An aircraft automatic driving perception network training data set construction method and system

Abstract

The application discloses a kind of aircraft autopilot perception network training data set construction method and system.The method comprises: determining the scene richness requirement of data set collection;Each individual scene requirement and expected data proportion are refined;Each sample is combined category statistics, determines data volume according to existing data set information and project scene requirement, and is sampled based on Monte Carlo method;Actual scene building and data collection are carried out according to the sampling result.The application is based on manned electric aircraft autopilot neural network training data set scene richness requirement, designs a kind of scene richness building method, combines each scene, determines the sample quantity that needs to be collected, finally collects a large amount of sample data, and overcomes the defect that intelligent driving data set is insufficient for low-altitude flight.

Description

Technical Field

[0001] This invention relates to the field of aircraft autopilot technology, and more specifically, to a method and system for constructing a training dataset for an aircraft autopilot perception network. Background Technology

[0002] With the rapid development of manned electric aircraft technology, the application of autopilot systems in aircraft is becoming increasingly widespread. The core of an autopilot system is the neural network model, and a high-quality training dataset is the foundation for building a high-performance neural network model. In the development and training of the EVTOLAI visual perception algorithm, the diversity and volume of scene datasets are enormous. Currently, one of the main challenges facing manned electric aircraft autopilot systems is how to construct a training dataset with sufficient scene richness to ensure that the model has good generalization ability in complex and ever-changing real-world flight environments. Therefore, how to construct training dataset scenarios for manned electric aircraft autopilot neural network models is a crucial issue that needs attention.

[0003] Traditional target detection datasets lack low-altitude flight data and scene diversity. Existing 2D low-altitude image datasets mainly focus on small and medium-sized UAVs, and existing dataset scene construction methods are mostly based on virtual scenes.

[0004] CN114546875A discloses a method for generating autonomous driving simulation scenarios. This method determines the target functional scenario based on scenario requirements in the testing needs, identifies scenario variable characteristics, selects target scenario variable samples from a pre-set scenario database, and performs sampling based on the data distribution characteristics of the scenario variable samples to finally generate the target simulation scenario. While this method can ensure the coverage of the simulation scenario, it is mainly aimed at ground-based autonomous driving systems and is difficult to directly apply to the specific scenario requirements of manned eVTOL autonomous driving.

[0005] CN118570667A proposes a dataset creation method based on image fusion technology and generative adversarial networks. This method can generate a large number of new datasets containing targets using a small amount of existing UAV target data and arbitrary backgrounds. However, this method mainly focuses on the generation of static images and is difficult to meet the needs of manned electric aircraft autopilot systems for dynamic scenes and complex interactions.

[0006] CN119964062A introduces a sample collection and training method. This method constructs a virtual scene or a 3D model of a real scene, simulates the changing parameters of factors in the model, drives the dynamic operation of the environment and target, and collects samples by controlling the movement of the acquisition lens to capture images of the target. Although this method considers the diversity of sample collection, it does not fully consider the unique flight environment and interaction scenarios of manned electric aircraft.

[0007] In summary, the current construction of training datasets for autonomous driving neural network models of manned electric aircraft mainly faces the following problems:

[0008] (1) Insufficient dataset sample richness leads to: the inability to effectively identify UAV targets during take-off and landing; low accuracy in identifying some unknown UAV targets, resulting in frame drops and other phenomena.

[0009] (2) Using virtual scenes to build dataset scenes results in insufficient realism and fidelity. Virtual scenes cannot fully and accurately simulate the complexity and randomness of the real world, and many details are biased. The verification of the dataset depends on real data. To evaluate the performance of the model trained on virtual data and to verify the fidelity of the virtual environment itself, we still need to rely on real-world data and tests. We cannot completely separate ourselves from real data to close the loop. Summary of the Invention

[0010] In view of the above problems, the purpose of this invention is to provide a method and system for constructing a training dataset for an autonomous flight perception network.

[0011] The first aspect of this invention provides a method for constructing a training dataset for an aircraft autopilot perception network, the method comprising:

[0012] S1: By deconstructing the core scene elements of eVTOL low-altitude flight, the dimensions and categories of the dataset samples are determined;

[0013] S2: Determine the expected data proportion for each sample category;

[0014] S3: Perform combined category statistics on each sample, and determine the total required combined scenario sample data volume based on existing dataset information and project scenario requirements;

[0015] S4: Based on the expected data proportion of each sample category, the sample data volume of the combined scenarios of total demand is sampled using the Monte Carlo method;

[0016] S5: Based on the sampling results of S4, build the actual scenario and collect data.

[0017] Preferably, the sample dimensions include: environment dimension, spatial location dimension, action dimension, and target dimension;

[0018] The environmental dimension includes the following sample categories: weather, background, background blending degree, and time.

[0019] The spatial location dimension includes the following sample categories: local height, target height, and target horizontal distance;

[0020] The action dimension includes the following sample categories: local maneuvers and target maneuvers;

[0021] The target dimension includes the following sample categories: target machine type.

[0022] Preferably, the rule for dividing the expected data proportions of each sample category is as follows:

[0023] In the environmental dimension: combined with the actual sightseeing and tourism operation scenarios of eVTOL, the weather scenario is mainly sunny, supplemented by cloudy, rainy, fog / haze, the time dimension covers morning, noon and afternoon, and the background category highlights the sky and evenly distributes mountains, buildings and forests;

[0024] Spatial location dimension: The vertical range of the machine and the target is 20-120 meters. The target height and horizontal distance are divided into 5 levels, but no preset ratio is used to adapt to dynamic tasks.

[0025] Action dimensions: Based on the actual scenario, the main actions of the aircraft are hovering and forward movement, with the addition of ascent, descent and low-speed horizontal movement. The target maneuvers are hovering, approaching and moving away from the aircraft, and crossing the path in front of the aircraft to the left and right.

[0026] Target aircraft dimensions: classified by size into small and medium-large, and by shape into rotor and fixed-wing.

[0027] Preferably, S4 includes:

[0028] S4.1: Let N be the total number of sample data for the combined scenarios of the total requirements. Initialize an empty result list and construct a scenario tree.

[0029] S4.2: Based on the scene tree, perform N independent random walks on the scene tree and record the complete scene combinations;

[0030] S4.3: Add the scene combinations from each walk to the results list.

[0031] Preferably, the construction of the scene tree includes the following steps:

[0032] S4.1.1: Define the sample categories and probability distributions based on the expected data proportions of each sample category;

[0033] S4.1.2: Create a node class and define the data structure of the tree node, including category, scenario, probability, and list of child nodes;

[0034] S4.1.3: Construct the scene tree structure and use a recursive method to build the complete scene tree.

[0035] Preferably, the scene tree is a hierarchical data structure used to represent all possible combinations of autonomous driving training scenarios and their probability distributions; the starting point of the scene tree is the root node; each layer corresponds to a sample dimension; each child node corresponds to a sample category and stores its probability value.

[0036] Preferably, S4.2 specifically includes:

[0037] Starting from the root node of the scene tree, randomly select a child node from all the child nodes of the current node according to its probability weight; move down along the selected path, repeat this process until a leaf node is reached, and record the complete scene combination formed by this walk.

[0038] Preferably, S5 specifically comprises:

[0039] Based on the scenario list obtained by S4 sampling, standardized actions were performed in real-world environments using at least two drone models to collect data containing real physical attributes, forming a high-fidelity training dataset.

[0040] A second aspect of the present invention provides a system for constructing a training dataset for an autonomous flight perception network, comprising a memory and a processor. The memory includes a program for constructing a training dataset for an autonomous flight perception network. When the program for constructing a training dataset for an autonomous flight perception network is executed by the processor, it implements the steps of a method for constructing a training dataset for an autonomous flight perception network.

[0041] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0042] (1) This invention systematically breaks through the bottleneck of low-altitude flight scene coverage by setting up a four-dimensional classification framework of "environment-spatial position-action-target" (covering 10 sub-categories). By finely defining the altitude layer of the aircraft, it specifically solves the problem of missing target detection during the take-off and landing phase of the aircraft and improves the aircraft's ability to identify unknown targets.

[0043] (2) Based on the combined statistical model, this invention quantifies the minimum dataset covering the entire scene, configures the priority according to the actual working conditions of urban low-altitude flight, and focuses on strengthening typical scenes such as sky background and building background, fundamentally solving the defects of scarcity and lack of diversity of low-altitude flight data.

[0044] (3) This invention avoids the domain gap problem of virtual simulation in terms of physical fidelity, behavioral complexity and sensor noise authenticity by using closed-loop acquisition of real scene. Attached Figure Description

[0045] Figure 1 This is a flowchart of a method for constructing a training dataset for an autonomous flight perception network as described in Example 1. Detailed Implementation

[0046] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0047] 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 therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0048] Example 1

[0049] like Figure 1 As shown in the figure, this embodiment discloses a method for constructing a training dataset for an aircraft autopilot perception network, the method comprising:

[0050] S1: By deconstructing the core scene elements of eVTOL low-altitude flight, the sample dimensions and sample categories of the dataset are determined.

[0051] In this embodiment, based on the scene richness requirements of the eVTOL low-altitude flight dataset, the requirements include four major sample dimensions and ten specific sample category requirements.

[0052] The four sample dimensions required include: environment, spatial location, action, and target. This demonstrates a scene classification framework for the autonomous driving training dataset for manned electric aircraft.

[0053] The four dimensions encompass a total of 10 specific sample category requirements:

[0054] (1) The environmental dimension includes the following sample categories: weather (meteorological conditions), background (static ground features), background blending degree (interaction between dynamic objects and the environment), and time (lighting changes).

[0055] (2) The sample categories included in the spatial position dimension are: aircraft height (aircraft vertical position), target height (obstacle / other aircraft vertical position) and target horizontal distance (obstacle / other aircraft horizontal proximity).

[0056] (3) The sample categories included in the action dimension are: self-maneuver (aircraft attitude adjustment and emergency operation) and target maneuver (sudden movement of other aircraft).

[0057] (4) The target dimension includes sample categories that only include target aircraft type (threat level classification such as UAV / civil aircraft).

[0058] This classification system covers the entire chain from external environment perception (weather, lighting, static background), three-dimensional spatial situational awareness (spatial relationship between itself and the target), dynamic behavior decision-making (maneuvering interaction between the local machine and other machines) to threat target identification, providing structured guidance for dataset construction and ensuring that training scenarios can systematically meet the perception, decision-making and safety control needs of eVTOL in urban low-altitude environments.

[0059] It should be noted that eVTOL stands for Electric Vertical Take-Off and Landing Manned Aircraft.

[0060] S2: Determine the expected proportion of data for each sample category.

[0061] In this embodiment, each individual scenario requirement is refined based on S1. The expected data ratio for each scenario requirement is determined. For the above 10 different sample category requirements, the refinement and expected data ratio for each individual sample category requirement are as follows:

[0062] (1) In the environmental dimension: combined with the actual sightseeing and tourism operation scenario of eVTOL, the weather scenario is mainly sunny, supplemented by cloudy, rainy, fog / haze, the time dimension covers morning, noon and afternoon, and the background category highlights the sky and evenly distributes mountains, buildings, forests, etc.

[0063] (2) Spatial position dimension: The vertical range of the machine and the target is 20-120 meters. The target height and horizontal distance are divided into 5 levels (10-200 meters), but no preset ratio is used to adapt to dynamic tasks;

[0064] (3) Action dimension: According to the actual scenario, the main actions of the aircraft are hovering (viewing) and moving forward (navigation), with the addition of ascent, descent and low-speed horizontal movement. The target maneuvers are hovering and approaching or moving away from the aircraft, as well as crossing the path in front of the aircraft to the left and right.

[0065] (4) Target aircraft dimensions: According to size, they are divided into small (outer envelope size less than 30*30*30cm) and medium-large, and the shapes are divided into rotor and fixed wing.

[0066] The above-mentioned dimensional allocation focuses on typical urban low-altitude operating conditions—such as high clear sky rate, building / sky background dominance, and mid-to-low altitude maneuvering. It optimizes data collection efficiency by quantifying priorities, while the blank space for the target height / distance ratio indicates that it needs to be dynamically adjusted in combination with the task. The inclusion of eVTOL aircraft and self-assembling drones ensures coverage of new threats, forming a structured data generation framework with clear primary and secondary priorities.

[0067] As a specific example, the expected data ratios for each sample category in this example are shown in Table 1.

[0068] Table 1 Expected Data Proportion for Each Sample Category

[0069]

[0070] This embodiment mainly uses the percentages of different scenarios in Table 1 to describe the weight allocation strategy for scenario categories in the construction of the autonomous driving dataset for manned electric aircraft. The core is to solve the problem that a single scenario cannot cover all requirement categories through differentiated data collection:

[0071] (1) Environmental dimension: highlighting the comfort and complexity of sightseeing at tourist attractions.

[0072] Weather: Sunny (60%) and cloudy (20%) are the dominant conditions. This is because sightseeing activities heavily rely on good weather conditions to provide the best experience, and the vast majority of flights are scheduled in good visibility conditions. Rain (10%) and fog / haze (10%) are allocated to ensure that the sensing system remains largely operational in adverse weather conditions, responding to sudden weather changes and ensuring safe diversions.

[0073] Time: Noon (40%) had the highest percentage because it is the peak tourist season with strong sunlight and obvious shadows, making it a typical scenario for testing the camera's perception capabilities. The morning and afternoon percentages were more evenly distributed, covering the effects of different lighting angles such as oblique sunlight and long shadows.

[0074] Background and background blending:

[0075] 1) The sky (30%), mountains (20%), and buildings (20%) are the most common background elements, which perfectly correspond to the natural environment (mountains and rivers, sky) and man-made environment (observation towers, distinctive buildings) commonly encountered in scenic spots. These are the backgrounds that aircraft most frequently encounter in scenic areas.

[0076] 2) Woodlands (10%) and farmland (10%) represent the background of country parks or rural attractions.

[0077] 3) In the background blending, no blending (50%) accounts for half, which is a basic capability. However, easily identifiable blending (25%) and difficult-to-identify blending (25%) are specially set because target drones can easily blend into complex mountain textures, building windows, or tree shadows, which is a major challenge for collision avoidance systems and must be focused on in training.

[0078] (2) Spatial position dimension and action dimension

[0079] Ego Motion: Forward movement (40%) and low-speed horizontal movement (20%) account for the largest proportion, which accurately describes the core operating mode of sightseeing aircraft: "smooth, slow cruise". Hovering (10%) corresponds to a short stop at a specific viewpoint, while climb / landing (10% each) corresponds to the stages of takeoff, landing, and changing altitude.

[0080] Local / Target Altitude & Horizontal Distance:

[0081] 1) The altitude is concentrated at 30m, 40m, and 60m (60% in total). This is the most common cruising altitude for low-altitude sightseeing in cities and natural scenic areas. It can ensure the view while avoiding complex obstacles on the ground.

[0082] 2) Scenarios with a horizontal distance of less than 100 meters account for 80%. This is a safety-critical area. At tourist attractions, multiple flight paths may intersect, and drones need to converge, follow, or avoid each other at close range. Therefore, the perception system must be very good at handling target recognition and tracking at medium to close range.

[0083] (3) Target characteristic category: targeting potential conflict parties in the scenic area

[0084] 1) Target relative maneuvering: The allocation is very balanced (20% or 10% each). This is because in shared airspace, the behavior of other drones or aircraft cannot be predicted. "Crossing left and right (20%)" is the highest risk scenario (e.g., intersecting flight paths), "approaching in all directions (20%)" poses a head-on or rear-end collision risk, and "hovering (20%)" (e.g., other drones filming) represents static obstacles. All must be adequately trained.

[0085] 2) Target aircraft types: Small rotorcraft (50%) and small fixed-wing aircraft (30%) account for the vast majority. This truly reflects that the current low-altitude market is dominated by various small consumer or industrial drones, which are most likely to appear in the airspace of tourist attractions. Medium and large aircraft account for a lower proportion because their operation is not as frequent, but coverage is still necessary.

[0086] It should be noted that this embodiment uses dynamic weights to drive priority allocation.

[0087] Introducing a scene weighting mechanism solves the problem that a single data collection cannot cover all dimensions:

[0088] High-weight scenarios should be prioritized for data collection, such as the machine's maneuverability and altitude, corresponding to high-risk scenarios like takeoff and landing obstacle avoidance.

[0089] Supplement as needed for low-weight scenarios: such as only collecting new threats for target machine types.

[0090] By allocating resources based on weights, resources are concentrated on core security needs, while secondary scenarios are dynamically adjusted proportionally.

[0091] S3: Perform combined category statistics on each sample, and determine the total required amount of combined scenario sample data based on existing dataset information and project scenario requirements.

[0092] It should be noted that this embodiment performs combined category statistics on each sample, for example, by designing according to the method in Table 1, in order to increase the richness and diversity of training data scenarios, improve the generalization performance of the model, and thus achieve good performance in more different types of scenarios.

[0093] S4: Based on the expected data proportion of each sample category, the sample data volume of the combined scenario of total demand is sampled using the Monte Carlo method.

[0094] As a specific embodiment, this embodiment requires 1,000,000 combined scene sample data. Then, using the Monte Carlo sampling method, sampling is performed according to the scene allocation ratio specified in Table 1 to obtain the final combined scene samples for actual scene construction and corresponding data collection. The specific sampling process is as follows:

[0095] S4.1: Assume the total required combined scenario sample data volume is 1,000,000, initialize an empty result list, and construct a scenario tree.

[0096] A scene tree is a hierarchical data structure used to represent all possible combinations of training scenarios for autonomous driving and their probability distributions. Each level of the tree corresponds to a sample dimension (such as weather, time, background, etc.), and each node represents a sample category, storing the probability value corresponding to the sample category within the node.

[0097] In this embodiment, the sample tree has the starting point as the root node, the sample category as the child node, and the sample dimension as the level.

[0098] The main steps in constructing a scene tree are as follows:

[0099] S4.1.1: Define the sample categories and probability distributions based on the expected data proportions of each sample category;

[0100] For example:

[0101] categories = {

[0102] Weather: [("Sunny", 0.6), ("Cloudy", 0.2), ("Rain", 0.1), ("Fog / Haze", 0.1)],

[0103] Background: [("Sky", 0.3), ("Mountains", 0.2), ("Buildings", 0.2), ("Forest", 0.1), ("Farmland", 0.1), ("Bare Ground", 0.1)],

[0104] Time: [("Morning", 0.3), ("Noon", 0.4), ("Afternoon", 0.3)],

[0105] "Background blending": [("No blending", 0.5), ("Easily discernible blending", 0.25), ("Difficult to discern blending", 0.25)],

[0106] "Target Height": [("20", 0.1), ("30", 0.2), ("40", 0.2), ("60", 0.2),("120", 0.1)],

[0107] "Target horizontal distance": [("0-10", 0.2), ("10-30", 0.2), ("30-50", 0.2), ("50-100", 0.2), ("100-200", 0.2)],

[0108] "Target Relative Maneuver": [("Hovering", 0.2), ("Distance in All Directions", 0.2), ("Approach in All Directions", 0.2), ("Crossing Left and Right", 0.2), ("Ascending", 0.1), ("Descending", 0.1)],

[0109] "In-vehicle Maneuvers": [("Hover", 0.1), ("Forward", 0.4), ("Climb", 0.1), ("Descent", 0.1), ("Low-speed Horizontal Movement", 0.2)],

[0110] "Target Aircraft Type": [("Small Rotorcraft", 0.5), ("Small Fixed-Wing Aircraft", 0.3), ("Medium-Large Rotorcraft", 0.1), ("Medium-Large Fixed-Wing Aircraft", 0.1)]

[0111] }

[0112] S4.1.2: Create a node class to define the data structure of the tree node, including category, scenario, probability, and list of child nodes.

[0113] For example:

[0114] class ScenarioNode:

[0115] def __init__(self, category, scene, probability):

[0116] self.category = category # Category name (e.g., "Weather")

[0117] self.scene = scene # Scene name (e.g., "Sunny")

[0118] self.probability = probability # The probability of reaching this node

[0119] self.children = [] # List of child nodes

[0120] S4.1.3: Construct the scene tree structure and use a recursive method to build the complete scene tree.

[0121] For example:

[0122] def build_scenario_tree(category_order=None):

[0123] """Build a complete scene tree"""

[0124] # Define the category order (if not provided)

[0125] if category_order is None:

[0126] category_order = [

[0127] "Weather", "Background", "Time", "Background Blending"

[0128] "Target Height", "Target Horizontal Distance"

[0129] "Target relative maneuver", "Aircraft maneuver", "Target aircraft type" ]

[0131] # Create root node

[0132] root = ScenarioNode("Root", None, 1.0)

[0133] # Recursive Tree Construction

[0134] current_level = [root]

[0135] for category in category_order:

[0136] next_level = []

[0137] for parent_node in current_level:

[0138] for scene, prob in categories[category]:

[0139] # Calculate the cumulative probability of reaching this node

[0140] cumulative_prob = parent_node.probability * prob

[0141] child_node = ScenarioNode(category, scene,cumulative_prob)

[0142] parent_node.children.append(child_node)

[0143] next_level.append(child_node)

[0144] current_level = next_level

[0145] return root

[0146] S4.2: Based on the scene tree, perform N independent random walks on the scene tree and record the complete scene combinations. Specifically:

[0147] Starting from the root node of the scene tree, randomly select a child node from all the child nodes of the current node according to its probability weight (i.e., the scene percentage in Table 1); move down along the selected path, repeat this process until the child node is reached, and record the complete scene combination formed by this walk.

[0148] S4.3: Add the scene combinations from each walk to the results list.

[0149] S5: Based on the sampling results of S4, build actual scenarios and collect data to obtain the training dataset for the aircraft's autopilot perception network.

[0150] In this embodiment, based on the sampled scenario list, standardized actions (fixed static parameters and varied dynamic parameters) are performed by multiple types of drones in real environments (such as building complexes on sunny days or mountainous areas in rain and fog) to collect data containing real physical attributes and form a high-fidelity training dataset.

[0151] It should be noted that this embodiment uses multiple drone models to perform standardized actions in real-world environments (such as building complexes in sunny weather or mountainous terrain in rain or fog):

[0152] Fixed variable combinations: such as "sunny morning + architectural background + camera hovering 40 meters";

[0153] Change key parameters: The target aircraft performs free flight (horizontal distance 20-40 meters);

[0154] Ultimately, through a closed loop of physical realism, we can directly obtain physical property data such as light and shadow reflection and material deformation, which cannot be simulated virtually.

[0155] This embodiment addresses the lack of low-altitude flight data and scene diversity in traditional target detection datasets by pioneering a four-dimensional classification framework of "environment-spatial location-action-target" (covering 10 subcategories). This systematically breaks through the bottleneck of low-altitude flight scene coverage. By finely defining the hierarchical layer of the aircraft's altitude, it specifically solves the problem of missing target detection during takeoff and landing. Furthermore, it proactively incorporates data from 20% of manned eVTOL aircraft and 10% of self-assembled UAVs to improve the ability to identify unknown targets, achieving full coverage of 46,656 scene combinations.

[0156] Meanwhile, this embodiment also addresses the problem that existing 2D low-altitude image datasets are mainly concentrated in the field of small and medium-sized UAVs. By quantifying the minimum dataset covering the entire scene based on a combined statistical model, prioritizing the configuration according to the actual working conditions of urban low-altitude areas, and focusing on strengthening typical scenes such as sky background and building background, it fundamentally solves the defects of scarcity and lack of diversity of low-altitude flight data.

[0157] Finally, this embodiment also addresses the issue that many existing dataset scene construction methods are based on virtual scenes. By using closed-loop acquisition of real-world scenes, it completely avoids the domain gaps in physical fidelity, behavioral complexity, and sensor noise realism inherent in virtual simulation.

[0158] Example 2

[0159] This embodiment discloses a system for constructing a training dataset for an autonomous flight perception network, including a memory and a processor. The memory includes a program for constructing a training dataset for an autonomous flight perception network as described in Embodiment 1. When the processor executes the program for constructing a training dataset for an autonomous flight perception network, it implements the steps of the method for constructing a training dataset for an autonomous flight perception network as described in Embodiment 1.

[0160] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0161] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0162] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

Claims

1. A method for constructing a training dataset for an aircraft autopilot perception network, characterized in that, The method includes: S1: By deconstructing the core scene elements of eVTOL low-altitude flight, the dimensions and categories of the dataset samples are determined; S2: Determine the expected data proportion for each sample category; S3: Perform combined category statistics on each sample, and determine the total required combined scenario sample data volume based on existing dataset information and project scenario requirements; S4: Based on the expected data proportion of each sample category, the sample data volume of the combined scenarios of total demand is sampled using the Monte Carlo method; S5: Based on the sampling results of S4, build the actual scenario and collect data; S4 includes: S4.1: Let the total number of sample data for the combined scenarios of the total requirements be N. Initialize an empty result list and construct a scenario tree. S4.2: Perform N independent random walks on the scene tree and record the complete scene combinations; S4.3: Add the scene combinations from each walk to the results list; The construction of the scene tree includes the following steps: S4.1.1: Define the sample categories and probability distributions based on the expected data proportions of each sample category; S4.1.2: Create a node class: Define the data structure of the tree node, including category, scenario, probability, and list of child nodes; S4.1.3: Constructing the scene tree structure: Constructing the complete scene tree using a recursive method; The scene tree starts at the root node; each layer corresponds to a sample dimension; each child node corresponds to a sample category and stores its probability value. Specifically, S4.2 is as follows: Starting from the root node of the scene tree, randomly select a child node from all the child nodes of the current node according to its probability weight; move down along the selected path, repeat this process until the child node is reached, and record the complete scene combination formed by this walk.

2. The method for constructing a training dataset for an aircraft autopilot perception network according to claim 1, characterized in that, The dimensions of the sample set include: environment dimension, spatial location dimension, action dimension, and target dimension.

3. The method for constructing a training dataset for an aircraft autopilot perception network according to claim 2, characterized in that, The environmental dimension includes the following sample categories: weather, background, background blending degree, and time. The spatial location dimension includes the following sample categories: local height, target height, and target horizontal distance; The action dimension includes the following sample categories: local maneuvers and target maneuvers; The target dimension includes the following sample categories: target machine type.

4. The method for constructing a training dataset for an aircraft autopilot perception network according to claim 3, characterized in that, The rules for dividing the expected data proportions of each sample category are as follows: In the environmental dimension: combined with the actual sightseeing and tourism operation scenarios of eVTOL, the weather scenario is mainly sunny, supplemented by cloudy, rainy, fog / haze, the time dimension covers morning, noon and afternoon, and the background category highlights the sky and evenly distributes mountains, buildings and forests; Spatial location dimension: The vertical range of the machine and the target is 20-120 meters. The target height and horizontal distance are divided into 5 levels, but no preset ratio is used to adapt to dynamic tasks. Action dimensions: Based on the actual scenario, the main actions of the aircraft are hovering and forward movement, with the addition of ascent, descent and low-speed horizontal movement. The target maneuvers are hovering, approaching and moving away from the aircraft, and crossing the path in front of the aircraft to the left and right. Target aircraft dimensions: classified by size into small and medium-large, and by shape into rotor and fixed-wing.

5. A method for constructing a training dataset for an aircraft autopilot perception network according to claim 1 or 4, characterized in that, Specifically, S5 is: Based on the scenario list obtained by S4 sampling, standardized actions were performed in real-world environments using at least two drone models to collect data containing real physical attributes, forming a high-fidelity training dataset.

6. A system for constructing a training dataset for an aircraft's autopilot perception network, characterized in that, The system includes a memory and a processor. The memory includes a program for constructing a training dataset for an autonomous driving perception network for an aircraft. When the program is executed by the processor, it implements the steps of constructing a training dataset for an autonomous driving perception network for an aircraft.

Images