Airspace flight height determination method and device

Abstract

The invention relates to a method and a device for determining the flight height of an airspace. The method comprises the following steps: receiving multi-height layer map information of the airspace; the multi-height-layer map information is preprocessed, the preprocessed multi-height-layer map information corresponds to three-dimensional image information of which the shape is (x, y, z), and the preprocessed multi-height-layer map information comprises x * y * z voxels and corresponding voxel values; the preprocessed multi-height-layer map information is input into a 3D-2D hybrid encoder-decoder model, a two-dimensional height suggestion map comprising x * y voxels is output, and the voxel values of the x * y voxels correspond to x * y height suggestion values respectively. According to the scheme, two-dimensional plane planning is upgraded into three-dimensional intelligent decision making, the airspace load balancing capacity in a complex airspace environment is remarkably improved, and the quality and dynamic adaptability of path planning are improved.

Description

Technical Field

[0001] This application relates to the field of low-altitude flight path planning technology, and in particular to a method and apparatus for determining airspace flight altitude. Background Technology

[0002] In traditional low-altitude flight path planning systems, path optimization is primarily based on two-dimensional maps, using obstacle avoidance algorithms and shortest path search methods such as A*, Dijkstra's algorithm, and Navmesh algorithm to generate navigation trajectories. However, with the rapid development of the low-altitude economy, especially the widespread adoption of urban air traffic and drone logistics, the limitations of two-dimensional path planning are becoming increasingly apparent. The main problems include:

[0003] (1) The airspace utilization efficiency is low and does not meet the actual flight environment requirements.

[0004] Existing systems typically simplify the problem by fixing altitudes. During environmental modeling, they assume the UAV's vertical altitude remains constant during the cruise phase, considering only horizontal movement and presetting flight altitudes to fixed values ​​or simple tiers (e.g., 50m / 100m / 150m). Such methods cannot dynamically adapt to complex airspace structures (e.g., gaps between tall buildings, high-voltage lines, temporary no-fly zones, and other three-dimensional obstacles), leading to a disconnect between planning results and actual flight requirements. Some existing systems use fixed-altitude path planning methods, which avoid power line areas by presetting altitudes but fail to dynamically adjust flight altitudes based on the three-dimensional features of the terrain (e.g., the uneven structure of canyon walls), resulting in blind spots in data acquisition. Furthermore, in dense flight scenarios, two-dimensional planning may lead to multiple UAVs intersecting paths on the same horizontal plane, ignoring the potential for vertical space reuse and wasting airspace resources.

[0005] (2) Insufficient response to dynamic environmental conditions, unable to meet real-time requirements.

[0006] Some existing methods rely on static map data, slicing the 3D space into multiple 2D planes (such as the OXY plane) along the height axis using a layered approach. Each layer independently performs 2D meshing and path selection, ultimately connecting the free meshes of different layers through curve fitting. However, these methods have limited dynamic obstacle avoidance capabilities in the vertical direction, making it difficult to respond in real-time to sudden obstacles (such as flocks of birds or temporary structures) or changes in airspace control. They also cannot adjust altitude in real-time according to airspace load, still requiring manual intervention to adjust altitude strategies. Furthermore, when these methods extend altitude as an independent variable to 3D path search, the state space grows exponentially (e.g., a 100m×100m area, after being meshed into a 3D grid at 1m resolution, results in millions of search nodes), making traditional algorithms unable to meet real-time requirements. Summary of the Invention

[0007] To address the problems existing in the prior art, this application provides a method for determining airspace flight altitude, which can autonomously perceive three-dimensional airspace characteristics and dynamically optimize altitude selection.

[0008] According to a first aspect of this application, a method for determining airspace flight altitude is provided, characterized in that it includes:

[0009] Receive multi-level map information of the airspace;

[0010] The multi-height layer map information is preprocessed, wherein the preprocessed multi-height layer map information corresponds to three-dimensional image information with shape (x, y, z), where x, y, and z correspond to the width, length, and height of the preprocessed multi-height layer map information, respectively. The preprocessed multi-height layer map information includes x*y*z voxels and their corresponding voxel values, where x, y, and z are positive integers; and

[0011] The preprocessed multi-height layer map information is input into a 3D-2D hybrid encoder-decoder model, and the output is a two-dimensional height suggestion map including x*y voxels, wherein the voxel values ​​of the x*y voxels correspond to x*y height suggestion values ​​respectively.

[0012] Optionally, the preprocessing of the multi-height-layer map information includes:

[0013] The multi-height-layer map information is scaled according to a preset resolution; and

[0014] According to the size requirements of the 3D-2D hybrid encoder-decoder model, the scaled multi-height layer map information is cropped.

[0015] Optionally, the step of inputting the preprocessed multi-height layer map information into a 3D-2D hybrid encoder-decoder model and outputting a two-dimensional height proposal map including x*y voxels includes:

[0016] The preprocessed multi-height layer map information is processed by 3D convolution through an encoder to obtain x*y*1 voxels, and the passage difficulty coefficient corresponding to each voxel in the x*y*1 voxels is determined.

[0017] The x*y*1 voxels are processed by a decoder using 2D transposed convolution and an attention mechanism to generate the height layer recommendation probability values ​​for the decoded x*y*z voxels; and

[0018] Based on the height layer recommendation probability value of the decoded x*y*z voxels, a two-dimensional height suggestion map of size x*y is determined, wherein the voxel value of each voxel in the two-dimensional height suggestion map corresponds to the height suggestion value at the corresponding position.

[0019] Optionally, before determining a two-dimensional height suggestion map of size x*y based on the height layer recommendation probability values ​​of the decoded x*y*z voxels, the step of inputting the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model and outputting a two-dimensional height suggestion map including x*y voxels further includes:

[0020] Real-time acquisition of dynamic airspace data through a pre-defined rules engine;

[0021] The height layer recommendation probability value of the decoded x*y*z voxels is adjusted according to the spatial dynamic data to obtain the adjusted height layer recommendation probability value.

[0022] Furthermore, determining a two-dimensional height suggestion map of size x*y based on the height layer recommendation probability value of the decoded x*y*z voxels includes:

[0023] Based on the adjusted height layer recommendation probability value, a two-dimensional height suggestion map with size x*y is determined.

[0024] Optionally, the step of processing the x*y*1 voxels through a decoder using 2D transposed convolution and attention mechanisms to generate the height layer recommendation probability value of the decoded x*y*z voxels includes:

[0025] The mid-level feature resolution feature map retained by the encoder is passed to the corresponding level of the decoder through a skip connection channel.

[0026] Optionally, before inputting the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model and outputting a two-dimensional height proposal map including x*y voxels, the method further includes:

[0027] The 3D-2D hybrid encoder-decoder model is trained using a neural network.

[0028] Optionally, the loss function used for neural network training of the 3D-2D hybrid encoder-decoder model includes one or more of the following: continuity loss function, height loss function, map complexity loss function, and dynamic equilibrium loss function.

[0029] Optionally, the continuity loss function includes:

[0030]

[0031] Among them, y i,j This represents the value at position [i,j] of the output image voxel, with the value ranging from the preset height layer.

[0032] Optionally, the height loss function includes:

[0033]

[0034] Among them, y i,j This represents the value at voxel position [i,j] in the output image, with the value ranging from the preset height layer, where n is the length of the image and m is the width of the image.

[0035] Optionally, the map complexity loss function includes:

[0036]

[0037] Among them, y i,j This represents the value at voxel position [i,j] in the output image, with the value ranging from the preset height layer, where n is the image length, m is the image width, and x[i,j,y] is the voxel position. i,j ] represents the input image position [i,j,y] i,j The value at [].

[0038] Optionally, the dynamic equilibrium loss function includes:

[0039] L 动态均衡损失 =MSE(y hist ,1)

[0040] Where MSE represents Mean Squared Error loss, y hist This represents the values ​​of the output image after histogram statistics.

[0041] Optionally, the x*y*z voxels correspond to x*y*z voxel values, and the voxel values ​​include the passage probability, aircraft density, and / or priority of the corresponding voxel.

[0042] According to a second aspect of this application, an airspace flight altitude determination device is provided, characterized in that it comprises:

[0043] The receiving module is used to receive multi-height layer map information in the airspace;

[0044] A preprocessing module is used to preprocess the multi-height layer map information, wherein the preprocessed multi-height layer map information corresponds to three-dimensional image information with shape (x, y, z), where x, y, and z correspond to the width, length, and height of the preprocessed multi-height layer map information, respectively. The preprocessed multi-height layer map information includes x*y*z voxels and their corresponding voxel values, where x, y, and z are positive integers; and

[0045] The output module is used to input the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model and output a two-dimensional height suggestion map including x*y voxels, wherein the voxel values ​​of the x*y voxels correspond to x*y height suggestion values ​​respectively.

[0046] According to a third aspect of this application, an electronic device is provided, comprising:

[0047] Processor; and

[0048] A memory storing computer instructions that, when executed by the processor, cause the processor to perform the method described in the first aspect.

[0049] According to a fourth aspect of this application, a non-transitory computer storage medium is provided, which stores a computer program that, when executed by a plurality of processors, causes the processors to perform the method described in the first aspect.

[0050] According to the airspace flight altitude determination method and apparatus provided in this application, the airspace altitude layer suggestion scheme based on neural networks upgrades traditional two-dimensional obstacle avoidance planning to a three-dimensional trajectory generation scheme that is deeply integrated with the airspace management system. This scheme can autonomously perceive three-dimensional airspace characteristics and dynamically optimize altitude selection. Upgrading two-dimensional planar planning to three-dimensional intelligent decision-making not only significantly improves airspace load balancing capabilities in complex airspace environments but also enhances the quality and dynamic adaptability of path planning. Furthermore, the loss function (optimization objective function) used for neural network training can simultaneously include obstacle avoidance and altitude-layer traffic balancing constraints, ensuring the optimal distribution of resource utilization at each airspace layer. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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 exceeding the scope of protection claimed by this application.

[0052] Figure 1 This is a flowchart of a method for determining airspace flight altitude according to an embodiment of this application.

[0053] Figure 2 This is a three-dimensional schematic diagram of multi-height layer map information according to one embodiment.

[0054] Figure 3 Based on one embodiment Figure 2 The image shows a two-dimensional height suggestion map output from multi-height layer map information.

[0055] Figure 4 This is a flowchart of a method for determining airspace flight altitude according to another embodiment of this application.

[0056] Figure 5 This is a schematic diagram of an airspace flight altitude determination device according to an embodiment of this application.

[0057] Figure 6 This is a schematic diagram of an airspace flight altitude determination device according to another embodiment of this application.

[0058] Figure 7 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] In general, this application preprocesses the acquired map information from multiple height layers using a 3D-2D hybrid encoder-decoder model to determine the suggested height values ​​for each voxel within the map information at each height layer. Specifically, the 3D-2D hybrid encoder-decoder model is trained using a pre-defined loss function in an unsupervised or self-supervised learning manner. The loss function can simultaneously incorporate obstacle avoidance and height layer traffic balance constraints to ensure optimal distribution of resource utilization across all airspace layers.

[0061] The solution proposed in this application can be applied to various scenarios. For example, it can be applied to drone logistics scenarios to generate obstacle avoidance height corridors in dense urban areas; it can be applied to urban air traffic control scenarios to assign height layers that minimize conflicts to multiple agents; and it can also be applied to emergency rescue scenarios to dynamically adjust the altitude of aircraft to avoid fire smoke areas.

[0062] According to one aspect of this application, a method for determining airspace flight altitude is provided. Figure 1 This is a flowchart of a method for determining airspace flight altitude according to an embodiment of this application. Figure 1 As shown, the method includes the following steps.

[0063] Step S101: Receive multi-height layer map information of the airspace.

[0064] In this application, multi-height layer map information of the airspace is input into a 3D-2D hybrid encoder-decoder model for inference and prediction to determine the suggested height values ​​for each location on the map. In one embodiment, the multi-height layer map information input into the 3D-2D hybrid encoder-decoder model is a 3D (three-dimensional) tensor, such as... Figure 2 As shown. The shape of this 3D tensor can be represented as (B, C, X, Y, Z) or B×C×X×Y×Z, where B represents the batch size, C represents the channel, X represents the width, Y represents the length, Z represents the height, and B, C, X, Y, and Z are positive integers.

[0065] Step S102: Preprocess the multi-height layer map information.

[0066] In one embodiment, the original map may be anisotropic (e.g., 5m x 5m x 10m), requiring the resolution of the X and Y axes to be adjusted to 10 meters (the same as the Z axis). Each XY plane layer is then scaled down proportionally using common image processing tools. For input multi-height layer map information, it can be divided into multiple voxels based on the resolution. For example, for an original map with a true physical value of 5000m x 5000m x 80m, scaling the map to a resolution of 10m results in a map with voxel dimensions of 500 x 500 x 8 (i.e., including 500 x 500 x 8 voxels), with each voxel measuring 10m x 10m x 10m.

[0067] Furthermore, to meet the size requirements of the 3D-2D hybrid encoder-decoder model input, the processed X, Y, and Z dimensions can be multiples of 8. Therefore, the rescaled input image can be cropped from 500*500*8 to 480*480*8. Now we have an isotropic 3D image with a shape of (480, 480, 8) and a resolution of 10 meters for each axis.

[0068] In one embodiment, the preprocessed multi-height layer map information corresponds to 3D image information with shape (x,y,z), where x, y, and z correspond to the width, length, and height of the preprocessed multi-height layer map information, respectively. The preprocessed multi-height layer map information includes x*y*z voxels and their corresponding voxel values, where x, y, and z are positive integers.

[0069] For each voxel in the multi-altitude map information, its corresponding voxel value can correspond to or encode obstacle avoidance (corresponding to passage probability), load balancing (corresponding to aircraft density) and / or priority. Among them, obstacle avoidance corresponds to the automatic mapping of high obstacle density areas in the input tensor to low passage probability; load balancing corresponds to the aircraft density statistics of each altitude layer (Z axis) input through an additional channel (C≥2); priority can correspond to emergency task areas (such as rescue channels) and can be marked as high weight areas through a binary mask.

[0070] In one specific embodiment, the voxel value can range from [0, 255]. For passage probability, a larger voxel value indicates a higher passage probability; for example, 255 represents no obstruction, and 0 represents complete obstruction. For aircraft density, a larger voxel value indicates a more crowded altitude at that location. For priority, the voxel value can have only two values: 0 or 255. For example, based on a binary no-fly / priority matrix issued by the control department, 255 represents "high priority / rescue channel," and 0 represents a normal area. In addition to obstacle avoidance, load balancing, and priority, voxel values ​​can also encode other factors, and a separate channel can be set for different factors. Furthermore, voxel values ​​can be normalized so that the values ​​are within the range of [0, 1] or [-1, 1].

[0071] In an optional embodiment, step S102 may include:

[0072] The multi-height-layer map information is scaled according to a preset resolution; and

[0073] According to the size requirements of the 3D-2D hybrid encoder-decoder model, the scaled multi-height layer map information is cropped.

[0074] Step S103: Input the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model and output a two-dimensional height suggestion map including x*y voxels, wherein the voxel values ​​of the x*y voxels correspond to x*y height suggestion values ​​respectively.

[0075] In one embodiment, preprocessed multi-height layer map information with shape (x, y, z) is input into a 3D-2D hybrid encoder-decoder model for inference and prediction, outputting a two-dimensional image with shape (x, y). The voxel values ​​of x*y voxels represent the suggested heights at the corresponding positions in the generated two-dimensional image, such as... Figure 3 As shown. For example, multi-height layer map information with shape (480,480,8) is input into a 3D-2D hybrid encoder-decoder model for inference and prediction, and outputs a two-dimensional image with shape (480,480). The voxel value at the voxel position (220,430) represents the suggested flight altitude value for that position.

[0076] In one embodiment, before inputting the preprocessed multi-height layer map information with shape (x,y,z) into the 3D-2D hybrid encoder-decoder model, the 3D multi-height layer map information can be standardized, for example, the voxel values ​​can be normalized so that the values ​​are in the range of [0,1] or [-1,1].

[0077] In one embodiment, the 3D-2D hybrid encoder-decoder model includes an encoder and a decoder, wherein the encoder can be used for 3D convolution and residual connection processing, while the decoder can be used for 2D transposed convolution and attention mechanism processing, and can also be used for multi-scale skip connection processing.

[0078] In one embodiment, the encoder receives input 3D multi-height-layer map information. The 3D convolutional layer can extract three-dimensional spatial features, such as the continuous obstacle shape of tall buildings along the z-axis and the suspended structure of high-voltage power lines. In one embodiment, for multi-height-layer map information with shape (x, y, z), the height layers of the preprocessed map are quantized into z discrete levels. The convolutional kernel of the 3D convolutional layer covers all z layers in the Z-axis direction, compressing the height information of the z layer corresponding to each (x, y) position into one feature value, generating x*y*1 voxels and determining the passage difficulty coefficient corresponding to each voxel. The passage difficulty coefficient represents the passage difficulty at the corresponding location.

[0079] In one specific embodiment, for multi-height layer map information with a shape of (480, 480, 8), the z-axis (height layer) of the preprocessed map is quantized into 8 discrete levels. A special convolutional kernel of a 3D convolutional layer Conv3D (kernel_size = (1, 1, 8), stride = 1, filters = 1) is used to cover all 8 levels in the Z-axis direction, compressing the 8-level height information corresponding to each (x, y) position into a single feature value, outputting a (480, 480, 1) feature map. For example, the 8-level height data of a voxel position (i, j) are [0.2, 0.8, 0.9, 0.3, 0.1, 0, 0, 0] (0 represents no obstacles), which, after convolution, generates a single value (e.g., 0.75), which can characterize the overall accessibility difficulty at this position. Although the z-axis structure is lost after processing with 3D convolutional layers (making it impossible to distinguish specific layers), key features in the height direction (such as the highest obstacle position and average passability) are preserved, which can avoid 3D computational explosion and maintain height perception capability.

[0080] Residual connections can be used to alleviate the vanishing gradient problem in deep neural network training. They can preserve shallow layer details and prevent canyons / small obstacles from being lost in deep networks.

[0081] After passing through the encoder's storage, the input 3D multi-height layer map information is processed into a 2.5D feature map or "overall accessibility map". The 2.5D feature map no longer represents "height" itself, but rather the comprehensive risk / access cost at all heights at the corresponding voxel location.

[0082] In one embodiment, the decoder processes the encoder's output. The decoder's 2D transposed convolution is used to upsample the feature map to its original resolution, reconstructing spatial details, such as building edge precision. An attention mechanism can be used to enhance the model's ability to focus on important features. By calculating the importance weight of each location in the feature map, the model can focus more on important feature regions, thereby improving the accuracy and robustness of the output. In a specific embodiment, a dynamically weighted approach can be used to calculate the contributions of different altitude layers; for example, if a temporary no-fly zone is detected at the 80m level, the weight of that layer can be reduced.

[0083] In one embodiment, multi-scale skip-connect processing can fuse mid-level features from the encoder to repair minor obstacles lost during the decoding process. The operations employed include mid-level feature extraction and feature fusion. When the encoder downsamples to a mid-level feature resolution map (not the deepest layer), the feature map at that scale is saved. Then, when upsampled to that scale, the mid-level feature resolution map extracted by the encoder is concatenated or added using skip-connect channels. For example, for multi-height layer map information with a shape of (480, 480, 8), when downsampled to a 240×240 resolution, the feature map at that scale is saved; when the decoder upsamples to 240×240, the mid-level feature resolution map extracted by the encoder is concatenated or added using skip-connect channels. Multi-scale skip connection processing can combine shallow high-resolution details (such as wire positions) with deep semantic information (such as airspace congestion areas) to repair obstacle edges. To address the problem of easily lost small obstacles at low altitudes, a cross-scale feature grafting mechanism is designed. Mid-resolution feature maps are extracted in the middle of the encoder and injected into the corresponding level of the decoder through skip connections, achieving synergistic optimization of detail repair and semantic enhancement.

[0084] The x*y*1 voxel output by the encoder is processed by the decoder to generate the decoded x*y*z voxel. Each voxel in the decoded x*y*z voxel includes the recommendation probability value for the corresponding height layer. For example, in the decoded 480*480*8 voxel, (20,15,5) represents the recommendation probability of the position at (20,15) in the 5th height layer.

[0085] Then, based on the recommended probability values ​​of the corresponding height layers, a two-dimensional height suggestion map of size x*y is determined, wherein the voxel value of each voxel in the two-dimensional height suggestion map corresponds to the height suggestion value at the corresponding position. In a specific embodiment, the maximum index value (argmax) algorithm can be used to determine the recommended height layer from the recommended probability values ​​of each height layer. For example, at the position (20,15), the recommended probabilities of height layers 1 to 8 are 0, 0.1, 0.8, 0.05, 0.03, 0.02, 0.01, and 0, respectively. Therefore, the height suggestion value at the position (20,15) is 3 (representing the third height layer).

[0086] In an optional embodiment, step S103 may include:

[0087] Sub-step 1: The preprocessed multi-height layer map information is processed by 3D convolution through an encoder to obtain x*y*1 voxels, and the passage difficulty coefficient corresponding to each voxel in the x*y*1 voxels is determined.

[0088] Sub-step 2: The x*y*1 voxels are processed by 2D transposed convolution and attention mechanism through the decoder to generate the height layer recommendation probability value of the decoded x*y*z voxels; and

[0089] Sub-step 3: Based on the height layer recommendation probability value of the decoded x*y*z voxels, determine a two-dimensional height suggestion map of size x*y, wherein the voxel value of each voxel in the two-dimensional height suggestion map corresponds to the height suggestion value at the corresponding position.

[0090] In one embodiment, unforeseen circumstances or changes may occur in the airspace or aircraft, necessitating real-time adjustments to flight altitude. In this embodiment, a pre-defined lightweight rules engine can access dynamic airspace data (such as no-fly zone coordinates and weather information) in real time to adjust the altitude layer recommendation probability value distribution output by the 3D-2D hybrid encoder-decoder model online, obtaining adjusted altitude layer recommendation probability values. Then, based on the adjusted altitude layer recommendation probability values, a two-dimensional altitude suggestion map is determined.

[0091] In one specific embodiment, online adjustment of the height layer recommendation probability value distribution output by the 3D-2D hybrid encoder-decoder model includes: mandatory avoidance of no-fly zones, when no-fly zone coordinates are detected, resetting the height value of that location to a safe height layer, and forcing the no-fly zone coordinate point to point to a safe height layer; traffic rebalancing, if the real-time traffic of a certain height layer exceeds a threshold, the recommendation probability of that layer can be reduced according to a preset gradient, or the recommendation probability of overloaded height layers can be suppressed according to the real-time traffic ratio, etc.

[0092] By adjusting the distribution of the height layer recommendation probability values ​​online, the final output height suggestion value conforms to the optimized height map with dynamic safety constraints, and the response speed is fast, reaching millisecond level as shown in experiments.

[0093] In an optional embodiment, before sub-step 3, step S103 may further include:

[0094] Sub-step 4: Obtain dynamic airspace data in real time through a preset rule engine;

[0095] Sub-step 5: Adjust the height layer recommendation probability value of the decoded x*y*z voxels according to the spatial dynamic data to obtain the adjusted height layer recommendation probability value.

[0096] Figure 4 This is a flowchart of a method for determining airspace flight altitude according to another embodiment of this application. Figure 1 compared to, Figure 4 Steps S401, S402, and S404 of the method are the same as those of the previous steps. Figure 1 The steps S101 to S103 of the method are the same or similar, except that before step S404... Figure 4 The method shown may also include:

[0097] Step S403: Perform neural network training on the 3D-2D hybrid encoder-decoder model.

[0098] In one embodiment, the 3D-2D hybrid encoder-decoder model can be trained using unsupervised or self-supervised learning methods without training data (ground-based height cues). The neural network of the 3D-2D hybrid encoder-decoder model is trained using a loss function. In one embodiment, the loss function can be a joint optimization function, which may be dominated by a dynamic equilibrium loss function. The loss function can be one or more of the following: continuity loss function, height loss function, map complexity loss function, and dynamic equilibrium loss function.

[0099] In one specific embodiment, the difference between the value of each voxel and its neighboring voxels should not be too large, because the aircraft altitude should not exhibit illegal oscillations. For continuous loss functions, there are many loss functions that can meet this requirement, such as threshold loss, adaptive neighborhood loss, edge-aware smoothness loss, etc. In one specific embodiment, the total variation loss function can be used, as shown in equation (1):

[0100]

[0101] Among them, y i,j This represents the value at position [i,j] of the output image voxel, with the value ranging from the preset height layer.

[0102] In one specific embodiment, assuming that the starting point and target point of the path planning are always on the ground, the energy loss function aims to minimize flight energy consumption: L = altitude * (takeoff power + landing power) + distance * flight power

[0103] Energy loss can include an altitude loss function and a map complexity loss function. The higher the suggested altitude, the greater the loss. The altitude loss function can be represented by equation (2):

[0104]

[0105] Among them, y i,j This represents the value at voxel position [i,j] in the output image, with the value ranging from the preset height layer, where n is the length of the image and m is the width of the image.

[0106] In one specific embodiment, the larger the obstacle area, the greater the loss. It is assumed that the more obstacles there are, the longer the path length and the greater the energy consumption. The map complexity loss function can be expressed as shown in equation (3):

[0107]

[0108] Among them, y i,j This represents the value at voxel position [i,j] in the output image, with the value ranging from the preset height layer, where n is the image length, m is the image width, and x[i,j,y] is the voxel position. i,j ] represents the input image position [i,j,y] i,j The value at [].

[0109] In one specific embodiment, to avoid too many aircraft flying at the same altitude level and to achieve load balancing, each input altitude level should be uniformly recommended. If all altitude levels are at the same level, the loss is greater; if the altitude levels are uniformly distributed, the loss is lower. The dynamic balancing loss function can be expressed as shown in equation (4):

[0110] L 动态均衡损失 =MSE(y hist ,1) (4)

[0111] Where MSE represents Mean Squared Error loss, y hist This represents the values ​​of the output image after histogram statistics.

[0112] According to another aspect of this application, an apparatus for determining airspace flight altitude is provided. Figure 5 This is a schematic diagram of an airspace flight altitude determination device according to an embodiment of this application. Figure 5As shown, the device includes a receiving module 501, a preprocessing module 502, and an output module 503. The receiving module 501 receives multi-height layer map information in the airspace. The preprocessing module 502 preprocesses the multi-height layer map information, wherein the preprocessed multi-height layer map information corresponds to three-dimensional image information with shape (x, y, z), where x, y, and z correspond to the width, length, and height of the preprocessed multi-height layer map information, respectively. The preprocessed multi-height layer map information includes x*y*z voxels and their corresponding voxel values, where x, y, and z are positive integers. The output module 504 inputs the preprocessed multi-height layer map information into a 3D-2D hybrid encoder-decoder model and outputs a two-dimensional height suggestion map including x*y voxels, wherein the voxel values ​​of the x*y voxels correspond to x*y height suggestion values.

[0113] In an optional embodiment, the preprocessing module 502 can be used to:

[0114] The multi-height-layer map information is scaled according to a preset resolution; and

[0115] According to the size requirements of the 3D-2D hybrid encoder-decoder model, the scaled multi-height layer map information is cropped.

[0116] In an alternative embodiment, the output module 503 may include:

[0117] The acquisition unit is used to perform 3D convolution processing on the preprocessed multi-height layer map information through an encoder to obtain x*y*1 voxels, and determine the passage difficulty coefficient corresponding to each voxel in the x*y*1 voxels.

[0118] The generation unit is used to process the x*y*1 voxels through a decoder using 2D transposed convolution and an attention mechanism to generate the height layer recommendation probability values ​​of the decoded x*y*z voxels; and

[0119] The determining unit is used to determine a two-dimensional height suggestion map of size x*y based on the height layer recommendation probability value of the decoded x*y*z voxels, wherein the voxel value of each voxel in the two-dimensional height suggestion map corresponds to the height suggestion value at the corresponding position.

[0120] In an optional embodiment, the output module 503 may further include:

[0121] The acquisition unit is used to acquire dynamic airspace data in real time through a preset rule engine;

[0122] The adjustment unit is used to adjust the height layer recommendation probability value of the decoded x*y*z voxels according to the spatial dynamic data, so as to obtain the adjusted height layer recommendation probability value.

[0123] In an optional embodiment, the generation unit can be used for:

[0124] The mid-level feature resolution feature map retained by the encoder is passed to the corresponding level of the decoder through a skip connection channel.

[0125] Figure 6 This is a schematic diagram of an airspace flight altitude determination device according to another embodiment of this application. Figure 5 compared to, Figure 6 Modules 601, 602, and 604 of the device and Figure 5 The modules 501 to 503 of the device are the same or similar, except that, Figure 6 The apparatus shown may also include:

[0126] The neural network training module 603 is used to train the neural network of the 3D-2D hybrid encoder-decoder model.

[0127] According to the airspace flight altitude determination method and apparatus provided in this application, the airspace altitude layer suggestion scheme based on neural networks upgrades traditional two-dimensional obstacle avoidance planning to a three-dimensional trajectory generation scheme that is deeply integrated with the airspace management system. This scheme can autonomously perceive three-dimensional airspace characteristics and dynamically optimize altitude selection. Upgrading two-dimensional planar planning to three-dimensional intelligent decision-making not only significantly improves airspace load balancing capabilities in complex airspace environments but also enhances the quality and dynamic adaptability of path planning. Furthermore, the loss function (optimization objective function) used for neural network training can simultaneously include obstacle avoidance and altitude-layer traffic balancing constraints, ensuring the optimal distribution of resource utilization at each airspace layer.

[0128] In addition, the solution of this application also has the following effects:

[0129] (1) Flexible allocation of three-dimensional spatial resources:

[0130] By employing neural implicit field technology and deeply integrating real-time occupancy data from the airspace management system, the atmospheric flow density, obstacle distribution, and two-dimensional plane path are jointly encoded into a continuously differentiable load balancing function. Through the resource utilization gradient backpropagation mechanism, the distribution of aircraft at each atmospheric level is dynamically optimized, achieving elastic allocation of three-dimensional airspace resources and avoiding airspace resource waste and local congestion caused by discrete layering.

[0131] (2) Dynamic height optimization mechanism:

[0132] An online learning module is introduced to integrate real-time dynamic airspace information (such as temporary no-fly zones, meteorological data, and the status of other aircraft). Altitude suggestions are dynamically adjusted through gradient optimization to ensure real-time safety and efficiency of the flight path.

[0133] (3) Lightweight network architecture:

[0134] Design a lightweight network architecture that, under conditions of limited computing resources, simultaneously optimizes the security (obstacle avoidance), economy (lowest energy consumption), and compliance (airspace rules) of highly selective systems, and outputs a Pareto optimal solution.

[0135] This application relates to a height-level recommendation system based on neural networks. Its core lies in utilizing neural network algorithms to efficiently model and accurately predict the complex occupancy states of three-dimensional airspace. Through deep learning technology, the system can automatically learn the inherent patterns of key information such as obstacle distribution and aircraft dynamics in the airspace, thereby providing intelligent height-level recommendations for aircraft. This system not only improves the efficiency of airspace resource allocation but also enhances flight safety, representing an indispensable technological innovation in modern airspace management.

[0136] This company has innovatively extended traditional two-dimensional path planning to three-dimensional space, breaking through the limitations of traditional two-dimensional path planning. By introducing altitude layer information, the system can generate more optimized flight trajectories in three-dimensional space. This three-dimensional path planning not only considers obstacle avoidance in the horizontal direction but also takes into account altitude layer optimization in the vertical direction, effectively improving the quality and dynamic adaptability of path planning. In complex and ever-changing airspace environments, this technology can significantly improve the utilization rate of airspace resources, providing aircraft with more flexible and efficient flight paths.

[0137] This application also aims to achieve a balanced distribution of airspace load. Through careful design of the objective function, the system simultaneously considers obstacle avoidance and altitude-level traffic balance constraints during path planning. This comprehensive optimization strategy ensures that the utilization rate of resources at each airspace level reaches the optimal distribution, avoiding local congestion or resource idleness. In a dynamic airspace environment, the system can respond in real time to changes in airspace occupancy status and dynamically adjust path planning, thereby achieving efficient utilization and load balancing of airspace resources, providing strong support for flight management in complex airspace environments.

[0138] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0139] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0140] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be an electrical connection or other forms.

[0141] See Figure 7 , Figure 7 An electronic device is provided, including a processor and a memory. The memory stores computer instructions or one or more programs, which, when executed by the processor, cause the processor to execute the computer instructions to achieve the following: Figure 1 and Figure 4 The method and its detailed scheme are shown.

[0142] It should be understood that the above-described device embodiments are merely illustrative, and the device disclosed in this invention can be implemented in other ways. For example, the division of units / modules described in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, integrated into another system, or some features may be ignored or not executed.

[0143] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of the present invention can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0144] If the integrated unit / module is implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor or chip can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the on-chip cache, off-chip memory, and storage can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), etc.

[0145] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer electronic device (which may be a personal computer, server, or network electronic device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0146] This application also provides a computer-readable storage medium storing one or more computer programs, which, when executed by multiple processors, cause the processors to perform the following actions: Figure 1 and Figure 4 The method and its detailed scheme are shown.

[0147] This application also provides a computer program product, which includes a computer program that, when run on a computer, causes the computer to perform the methods of any of the above embodiments.

[0148] References to features, advantages, or similar language in this specification do not imply that all features and advantages achievable with this solution should be included or included in any single implementation thereof. Rather, references to features and advantages are understood to mean that a particular feature, advantage, or characteristic described in connection with an embodiment is included in at least one embodiment of this solution. Therefore, discussions of features, advantages, and similar language throughout this specification may, but do not necessarily, refer to the same embodiments.

[0149] Furthermore, the features, advantages, and characteristics described herein can be combined in any suitable manner in one or more embodiments. Based on the description herein, those skilled in the art will recognize that this solution can be implemented without one or more specific features or advantages of a particular embodiment. In other instances, additional features and advantages can be appreciated in specific embodiments not presented in all embodiments of this solution.

[0150] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. Furthermore, any changes or modifications made by those skilled in the art based on the ideas of this application, and on the specific implementation methods and application scope of this application, are all within the scope of protection of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for determining the altitude of an aerial vehicle, characterized in that, The method comprises: receiving multi-height layer map information of an airspace; preprocessing the multi-height layer map information, wherein the preprocessed multi-height layer map information corresponds to three-dimensional image information in the shape of (x, y, z), wherein x, y, and z respectively correspond to the width, length, and height of the preprocessed multi-height layer map information, and the preprocessed multi-height layer map information comprises x*y*z voxels and corresponding voxel values, wherein x, y, and z are positive integers; and inputting the preprocessed multi-height layer map information into a 3D-2D hybrid encoder-decoder model to output a two-dimensional height proposal map comprising x*y voxels, wherein the voxel values of the x*y voxels respectively correspond to x*y height proposal values.

2. The method of claim 1, wherein, The preprocessing of the multi-height layer map information comprises: scaling the multi-height layer map information according to a preset resolution; and cropping the scaled multi-height layer map information according to the size requirements of the 3D-2D hybrid encoder-decoder model.

3. The method of claim 1, wherein, The inputting of the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model to output a two-dimensional height proposal map comprising x*y voxels comprises: performing 3D convolution processing on the preprocessed multi-height layer map information through an encoder to obtain x*y*1 voxels and determine the passing difficulty coefficient corresponding to each voxel in the x*y*1 voxels; performing 2D transpose convolution and attention mechanism processing on the x*y*1 voxels through a decoder to generate height layer recommendation probability values of the decoded x*y*z voxels; and determining a two-dimensional height proposal map with a size of x*y according to the height layer recommendation probability values of the decoded x*y*z voxels, wherein the voxel values of each voxel in the two-dimensional height proposal map correspond to height proposal values at corresponding positions.

4. The method of claim 3, wherein, Before determining a two-dimensional height proposal map with a size of x*y according to the height layer recommendation probability values of the decoded x*y*z voxels, the inputting of the preprocessed multi-height layer map information into the 3D-2D hybrid encoder-decoder model to output a two-dimensional height proposal map comprising x*y voxels further comprises: obtaining airspace dynamic data in real time through a preset rule engine; adjusting the height layer recommendation probability values of the decoded x*y*z voxels according to the airspace dynamic data to obtain adjusted height layer recommendation probability values; and wherein the determining of the two-dimensional height proposal map with a size of x*y according to the height layer recommendation probability values of the decoded x*y*z voxels comprises: determining the two-dimensional height proposal map with a size of x*y according to the adjusted height layer recommendation probability values.

5. The method of claim 3, wherein, The 2D transpose convolution and attention mechanism processing on the x*y*1 voxels through the decoder to generate height layer recommendation probability values of the decoded x*y*z voxels comprises: passing the middle-level feature resolution feature map reserved by the encoder to the corresponding level of the decoder through a skip connection channel.

6. The method according to any one of claims 1 to 5, characterized in that, Before inputting the preprocessed multi-height layer map information into a 3D-2D hybrid encoder-decoder model and outputting a two-dimensional height proposal map including x*y voxels, the method further comprises: neural network training is performed on the 3D-2D hybrid encoder-decoder model.

7. The method of claim 6, wherein, The loss function used for neural network training of the 3D-2D hybrid encoder-decoder model comprises one or more of a continuity loss function, a height loss function, a map complexity loss function, and a dynamic balance loss function.

8. An apparatus for determining the spatial flight height, characterized by Comprise: a receiving module configured to receive multi-height layer map information of a space; a preprocessing module configured to preprocess the multi-height layer map information, wherein the preprocessed multi-height layer map information corresponds to three-dimensional image information with a shape of (x, y, z), wherein x, y, and z respectively correspond to the width, length, and height of the preprocessed multi-height layer map information, the preprocessed multi-height layer map information comprises x*y*z voxels and their corresponding voxel values, wherein x, y, and z are positive integers; and an output module configured to input the preprocessed multi-height layer map information into a 3D-2D hybrid encoder-decoder model and output a two-dimensional height proposal map including x*y voxels, wherein the voxel values of the x*y voxels respectively correspond to x*y height proposal values.

9. An electronic device, comprising: comprise a memory and a processor, the memory having a computer program stored thereon, and the processor, when executing the computer program stored on the memory, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-7. The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-7.

Images