Unmanned aerial vehicle path planning system and method based on fusion of apf-rrt* and genetic algorithm

Abstract

The application provides a UAV path planning system and method based on fusion of APF-RRT* and genetic algorithm, and belongs to the technical field of autonomous flight of UAVs. An improved APF-RRT* algorithm is used for path searching to generate a collision-free initial path from a starting point to a target point. The improved APF-RRT* algorithm includes intelligent sampling by fusing a region probability sampling strategy and a dynamic target bias strategy to guide the growth of a random tree; and adaptively adjusting an expansion step length according to the obstacle density around a current node. Path points of the initial path are used as an initial population of a genetic algorithm, and a final path meeting the dynamic constraint of a UAV is generated through iterative optimization of the genetic algorithm. The fitness function of the genetic algorithm considers path length, collision conditions and whether a path turning angle exceeds a constraint. The application improves planning efficiency and path quality, guarantees the executability of the final path, and balances optimization speed and path quality.

Description

Technical Field

[0001] This application relates to the field of autonomous flight technology for unmanned aerial vehicles (UAVs), and in particular to a UAV path planning system and method that integrates APF-RRT* and genetic algorithm. Background Technology

[0002] In autonomous flight missions of unmanned aerial vehicles (UAVs), path planning is one of the core technologies. Its goal is to plan an optimal or suboptimal collision-free route from the starting point to the target point in a given flight environment. Among numerous path planning algorithms, the sampling-based fast expanding random tree algorithm and its asymptotically optimal variant algorithm are widely studied and applied due to their good adaptability to high-dimensional spaces and probabilistic completeness.

[0003] However, the standard asymptotically optimal fast expanding random tree algorithm uses a completely random sampling method when exploring the space. When the UAV faces a complex environment with dense obstacles or narrow passages, this random exploration method generates a large number of invalid sampling points, resulting in low expansion efficiency of the random tree, excessively long path search time, and difficulty in meeting real-time requirements. To improve this problem, a two-stage framework of "coarse search-fine optimization" has emerged in the prior art. For example, some schemes first use fast expanding random tree algorithms to quickly generate a collision-free initial path connecting the starting point and the target point, and then use optimization algorithms such as genetic algorithms to smooth and optimize the initial path to meet the dynamic constraints of the UAV. Although this framework improves the path quality to some extent, it still has inherent defects. First, the efficiency problem of the first stage coarse search is not fundamentally solved. If only the standard or simply improved fast expanding random tree algorithm is used, its random exploration nature will still lead to inconsistent quality of the generated initial paths, and even excessive time consumption, which brings a heavy computational burden to the second stage optimization. Secondly, the low quality of the initial population provided to the genetic algorithm (i.e., the initial path generated in the first stage) leads to a longer evolutionary cycle before converging to a satisfactory solution, limiting the overall efficiency improvement. Therefore, significantly improving the efficiency of the first-stage coarse search and the quality of the initial path is a critical technical bottleneck that needs to be addressed in this approach. Summary of the Invention

[0004] The purpose of this application is to provide a UAV path planning system and method that integrates APF-RRT* and genetic algorithm, aiming to solve the problems of low search efficiency and poor initial path quality caused by random exploration in the existing technology, thereby resolving the contradiction between search efficiency and path feasibility.

[0005] To achieve the above objectives, this application provides a UAV path planning method that integrates APF-RRT* and genetic algorithm, including: a data loading step: obtaining environmental map information, task parameter information, and UAV performance constraint parameters;

[0006] The efficient coarse search step is performed: based on the environmental map information and task parameter information, the improved APF-RRT* algorithm is used to perform path search and generate a collision-free initial path from the starting point to the target point. The improved APF-RRT* algorithm includes intelligent sampling that integrates regional probability sampling strategy and dynamic target bias strategy to guide the growth of random tree; and adaptively adjusting the expansion step size according to the obstacle density around the current node.

[0007] The constraint-based optimization step is as follows: the path points of the initial path are used as the initial population of the genetic algorithm. The genetic algorithm iteratively optimizes to generate the final path that satisfies the UAV dynamics constraints. The fitness function of the genetic algorithm at least considers the path length, collision situation and whether the path turning angle exceeds the constraints.

[0008] Step to output the final path: Output the final path.

[0009] Optionally, the intelligent sampling specifically includes: employing the regional probability sampling strategy to prioritize sampling in the region close to the line connecting the starting point and the target point; and employing the dynamic target bias strategy to dynamically adjust the probability of sampling toward the target point based on the distance between the current node and the target point and the density of surrounding obstacles.

[0010] Optionally, the adaptive adjustment of the expansion step size specifically includes: calculating the expansion step size using a function related to the local obstacle density, such that the expansion step size decreases in areas with dense obstacles and increases in areas with open areas.

[0011] Optionally, the genetic algorithm includes performing evolutionary operations such as tournament selection, crossover, and mutation on the initial population.

[0012] Optionally, the fitness function is expressed as:

[0013]

[0014] in, This is the total path length item. This is a comprehensive penalty item;

[0015] The comprehensive penalty items At least including:

[0016] Collision penalty is used to impose a penalty on the path that collides with an obstacle;

[0017] Angle constraint penalty term is used to impose a penalty on paths where the turning angle formed by two consecutive path segments exceeds the limit.

[0018] Optionally, in the dynamic target bias strategy, the target bias probability Calculated using the following formula:

[0019]

[0020] in, It is the Euclidean distance between the newly expanded node in the current random tree and the target point; It is the local obstacle density within a specified range around the node; The weights for the distance to the target point; To control the decay rate of the distance to the target point; The weights for the local obstacle density; To control the decay rate of local obstacle density.

[0021] Optionally, in the adaptive adjustment of the expansion step size, the new expansion step size The calculation is performed using the following formula:

[0022]

[0023] in, It is the set minimum expansion step size; This is the step size adjustment coefficient; To control the steepness of the hyperbolic tangent function; The local obstacle density varies from 0 to 1. This is the center point of the step size change.

[0024] Optionally, the efficient coarse search step further includes a tree construction and optimization process, specifically including:

[0025] Step A: In the currently generated random tree, find the node closest to the sampling point;

[0026] Step B: Starting from the nearest node, expand a new node along the direction toward the sampling point with the calculated adaptive step size;

[0027] Step C: Perform collision detection on the path segments connecting the nearest node to the new node;

[0028] Step D: If there is no collision, add the new node to the random tree.

[0029] Optionally, after step D, a path optimization step is also included:

[0030] Step E: Reselect the parent node, check the neighboring nodes within a set radius around the new node, if connecting to the new node through one of the neighboring nodes can reduce the path cost of the new node, then update that neighboring node as the parent node of the new node.

[0031] Step F: Reconnection operation. Check if a shorter path can be provided to the neighboring node through the new node. If so, update the new node as the parent node of the neighboring node.

[0032] To achieve the above objectives, this application also provides a UAV path planning system that integrates APF-RRT* and a genetic algorithm, comprising:

[0033] Input module: Configured to acquire environmental map information, mission parameter information, and UAV performance constraint parameters;

[0034] The first-stage path search module is configured to perform path search using an improved APF-RRT* algorithm based on the environmental map information and task parameter information, generating a collision-free initial path from the starting point to the target point. The improved APF-RRT* algorithm includes intelligent sampling that integrates a regional probability sampling strategy and a dynamic target bias strategy to guide the growth of the random tree; and adaptively adjusting the expansion step size according to the obstacle density around the current node.

[0035] The second-stage path optimization module is configured to use the path points of the initial path as the initial population of the genetic algorithm, and generate a final path that satisfies the dynamic constraints of the UAV through iterative optimization of the genetic algorithm. The fitness function of the genetic algorithm at least considers the path length, collision situation and whether the path turning angle exceeds the constraints.

[0036] Output module: configured to output the final path.

[0037] The UAV path planning method integrating APF-RRT* and genetic algorithm provided by this invention has the following significant advantages compared to existing technologies:

[0038] I. Significantly Improved Path Planning Efficiency and Real-Time Performance: By integrating regional probability sampling and dynamic target bias strategies for intelligent sampling in the first stage (coarse search), and introducing an adaptive expansion step size mechanism based on local obstacle density, the guidance and environmental adaptability of random tree expansion are significantly enhanced. This method can concentrate search computational resources on more promising areas, effectively reducing invalid sampling and redundant exploration in obstacle-dense or narrow passage environments. Simulation experiments show that in complex and narrow scenarios, the path search time of this method is significantly reduced compared to traditional RRT and APF-RRT algorithms (e.g., in narrow scenarios, the time is reduced by approximately 84.2% and 64.3%, respectively), greatly improving planning efficiency and better meeting the needs of real-time online planning for UAVs.

[0039] Second, generating a high-quality initial path lays a solid foundation for subsequent optimization: The improved APF-RRT* algorithm not only has a fast search speed, but also, thanks to its intelligent guidance mechanism, can generate a shorter, less node-intensive, and closer-to-the-global optimum initial collision-free path. This high-quality initial path is directly used as the initial population for the second-stage genetic algorithm, providing an excellent starting point for the optimization process and avoiding the problem of blind searching and slow convergence caused by poor initial path quality in traditional methods.

[0040] III. Efficient Path Smoothness and Dynamic Constraint Satisfaction: In the second stage (refinement optimization), the genetic algorithm iteratively optimizes the high-quality initial path generated in the previous stage. By designing a fitness function that includes path length, collision penalty, and key angle constraint penalties, the evolutionary direction can be effectively guided, quickly optimizing potentially sharp inflection points on a broken path into a smooth, continuous, and feasible flight path that fully meets the dynamic constraints such as the UAV's maximum yaw angle. Experimental data show that this method requires only a small number of genetic algorithm iterations to output a final path that satisfies the constraints.

[0041] IV. Effectively Resolving the Contradiction Between Search Efficiency and Path Feasibility: This invention creatively combines guided rapid search with constraint-driven optimization, forming a collaborative framework of "efficient coarse search - constraint fine optimization." This framework fundamentally resolves the contradiction between the high randomness and low efficiency of standard sampling planning algorithms and the inconsistency of paths with the physical constraints of the aircraft. The first stage focuses on rapidly and guidedly discovering feasible paths, while the second stage focuses on fine-tuning within feasible paths to meet flight quality requirements. The two stages complement each other, thus achieving a path planning effect that is both "fast" and "good" overall.

[0042] V. Enhancing the Robustness and Stability of the Algorithm in Complex Environments: The adaptive step-size expansion strategy enables the algorithm to dynamically adjust its exploration behavior based on environmental complexity (obstacle density), allowing for rapid advancement in open areas and precise navigation in narrow areas, thus improving the algorithm's robustness against different obstacle layouts. Furthermore, due to the high quality of the initial path and the clear optimization objective, the performance metrics (such as path length and time) of this method are relatively small in multiple runs across different scenarios (simple, complex, narrow), demonstrating good stability and reliability, which is beneficial for the practical deployment and application of UAVs.

[0043] In summary, this invention, through the deep integration and synergistic improvement of the APF-RRT* algorithm and the genetic algorithm, effectively overcomes the technical bottlenecks in the prior art, such as low path planning efficiency, low initial path quality, and difficulty in meeting the dynamic constraints of aircraft dynamics. It provides an efficient, reliable, and practical path planning solution for UAVs to perform autonomous flight missions in complex environments. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the structure of an unmanned aerial vehicle (UAV) path planning system provided in an embodiment of this application;

[0046] Figure 2 A flowchart of a UAV path planning method provided in this application embodiment;

[0047] Figure 3 This is a schematic diagram of the regional probability sampling strategy in an embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the adaptive expansion step size strategy in the embodiments of this application;

[0049] Figure 5 This is a schematic diagram of the path planning simulation results of the existing RRT* algorithm in different scenarios.

[0050] Figure 6 This is a schematic diagram of the path planning simulation results of the existing APF-RRT* algorithm in different scenarios.

[0051] Figure 7 These are schematic diagrams illustrating the path planning simulation results of the algorithm in different scenarios according to embodiments of this application.

[0052] Figure 8 This table compares experimental data on the performance metrics of various path planning algorithms in different scenarios.

[0053] Figure labeling: 10, Input module; 20, Output module; 110, First-stage path search module; 111, Intelligent sampling unit; 112, Adaptive expansion unit; 113, Tree construction and reconnection unit; 120, Second-stage path optimization module; 121, Population initialization unit; 122, Fitness evaluation unit; 123, Evolutionary operation unit; A1, Initial path; A2, Final path. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] Example 1

[0056] This embodiment details the implementation process of a UAV path planning method that integrates APF-RRT* and a genetic algorithm. APF-RRT* is an asymptotically optimal fast expanding random tree algorithm that integrates an artificial potential field. This embodiment aims to comprehensively demonstrate the basic architecture and core workflow of the technical solution of this application.

[0057] See Figure 1 This is a schematic diagram of the structure of a UAV path planning system provided in an embodiment of this application. The system can be deployed on the UAV's onboard computing platform or a ground station's planning server. Its hardware implementation may include a processor, memory, and communication interfaces for receiving and sending data. The memory stores computer program instructions, which, when executed by the processor, are used to implement... Figure 1 The various functional modules are shown.

[0058] Specifically, the system adopts a two-stage modular design of "efficient coarse search - constrained fine optimization", including input module 10, first-stage path search module 110, second-stage path optimization module 120 and output module 20.

[0059] Input module 10 is used to receive the basic data required for the path planning task. This data may include: environmental map information, such as a two-dimensional or three-dimensional grid map showing the distribution of obstacles; task parameter information, such as the starting coordinates and target coordinates of the UAV; and the UAV's own performance constraint parameters, such as its maximum permissible yaw angle, maximum speed, and maximum acceleration. In this embodiment, a 100m × 100m two-dimensional grid map environment is set, which is a complex scene containing 14 irregular obstacles, such as... Figure 7 As shown in (b). The mission parameters are set with the starting point at coordinates (0,0), the target point at coordinates (100,100), and the maximum allowable yaw angle for the UAV at 30 degrees.

[0060] The first-stage path search module 110, as the core of the efficient coarse search, is used to quickly generate a collision-free initial path A1 from the starting point to the target point based on the received environmental and task information using an improved APF-RRT* algorithm. It should be noted that this module integrates multiple cooperating units to significantly improve search efficiency, including: an intelligent sampling unit 111, an adaptive expansion unit 112, and a tree construction and reconnection unit 113.

[0061] The intelligent sampling unit 111 aims to replace the completely random sampling in the traditional Rapidly Expanding Random Tree (RRT*) algorithm, generating sampling points in a more guided manner to guide the random tree to grow towards more promising regions. In this embodiment, the unit integrates two sampling strategies: a region probability sampling strategy and a dynamic target bias strategy.

[0062] The regional probability sampling strategy is based on the idea that the optimal path is highly likely to appear near the straight line connecting the starting point and the target point. Therefore, this strategy prioritizes sampling within this high-value region. Figure 3 As shown, a large number of sampling points are densely distributed near the line connecting the starting point and the target point, while sampling points further away from this area are sparse. In practical implementation, the probability of a sampling point being selected can be associated with its perpendicular distance to the line connecting the starting point and the target point. For example, a probability model based on a Gaussian distribution can be used to calculate the probability of a sampling point. Probability of being selected :

[0063]

[0064] in, Sampling points The perpendicular distance from the starting point to the target point on a straight line. It is the standard deviation of the Gaussian distribution, used to control the concentration of sampling points. The smaller the value, the more concentrated the sampling points are near the center line. In this way, the intelligent sampling unit 111 can use most of its computing resources to explore the region most likely to contain the optimal solution, thereby effectively reducing redundant exploration in irrelevant regions.

[0065] Accordingly, the dynamic target bias strategy provides stronger directionality for the growth of the random tree while avoiding getting trapped in local optima in complex environments. This strategy directly selects the target point as the sampling point with a dynamically adjusted probability. Specifically, the target bias probability... It can be calculated using the following formula:

[0066]

[0067] in, It is the Euclidean distance between the newly expanded node in the current random tree and the target point; It is the local obstacle density within a specified range around the node; The weights for the distance to the target point; To control the decay rate of the distance to the target point; The weights for the local obstacle density; To control the decay rate of local obstacle density; These are adjustable weight parameters. The first part of the formula helps accelerate exploration towards the target point when it is far away, while reducing the bias near the target point to prevent the random tree from being hindered by obstacles near the target point during expansion. The second part of the formula adjusts the bias probability based on the local obstacle density, which helps reduce the number of collision detection failures.

[0068] The adaptive expansion unit 112 is used to dynamically adjust the expansion step size when the random tree expands from existing nodes towards new sampling points. Traditional fast random tree expansion algorithms typically use a fixed expansion step size, which is poorly adaptable to complex and changing environments. Figure 4 As shown in the figure, X start Indicates the starting position of the path planning; X goal The target point location is represented by the symbol 'target'. Obstacles represent the obstacle distribution area around the current node used to calculate the local obstacle density. The figure shows a circular sliding window centered on the current node, used to evaluate the complexity of the local environment (obstacle density). The adaptive extension unit 112 of this application evaluates the complexity of the current environment by setting a virtual sliding window around the current path node and calculating the obstacle density within that window. The local obstacle density... This can be obtained by the ratio of the total area of ​​obstacles to the area of ​​the circular window. It is calculated by determining the proportion of the obstacle grid within the total window area, centered at the current node and with radius R. Then, based on this density value... To calculate the new expansion step size In this embodiment, a formula based on the hyperbolic tangent function is used for calculation:

[0069]

[0070] in, It is the set minimum expansion step size, ensuring that even the densest areas can make small movements; This is the step size adjustment coefficient, which determines the magnitude of the step size change; To control the steepness of the hyperbolic tangent function, the larger the value, the faster the step size decreases; The local obstacle density varies from 0 to 1. The center point of the step size change, when near The change in step size is most significant when the local obstacle density is [value missing]. The characteristic of this function is that when the local obstacle density [value missing]... When it is very small (i.e., in an open area), The function value is close to -1. Approaching a large value ( Thus, large strides and rapid progress are achieved; when When it is very large (i.e., in areas with dense obstacles), The function value is close to 1. Approaching the minimum step size This allows for precise exploration in small steps, improving path safety and making it easier to navigate narrow passages.

[0071] The tree construction and reconnection unit 113 is responsible for executing the standard core process of the fast expansion random tree algorithm. After the intelligent sampling unit 111 generates a sampling point, it first finds the node closest to the sampling point in the currently generated random tree. Then, the adaptive expansion unit 112 starts from the nearest node and expands a new node along the direction towards the sampling point with a calculated adaptive step size. Next, the unit performs collision detection on the path segment connecting the nearest node to the new node. In this embodiment, the collision detection adopts a path discretization detection method based on a grid map, that is, the path segment is discretized into a series of sufficiently dense intermediate points, and each intermediate point's grid cell is checked one by one to see if it is marked as an obstacle. If any intermediate point is located in an obstacle grid cell, the path segment is determined to have a collision, and the expansion fails. If there is no collision, the new node is successfully added to the random tree. Subsequently, the unit executes the optimization steps of the fast expanding random tree algorithm: First, it examines other neighboring nodes within a certain radius around the new node and attempts to find a parent node whose total path cost (distance from the starting point to the new node) is lower when connected to the new node (i.e., reselecting the parent node); second, it checks whether a shorter path can be provided to the surrounding neighboring nodes through the new node. If so, the parent nodes of these neighboring nodes are updated to the new node (i.e., reconnection operation). These operations ensure the asymptotic optimality of the generated path.

[0072] Through the collaborative work of the three units mentioned above, the first-stage path search module 110 iterates continuously until a node in the random tree successfully expands into the target point region, at which point a collision-free initial path A1 from the starting point to the target point is generated. This path consists of a series of path point sequences, such as... Figure 7 As shown by the thick green line in (b), thanks to the synergy of intelligent sampling and adaptive step size, the generated initial path A1 has higher quality and is closer to the optimal solution compared to traditional algorithms.

[0073] Subsequently, the high-quality initial path A1 is passed to the second-stage path optimization module 120. This module is used to perform constrained fine optimization, aiming to optimize the polygonal path generated in the first stage, which may have sharp inflection points, into a smooth final path A2 that fully satisfies the UAV dynamics constraints (especially the maximum yaw angle constraint). This module is implemented based on a genetic algorithm and also contains several key units: a population initialization unit 121, a fitness evaluation unit 122, and an evolutionary operation unit 123.

[0074] A key feature of the population initialization unit 121 is that, unlike traditional genetic algorithms that randomly generate the initial population, it directly uses the path point sequence of the initial path A1 output by the first-stage path search module 110 as the initial population for the genetic algorithm. In some implementations, this high-quality path can be used as the sole initial individual, or it can be slightly perturbed to generate a small initial population surrounding that high-quality path. This approach provides an excellent starting point for the subsequent optimization process, avoiding blind searching from scratch, and is a crucial step in improving the overall planning efficiency.

[0075] The fitness evaluation unit 122 is used to evaluate the merits of each individual (i.e., each candidate path) in the population and calculate a fitness value for it. The design of the fitness function is crucial, as it directly determines the direction of optimization. In this embodiment, a fitness function is designed. Its expression is as follows:

[0076]

[0077] in, The total path length is one of the main optimization objectives; This is a comprehensive penalty term used to handle path constraint violations. It includes at least two parts: a collision penalty term and an angle constraint penalty term. The logic behind the collision penalty term is that during the crossover and mutation process of the genetic algorithm, new paths that collide with obstacles may be generated. The fitness evaluation unit 122 performs collision detection on each path. If a path collides with an obstacle, its fitness value is increased by a very large penalty number, causing it to be eliminated in subsequent selections.

[0078] More importantly, this application introduces an angle constraint penalty term. The fitness evaluation unit 122 traverses each intermediate path point (i.e., inflection point) on the path and calculates the turning angle formed by the preceding and following path segments. If the turning angle is less than the minimum turning angle allowed by UAV dynamics (determined by the minimum turning radius), or greater than the preset maximum allowable turning angle (determined by the maximum yaw angle change rate), then a penalty value proportional to the degree of exceeding the range is added to the fitness value of that path. In this way, paths that do not meet the dynamic constraints will obtain poor fitness values ​​and will be gradually eliminated during the evolutionary process.

[0079] The evolutionary operation unit 123 performs standard genetic algorithm operations on the population to iteratively generate a better offspring population. In this embodiment, this unit performs the following operations:

[0080] 1. Selection Operation: A tournament selection strategy is employed. Each time, a number of individuals (e.g., 2 or 4) are randomly selected from the current population for comparison, with the one having the best fitness value (i.e., the one with the best fitness value) being selected. The individual with the lowest value wins and is replicated in the next generation of the population. This process is repeated until the size of the new population reaches the preset value.

[0081] 2. Crossover operation: Pairing individuals in the new population. For each pair of parent individuals, a single-point crossover operation is performed with a certain crossover probability (e.g., 0.8). Specifically, a position in the path point sequence is randomly selected as the crossover point, and then all path points after the crossover point of the two parent individuals are swapped, thereby generating two new offspring individuals.

[0082] 3. Mutation Operation: For the offspring individuals generated after crossover, a mutation operation is performed with a low mutation probability (e.g., 0.1). Specifically, one or more intermediate path points in the path are randomly selected, and their coordinates are randomly perturbed within a specified range to increase the diversity of the population and help escape local optima.

[0083] Understandably, the second-stage optimization process is an iterative loop. Each generation of the population undergoes fitness evaluation, selection, crossover, and mutation to generate a new offspring population, which replaces the parent generation, and so on. This iterative process continues until a preset termination condition is met, such as reaching the maximum number of iterations (which can be set to 50 in this embodiment), or the fitness value of the best individual in the population has not significantly improved for several consecutive generations.

[0084] When the optimization process terminates, the second-stage path optimization module 120 will select the individual with the best fitness value in the current population—that is, the path with the shortest total length, no collisions, and all inflection angles less than 30 degrees—as the final path A2, and output it through the output module 20. This path is as follows: Figure 7As shown by the thick red line in (b), it is smooth, safe, and efficient, and can be directly used by drones to perform flight missions.

[0085] See Figure 2 The method described in this embodiment is summarized in the form of a flowchart. The method starts at "Start," first executing the data loading step S10 to obtain map, task, and UAV parameters. Next, it enters the efficient coarse search step S20, which is executed cyclically, constructing a random tree through intelligent sampling and adaptive expansion, and checking whether the target point has been reached in step S21. If "No," the search continues; if "Yes," an initial path is generated and the process enters the constraint fine optimization step S30. This step is also a loop, internally executing iterative optimization using a genetic algorithm, and checking whether the termination condition is met in step S31. If "No," the evolution continues; if "Yes," the optimization is complete, and finally, the final path output step S40 is executed, outputting the optimal path and the process "Ends."

[0086] Figure 5 The diagram shows the simulation results of path planning in different scenarios using the existing RRT* algorithm, where (a) is a simple scenario, (b) is a complex scenario, and (c) is a narrow scenario.

[0087] Figure 6 The diagram shows the simulation results of path planning in different scenarios using the existing APF-RRT* algorithm, where (a) is a simple scenario, (b) is a complex scenario, and (c) is a narrow scenario.

[0088] Figure 7 The diagram shows the simulation results of path planning in different scenarios of the algorithm of this application embodiment, where (a) is a simple scenario, (b) is a complex scenario, and (c) is a narrow scenario.

[0089] like Figures 5-7 As shown in the diagram. Red dots represent sampling points, thin blue lines represent random trees, and thick red lines represent paths. In this application, green lines represent paths searched by the improved APF-RRT* algorithm, and red lines represent paths optimized by the genetic algorithm. Different algorithms were tested 20 times in different scenarios, and the average and variance of running time, path length, and number of traversed nodes were recorded. The data were then filled into [the database / system / database]. Figure 8 The experimental data comparison table is shown.

[0090] like Figure 5 As shown, although RRT* performs a large number of samplings in the map scene, the number of effective nodes that meet the yaw angle constraint is relatively small. Moreover, a large number of invalid samples increase the time cost of calculating the optimal path. The coordinates of the effective nodes in the middle part of the traversal also deviate from the target point, resulting in an excessively long explored path.

[0091] like Figure 6 As shown, in scenarios with an appropriate number of obstacles, APF-RRT* plays a certain guiding role in the generation direction of tree nodes due to the repulsive force of the obstacles and the attraction of the target point and sampling point. However, in narrow scenarios (with a large number of obstacles), a large number of invalid samples are also performed.

[0092] The APF-RRT* algorithm itself performs well, but the inherent random sampling characteristic of RRT* contradicts the yaw angle constraint, and its performance is poor in narrow scenarios, resulting in time consumption and wasted sampling points. Therefore, this application does not consider angle constraints during APF-RRT* search, and combines target bias strategy and region probability sampling strategy to guide sampling points, as well as an adaptive step size strategy to reduce the number of times the random tree is expanded, so that the improved APF-RRT* can efficiently search for a less costly path. Since the initial path searched is of high quality, and the tournament selection strategy retains high-quality parents, after path points are crossed and mutated, the genetic algorithm can output a safe flight path for the UAV that satisfies the angle constraint with fewer iterations. See Figure 7 and Figure 8 As shown, the algorithm of this application, which combines and improves upon the APF-RRT* and genetic algorithms, performs better overall than RRT* and APF-RRT*, reducing the number of nodes traversed, shortening the path search time, and improving and stabilizing path quality. Especially in confined spaces, compared to RRT*, the proposed solution reduces runtime by 84.2%, path length by 18.2%, and effective node traversal by 67.5%; compared to APF-RRT*, the proposed solution reduces runtime by 64.3%, path length by 1.6%, and effective node traversal by 70.3%.

[0093] Simulation results show that, Figure 7 In the complex scenario shown in (b), using the technical solution of this embodiment, the total time from initial planning to outputting the final path is approximately 1.20 seconds, and the final path length is approximately 144.99 meters. In comparison, as Figure 5 (b) and Figure 6 As shown in (b), the traditional Fast Expanding Random Tree algorithm and APF-RRT* algorithm not only take longer in this scenario, but also generate poorer paths, filled with unnecessary random exploration nodes. This fully demonstrates the significant advantages of the solution provided in this embodiment in terms of planning efficiency and path quality.

[0094] Example 2

[0095] This embodiment is another variation of Embodiment 1, intended to illustrate that the specific implementation of the adaptive extension unit 112 is flexible and diverse. Its core idea of ​​"adaptively adjusting the extension step size according to obstacle density" is not limited to the specific formula based on the hyperbolic tangent function proposed in Embodiment 1. This embodiment will demonstrate other feasible functional forms to support a broader protection scope.

[0096] In this embodiment, the step size calculation logic of the adaptive expansion unit 112 in the first-stage path search module 110 is replaced. The structure and function of the rest of the system remain consistent with those in Embodiment 1.

[0097] As an alternative implementation, the step size calculation can employ a piecewise linear function. For example, a threshold for the local obstacle density can be preset. (e.g., 0.3). When the adaptive extension unit 112 calculates the local obstacle density of the current node... Less than this threshold When the system is in a relatively open area, it indicates that the system is using a pre-set, relatively large, fixed step size. Expand. When Greater than or equal to the threshold When this indicates that the system has entered an area with dense obstacles or a narrow passage, it immediately switches to a preset, smaller, fixed step size. To conduct a detailed exploration. Its mathematical expression can be written as:

[0098]

[0099] The advantage of this implementation is that the calculation is extremely simple, the response is very direct, and it can quickly switch between the two exploration modes.

[0100] As an alternative implementation, the expansion step size can also be a continuous function linearly related to the obstacle density. For example, a linearly decreasing function could be used:

[0101]

[0102] in, It is the preset maximum expansion step size, corresponding to the ideal situation where there are no obstacles at all. ). It is a positive linear coefficient used to control the rate at which the step size decreases with increasing density. To prevent the step size from becoming negative or too small, a lower limit can be added, i.e. ,in It is the minimum step size.

[0103] When using any of the above-mentioned variant functions, the system uses a larger step size when the random tree expands in an open area to quickly advance towards the target area. However, when a node of the tree approaches or enters an area with dense obstacles, the system automatically calculates and uses a smaller step size, making the exploration behavior more cautious and refined, thus making it easier to find collision-free paths in complex obstacle layouts.

[0104] It is understandable that although the above-mentioned variant functions differ from the hyperbolic tangent function in Example 1 in terms of smoothness or specific calculation methods, they all achieve the core function of "using a smaller step size in areas with dense obstacles and a larger step size in open areas." Therefore, this embodiment fully demonstrates that the technical feature of "calculating the expansion step size through a function related to the local obstacle density" covers a variety of specific mathematical implementations and has good universality.

[0105] Example 3

[0106] This embodiment is another variation of Embodiment 1, and its purpose is to illustrate that the specific optimization algorithm and its operators in the second-stage path optimization module 120 also have multiple feasible implementation methods. The "efficient coarse search-constrained fine optimization" framework proposed in this application does not strictly depend on a specific combination of genetic algorithm operators.

[0107] In this embodiment, the specific operator used by the evolutionary operation unit 123 in the second-stage path optimization module 120 is replaced. The overall framework of the system, the first-stage module, and the population initialization unit 121 and fitness evaluation unit 122 in the second-stage module are all consistent with those in Embodiment 1.

[0108] In Example 1, the selection operation employs a "tournament selection" strategy. As an optional implementation, in this example, it can be replaced with a "roulette wheel selection" strategy. At the beginning of each generation of evolution, the fitness evaluation unit 122 first calculates the fitness value of each individual (path) in the population. Since the optimization objective is to minimize Therefore, the fitness value needs to be converted into a selection probability. Generally, the smaller the fitness value, the higher the probability of being selected. A common approach is to calculate the selection probability. ,in It is the first The fitness value of each individual It refers to the population size. Then, the system randomly selects individuals to enter the next generation based on these probabilities.

[0109] In Example 1, the crossover operation used a "single-point crossover". In this example, it can be replaced with a "two-point crossover" strategy. For a pair of parent paths selected for crossover, the system randomly selects two different crossover point positions (e.g., the first...). The path point and the first There are path points, among which Then, the path subsequences of the two parent individuals between these two intersection points are swapped, thereby generating two new child paths. For example, if parent 1 is (A,B,C,D,E) and parent 2 is (a,b,c,d,e), and the intersection points are the 2nd and 4th positions, then child 1 becomes (A,b,c,D,E) and child 2 becomes (a,B,C,d,e). This crossing method can exchange more complex path structure information.

[0110] The mutation operation can remain the same as in Example 1, that is, randomly change the position of a certain intermediate point in the path with a certain probability.

[0111] Using the replaced evolutionary operator, the optimization process in the second stage remains the same: the high-quality initial path A1 generated in the first stage is used as the initial population, and then multiple generations of iterations are performed. In each generation, fitness evaluation is performed, and then superior individuals are selected using a roulette wheel selection strategy. These individuals are then subjected to two-point crossover and mutation operations to generate a new generation of population. This process continues until the termination condition is met.

[0112] Experimental results show that the combination of "roulette wheel selection" and "two-point intersection" can effectively optimize the initial path, gradually eliminating collisions and angle overruns, shortening the path length, and ultimately outputting a smooth path that satisfies all constraints. This result demonstrates the universality of the second-stage optimization framework of this application; that is, it does not depend on specific genetic algorithm operators. Any evolutionary operation capable of achieving survival of the fittest and information exchange can be effectively combined with the framework of this application to achieve constrained fine-tuning of the path.

[0113] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A UAV path planning method integrating APF-RRT* and genetic algorithm, characterized in that, include: Data loading steps: Obtain environmental map information, mission parameter information, and UAV performance constraint parameters; Perform an efficient coarse search step: Based on the environmental map information and task parameter information, use the improved APF-RRT* algorithm to perform path search and generate a collision-free initial path from the starting point to the target point. The improved APF-RRT* algorithm includes intelligent sampling that integrates regional probability sampling strategy and dynamic target bias strategy to guide the growth of the random tree. And adaptively adjust the expansion step size based on the obstacle density around the current node; The intelligent sampling specifically includes: The aforementioned regional probability sampling strategy is adopted to prioritize sampling in the region close to the line connecting the starting point and the target point; The dynamic target bias strategy is adopted to dynamically adjust the probability of sampling toward the target point based on the distance between the current node and the target point and the density of surrounding obstacles; In the dynamic target bias strategy, the target bias probability Calculated using the following formula: ； in, It is the Euclidean distance between the newly expanded node in the current random tree and the target point; It is the local obstacle density within a specified range around the node; The weights for the distance to the target point; To control the decay rate of the distance to the target point; The weights for the local obstacle density; To control the decay rate of local obstacle density; The constraint-based optimization step is as follows: the path points of the initial path are used as the initial population of the genetic algorithm. The genetic algorithm iteratively optimizes to generate the final path that satisfies the UAV dynamics constraints. The fitness function of the genetic algorithm at least considers the path length, collision situation and whether the path turning angle exceeds the constraints. Step to output the final path: Output the final path.

2. The method according to claim 1, characterized in that, The adaptive adjustment of the expansion step size specifically includes: The expansion step size is calculated using a function related to the local obstacle density, such that the expansion step size decreases in areas with dense obstacles and increases in areas with open obstacles.

3. The method according to claim 1, characterized in that, The genetic algorithm includes performing evolutionary operations such as tournament selection, crossover, and mutation on the initial population.

4. The method according to claim 1, characterized in that, The fitness function is expressed as follows: in, This is the total path length item. This is a comprehensive penalty item; The comprehensive penalty items At least including: Collision penalty is used to impose a penalty on the path that collides with an obstacle; Angle constraint penalty term is used to impose a penalty on paths where the turning angle formed by two consecutive path segments exceeds the limit.

5. The method according to claim 2, characterized in that, In the adaptive adjustment of the expansion step size, New extended step size The calculation is performed using the following formula: ； in, It is the set minimum expansion step size; This is the step size adjustment coefficient; To control the steepness of the hyperbolic tangent function; The local obstacle density varies from 0 to 1. This is the center point of the step size change.

6. The method according to claim 1, characterized in that, The efficient coarse search step also includes a tree construction and optimization process, specifically including: Step A: In the currently generated random tree, find the node closest to the sampling point; Step B: Starting from the nearest node, expand a new node along the direction toward the sampling point with the calculated adaptive step size; Step C: Perform collision detection on the path segments connecting the nearest node to the new node; Step D: If there is no collision, add the new node to the random tree.

7. The method according to claim 6, characterized in that, Following step D, a path optimization step is also included: Step E: Reselect the parent node, check the neighboring nodes within a set radius around the new node, if connecting to the new node through one of the neighboring nodes can reduce the path cost of the new node, then update that neighboring node as the parent node of the new node. Step F: Reconnection operation. Check if a shorter path can be provided to the neighboring node through the new node. If so, update the new node as the parent node of the neighboring node.

8. A UAV path planning system integrating APF-RRT* and genetic algorithm, used to implement the method as described in any one of claims 1-7, characterized in that, include: Input module: Configured to acquire environmental map information, mission parameter information, and UAV performance constraint parameters; The first-stage path search module is configured to perform path search using an improved APF-RRT* algorithm based on the environmental map information and task parameter information, generating a collision-free initial path from the starting point to the target point. The improved APF-RRT* algorithm includes intelligent sampling that integrates a regional probability sampling strategy and a dynamic target bias strategy to guide the growth of the random tree; and adaptively adjusting the expansion step size according to the obstacle density around the current node. The second-stage path optimization module is configured to use the path points of the initial path as the initial population of the genetic algorithm, and generate a final path that satisfies the dynamic constraints of the UAV through iterative optimization of the genetic algorithm. The fitness function of the genetic algorithm at least considers the path length, collision situation and whether the path turning angle exceeds the constraints. Output module: configured to output the final path.

Images