An airport scene trajectory prediction method based on multi-scale S2 space coding

Abstract

The application discloses an airport scene trajectory prediction method based on multi-scale S2 space coding, comprising the following steps: step 1, obtaining the historical trajectory state of a target agent, the state of surrounding neighbor agents and local high-precision map information, and constructing input features; step 2, using a global discrete grid system to perform multi-scale S2 space discretization coding on the continuous longitude and latitude coordinates of the target agent, and mapping the longitude and latitude coordinates into a space feature embedding vector with a hierarchical structure; step 3, further processing the input features to obtain time series dynamics features and map topological structure features; step 4, splicing the multi-scale S2 space feature embedding vector, the time series dynamics features, the map topological structure features and the agent type features, and performing global feature interaction and updating; and step 5, inputting the feature sequence after the interaction and updating into a multi-modal Gaussian mixture decoder, and predicting and outputting the probability trajectory distribution of the target agent in a future prediction time domain.

Description

Technical Field

[0001] This invention relates to the field of trajectory prediction, and more particularly to an airport surface trajectory prediction method based on multi-scale S2 spatial coding. Background Technology

[0002] With the continued surge in global air transport demand, the airport surface of large hub airports has evolved into a complex system characterized by high density, strong dynamics, and tight physical coupling. Frequent interactions between aircraft and ground support vehicles within extremely limited spatiotemporal resources increase the safety risks of runway incursions and pose a severe challenge to the scheduling efficiency of ground operations control systems. Against this backdrop, accurately predicting the future trajectory and multimodal intentions of moving targets in specific complex airport environments has become a core supporting technology for achieving refined taxiing guidance, early conflict warning, and next-generation airport collaborative decision-making (A-CDM) systems.

[0003] The field of trajectory prediction has undergone a paradigm shift from traditional kinematic models (such as Kalman filtering) to deep learning models. In recent years, advanced deep learning prediction models have significantly improved short-term prediction accuracy by encoding historical state sequences, multi-agent neighborhood interactions, and vectorized high-precision maps. However, in order to pursue cross-scene generalization capabilities across different road network structures, current mainstream trajectory prediction methods generally adopt a "de-absolute" data processing approach, that is, normalizing all input data to a local coordinate system centered on the target agent.

[0004] This approach, based on a local relative coordinate system, has achieved good results in autonomous driving on open urban roads, but it reveals fundamental flaws in the relatively closed and strictly regulated environment of airports. Unlike open roads, the layout of large airports has long-term stability, and the taxiing logic of aircraft is highly coupled to specific absolute geographical locations (for example, specific runway intersections imply fixed yield priorities, and specific apron entrances mandate specific deceleration modes). The local coordinate system strips away these key absolute geographical semantics, making it difficult for the predictive model to establish a direct "location-rule" mapping relationship.

[0005] Because of the loss of spatial semantics, existing models cannot effectively "memorize" the hard topological constraints specific to a particular airport. When existing general prediction models are directly transferred to complex airport scene scenarios, they are prone to producing prediction results that seriously violate physical boundaries, such as cutting corners, ignoring stop lines, or even running off the runway, when faced with large curvature turns or complex intersections. Especially in long-term prediction tasks, this relative geometric feature without absolute spatial "anchor" constraints can lead to a rapid amplification of accumulated errors, causing severe path drift.

[0006] Therefore, there is an urgent need for a surface trajectory prediction method that can effectively integrate absolute geographic information with complex spatiotemporal interaction features to compensate for the spatial semantic loss caused by the existing local relative coordinate system, thereby improving the accuracy and physical compliance of long-term airport surface trajectory prediction. Summary of the Invention

[0007] Purpose of the invention: The technical problem to be solved by the present invention is to provide an airport surface trajectory prediction method based on multi-scale S2 spatial coding, which addresses the shortcomings of the existing technology.

[0008] To address the aforementioned technical problems, this invention discloses an airport surface trajectory prediction method based on multi-scale S2 spatial coding, comprising the following steps:

[0009] Step 1: Obtain the historical trajectory state of the target agent, the state of surrounding neighboring agents, and local high-precision map information to construct input features. The input features specifically include:

[0010] Target historical state sequence: contains the sequence of the target in the observation time domain. Each state vector is represented as a vector containing two-dimensional absolute coordinates, velocity, heading angle, acceleration and agent type.

[0011] Neighbor interaction features: Includes historical state sequences of multiple nearest neighbors, with neighbor coordinates transformed to a local coordinate system with the target's current location as the origin.

[0012] Local high-precision map information: Local high-precision map information is represented as a set of map fragments. Each map fragment consists of a sequence of centerline points obtained by discrete sampling at fixed intervals and carries corresponding semantic attributes.

[0013] Specifically, at the current moment Based on the current position of the target agent Centered on a high-precision map of the airport, a local map area within a preset spatial range is extracted; preferably, the local map area is centered on... The centered circular or rectangular region has a radius of . Or the side length is .

[0014] Within the local map area, all centerlines associated with runway centerlines, taxiway centerlines, apron entrance / exit guide lines, and stop lines intersecting with that area are extracted and divided into multiple map segments according to topological breakpoints. These breakpoints include intersections, runway / taxiway connection points, stop / wait locations, runway entrances / exits, and centerline endpoints. Each map segment consists of a sequence of centerline points obtained through discrete sampling at fixed intervals and carries corresponding semantic attributes. These semantic attributes include at least segment type, direction of travel, whether it is adjacent to a stop / wait line, whether it is adjacent to a runway, and its corresponding functional area. For map segments partially falling within the local map area, their centerline points within that area are retained. Preferably, adjacent map segments directly connected to the target's current position along the topological network within a preset number of hops forward and backward are retained to ensure that all possible turns, waiting areas, and crossing areas involved in the target's future trajectory are included in this set of map segments.

[0015] Step 2: Employ the Discrete Global Grid System (DGGS) to perform multi-scale S2 spatial discretization encoding on the continuous latitude and longitude coordinates of the target agent, mapping the latitude and longitude coordinates into hierarchical spatial feature embedding vectors. This specifically includes macroscopic regional encoding and microscopic relative encoding.

[0016] Step 2-1: Macro-region coding. Select a Level 16 grid with a side length of approximately L to capture specific functional zones within the airport. For the Level 16 S2 grid ID to which the target coordinates belong, a feature hashing strategy is used to map it to a fixed-size vocabulary index. The calculation formula is:

[0017]

[0018] in, For hash functions, This refers to the size of the macro-vocabulary.

[0019] Step 2-2: Microscopic Relative Encoding. Based on the quadtree hierarchical structure of the S2 grid, the specific Level 16 parent node grid determined in Step 2-1 is further subdivided, selecting a Level 20 sub-grid with a side length of approximately L2 to precisely locate the agent's specific position within that functional area (parent node). Its position within the parent node is calculated. Internal relative Hilbert index The formula is:

[0020]

[0021] in, This represents the linear positional mapping of the S2 mesh on the Hilbert space fill curve.

[0022] Steps 2-3: Spatial Feature Fusion. After concatenating the macroscopic region encoding and the microscopic relative encoding, a multilayer perceptron is used for feature interaction and dimensionality reduction extraction to obtain the final spatial feature embedding vector. The calculation formula is:

[0023]

[0024] Step 3: Further process the input features to obtain temporal dynamic features and map topology features.

[0025] Temporal feature encoding: The historical state sequences of the target agent and its neighboring agents are processed by a temporal encoder, which consists of a linear layer, a positional encoder, and two Transformer Encoder layers to extract temporal dynamic features. .

[0026] Map feature encoding: For each map segment, the relative displacement between point pairs is first calculated, local features are extracted via the PointNet network, and then aggregated using max pooling to obtain line segment-level map features. .

[0027] Step 4: Concatenate the multi-scale S2 spatial feature embedding vector, temporal dynamics features, map topology features, and agent type features, and input them into the global attention interaction Transformer module. Global feature interaction and updates are performed through a self-attention mechanism. The specific processing procedure is as follows:

[0028] Step 4-1: Deeply fuse historical trajectory features Current moment absolute space embedding and type embedding Construct the target token;

[0029] Step 4-2: Extract the line segment-level map features The projection is a set of map tokens, and the temporal features of neighbors are projected into a set of neighbor tokens. The target token, the set of map tokens, and the set of neighbor tokens are concatenated to form a complete input sequence for global interaction with the Transformer. ;

[0030] Step 4-3: Input sequence The data is fed into a multi-head self-attention module, and a dynamic mask matrix is ​​introduced to block out invalid neighbors or map nodes, thus completing the global feature interaction update.

[0031] Step 5: Input the interactively updated feature sequence into the multimodal Gaussian mixture decoder to predict and output the probability trajectory distribution of the target agent in the future prediction time domain.

[0032] Step 5-1: Output probability trajectory distribution The Gaussian mixture model is used for parameterization, and the formula is as follows:

[0033]

[0034] in, This represents the comprehensive input features constructed at the current moment. This represents the trajectory state of the target intelligent agent in the future prediction time domain. For the first The confidence probability of each modality To predict the positional mean, It is a covariance matrix that includes variance and correlation coefficient.

[0035] Step 5-2: During model training, a "winner-takes-all" optimization strategy is adopted, targeting only the optimal mode with the smallest error to the true trajectory. Perform parameter updates. Total loss function. It consists of three parts:

[0036]

[0037] in, To maximize the negative log-likelihood loss of the true trajectory under the best predicted Gaussian distribution, To use the classification loss optimized by cross-entropy, Smooth-L1 regression loss is applied to the mean of the optimal mode.

[0038] Beneficial effects:

[0039] 1. Effectively overcomes physical boundary violations and significantly improves the physical compliance of predicted trajectories: This invention innovatively introduces hierarchical multi-scale S2 spatial discretization coding (combining macroscopic hash features and microscopic relative indexes) to inject absolute geographic location features as explicit spatial inductive biases into the prediction model. This design is like setting "digital anchors" for the model, enabling the model to establish a direct mapping between "absolute position and specific rules," effectively overcoming the phenomena of cutting corners or ignoring stop lines that easily occur in traditional local relative coordinate systems, thus violating physical boundaries.

[0040] 2. Fundamentally correcting path drift and significantly improving long-term prediction accuracy: This invention deeply integrates multi-scale S2 spatial embedding with a Transformer-based global interactive network, compensating for the semantic information lost by pure vector maps in local coordinate systems. Experiments based on the large-scale real airport operations dataset (Amelia) show that this invention performs exceptionally well in handling high-density, strongly coupled taxiing scenarios. Compared to classic pure time-series prediction models (such as Vanilla LSTM), the final displacement error (FDE) is significantly reduced by approximately 69.3% (down to 22.68m) in a 50-second long-term prediction window; even compared to a strong baseline model without S2 encoding, the final displacement error of its 50-second long-term prediction is improved by 17.1%, effectively correcting the divergence and drift phenomena in long-term prediction.

[0041] 3. Serving as a strong regularizer to improve model training convergence and generalization stability: The absolute geospatial information introduced in this invention acts as a strong regularizer during the dynamic model training process. This not only allows the model to exhibit a steeper error descent gradient and converge to a better error limit in the early stages of training (e.g., converging to around 11.5m, better than the benchmark of 13.0m), but also effectively reduces the error fluctuation variance of the model when facing unseen samples. This helps the model quickly lock onto the correct motion mode using location priors, avoiding confusion in similar geometric structures (such as parallel taxiways).

[0042] 4. Supporting Airport Collaborative Decision-Making Systems with Significant Engineering Application Value: The method of this invention can provide high-precision multimodal intent and location prediction support for complex operations at large hub airports. This not only verifies the effectiveness of absolute geocoding in specific closed scenarios but also provides solid technical support for building a new generation of Airport Collaborative Decision-Making (A-CDM) systems that support high-precision location prediction. It has significant engineering application value for improving operational efficiency, achieving refined taxiway guidance, and preventing runway incursion risks. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall architecture of an airport surface trajectory prediction model (S2-ViT) based on multi-scale S2 spatial coding provided in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the mapping of the hierarchical S2 spatial coding mechanism provided in an embodiment of the present invention in a local area of ​​an airport (KMDW airport).

[0045] Figure 3This is a comparison chart of the convergence change curves of the minimum average displacement error (MinADE) of the model provided in the embodiments of the present invention during the training process on the validation set.

[0046] Figure 4 is a schematic diagram illustrating the verification effect of the qualitative comparison between the predicted trajectory and the actual trajectory in a typical airport taxiway turning scenario provided by an embodiment of the present invention. Detailed Implementation

[0047] This invention addresses the shortcomings of existing airport surface trajectory prediction methods, which generally employ local relative coordinate systems. These methods often result in the loss of crucial absolute geographic semantics, difficulty in establishing direct "location-rule" mappings, and consequently, susceptibility to issues such as cornering, ignoring stop lines, and long-term prediction path drift under complex surface constraints. This invention provides a multi-scale S2 spatial coding-based airport surface trajectory prediction method. By introducing a Discrete Global Grid System (DGGS), this invention injects multi-scale absolute geographic information as an explicit spatial inductive bias into the prediction model, thereby endowing the model with the ability to "memorize" the functions of specific regions and significantly improving the accuracy and physical compliance of long-term surface trajectory prediction. The S2 algorithm is the Google S2 Geometry Library, often referred to in engineering as the "S2 Spherical Geometry Library" or "S2 Spatial Indexing Algorithm."

[0048] Example 1:

[0049] This embodiment provides an airport surface trajectory prediction method based on multi-scale S2 spatial coding. The implementation process of the present invention will be described in detail below with reference to specific figures and implementation steps.

[0050] like Figure 1 The diagram shown illustrates the overall architecture of the airport surface trajectory prediction model (S2-ViT) based on multi-scale S2 spatial coding provided by this invention. The specific implementation process and steps are as follows:

[0051] Step 1: Analyze the airport surface operating environment and construct the input representation and dataset.

[0052] This embodiment uses the large-scale real-world dataset (Amelia) from Chicago Midway International Airport (KMDW) for implementation. The data is strictly divided into training, validation, and test sets in an 8:1:1 ratio, with a sampling frequency of 1Hz. Let the current time be... The model acquires the historical state of the target agent over the past 10 seconds (10 steps) and predicts its multimodal trajectory for the next 50 seconds (50 steps). The model's input features... It includes the following three parts:

[0053] 1) Target historical state sequence: Contains the sequence of the target agent in the observation time domain, with each state vector... Where x and y represent two-dimensional absolute coordinates, Indicates speed, Indicates heading angle, Indicates acceleration and Indicates the type of intelligent agent.

[0054] 2) Neighbor interaction features: Extract the historical state sequence (sequence of state vectors) of the nearest neighbor aircraft around the target. To ensure the translation invariance of local interactions, the neighbor coordinates need to be transformed into a local coordinate system with the target's current position as the origin.

[0055] 3) Local high-precision map information: Local high-precision map information is represented as a set of map fragments (local map).

[0056] Specifically, at the current moment Based on the current position of the target agent Centered on a high-precision map of the airport, a local map area within a preset spatial range is extracted; preferably, the local map area is centered on... The centered circular or rectangular region has a radius of . Or the side length is In one embodiment, the radius is taken. A local region of km; in other embodiments, A range of 0.8 to 1.5 km is acceptable.

[0057] Within the local map area, all centerlines associated with runway centerlines, taxiway centerlines, apron entrance / exit guide lines, and stop lines intersecting with that area are extracted and divided into multiple map segments according to topological breakpoints. These breakpoints include intersections, runway / taxiway connection points, stop / wait locations, runway entrances / exits, and centerline endpoints. Each map segment consists of a sequence of centerline points obtained through discrete sampling at fixed intervals and carries corresponding semantic attributes. These semantic attributes include at least segment type, direction of travel, whether it is adjacent to a stop / wait line, whether it is adjacent to a runway, and its corresponding functional area. For map segments partially falling within the local map area, their centerline points within that area are retained. Preferably, adjacent map segments directly connected to the target's current position along the topological network within a preset number of hops forward and backward are retained to ensure that all possible turns, waiting areas, and crossing areas involved in the target's future trajectory are included in this set of map segments.

[0058] Step 2: Perform multi-scale S2 spatial encoding (feature encoding) using a discrete global grid system.

[0059] like Figure 2 As shown, the continuous latitude and longitude coordinates of the target agent at each time step within the historical observation time window (past 10 seconds) are displayed. Mapped to spatial embedding vectors with hierarchical structure (target fusion encode) is used to compensate for the lack of geographic semantics in local coordinate systems.

[0060] Step 2-1 (Level 16 Macro-region Coding): Select a Level 16 grid with a side length of approximately 150m to capture specific functional zones within the airport (such as runway intersections, holding points, etc.). Let... target coordinates Level 16 S2 grid ID, coordinates Falling into space Within the corresponding grid range, that is, through Established coordinate position Mapping relationships with specific functional zones within the airport. Due to the extremely large number of global S2 grids, a global embedding vocabulary cannot be directly created; therefore, a custom embedding vocabulary is constructed. A feature hashing strategy is employed to map it to a fixed-size vocabulary index. The formula is:

[0061]

[0062] in, The size of the macro-vocabulary, which is a hash function and is constructed by the user. Setting it to 50,000 means mapping the massive grid IDs to this fixed-size vocabulary space using hash modulo;

[0063] Step 2-2 (Level 20 Microscopic Relative Encoding): Based on the quadtree hierarchical structure of the S2 mesh, the specific Level 16 parent node mesh determined in Step 2-1 is further subdivided, selecting a Level 20 sub-mesh with a side length of approximately 10m to precisely locate the agent's specific position within the functional area (parent node). Its position within the parent node is calculated. Internal relative Hilbert index The formula is:

[0064]

[0065] in, target coordinates The corresponding Level 20 S2 grid ID; This represents the linear positional mapping of the S2 mesh on the Hilbert space-filling curve. Due to the quadtree hierarchical structure of S2, each Level 16 mesh strictly contains... Each Level 20 subgrid, therefore .

[0066] Steps 2-3 (Inter-room Feature Fusion): After stitching together macroscopic and microscopic features, feature interaction and dimensionality reduction extraction are performed using a multilayer perceptron (MLP) to obtain the final spatial embedding. :

[0067]

[0068] in, This indicates a vector concatenation operation.

[0069] Step 3: Extract heterogeneous features of temporal dynamics and map topology

[0070] Temporal encoding: The historical state sequences of the target agent and its neighboring agents are processed by a temporal encoder, which consists of a linear layer, a positional encoder, and two Transformer Encoder layers to extract temporal dynamic features. .

[0071] Map feature encoding (VectorNet map encoding): For each map segment, the relative displacement between point pairs is first calculated, local features are extracted via the PointNet network, and then aggregated using max pooling to obtain line segment-level map features. .

[0072] Step 4: Use Transformer for global attention interaction.

[0073] Step 4-1: Project each modal feature onto the same dimension And concatenate them into the input sequence Tokens. First, obtain the agent type embedding, based on the type label of the target agent. (Such as aircraft, ground vehicles, etc.), which are mapped to type feature vectors through a learnable type embedding layer. ;

[0074] Next, by deeply integrating historical trajectory features Current moment absolute space embedding and type embedding , Construct the target token The formula is:

[0075]

[0076] Step 4-2: Extract the line segment-level map features Projection as a collection of map tokens Projecting the temporal features of neighbors into a set of neighbor tokens Finally, these are concatenated into a complete input sequence for global interaction with the Transformer.

[0077]

[0078] Step 4-3: Feed the above sequence into the Multi-Head Self-Attention (MHSA) module. Introduce a dynamic mask matrix. To filter out invalid neighbors or map nodes, global feature interaction updates are completed. To prevent overfitting, Dropout of 0.1 is used in the Transformer interaction layer.

[0079] Step 5: Perform multi-task decoding and model training on the interactively updated feature sequence multimodal Gaussian mixture decoder to predict and output the probability trajectory distribution of the target agent in the future prediction time domain:

[0080] Step 5-1: Output probability trajectory distribution The Gaussian mixture model is used for parameterization, and the formula is as follows:

[0081]

[0082] In this embodiment, The representative model constructs the comprehensive input features at the current moment, specifically referring to the historical state sequence of the target agent in the past 10 seconds, neighbor interaction features, and local high-precision map information obtained in step 1. It represents the trajectory state of the target intelligent agent in the future prediction time domain, specifically referring to its continuous position coordinate sequence in the next 50 seconds.

[0083] In the formula, In this embodiment, the total number of modes is [number]. In the airport surface operation scenario of this invention, "modality" specifically corresponds to the macroscopic taxiing intention or specific path selection that the target intelligent agent (such as an aircraft or support vehicle) may take in the future. Because the surface environment (such as complex intersections) has multiple legal driving routes, this... Each modality represents the most likely future of the intelligent agent. Different taxiing decision routes (e.g., maintaining a straight course along the current taxiway, turning left into an intersecting taxiway, turning right off the runway, or slowing down and stopping before the stop line).

[0084] For the first The mode (i.e. the th mode) The confidence probability of a path selection intention. This represents the mean of the predicted location under this modal intent. The covariance matrix includes variance and correlation coefficient (including variance) and correlation coefficient ).

[0085] Step 5-2: During model training, a "winner-takes-all" optimization strategy is adopted, targeting only the optimal mode with the smallest error to the true trajectory. Update the parameters.

[0086] Model training: Implemented using the PyTorch framework, trained on a single NVIDIA RTX 4090 GPU. The Adam optimizer was used, with an initial learning rate set to... The system employs a cosine annealing strategy for attenuation, with a batch size of 64 and a total of 150 epochs. Training utilizes a winner-take-all (WTA) optimization strategy, updating parameters only for the optimal mode that minimizes the error with the true trajectory.

[0087] Total loss function for:

[0088]

[0089] The parameters and loss terms are explained below:

[0090] : Represents the true ground truth trajectory. This true trajectory data comes directly from the Amelia airport surface real-world operation dataset described in step 1, and is a continuous sequence of position coordinates of the target agent that is actually observed and recorded in the future prediction time domain (i.e., the next 50 seconds).

[0091] : Represents the actual trajectory Index of the optimal mode (Winner) with the smallest error.

[0092] : Represents the optimal mode The predicted trajectory follows a Gaussian distribution.

[0093] : Represents the set of model outputs containing the confidence probabilities of each modality.

[0094] : Represents the optimal mode The predicted mean of the Gaussian distribution location.

[0095] : This is a hyperparameter used to balance the weights of various losses. In this embodiment, and The preferred value range is [0.1, 10]; more specifically, in this embodiment, the default value is [0.1, 10]. .

[0096] : Negative Log-Likelihood Loss, used to maximize the likelihood probability of the true trajectory under the optimal predicted Gaussian distribution.

[0097] Cross-entropy classification loss is used to optimize output probabilities and improve the optimal mode. The confidence level.

[0098] : This is a Smooth-L1 based regression loss that directly imposes a hard constraint on the mean of the optimal mode to accelerate model convergence.

[0099] Effect verification:

[0100] like Figure 3 As shown, the minimum average displacement error (MinADE) convergence curve of the validation set during the training process indicates that the model of this invention (red curve) after introducing multi-scale S2 encoding converges to around 11.5m at Epoch 100, which is significantly better than the baseline model (blue curve) without introducing S2 encoding. Moreover, the variance of the validation error fluctuation is smaller, which proves that absolute geographic information acts as a strong regularization term.

[0101] Figure 4 shows a comparison of trajectory predictions under a typical taxiway turning scenario at KMDW Airport. Specifically, focus on the sharp turn section slightly above the center of the figure. In Figure 4a (the comparison model without S2 spatial coding), it can be observed that in the middle and later stages of the turn, the predicted trajectory (red dashed line) deviates significantly from the actual trajectory (green solid line), shifting towards the inside of the curve (i.e., the "angle-cutting phenomenon"). This causes the predicted trajectory to deviate from the taxiway centerline, posing a risk of violating physical boundaries.

[0102] Conversely, as shown in Figure 4b, the method provided by this invention, by incorporating absolute geographic semantics, ensures that its predicted trajectory (red dashed line) and the actual trajectory (green solid line) remain highly consistent and overlap throughout the entire large-angle turn, effectively overcoming the tangential phenomenon caused by a purely local relative coordinate system and significantly improving the physical compliance of the predicted trajectory. Quantitative tests show that, under a 50-second prediction window, the final displacement error (FDE) of this invention reaches 22.68m, a reduction of 69.3% compared to the classic LSTM model.

[0103] This invention provides a method for predicting airport surface trajectories based on multi-scale S2 spatial coding. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting airport surface trajectories based on multi-scale S2 spatial coding, characterized in that, Includes the following steps: Step 1: Obtain the historical trajectory state of the target agent, the state of surrounding neighboring agents, and local high-precision map information to construct input features; Step 2: Use a global discrete grid system to perform multi-scale S2 spatial discretization encoding on the continuous latitude and longitude coordinates of the target intelligent agent, and map the latitude and longitude coordinates into spatial feature embedding vectors with hierarchical structure; Step 3: Further process the input features to obtain temporal dynamic features and map topology features; Step 4: Concatenate the multi-scale S2 spatial feature embedding vector, temporal dynamic features, map topology features, and agent type features, and input them into the global attention interaction Transformer module to perform global feature interaction and update through the self-attention mechanism; Step 5: Input the interactively updated feature sequence into the multimodal Gaussian mixture decoder to predict and output the probability trajectory distribution of the target agent in the future prediction time domain.

2. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 1, characterized in that, The input features mentioned in step 1 specifically include: Target historical state sequence: contains the sequence of the target in the observation time domain. Each state vector is represented as a vector containing two-dimensional absolute coordinates, velocity, heading angle, acceleration, and agent type. Neighbor interaction features: Includes historical state sequences of multiple nearest neighbors, with neighbor coordinates transformed into a local coordinate system with the target's current location as the origin; Local high-precision map information: Local high-precision map information is represented as a set of map fragments. Each map fragment consists of a sequence of centerline points obtained by discrete sampling at fixed intervals and carries corresponding semantic attributes.

3. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 2, characterized in that, The acquisition of the local high-precision map information specifically involves: Step 1-1, at the current moment Based on the current position of the target agent Centered on the airport scene, a local map area within a preset spatial range is extracted from the high-precision map of the airport scene; Steps 1-2: Within the local map area, extract all runway centerlines, taxiway centerlines, apron entrance / exit guide lines, stop lines, and associated centerlines that intersect with the area, and divide them into multiple map segments according to topological breakpoints; for map segments that partially fall within the local map area, retain their centerline points located within the local map area.

4. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 3, characterized in that, The topological breakpoints mentioned in steps 1-2 include intersections, runway-task connection points, stop and wait positions, runway entrances / exits, and centerline endpoints.

5. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 1, characterized in that, The multi-scale S2 spatial discretization coding described in step 2 specifically includes macroscopic regional coding and microscopic relative coding: Step 2-1: Macro-region coding: Select a Level 16 grid with a side length of approximately L to capture specific functional zones within the airport, and set... For the Level 16 S2 grid ID to which the target coordinates belong, a feature hashing strategy is used to map it to a fixed-size vocabulary index. ; Step 2-2: Microscopic Relative Encoding: Based on the quadtree hierarchical structure of the S2 grid, the specific Level 16 parent node grid determined in Step 2-1 is further subdivided, and a Level 20 sub-grid with a side length of approximately L2 is selected to precisely locate the agent's specific position within the functional area and calculate its position relative to the parent node. Internal relative Hilbert index ; Steps 2-3: Spatial Feature Fusion: After concatenating the macroscopic region encoding results with the microscopic relative results, feature interaction and dimensionality reduction extraction are performed using a multilayer perceptron to obtain the final spatial feature embedding vector. The calculation formula is: in, This indicates a vector concatenation operation.

6. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 5, characterized in that, Step 3 details include: Temporal feature encoding: The historical state sequences of the target agent and its neighboring agents are processed by a temporal encoder, which consists of a linear layer, a positional encoder, and two Transformer Encoder layers to extract temporal dynamic features. ; Map feature encoding: For each map segment, first calculate the relative displacement between point pairs, extract local features through the PointNet network, and then aggregate them using max pooling to obtain line segment-level map features. .

7. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 6, characterized in that, The specific processing procedure of the global attention interaction module mentioned in step 4 is as follows: Step 4-1: Deeply fuse historical trajectory features Current moment absolute space embedding and type embedding Construct the target token; Step 4-2: Extract the line segment-level map features The projection is a set of map tokens, and the temporal features of neighbors are projected into a set of neighbor tokens. The target token, the set of map tokens, and the set of neighbor tokens are concatenated to form a complete input sequence for global interaction with the Transformer. ; Step 4-3: Input sequence The data is fed into a multi-head self-attention module, and a dynamic mask matrix is ​​introduced to block out invalid neighbors or map nodes, thus completing the global feature interaction update.

8. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 1, characterized in that, The probability trajectory distribution output by the multimodal Gaussian mixture decoder described in step 5 The Gaussian mixture model is used for parameterization, and the formula is as follows: in, This represents the comprehensive input features constructed at the current moment. This represents the trajectory state of the target intelligent agent in the future prediction time domain. The total number of modes, For the first The confidence probability of each modality To predict the positional mean, It is a covariance matrix that includes variance and correlation coefficient.

9. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 8, characterized in that, The modality specifically corresponds to the macroscopic gliding intention or specific path selection that the target intelligent agent may take in the future.

10. The airport surface trajectory prediction method based on multi-scale S2 spatial coding according to claim 9, characterized in that, Training only targets the optimal mode that minimizes the error with the true trajectory. Perform parameter updates, total loss function for: in, To maximize the negative log-likelihood loss of the true trajectory under the best predicted Gaussian distribution, To use the classification loss optimized by cross-entropy, The Smooth-L1 regression loss is applied to the mean of the optimal mode. Represents the true ground truth trajectory. Representation and actual trajectory The index of the optimal mode with the minimum error. Represents the optimal mode The predicted trajectory follows a Gaussian distribution. : Represents the set of model outputs containing the confidence probabilities for each modality. : Represents the optimal mode, The predicted mean of the Gaussian distribution location. Hyperparameters are used to balance the weights of various losses.

Images