it can handle dynamic obstacles, improving the safety and reliability of emergency landings.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a dynamic obstacle avoidance and emergency landing method for UAVs. Background Technology

[0002] As the global low-altitude economy enters a new stage of large-scale development, drones, as core carriers, are widely used in logistics, inspection, emergency rescue and other fields. The ability to autonomously avoid obstacles and make emergency landings in complex environments has become a key safety requirement.

[0003] Traditional drones rely heavily on GPS positioning and preset landing points. However, in urban canyons, forests, or dynamic obstacle scenarios, GPS signals are easily interfered with, and preset landing points may be temporarily occupied or pose safety hazards. Existing vision-based obstacle avoidance solutions still face significant technical bottlenecks in terms of real-time performance, dynamic environment adaptability, and multi-sensor fusion efficiency, resulting in a low success rate for emergency landings and severely restricting the reliable deployment of drones in high-risk scenarios. For example, existing drone obstacle avoidance patents mostly focus on obstacle detection and avoidance, employing technologies such as ultrasonic sensors, radar, and deep learning. (Publication number C) Chinese patent N119126824A discloses a trajectory planning method for obstacle avoidance in unmanned aerial vehicles (UAVs). This method combines three-dimensional and two-dimensional Gaussian-Markov models to generate random trajectories, uses ultrasonic sensors to detect obstacles, and adjusts the heading and pitch angles based on a repulsive field to achieve obstacle avoidance. Another example is Chinese patent CN202410647520.5, which discloses a UAV obstacle avoidance method, system, and medium based on deep reinforcement learning. This method utilizes panoramic images and a deep reinforcement learning model to extract obstacle features, predict collision time, and dynamically adjust flight parameters to achieve obstacle avoidance.

[0004] Based on the above materials, it can be seen that existing drone obstacle avoidance patents mainly focus on obstacle detection and avoidance, employing technologies such as ultrasonic sensors, radar, and deep learning. However, these solutions generally suffer from the following shortcomings:

[0005] 1. Lack of intelligent identification and assessment of landing target points: Most solutions do not address how to select a safe landing point in an emergency, thus failing to meet the needs of emergency landing.

[0006] 2. Some solutions use simple repulsive fields or height adjustments, which are difficult to cope with dynamic changes in complex environments.

[0007] 3. High computational resource requirements and poor real-time performance: Deep learning and other methods have high computational resource requirements, which may affect real-time performance and energy consumption control. Summary of the Invention

[0008] The purpose of this invention is to provide a dynamic obstacle avoidance and emergency landing method for unmanned aerial vehicles (UAVs) to solve the problems of lack of safe landing points, insufficient dynamic adaptability to complex environments, and limited computing resources in emergency landing scenarios mentioned above.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a dynamic obstacle avoidance and emergency landing method for unmanned aerial vehicles (UAVs), comprising the following steps:

[0010] S1. Input and Potential Map Construction: Using an image captured by a downward-looking camera, with a size of M×N, potential values ​​are assigned to each region through image preprocessing, resulting in an M×N matrix. The value of each element represents the potential field value of the region where the corresponding pixel is located;

[0011] S2. Dynamic estimation of landing target point: Constructing a binary security mask → Distance transformation → Connected component analysis → Multidimensional security scoring → Optimal landing point selection;

[0012] S3. Optimize artificial potential field path planning: Initialization → Iterative loop → Output path;

[0013] S4. Trap Status Identification and Handling Mechanism: Trap Judgment Logic → Virtual Repulsive Force Disturbance Escape Mechanism;

[0014] S5, Path Smoothing and Output: Smooth the obtained route and then output it.

[0015] Preferably, the potential field value function in step S1 is: Segment is an image segmentation model that splits regions in an image. Regions with higher risk have larger potential values.

[0016] Preferably, in step S2:

[0017] Construct a binary security mask: ,

[0018] in Represents the set of pixels in the safe region. Representing the potential field map in pixels The potential energy value at that point, Indicates the obstacle threshold;

[0019] Distance transformation: for mask Perform a Euclidean distance transformation on the complement of the set to obtain a distance map. ,

[0020] ;

[0021] Connected component analysis: In Perform connected component labeling within the component and identify the largest connected component. ;

[0022] Multidimensional security score: for each Calculate the safety score.

[0023]

[0024] in These represent the potential field, distance, and region penalty weights in the security scoring function, respectively. This is the maximum distance score.

[0025] Optimal landing point selection: Choose the point with the lowest score.

[0026]

[0027] in To determine the final safe landing target point, once the target point is determined, path planning begins using the target point as input.

[0028] Preferably, in step S3:

[0029] When initializing the iteration counter and path list,

[0030] Iteration counter t:

[0031]

[0032] Path list:

[0033]

[0034] When in the loop iteration If the termination condition is not triggered, repeat the sub-step;

[0035] After the sub-step, The planned path.

[0036] Preferably, the sub-step includes:

[0037] Calculate the gravitational vector:

[0038] Target For the current position Generate attraction

[0039]

[0040] in For gravity, For gravity weighting coefficients, The target point location, Current position;

[0041] Calculate the repulsion vector:

[0042] In radius Within the area, for each obstacle pixel Accumulation

[0043]

[0044] in As a repulsive force, The repulsive force intensity coefficient, To prevent division by zero of small constants;

[0045] Resultant force and adaptive step size:

[0046] Calculate the resultant force. ,

[0047] Calculate the step size. ,

[0048] in The resultant force vector, As the weights of attraction and repulsion, Step size, The base step size is proportional to the image size. The minimum allowable step size, This is the step size scaling factor;

[0049] Location update:

[0050]

[0051] Will After cropping to the boundary, append it to the path list. ;

[0052] Termination decision:

[0053] After each iteration, immediately exit the loop if any of the following conditions are met.

[0054] Reach the destination

[0055] The combined force approaches zero.

[0056] Trap state, see step S4

[0057] The maximum number of iterations has been reached. ;

[0058] Iteration count accumulation:

[0059] .

[0060] Preferably, in step S4:

[0061] In the "termination determination" phase of each iteration, if a trap state is detected, a virtual repulsion perturbation and jump are performed to move to a new position. Recorded Then, immediately revert the sub-step, increment the iteration count t, and continue iterating. The specific trap identification and escape mechanism is as follows:

[0062] The trap detection logic evaluates the following three metrics in each iteration:

[0063] a) A small potential field variance indicates that the area has entered a flat region.

[0064]

[0065] in Indicates recent The potential field value sequence of the step, The threshold for determining potential field fluctuations;

[0066] b. The positional movement is extremely small, indicating no significant advancement.

[0067]

[0068] in The threshold for determining the range of motion;

[0069] c. Path cyclical oscillation, if it exists Make If it is a periodic oscillation, it is considered to satisfy any two of the above conditions. ;

[0070] The virtual repulsive perturbation escape mechanism triggers the following escape operation when a trap state is detected:

[0071] Construct a disturbance steering mechanism at the current location Add a virtual repulsion term to the current position.

[0072]

[0073] in The virtual repulsive disturbance intensity coefficient. This is a unit-direction perturbation vector used to construct the jump direction;

[0074] Increase the search jump step size. ,in This represents the jump increase coefficient;

[0075] Reconstructing the current direction of progress ;

[0076] This disturbance operation is triggered only once. If the user escapes the trap afterward, the process will return to normal.

[0077] Preferably, in step S5:

[0078] The path list generated during the smoothing process:

[0079]

[0080] Apply window size to path Moving average:

[0081]

[0082] Obtain the smoothed path .

[0083] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0084] 1. Intelligently identify landing target points, combine visual information and altitude data to dynamically assess the environment, and select safe landing points.

[0085] 2. Optimized path planning: By using an improved artificial potential field method, the problem of local minima is avoided, and efficient path planning is achieved.

[0086] 3. High computational efficiency: Compared with deep learning methods, it has lower computational resource requirements and is suitable for real-time use.

[0087] 4. Adaptable to dynamic environments: It can handle dynamic obstacles, improving the safety and reliability of emergency landings. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the overall system flow of the present invention. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Example 1: Please refer to Figure 1 This embodiment specifically discloses:

[0091] (1) Input and potential field diagram construction

[0092] (1.1) Image taken using a downward-facing camera, size By preprocessing the image, a potential field value is assigned to each region, resulting in a... matrix Each element's value represents the potential field value of the region where the corresponding pixel is located. Specifically,

[0093]

[0094] Segmentation is an image segmentation model that splits regions in an image. Regions with higher risk have larger potential field values.

[0095] (2) Dynamic estimation of landing target point

[0096] (2.1) Constructing a binary security mask:

[0097]

[0098] in Represents the set of pixels in the safe region. Representing the potential field map in pixels The potential energy value at that point, This indicates the obstacle threshold.

[0099] (2.2) Distance Transformation

[0100] For the mask Perform a Euclidean distance transformation on the complement of the set to obtain a distance map.

[0101]

[0102] (2.3) Connected Component Analysis

[0103] exist Perform connected component labeling within the component and identify the largest connected component.

[0104] (2.4) Multidimensional security score

[0105] For each Calculate the safety score:

[0106]

[0107] in These represent the potential field, distance, and region penalty weights in the security scoring function, respectively. Maximum distance rating

[0108] (2.5) Optimal landing point selection

[0109] Take the lowest score:

[0110]

[0111] in This is the final selected safe landing target point. Once the target point is determined, path planning begins using the target point as input.

[0112] (3) Optimize the path planning of artificial potential fields

[0113] This invention addresses the problem that traditional artificial potential field methods are prone to getting trapped in local minima during path planning. It proposes a multimodal escape mechanism and active judgment strategy that can intelligently identify trap situations and guide the path to break through inefficient oscillation regions, thereby improving the success rate of planning and the quality of the path.

[0114] This method does not rely on map reconstruction or global path planning. Instead, it designs a dynamic discrimination mechanism based on multi-source features (potential field change rate, position movement amplitude, path history backtracking, etc.) and is supplemented by a local jump perturbation + virtual repulsion reconstruction mechanism to escape traps. It is suitable for real-world scenarios with no prior map and dynamic changes, and has high real-time performance and high adaptability.

[0115] (3.1) Initialization

[0116] Initialize the iteration counter and path list, where

[0117] Iteration counter t:

[0118]

[0119] Path list:

[0120]

[0121] (3.2) Iterative Loop

[0122] when If the termination condition is not triggered, repeat the following sub-steps:

[0123] (3.2.1) Calculate the gravitational vector:

[0124] Target For the current position Generate attraction:

[0125]

[0126] in For gravity, For gravity weighting coefficients, The target point location, Current location

[0127] (3.2.2) Calculate the repulsion vector:

[0128] In radius Within the area, for each obstacle pixel Accumulation:

[0129]

[0130] in As a repulsive force, The repulsive force intensity coefficient, To prevent small constants from being divided by zero

[0131] (3.2.3) Resultant force and adaptive step size

[0132] Calculate the resultant force:

[0133]

[0134] Calculate the step size:

[0135]

[0136] in The resultant force vector, As the weights of attraction and repulsion, Step size, The base step size is proportional to the image size. The minimum allowable step size, Step scaling factor

[0137] (3.2.4) Location Update

[0138]

[0139] Will After cropping to the boundary, append to the path list.

[0140]

[0141] (3.2.5) Termination Judgment

[0142] After each iteration, immediately exit the loop if any of the following conditions are met:

[0143] 1. Reaching the destination:

[0144] 2. The net force approaches zero:

[0145] 3. Trap state (see the judgment logic in step 4)

[0146] 4. The maximum number of iterations has been reached:

[0147] (3.2.6) Iteration count accumulation

[0148]

[0149] (3.3) Output path

[0150] After the loop ends, For the planned path

[0151] (4) Trap State Identification and Handling Mechanism

[0152] In the "termination determination" phase of each iteration, if a trap state is detected, a virtual repulsion perturbation and jump are performed to move to a new position. Recorded Then immediately revert to step 3.2.6, increase the iteration count t, and continue iterating.

[0153] Traditional artificial potential field methods are prone to getting trapped in local minima in complex environments. To address this issue, this invention introduces a trap identification and escape mechanism, as detailed below:

[0154] (4.1) Trap detection logic

[0155] The following three metrics are evaluated in each iteration:

[0156] - The potential field variance is too small (indicating that it has entered a flat region):

[0157]

[0158] in Indicates recent The potential field value sequence of the step, Threshold for determining potential field fluctuations

[0159] - Minimal positional movement (indicating no significant advancement):

[0160]

[0161] in Threshold for determining the range of motion

[0162] - Path cyclical oscillation: If it exists Make This indicates a periodic oscillation.

[0163] If any two of the above conditions are met, then it is considered that:

[0164]

[0165] (4.2) Virtual repulsive disturbance escape mechanism

[0166] When a trap is detected, the following escape operation is triggered:

[0167] - Construct a disturbance steering mechanism at the current location Add a virtual repulsion term to the current position:

[0168]

[0169] in The virtual repulsive disturbance intensity coefficient. The unit direction perturbation vector is used to construct the jump direction.

[0170] - Increase the search jump step size:

[0171]

[0172] in Jump Amplification Coefficient

[0173] - The current direction of restructuring is as follows:

[0174]

[0175] - This disturbance operation is triggered only once. If the trap is escaped afterward, normal progress will resume.

[0176] (5) Path smoothing and output

[0177] (5.1) Smoothing

[0178] The path list generated by iteration

[0179]

[0180] Apply window size to path Moving average:

[0181]

[0182] Obtain the smoothed path .

[0183] Example 2: This example provides the following specific application scenario (in an urban environment, drones need to cope with complex scenarios such as building obstructions, dynamic vehicles, and temporary obstacles, such as pedestrians and construction equipment):

[0184] 1. Construction of potential field diagram:

[0185] Real-time ground images are captured using a downward-looking camera (1920×1080 resolution) and input into a pre-trained SegNet segmentation model (Segment) to generate a potential field matrix. The potential field value of hazardous areas (such as vehicles and obstacles) is set to... ≥150, safe zone <50, parameter settings: =50, danger zone threshold.

[0186] 2. Dynamic estimation of landing point:

[0187] Constructing a binary security mask Perform Euclidean distance transformation on the unsafe region (U≥50) to generate distance map D;

[0188] Perform a connected component analysis on S and select the largest connected component Smax (with an area exceeding 60%).

[0189] Calculate security score ,in =200 pixels;

[0190] Select the coordinate with the lowest score. = (i,j) is used as the landing point (e.g., the lowest score point is located on an open sidewalk).

[0191] 3. Path planning and trap handling:

[0192] Initialize the gravitational weight k=2.0, the repulsive force strength η=1.5, and the step size. base=10 pixels;

[0193] When a trap state (potential field variance) is detected )5 And the magnitude of positional movement (5 pixels), triggering virtual repulsive perturbation, direction vector λ=0.8, step size increment. =0.5;

[0194] Output a smooth path (moving average window k=5), avoiding dynamic vehicles and landing. .

[0195] After verification, in a simulated urban environment, the drone successfully identified the temporarily occupied original preset landing point, dynamically switched to the safe area, the path planning time was less than 200ms, and the obstacle avoidance success rate was improved to 98%.

[0196] Example 3: This example provides the following specific application scenario (the forest environment has complex terrain, with static obstacles such as trees and rocks, and sudden weather interference, such as strong winds):

[0197] 1. Construction of potential field diagram:

[0198] A downward-looking camera (infrared mode) captures ground images, and a segmentation model identifies vegetation (potential field value). =120), Rock (potential field value) =180), safe zone <30.

[0199] 2. Landing point selection:

[0200] The largest connected component in the safety mask S is the forest clearing, and the weights of the scoring function are adjusted to ωv=0.4, ωd=0.5, ωr=0.1 (emphasizing distance from obstacles).

[0201] Optimal landing point Located in the center of an open space 150 pixels from the nearest tree;

[0202] 3. Path optimization and getting out of trouble:

[0203] Set the repulsive force radius rinf = 100 pixels, the attraction weight α = 1.2, and the repulsive force weight β = 0.8;

[0204] When the path experiences local oscillations due to dense trees, a perturbation mechanism is triggered, with the virtual repulsive force direction along the gaps between trees δi=[0.7,0.3], and the jump step size increasing to Δs′=15 pixels;

[0205] After smoothing the path, it meanders down along the gaps in the tree canopy, avoiding areas of strong wind.

[0206] After verification, in a simulated forest scenario, the drone successfully avoided dense trees and landed in a safe area. The path length was shortened by 30% compared to the traditional potential field method, and the calculation time was reduced by 40%.

[0207] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An unmanned vehicle dynamic obstacle avoidance emergency landing method, characterized in that, Comprising the following steps: S1, input and potential field map construction: using the image image captured by the downward-looking camera, size MxN, through image preprocessing, each region is assigned a potential field value, obtaining a MxN matrix , the value of each element represents the potential field value of the region where the corresponding pixel is located; S2, dynamic estimation of landing target point: construct binary safety mask → distance transform → connected component analysis → multi-dimensional safety scoring → optimal landing point selection; S3, optimization of artificial potential field path planning: initialization → loop iteration → output path; In the step S3, when initializing the iteration counter and the path list, Iteration counter t: ； Path list: ； When the loop iteration is in the , and the termination condition is not triggered, repeat the sub-step; sub-step, for the planned path; The sub-steps include: Calculation of attractive force vector: Objectives To the current location Generating attraction ； wherein is the gravitational force, is the gravitational weight coefficient, is the target point position, is the current position; Calculation of repulsive force vector: Within a radius For each obstacle pixel Accumulate ； wherein is the repulsive force, is the repulsive force strength coefficient, is a small constant to prevent division by zero; Combined force and adaptive step size: calculating the resultant force, , Computing the step length, , wherein is the resultant vector, is the weight of the attractive and repulsive forces, is the step size, is the base step size, proportional to the image size, is the minimum step size allowed, is the step size scaling factor; Position update: ； cut to the border, append to path list, cut to the border, append to path list, ; Termination criterion: After each iteration, check any of the following to immediately exit the loop, Arrive at target, ; the resultant force tends to zero, ; Trap state, see step S4 the number of iterations reaches an upper limit, ; Iteration count accumulation: ； S4, trap state identification and processing mechanism: trap discrimination logic → virtual repulsive force disturbance escape mechanism; In the step S4, At the "termination decision" stage of each iteration, if a trap state is detected, a virtual repulsive perturbation and jump is performed, with the new position recorded to , and immediately followed by a sub-step that increments the iteration count t and continues the iteration, with the specific trap identification and escape mechanism being: Trap discrimination logic, evaluate the following three indicators in each iteration: a. The potential field variance is too small, indicating that it is trapped in a flat area ； wherein represents the most recent the sequence of potential field values of the step, is a potential field fluctuation decision threshold; b. The position movement amplitude is extremely small, indicating that it has not made significant progress ； wherein is a movement amplitude determination threshold value; c. Path loop oscillation, if exists such that is periodic oscillation, mainly satisfying any two of the above conditions, i.e. ; Virtual repulsive force disturbance escape mechanism, when the trap state is detected, trigger the following escape operation: Disturbance direction machine for current position construction , virtual repulsive force term to current position ； wherein is a virtual repulsive force perturbation strength coefficient, is a unit directional perturbation vector used to construct the jump direction; increase the search hop step size, wherein is a hop increment factor; Reconfigure the current thrust direction to ; This disturbance operation is triggered only once, and if it escapes from the trap, it will return to normal propulsion; S5, path smoothing and output: the obtained route is smoothed and output.

2. The dynamic obstacle avoidance emergency landing method of the UAV according to claim 1, characterized in that: The potential field value function in step S1 is where Segment is an image segmentation model that splits regions in the image, and the higher the danger level of a region, the greater the potential field value.

3. The dynamic obstacle avoidance emergency landing method of the UAV according to claim 1, characterized in that: In step S2, Constructing a binary security mask: , wherein represents a set of safety region pixels, represents a potential energy value of the potential field map at a pixel , represents an obstacle threshold value; Distance transform: Perform a Euclidean distance transform on the complement of the mask to get a distance map ,​ ； Connected component analysis: In connected component labeling is performed to identify the largest connected component ; Multi-dimensional security score: for each computing a security score, ； wherein respectively represent the potential field, distance, and zone penalty weight in the safety score function, is the distance score upper limit; Optimal landing point selection: take the point with the minimum score, ； wherein is the final selected safe landing target point, after the target point is determined, the target point is taken as an input, and path planning is started.

4. The dynamic obstacle avoidance emergency landing method of the UAV according to claim 1, characterized in that: In step S5, During the smoothing process, the path list generated by iteration: ； Applying a window size of 5 to the path path sliding average: ； smoothed path .

Images