Vehicle trajectory prediction method and system based on target region feature aggregation

Abstract

The application relates to the field of intelligent transportation, and discloses a vehicle trajectory prediction method and system based on target region feature aggregation. The method comprises the following steps: acquiring a vehicle historical trajectory and map information and performing coordinate normalization; extracting scene context features and global abstract features; generating coarse-grained future trajectory hypotheses and compressing the future trajectory hypotheses into future abstract information, and performing feature back-annotation to obtain an enhanced scene representation; extracting intention features, scoring candidate anchor points, and performing two-dimensional offset regression to obtain refined anchor points; constructing a target region around the refined anchor points, aggregating elements in the region to obtain target region enhanced features; generating an initial trajectory proposal and performing residual correction to obtain a final prediction trajectory. The system comprises a preprocessing module, an encoding module, an enhancement module, a positioning module, a feature aggregation module, and a trajectory generation module for executing the above method. The application effectively improves the positioning accuracy and scene consistency of vehicle trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a vehicle trajectory prediction method and system based on target area feature aggregation. Background Technology

[0002] Accurate vehicle trajectory prediction is a prerequisite for safe decision-making and path planning in autonomous driving systems. Existing trajectory prediction methods typically extract the target vehicle's historical trajectory sequence, the motion state of surrounding traffic participants, and road network structure features to map and output the future driving trajectory. However, this conventional one-way prediction paradigm relies solely on historical observation information during the feature encoding stage, failing to incorporate the perception of potential future evolution trends in advance. This results in limitations in the model's inference of future directions when dealing with long-term and multimodal prediction tasks.

[0003] Meanwhile, existing technologies often employ a scoring strategy of discrete sampling along lane lines when predicting the trajectory endpoint. While this approach ensures the predicted endpoint lies within the road network, it inevitably introduces quantization errors due to limitations in sampling point density, affecting the final positioning accuracy. Furthermore, traditional methods often rely on global scene features or local features centered solely on the vehicle's current position when generating specific trajectories. When the predicted endpoint is far from the current position, this feature representation struggles to effectively impose physical constraints on the microscopic road network topology and dynamic interactions surrounding the endpoint. Without precise guidance from the endpoint's local environment, the generated predicted trajectory is prone to deviating from lane boundaries, exhibiting reduced smoothness, or displaying poor scene consistency at the endpoint. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a vehicle trajectory prediction method and system based on target region feature aggregation. This solves the problems of existing vehicle trajectory prediction methods, such as the lack of prior perception of future trends in the feature encoding stage, the limitation of endpoint positioning accuracy due to single discrete sampling, and the poor consistency of predicted trajectory scenarios due to the lack of local environmental constraints at the endpoint.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a vehicle trajectory prediction method based on target region feature aggregation, comprising the following steps: The system acquires the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene. It then performs a unified coordinate transformation to obtain the normalized input scene. Context encoding is performed on the normalized input scene to output context features and generate global summary features corresponding to the target vehicle. Based on the scene context features and the global summary features, a future cue vector and a coarse-grained future trajectory hypothesis are generated and compressed into future summary information; The future summary information is injected back into the scene context features using a gated update method to obtain an enhanced scene representation; Modal intent features are extracted based on the enhanced scene representation, and candidate anchor points are scored and subjected to two-dimensional offset regression based on the modal intent features to obtain refined anchor points; A target region is constructed around the refined anchor point, and local feature aggregation is performed within the target region to obtain enhanced features of the target region. An initial trajectory proposal is generated based on the enhanced features of the target region, and the corresponding confidence score is calculated. The initial trajectory proposal is then modified by a refinement network to obtain the final predicted trajectory after refinement.

[0006] Preferably, the step of acquiring the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene, and performing a unified coordinate transformation to obtain the normalized input scene, specifically includes: Acquire the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene; The current position of the target vehicle is taken as the origin of the local coordinate system, and the current orientation of the target vehicle is taken as the reference direction. A unified coordinate transformation is performed on all input data. For trajectory sequences whose length is less than the set number of historical observation time steps, zero-padding alignment is performed at the front of the time series, and validity markers are constructed to obtain the normalized input scene.

[0007] Preferably, the step of performing scene context encoding on the normalized input scene, outputting scene context features, and generating global summary features corresponding to the target vehicle specifically includes: The trajectory sequence and the historical trajectory are encoded using kinematic features, and the vectorized map information is encoded using vectorized structure. The system performs interactive modeling and calculation of relationships between intelligent agents and between intelligent agents and the map, and collaboratively integrates motion information, road information, and interaction information to output the scene context features. The scene context features are compressed and summarized to generate global summary features corresponding to the target vehicle.

[0008] Preferably, the step of generating future cue vectors and coarse-grained future trajectory hypotheses based on the scene context features and the global summary features, and compressing them into future summary information, specifically includes: For each prediction mode, a corresponding learnable pattern embedding is independently constructed. The learnable pattern embedding is linearly mapped and combined with the global summary features to generate the future hint vector; The future cue vector is converted into a spatial coordinate sequence through a decoding operation to form the coarse-grained future trajectory hypothesis of the corresponding modality; The coarse-grained future trajectory hypothesis and the corresponding future cue vector are subjected to feature aggregation processing, which is then compressed and aggregated into the future summary information.

[0009] Preferably, the step of injecting the future summary information back into the scene context features using a gated update method to obtain an enhanced scene representation specifically includes: Configure a gating update mechanism; The gating update mechanism learns gating weights and filters future evolution clues that are highly correlated with the current traffic environment. Using the gating weights, the extracted future evolutionary cues are nonlinearly fused with the scene context features to obtain the enhanced scene representation.

[0010] Preferably, the step of extracting modal intent features based on the enhanced scene representation, and scoring candidate anchor points and performing two-dimensional offset regression based on the modal intent features to obtain refined anchor points specifically includes: Assign a corresponding query vector to each prediction modality; The matching weight between the query vector and the enhanced scene representation is calculated using an attention mechanism, and the modal intent features are extracted. A set of candidate anchor points is formed by discrete sampling points on the center line of the lane. Each candidate anchor point is scored according to the modal intent features, and the candidate anchor point with the highest score is selected. The refined anchor point is calculated by performing two-dimensional offset regression on the candidate anchor point with the highest score in a continuous space to output the anchor point offset.

[0011] Preferably, the step of constructing a target region around the refined anchor point and performing local feature aggregation within the target region to obtain enhanced features of the target region specifically includes: Based on the set target area radius parameters, the target area is constructed with the corresponding refined anchor point as the geometric center; Extract local map elements and relevant traffic participants located within the target area from the enhanced scene representation; Calculate the relative positional relationship, relative directional relationship, and interaction relationship between the refined anchor point and the extracted local features; Based on the calculated relative and interaction relationships, the map element information and dynamic volume information within the target area are weighted and aggregated to obtain the enhanced features of the target area.

[0012] Preferably, the step of generating an initial trajectory proposal based on the enhanced features of the target region, calculating the corresponding confidence score, and then refining the initial trajectory proposal using a refinement network to obtain the refined final predicted trajectory specifically includes: Multiple initial trajectory proposals are decoded based on the enhanced features of the target region; The target region enhancement features are mapped and normalized using a linear mapping network, and the confidence scores corresponding to each prediction mode are calculated. The initial trajectory proposal is subjected to trajectory coordinate embedding and temporal feature modeling, and the corresponding trajectory residual is calculated in combination with the target region enhancement features; The trajectory residual is superimposed on the initial trajectory proposal for local deviation correction, resulting in the refined final predicted trajectory.

[0013] Preferably, the method, during the training phase, specifically includes: The generated coarse-grained future trajectory hypothesis, initial trajectory proposal, refined final predicted trajectory, and confidence scores for each prediction mode are integrated. Construct a joint loss function that includes coarse future trajectory supervision, proposed trajectory supervision, refined trajectory supervision, and modality classification supervision; The model is jointly supervised and optimized based on the joint loss function.

[0014] A second aspect of the present invention provides a vehicle trajectory prediction system based on target region feature aggregation, comprising: The scene input and preprocessing module is used to receive the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and vectorized map information, and to complete the local coordinate normalization process to obtain the normalized input scene; The encoding module is used to output scene context features that integrate historical movement, map structure, and interaction relationships, and generate global summary features; The enhancement module is used to generate coarse-grained future trajectory hypotheses and future summary information based on the scene context features and the global summary features, and to back-inject and update the future summary information to obtain an enhanced scene representation; The localization module is used to extract modal intent features and perform two-dimensional offset regression to output refined anchor points corresponding to each modality; The feature aggregation module is used to aggregate map and dynamic volume information within the target area around the refined anchor point to obtain enhanced features of the target area; The trajectory generation and scoring output module is used to generate an initial trajectory proposal based on the enhanced features of the target region and correct the trajectory residual, and output multiple candidate future trajectories and their corresponding confidence scores.

[0015] This invention provides a vehicle trajectory prediction method and system based on target region feature aggregation. It has the following beneficial effects: 1. This invention significantly improves the model's ability to predict long-term trajectory evolution through a coarse-grained future generation and future-aware scene enhancement mechanism. In the feature encoding stage, the system utilizes global summary features and modal priors to pre-decode coarse-grained future trajectory hypotheses and compresses them into future summary information, which is then injected back into the scene context. This gated update feature injection method ensures that the encoded enhanced scene representation not only contains rich historical dynamics and road network structure but also incorporates potential future trajectory clues in advance. This effectively avoids the limitations of models relying solely on historical observations for inference and improves the accuracy of multimodal trajectory prediction.

[0016] 2. This invention employs an intent prediction and two-stage anchor point localization structure, significantly improving the accuracy of trajectory endpoint localization. The system first extracts modal intent features through an attention mechanism and performs a coarse scoring and filtering of discrete sampling points on the lane centerline. Then, it performs two-dimensional offset regression on the candidate anchor point with the highest score in continuous space. This dimensionality reduction approach, from discrete space search to continuous space correction, effectively compensates for the quantization error caused by single discrete sampling, achieving high-precision endpoint coordinate prediction while ensuring the reasonableness of endpoint generation.

[0017] 3. This invention enhances the scene consistency and smoothness of predicted trajectories by combining target region feature aggregation with a two-stage trajectory generation strategy. A local target region is constructed around high-precision refined anchor points. The system explicitly extracts road topology and dynamic interaction elements within the endpoint's neighborhood and performs weighted aggregation. Using the enhanced features of this target region as conditions, an initial trajectory proposal is first generated, and then the trajectory residual is calculated through a refined network to correct local deviations. This design fully utilizes the local physical constraints of the endpoint, reduces interference from globally irrelevant environmental information, and makes the final generated trajectory more consistent with the real road network structure and traffic flow patterns. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a system module architecture diagram of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 This invention provides a vehicle trajectory prediction method based on target region feature aggregation, comprising the following steps: S1. Obtain the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the target vehicle's neighborhood, and the vectorized map information corresponding to the scene. Using the target vehicle's current position as the origin of the local coordinate system and the target vehicle's current orientation as the reference direction, perform a unified coordinate transformation on the historical trajectory, map elements, and candidate future targets to obtain the normalized input scene.

[0021] S2. Perform scene context encoding on the normalized input scene. Encode the kinematic features of the target vehicle and nearby traffic participants' historical trajectories to extract displacement, velocity, and direction information. Perform vectorized structural encoding on lane centerlines, boundary lines, and semantic region map elements to extract lane geometry, topological connectivity, and semantic region information. Establish a unified scene representation through agent-map interaction and agent-to-agent interaction, outputting scene context features that integrate historical motion, map structure, and interaction relationships. Compress and summarize the scene context features to generate global summary features corresponding to the target vehicle.

[0022] S3. Generate coarse future trajectories and future cue vectors based on scene context features. For each predicted modality, construct a learnable pattern embedding, and combine the target vehicle's historical motion representation and scene summary features to obtain a modality-level future cue vector. Decode the future cue vector into coarse-grained future trajectory hypotheses. Compress the coarse future trajectory and future cue vector into future summary information.

[0023] S4. Perform future-aware scene enhancement. A gated update method is used to inject future summary information back into agent features and map features respectively, resulting in an enhanced scene representation that includes historical states and future cues.

[0024] S5. Perform intent prediction and two-stage anchor point localization based on enhanced scene representation. A query vector is assigned to each predicted modality. A high-level semantic pattern is extracted from the enhanced scene representation using an attention mechanism to obtain modal intent features. Lane centerline sampling points are used as a candidate anchor point set, and the candidate anchor points are scored based on the intent features. Two-dimensional offset regression is performed on the candidate anchor point with the highest score to obtain refined anchor points.

[0025] S6. Construct a target region with a defined spatial range around each refined anchor point, and perform local feature aggregation within the target region. Select lane polyline segments from the map side that are within a set threshold distance from the refined anchor point, and select relevant traffic participants located in the neighborhood of the target region from the dynamic volume side. Combine the relative positional relationship, relative directional relationship, and interaction relationship between the anchor point and regional elements for weighted aggregation to obtain the enhanced features of the target region.

[0026] S7. Initial trajectory proposals are generated based on the enhanced features of the target region, and the trajectory residuals are further corrected through a refinement network to obtain multiple candidate future trajectories and their corresponding confidence scores. Initial trajectory proposals are generated based on the enhanced features of the target region, and the confidence scores for each prediction modality are calculated. The initial trajectories are then refined through a refinement network to correct the residuals, resulting in the refined final predicted trajectory. During the training phase, a joint loss function is constructed based on the coarse future trajectory, the initial trajectory proposals, the refined predicted trajectory, and the confidence classification results to optimize the model.

[0027] Please see the appendix Figure 2 This invention provides a vehicle trajectory prediction system based on target region feature aggregation, comprising: The scene input and preprocessing module is used to receive the target vehicle's historical trajectory, the trajectories of nearby traffic participants, and vectorized map information, and to complete the local coordinate normalization process. The encoding module is used to output an initial scene representation that integrates historical movement, map structure, and multi-agent interaction relationships; An enhancement module is used to generate coarse future trajectories and future cue vectors, and to update the initial scene representation; The localization module is used to extract modal intent features and output refined anchor points corresponding to each modality; The feature aggregation module is used to aggregate map and dynamic volume information within the target neighborhood around the refined anchor point; The trajectory generation and scoring output module is used to generate multiple candidate future trajectories and output the corresponding confidence scores.

[0028] In this embodiment of the invention, step S1 performed by the scene input and preprocessing module may further include the following sub-steps: S101. Obtain the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood of the target vehicle, and the vectorized map information corresponding to the current scene.

[0029] For acquiring the historical trajectories of target vehicles and other traffic participants, those skilled in the art can use sensors such as vehicle-mounted LiDAR and cameras for environmental perception and multi-target tracking. The information acquisition process is well-known in the field and will not be elaborated upon here. The vectorized map information corresponding to the current scene includes at least one or more of the following: lane centerline, lane boundary line, drivable area, and semantic region. This map information consists of polyline segments or polygons formed by connecting a series of discrete two-dimensional coordinate points.

[0030] The trajectory sequence of the target vehicle within the historical observation window is represented by the following formula: ; In the formula, The trajectory sequence of the target vehicle within the historical observation window; time Two-dimensional position coordinates; This represents the number of historical observation time steps.

[0031] S102. Using the current position of the target vehicle as the origin of the local coordinate system and the current orientation of the target vehicle as the reference direction, perform a unified coordinate transformation on the historical trajectory of the target vehicle, the trajectory of nearby traffic participants, and map elements.

[0032] The process of performing coordinate transformation on historical trajectory data and vectorized map coordinate points in the global coordinate system is shown in the following equation: ; In the formula, These are the normalized coordinates of the trajectory points; The global two-dimensional position coordinates before transformation; The coordinates of the target vehicle at the current moment; For the target vehicle's current orientation Constructed coordinate rotation matrix; Let be the transpose of this rotation matrix.

[0033] Performing coordinate transformation converts all input data to a local coordinate system based on the current state of the target vehicle, thus reducing the impact of differences in global coordinate space scale and absolute orientation differences in different scenarios.

[0034] S103. Align the trajectory sequences that are not long enough and construct validity markers.

[0035] When the length of the obtained trajectory sequence of some traffic participants is less than the set number of historical observation time steps, a completion operation is required.

[0036] For the alignment of trajectory sequences, those skilled in the art can perform alignment by zero-padding at the beginning of the time series and use a Boolean mask matrix as a validity marker to distinguish between real trajectory observation points and filled points. The filling calculation process is a well-known technique in the art and will not be described in detail here.

[0037] This step ensures consistency in the dimensions of input data across different real-world traffic scenarios. The scenario input and preprocessing module ultimately outputs the normalized historical trajectory of the target vehicle, the historical trajectories of nearby traffic participants, and a vectorized map representation, which are then provided to the encoding module.

[0038] In this embodiment of the invention, step S2 performed by the encoding module may further include the following sub-steps: S201. Perform kinematic feature encoding on the normalized historical trajectory of the target vehicle and the historical trajectories of nearby traffic participants.

[0039] The encoding module receives the scene input and the normalized data output by the preprocessing module, and extracts the dynamic features of displacement, velocity and direction of the target vehicle and nearby traffic participants.

[0040] For the extraction and encoding of historical trajectory temporal features, those skilled in the art can use recurrent neural networks or self-attention mechanisms for sequence modeling. The sequence modeling and feature extraction process is a well-known technology in this field and will not be elaborated here.

[0041] S202. Perform map structure encoding and propagation on vectorized map information.

[0042] For map elements containing lane centerlines, lane boundary lines, and semantic regions, the encoding module performs vectorized structural encoding. During map structure encoding, local geometry and topological semantics are propagated between point-level map elements and polygon-level road structures. Specifically, point-level representations are obtained by extracting local spatial features from the discrete coordinate points constituting the map elements. Feature aggregation operations are then performed along the lane extension direction or regional topological connectivity to form a polygon-level road structure representation that includes local geometry and topological connectivity.

[0043] S203, Perform interactive modeling of agent-to-agent interactions and agent-to-map interactions.

[0044] After obtaining historical motion encoding representations and map structure representations, the encoding module establishes relationships between vehicles and between vehicles and map features. The interaction modeling process calculates the mutual influence weights between different traffic participants through an attention mechanism, and calculates the dependencies between traffic participants and surrounding lanes and semantic regions, establishing a unified scene representation. The above encoding process collaboratively integrates motion information, road information, and interaction information within the same high-dimensional representation space, and outputs time-aligned agent feature sequences and map feature sequences, providing a shared context for subsequent processing.

[0045] Combining the above sub-steps, the encoding module outputs scene context features that integrate historical movement, map structure, and interaction relationships. The specific calculation of the scene context features is shown in the following formula: ; In the formula, For scene context features; Encoding function for scene context; The trajectory sequence of the target vehicle within the historical observation window; For nearby traffic participants to gather; To vectorize map information.

[0046] Based on the calculated scene context features, the encoding module further performs a feature aggregation operation to compress and summarize the scene context features, generating a global summary feature corresponding to the target vehicle. The calculation of the global summary feature is shown in the following formula: ; In the formula, The global summary features corresponding to the target vehicle; This is a feature aggregation operation used to compress and summarize scene context features; These are scene context features.

[0047] In this embodiment of the invention, step S3 performed by the enhancement module may further include the following sub-steps: S301. Construct a learnable pattern embedding for each prediction modality and generate a modality-level future cue vector by combining scene summary features.

[0048] The enhancement module receives the scene context features and the global summary features corresponding to the target vehicle from the encoding module. In order to distinguish different potential futures in the multimodal prediction task, the system independently constructs a corresponding learnable pattern embedding for each prediction modality.

[0049] The aforementioned learnable patterns are embedded and linearly mapped to the global summary features corresponding to the target vehicle. The mapped features are then combined to generate a future cue vector that includes modal priors and historical motion summaries.

[0050] The specific calculation of the future hint vector is shown in the following formula: ; In the formula, For the first Future cue vectors for each predicted modality; For the first Learnable pattern embeddings corresponding to each prediction modality; The linear mapping parameters are embedded for the corresponding learnable patterns; These are the linear mapping parameters corresponding to the global summary features; This refers to the global summary features corresponding to the target vehicle.

[0051] S302. Decode the future cue vector into a coarse-grained future trajectory hypothesis for the corresponding modality.

[0052] Based on the future cue vectors obtained from the above steps, the enhancement module converts them into a spatial coordinate sequence through decoding, forming a coarse-grained hypothesis about the future trajectory. This step is not intended to directly output the final predicted trajectory that meets the requirements of high-precision regression, but rather to establish coarse-grained conditional information for multiple possible futures in advance.

[0053] This design allows the model to be exposed to potential future evolutionary trends in advance during the feature encoding stage, avoiding the subsequent prediction module from relying entirely on historical observations to infer the future, thus providing modality-level future priors for subsequent future perception enhancement and anchor point localization.

[0054] For network operations that decode high-dimensional feature vectors into trajectory space sequences, those skilled in the art can use multilayer perceptrons to build corresponding decoding networks to achieve this. The feature decoding and mapping process is a well-known technology in this field and will not be described in detail here.

[0055] The specific calculation for generating the coarse-grained future trajectory hypothesis is shown in the following formula: ; In the formula, For the first The coarse future trajectory corresponding to each mode; This is a coarse future trajectory decoding function; For the first A future cue vector for each prediction modality.

[0056] In this embodiment of the invention, step S4 performed by the enhancement module may further include the following sub-steps: S401. Compress and aggregate the coarse future trajectory and the corresponding future hint vector into future summary information.

[0057] After acquiring initial future cues for each prediction modality, the enhancement module performs feature convergence processing on the generated coarse-grained future trajectory assumptions and corresponding future cue vectors. This operation reduces and compresses the multi-dimensional spatial trajectory information and high-dimensional cue information at the feature level, extracting a compact representation that can comprehensively characterize the potential future trend, and using this as the conditional input for subsequent feature updates.

[0058] For computational operations involving multidimensional feature compression and aggregation, those skilled in the art can use multilayer perceptrons or pooling layer structures to achieve feature dimensionality reduction extraction. The network construction and computation process are well-known technologies in this field and will not be elaborated here.

[0059] The calculation for generating future summary information is shown in the following formula: ; In the formula, This is the future summary information formed by the coarse future trajectory and the future cue vector; For feature aggregation operations; For the first The coarse future trajectory corresponding to each mode; For the first A future cue vector for each prediction modality.

[0060] S402. The future summary information is injected back into the scene representation using a gating update method to obtain the enhanced scene representation.

[0061] After obtaining future summary information, the enhancement module injects it into the already generated agent features and map features. During the injection process, the system configures a gating update mechanism to control the information flow of future summary information to scene features. This mechanism learns gating weights, filters future evolution clues highly relevant to the current traffic environment, and non-linearly fuses the extracted future clues with the original scene features.

[0062] Through the aforementioned feature back-injection operation, the system injects conditional information related to the future destination and future trajectory into the encoded features before formally entering the trajectory generation step. The enhanced scene representation after fusion and updating simultaneously covers the historical state of the target vehicle and its potential future movement trends.

[0063] For the specific network implementation of the gating update mechanism, those skilled in the art can use gating recurrent units or attention network structures based on gating units to filter features and fuse states. The construction of the gating network and the forward computation process are well-known technologies in the field and will not be described in detail here.

[0064] The specific calculation of the enhanced scene representation is shown in the following formula: ; In the formula, For future-oriented scene representation with enhanced perception; This is a future-aware augmentation function used to implement gated updates; To integrate scene context features including historical movements, map structure, and interaction relationships; This is the future summary information formed by the coarse future trajectory and the future cue vector.

[0065] In this embodiment of the invention, step S5 performed by the positioning module may further include the following sub-steps: S501. Assign a query vector to each predicted modality and extract modal intent features from the enhanced scene representation.

[0066] The localization module receives the enhanced scene representation output by the enhancement module. For multiple pre-defined prediction modalities, the localization module independently assigns a corresponding query vector to each modality. By configuring an attention mechanism, it calculates the matching weight between the query vector and the enhanced scene representation, extracts pattern-level semantic features from the enhanced scene representation, and generates the intent features for the corresponding modality.

[0067] For the specific network implementation of the attention mechanism, those skilled in the art can use a multi-head attention network structure to extract and calculate feature information. The network construction and forward inference process are well-known technologies in this field and will not be described in detail here.

[0068] The specific calculation of modal intent features is shown in the following formula: ; In the formula, For the first The intent features corresponding to each modality; This is an attention extraction operation used to extract pattern-level semantic information from the enhanced scene representation; For the first A query vector for each predicted modality assignment; For future scene representation with enhanced perception.

[0069] S502. Obtain a set of candidate anchor points and score the candidate anchor points based on modal intent features.

[0070] The localization module acquires a set of candidate anchor points, which consists of discrete sampling points along the lane centerline, serving as candidate locations for the trajectory endpoint. Based on the extracted modal intent features and the features of each candidate anchor point, the localization module calculates the preference degree of each predicted modality for each candidate anchor point, obtaining the candidate anchor point score result.

[0071] The score calculation for candidate anchor points is shown in the following formula: ; In the formula, For the first The modality pair of the first Scoring of each candidate anchor point; It is a normalized exponential function; For the first The transpose matrix of the intent features corresponding to each modality; For the first One candidate anchor point feature.

[0072] S503. Select the candidate anchor points with the highest scores and perform two-dimensional offset regression to obtain refined anchor points in continuous space.

[0073] Based on the candidate anchor point scoring results, the localization module selects the most probable target location for each mode, i.e., the candidate anchor point with the highest score. To address the quantization error caused by discrete sampling points, the localization module performs two-dimensional offset regression on the selected candidate anchor points in continuous space, outputs the corresponding anchor point offset, and then calculates the corrected refined anchor points.

[0074] For the specific implementation of the offset regression network, those skilled in the art can use a multilayer perceptron or a linear regression layer to map the features and output continuous coordinate offset values. The network construction process is a well-known technology in this field and will not be described in detail here.

[0075] The specific calculation for refining the anchor point is shown in the following formula: ; In the formula, For the first Refined anchor points corresponding to each mode; The candidate anchor point with the highest score; This represents the predicted anchor point offset.

[0076] Through the two-stage anchor point positioning structure constructed by the candidate anchor point scoring stage and the anchor point offset refinement stage, the positioning module establishes a connection between the discrete sampling search space and the continuous endpoint representation, which makes up for the insufficient accuracy caused by single discrete sampling and improves the accuracy of the endpoint position representation.

[0077] In this embodiment of the invention, step S6 performed by the feature aggregation module may further include the following sub-steps: S601. Construct a target area with a limited spatial range around the refined anchor point.

[0078] The feature aggregation module receives the refined anchor points for each modality output by the localization module. Based on the set target region radius parameter, the system constructs a local circular spatial range with the corresponding refined anchor point as the geometric center. This spatial range is used to delineate the geographical boundary for subsequent local feature extraction, avoiding interference from global scene information on the endpoint local constraints.

[0079] The specific construction conditions for the target region are shown in the following formula: ; In the formula, To focus on refining the anchor point The target area to be constructed; Let be the coordinates of any point within the target area; For the first Refined anchor points corresponding to each mode; The radius of the target area is set; The distance is the L2 norm.

[0080] S602. Extract local map elements and relevant traffic participants within the target area from the enhanced scene representation.

[0081] After determining the boundary of the target area, the feature aggregation module receives the scene representation enhanced by future perception. From the map side, the feature aggregation module selects lane polyline segments and semantic region elements whose distance from the refined anchor point is within the radius of the target area as defined above; from the dynamic volume side, the feature aggregation module selects relevant traffic participants whose spatial location is within the neighborhood of the target area. This step explicitly extracts the local road network structure and dynamic environmental state within the neighborhood of the endpoint.

[0082] S603. Combine the relative relationship between anchor points and regional elements to perform weighted aggregation of information within the region.

[0083] For the extracted local features, the feature aggregation module further calculates the relative positional and directional relationships between the refinement anchor point and these regional features, and extracts the potential interaction relationships between vehicles and local features. Based on the above relative and interaction relationships, the system calculates the influence weight of each local feature on the current refinement anchor point, and uses this weight to perform a weighted aggregation operation on the map element information and dynamic volume information within the target area, finally outputting the enhanced features of the target area.

[0084] By performing information aggregation within the target area, the system explicitly incorporates road constraints, local lane topology, and neighboring dynamic volume interaction constraints within the endpoint neighborhood into the feature representation, overcoming the deficiency of insufficient local constraints when using only a single point-level target as a condition.

[0085] For the specific network implementation of feature weighted aggregation, those skilled in the art can use graph attention networks or cross-attention mechanisms based on spatial relative coding to calculate fusion weights and extract aggregation features. The construction and forward computation process of the spatial feature aggregation network are well-known technologies in this field and will not be elaborated here.

[0086] The specific calculation of the target region enhancement features is shown in the following formula: ; In the formula, For the first Enhancement features of the target region corresponding to each modality; This is a function for regional aggregation operations, used to perform weighted aggregation of map and dynamic volume information within a target area; The target area is constructed around the refined anchor point; For future scene representation with enhanced perception.

[0087] In this embodiment of the invention, step S7 performed by the trajectory generation and scoring output module may further include the following sub-steps: S701. Generate an initial trajectory proposal based on the enhanced features of the target region, and calculate the confidence score of the predicted mode.

[0088] The trajectory generation and scoring output module receives the target region enhancement features output by the feature aggregation module and uses them as conditional input for trajectory decoding. During the trajectory generation proposal stage, the module decodes multiple coarse initial future trajectory hypotheses based on the target region enhancement features. Simultaneously, a linear mapping network is used to map and normalize the target region enhancement features, calculating the prediction probability, i.e., the confidence score, for each prediction mode.

[0089] For the specific implementation of the trajectory decoding network and the scoring calculation network, those skilled in the art can use a multilayer perceptron to build a trajectory proposal generation network and use a fully connected layer structure to implement probability calculation. The forward computation process of the network is a well-known technology in this field and will not be described in detail here.

[0090] The specific calculation for the initial trajectory proposal is shown in the following formula: ; In the formula, For the first Initial trajectory proposals corresponding to each mode; For the proposed trajectory generation function; For the first Enhancement features for the target region corresponding to each modality.

[0091] The confidence score is calculated using the following formula: ; In the formula, For the first The prediction confidence level corresponding to each modality; It is a normalized exponential function; This is the transpose matrix of the linear mapping parameters used to calculate the confidence level; For the first Enhancement features for the target region corresponding to each modality.

[0092] S702. Perform residual correction on the initial trajectory proposal to obtain multiple refined candidate future trajectories.

[0093] After generating the initial trajectory proposal, the trajectory generation and scoring output module enters the refinement stage. Trajectories are embedded into coordinates and temporal features are modeled on the initial trajectory proposal, and the corresponding trajectory spatial residuals are calculated by combining the enhanced features of the target region.

[0094] The calculated trajectory residuals are superimposed on the initial trajectory proposal to correct local biases, resulting in a refined final predicted trajectory. Multiple candidate future trajectories are then ranked according to the previously calculated confidence scores, and the final trajectory prediction result is output.

[0095] The system adopts a two-stage trajectory generation structure of proposal and refinement. First, a rough trajectory proposal is generated, and then the endpoint positioning and local spatial deviation are refined and corrected, which improves the smoothness of the overall trajectory, the accuracy of endpoint positioning, and the consistency of the scene.

[0096] For the specific network implementation of temporal feature modeling and residual calculation, those skilled in the art can use gated recurrent units or temporal networks based on self-attention mechanisms to extract sequence features and output residual coordinates. The network construction process is a well-known technology in this field and will not be described in detail here.

[0097] The specific calculation of the refined final predicted trajectory is shown in the following formula: ; In the formula, This is the final predicted trajectory after refinement; For the first Initial trajectory proposals corresponding to each mode; This represents the trajectory residual predicted by the network.

[0098] S703. Construct a joint loss function during the training phase to perform joint supervised optimization of each module.

[0099] During system model training, the trajectory generation and scoring output module integrates the outputs from the coarse future generation stage, the trajectory proposal stage, and the trajectory refinement stage to construct a joint training objective that includes coarse future trajectory supervision, proposed trajectory supervision, refined trajectory supervision, and modality classification supervision. This step ensures the consistency of coarse future prompt quality, trajectory generation accuracy, and modality probability allocation during the backpropagation update process.

[0100] The specific calculation of the joint loss function is shown in the following formula: ; In the formula, This is the joint loss function during the training phase; The loss term is for the coarse future trajectory; The loss term for the initial trajectory proposal; The loss term for refining the trajectory; The loss term for modality classification; , , , These are the weighting coefficients for the corresponding loss items.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A vehicle trajectory prediction method based on target region feature aggregation, characterized in that, Includes the following steps: The system acquires the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene. It then performs a unified coordinate transformation to obtain the normalized input scene. Context encoding is performed on the normalized input scene to output context features and generate global summary features corresponding to the target vehicle. Based on the scene context features and the global summary features, a future cue vector and a coarse-grained future trajectory hypothesis are generated and compressed into future summary information; The future summary information is injected back into the scene context features using a gated update method to obtain an enhanced scene representation; Modal intent features are extracted based on the enhanced scene representation, and candidate anchor points are scored and subjected to two-dimensional offset regression based on the modal intent features to obtain refined anchor points; A target region is constructed around the refined anchor point, and local feature aggregation is performed within the target region to obtain enhanced features of the target region. An initial trajectory proposal is generated based on the enhanced features of the target region, and the corresponding confidence score is calculated. The initial trajectory proposal is then modified by a refinement network to obtain the final predicted trajectory after refinement.

2. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The process of acquiring the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene, and performing a unified coordinate transformation to obtain the normalized input scene, specifically includes: Acquire the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and the vectorized map information corresponding to the current scene; The current position of the target vehicle is taken as the origin of the local coordinate system, and the current orientation of the target vehicle is taken as the reference direction. A unified coordinate transformation is performed on all input data. For trajectory sequences whose length is less than the set number of historical observation time steps, zero-padding alignment is performed at the front of the time series, and validity markers are constructed to obtain the normalized input scene.

3. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The step of performing scene context encoding on the normalized input scene, outputting scene context features, and generating global summary features corresponding to the target vehicle specifically includes: The trajectory sequence and the historical trajectory are encoded using kinematic features, and the vectorized map information is encoded using vectorized structure. The system performs interactive modeling and calculation of relationships between intelligent agents and between intelligent agents and the map, and collaboratively integrates motion information, road information, and interaction information to output the scene context features. The scene context features are compressed and summarized to generate global summary features corresponding to the target vehicle.

4. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The process of generating future cue vectors and coarse-grained future trajectory hypotheses based on the scene context features and the global summary features, and compressing them into future summary information, specifically includes: For each prediction mode, a corresponding learnable pattern embedding is independently constructed. The learnable pattern embedding is linearly mapped and combined with the global summary features to generate the future hint vector; The future cue vector is converted into a spatial coordinate sequence through a decoding operation to form the coarse-grained future trajectory hypothesis of the corresponding modality; The coarse-grained future trajectory hypothesis and the corresponding future cue vector are subjected to feature aggregation processing, which is then compressed and aggregated into the future summary information.

5. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The future summary information is injected back into the scene context features using a gated update method. The enhanced scene representation is obtained, specifically including: Configure a gating update mechanism; The gating update mechanism learns gating weights and filters future evolution clues that are highly correlated with the current traffic environment. Using the gating weights, the extracted future evolutionary cues are nonlinearly fused with the scene context features to obtain the enhanced scene representation.

6. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The process of extracting modal intent features based on the enhanced scene representation, and scoring candidate anchor points and performing two-dimensional offset regression based on the modal intent features to obtain refined anchor points, specifically includes: Assign a corresponding query vector to each prediction modality; The matching weight between the query vector and the enhanced scene representation is calculated using an attention mechanism, and the modal intent features are extracted. A set of candidate anchor points is formed by discrete sampling points on the center line of the lane. Each candidate anchor point is scored according to the modal intent features, and the candidate anchor point with the highest score is selected. The refined anchor point is calculated by performing two-dimensional offset regression on the candidate anchor point with the highest score in a continuous space to output the anchor point offset.

7. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The process of constructing a target region around the refined anchor point and performing local feature aggregation within the target region to obtain enhanced features for the target region specifically includes: Based on the set target area radius parameters, the target area is constructed with the corresponding refined anchor point as the geometric center; Extract local map elements and relevant traffic participants located within the target area from the enhanced scene representation; Calculate the relative positional relationship, relative directional relationship, and interaction relationship between the refined anchor point and the extracted local features; Based on the calculated relative and interaction relationships, the map element information and dynamic volume information within the target area are weighted and aggregated to obtain the enhanced features of the target area.

8. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The process of generating an initial trajectory proposal based on the enhanced features of the target region, calculating the corresponding confidence score, and then refining the initial trajectory proposal using a refinement network to obtain the refined final predicted trajectory specifically includes: Multiple initial trajectory proposals are decoded based on the enhanced features of the target region; The target region enhancement features are mapped and normalized using a linear mapping network, and the confidence scores corresponding to each prediction mode are calculated. The initial trajectory proposal is subjected to trajectory coordinate embedding and temporal feature modeling, and the corresponding trajectory residual is calculated in combination with the target region enhancement features; The trajectory residual is superimposed on the initial trajectory proposal for local deviation correction, resulting in the refined final predicted trajectory.

9. The vehicle trajectory prediction method based on target region feature aggregation according to claim 1, characterized in that, The method, in the training phase, specifically includes: The generated coarse-grained future trajectory hypothesis, initial trajectory proposal, refined final predicted trajectory, and confidence scores for each prediction mode are integrated. Construct a joint loss function that includes coarse future trajectory supervision, proposed trajectory supervision, refined trajectory supervision, and modality classification supervision; The model is jointly supervised and optimized based on the joint loss function.

10. A vehicle trajectory prediction system based on target region feature aggregation, applied to the vehicle trajectory prediction method based on target region feature aggregation as described in any one of claims 1-9, characterized in that, include: The scene input and preprocessing module is used to receive the trajectory sequence of the target vehicle within the historical observation window, the historical trajectories of other traffic participants in the neighborhood, and vectorized map information, and to complete the local coordinate normalization process to obtain the normalized input scene; The encoding module is used to output scene context features that integrate historical movement, map structure, and interaction relationships, and generate global summary features; The enhancement module is used to generate coarse-grained future trajectory hypotheses and future summary information based on the scene context features and the global summary features, and to back-inject and update the future summary information to obtain an enhanced scene representation; The localization module is used to extract modal intent features and perform two-dimensional offset regression to output refined anchor points corresponding to each modality; The feature aggregation module is used to aggregate map and dynamic volume information within the target area around the refined anchor point to obtain enhanced features of the target area; The trajectory generation and scoring output module is used to generate an initial trajectory proposal based on the enhanced features of the target region and correct the trajectory residual, and output multiple candidate future trajectories and their corresponding confidence scores.

Images