Vehicle track prediction method for scene without high-precision map coverage

Abstract

The invention relates to the technical field related to automatic driving, in particular to a vehicle trajectory prediction method for a scene without high-precision map coverage, and the method comprises the steps: designing an end-to-end dynamic graph model, extracting spatial features through constructing a dynamic graph structure, modeling historical trajectory time sequence dependence through employing a time sequence attention module with a mask, and predicting the trajectory of a vehicle. A trajectory query vector is generated through multi-stage time interval sampling and cross attention fusion, a trajectory is gradually generated by using a segmented iterative decoder, error accumulation is relieved, and a trajectory query distillation method is further provided: a teacher network fuses map features, a student network aligns trajectory query through two stages of distillation (encoder and decoder stages), and a trajectory query result is obtained. Teacher network priori knowledge is learned, the intensity is dynamically adjusted in combination with a self-adaptive distillation mechanism, experiments prove that the prediction precision of the method is close to that of a map dependence method in a map-free scene, and the vehicle track prediction robustness in a high-precision-free map is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a vehicle trajectory prediction method for scenarios without high-precision map coverage. Background Technology

[0002] In autonomous driving systems, trajectory prediction, as a key link connecting perception and planning modules, is responsible for reasoning and predicting the future movement state of surrounding traffic participants. The accuracy and robustness of trajectory prediction are directly related to the safety and stability of the entire autonomous driving vehicle system.

[0003] Currently, many advanced trajectory prediction methods incorporate high-precision maps as an important auxiliary information source. These maps provide detailed semantic information, helping models understand scene constraints and improving the accuracy and interpretability of trajectory prediction. However, the construction of high-precision maps often relies on the combination of high-resolution LiDAR and SLAM technology, accompanied by extensive manual annotation and data calibration, resulting in long production cycles, high costs, and difficulty in achieving high-frequency real-time updates. In practical deployments, the acquisition of high-precision maps may be constrained by factors such as policies and regulations, safety restrictions, and the complexity of geographical areas, especially in remote areas, urban fringe areas, or dynamic construction scenarios, where map coverage and timeliness may fail to meet the system's requirements for stability and reliability. These issues significantly limit the applicability and adaptability of trajectory prediction methods relying on high-precision maps in real-world autonomous driving scenarios. Therefore, researching trajectory prediction models that can achieve stable and accurate predictions even in scenarios without high-precision maps has become an important topic in the field of autonomous driving. Summary of the Invention

[0004] The purpose of this invention is to provide a deblurring method for event-based spatiotemporal motion perception in real dynamic scenarios, in order to solve the problem of vehicle trajectory prediction in scenarios without high-precision maps. It aims to improve the model's prediction performance of trajectory distribution in mapless scenarios by improving graph attention mechanism, hierarchical aggregation and trajectory query distillation strategy, and to overcome the shortcomings of existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a vehicle trajectory prediction method for scenarios without high-precision map coverage, comprising the following steps:

[0006] S1. Constructing a dynamic graph structure and extracting spatial features: Construct a dynamic graph based on relative coordinates, with the target vehicle and neighboring vehicles as nodes. Filter adjacent nodes through Euclidean distance threshold, and use an improved graph attention network to fuse vehicle interaction features. The graph attention network uses a scaled dot product attention mechanism to calculate node weights, and adaptively fuses the original features and attention output through gated residual connections.

[0007] S2. Temporal Feature Extraction and Temporal Dependency Modeling: A temporal attention module with a masking mechanism is adopted. Causality is ensured through a lower triangular mask, and the temporal modeling capability is enhanced by relative time position embedding to extract the temporal dependency features of historical trajectories.

[0008] S3. Hierarchical spatiotemporal feature fusion: Through multi-level time interval sampling and cross-attention mechanism, a hierarchical spatiotemporal feature representation is constructed to generate a multimodal trajectory query vector. The multi-level time interval sampling realizes feature aggregation at different time scales through a hierarchical aggregation module.

[0009] S4. Segmented Iterative Decoding and Trajectory Generation: A segmented iterative decoder with a fixed time window is used to gradually generate and stitch together future trajectories. Self-attention is used to optimize cross-modal trajectory information and alleviate the accumulation of errors in long-term prediction.

[0010] Preferably, the dynamic graph structure and spatial feature extraction processing in S1 includes:

[0011] S101. Using the target vehicle as the center, filter adjacent vehicle nodes based on the Euclidean distance threshold of 100 meters, and construct a directed edge τ. i,j =(A j →A i ), where A i For the target vehicle, A j For vehicles in the neighboring area;

[0012] S102, Obtain the global coordinates of neighboring vehicles. Convert to relative coordinates centered on the target vehicle:

[0013]

[0014] in, Used to characterize the spatial interaction relationships between vehicles;

[0015] S103, Calculate A for each node. i The feature vector contains the position increment, instantaneous velocity, and heading angle:

[0016]

[0017] in The position increment of adjacent frames, Instantaneous velocity For heading angle;

[0018] S104. Normalize the node features and map features such as coordinates and velocity to the [0,1] interval to eliminate the influence of dimensional differences on the model.

[0019] In addition, scaled dot product attention is used, which integrates the attention results of the lower triangular mask and the relative temporal position embedding, as well as the nonlinear mapping of the feedforward network. The calculation steps are as follows:

[0020] S105, Calculate the value for node A j To A i Attention weight a j :

[0021]

[0022] Where D is the number of channels, Q i K j This represents the channel-wise dot product operation between the Query and the Key;

[0023] S106. The normalized attention weights are used to calculate the weighted sum of the features of neighboring nodes, thereby fusing the information of neighboring nodes into the current node A. i attention characteristics

[0024]

[0025] S107, Obtaining attention features Subsequently, a residual connection mechanism was introduced to learn residual characteristics. To adjust the original feature e i and attention characteristics The ratio allows the attention mechanism to adaptively adjust the degree of feature fusion in different scenarios. Ultimately, node A is obtained. i New features

[0026]

[0027] g i It is a gating coefficient obtained through neural network learning, used to control the fusion ratio of residual information and attention features.

[0028] Preferably, the temporal attention module in S2 is based on scaled dot product attention, combined with a lower triangular masking mechanism to ensure causality, thereby extracting temporal dependency information from historical trajectories. Its calculation steps include:

[0029] S201. The trajectory features of all time steps constitute the trajectory feature matrix:

[0030]

[0031] S202, Define a mask for each time step

[0032]

[0033] The mask takes a value of 1 if and only if the historical positions at times t and t-1 are both valid. To ensure causality, the model introduces a lower triangular mask M∈R when calculating the attention score. T×T Its definition is as follows:

[0034]

[0035] S203. After obtaining the attention output, operations such as feedforward network and layer normalization are added to further enhance feature representation and stabilize the training process. The calculation process is as follows:

[0036] f = LN(Q + Attention(Q,K,V))

[0037] I(K,Q,V)=LN(f+FFN(f))

[0038] FFN(·) typically consists of two fully connected layers and an activation function, which effectively integrates input features through residual connections and layer normalization.

[0039] S204, Finally, the time characteristics are obtained:

[0040]

[0041] Preferably, the hierarchical spatiotemporal feature fusion module in S3 adopts a multi-level time interval sampling and cross-attention fusion method to gradually aggregate short-term maneuvering and long-term motion intentions, thereby generating a more discriminative trajectory query vector. The calculation process is as follows:

[0042] S301, initialize H different hierarchical queries in the HA module, where the h-th hierarchical query is used to aggregate at a time interval of 2. h-1 The spatiotemporal characteristics of can be expressed by the following formula:

[0043]

[0044] Among them W h Let h be the learnable weight matrix. An indicator function is used to filter time steps that satisfy the sampling interval condition, when time step t satisfies t mod 2. h-1 The value is 1 when the value is 0, otherwise it is 0.

[0045] S302. After concatenating all H hierarchical queries, input them into a multilayer perceptron for unified mapping to obtain a fused single trajectory query vector:

[0046] Q i =MLP(cat(q)i,1 ,q i,2 ,…,q i,H ))

[0047] S303. In the cross-attention calculation, the trajectory query Q... i As a query vector, the fused spatiotemporal features As keys and values, the trajectory query is progressively fused from key information from different times and spatial locations through an attention mechanism. After this cross-attention update, each agent vehicle ultimately obtains K fused trajectory queries {Q}. i,1 Q i,2 ,…,Q i,K}, where each vector will serve as the initial query for the decoder, corresponding to a future predicted trajectory.

[0048] Preferably, in S4, the segmented iterative decoder generates the trajectory segment by segment through a fixed time window, performs self-attention optimization in each iteration, and generates the absolute trajectory coordinates by accumulating the displacement increment. The calculation process is as follows:

[0049] S401. To optimize the multimodal trajectory distribution, self-attention calculation is performed on different trajectory queries for the same target vehicle in each iteration:

[0050]

[0051] S402. After completing the cross-modal optimization of trajectory query, a simple multilayer perceptron is used to predict future trajectories:

[0052]

[0053] in, This represents the predicted trajectory displacement increment (Δx, Δy) and its corresponding Gaussian distribution variance (σx, σy);

[0054] S403. To obtain the final absolute trajectory point, the increment is accumulated. This calculation process is performed in L. T This process is executed in each iteration. Each iteration predicts a trajectory segment within a fixed time window and concatenates it with previous predictions. Finally, after L... T After the first iteration, for the target vehicle i, output K future trajectories and their confidence levels:

[0055]

[0056] Where, N L N represents the total prediction time step. A This represents the total number of vehicles handled by the agent.

[0057] Preferably, the trajectory prediction model uses negative log-likelihood loss and classification loss to optimize the trajectory prediction accuracy and multimodal trajectory selection ability.

[0058] A map-free vehicle trajectory prediction method based on trajectory query distillation is proposed, which constructs a teacher-student network knowledge distillation architecture, specifically including:

[0059] S5. The teacher network includes a map encoder and a map-agent interaction fusion module, which are used to extract the topological and semantic features of high-precision maps and fuse them with vehicle spatiotemporal features.

[0060] S6. The student network adopts the end-to-end dynamic graph model architecture described in claim 1. During the training phase, it learns the map prior knowledge of the teacher network through distillation, and no map input is required during the testing phase.

[0061] Preferably, the implementation steps of the map encoder and interactive fusion module are as follows:

[0062] S501, the map encoder extracts features such as lane centerline coordinates, curvature, and semantic type from the high-precision map through scaling dot product attention, and outputs a high-dimensional map feature matrix f. m ;

[0063] S502, Map-Agent Interaction Fusion Module uses agent spatiotemporal features f st For querying, map feature f m Using keys and values, the relationship between vehicles and the map is modeled through a multi-head attention mechanism:

[0064] f f =HA(MHA(f) st ,f m ,f m ))

[0065] Among them, MHA is a multi-head attention operation, and HA is a hierarchical aggregation module used to fuse multi-scale spatiotemporal and map features.

[0066] The preferred implementation of the two-stage trajectory query distillation strategy is as follows:

[0067] S601, Encoder Stage Distillation: Aligning Encoder Output Trajectory Query for Teacher and Student Networks and The loss function is:

[0068]

[0069] S602, Decoder Stage Distillation: Aligning Trajectory Queries for Each Iteration Step and The keyframe iteration steps are covered, and the loss function is:

[0070]

[0071] Preferably, the adaptive distillation mechanism dynamically adjusts the distillation intensity through exponential decay, and the specific calculation steps are as follows:

[0072] Set the encoder distillation coefficient to α(t) = α0·e -γt The initial weight α0 = 1.0, and the attenuation rate γ = 0.01 are set; the decoder distillation coefficient is set as: β(t) = β0·e -γt Where the initial weight β0 = 0.5, and t is the number of training iterations; calculate the total distillation loss L. tqd =α(t)L enc +β(t)L dec The dynamic weighting balances teacher guidance with students' self-directed learning abilities.

[0073] Compared with the prior art, the beneficial effects of the present invention are:

[0074] 1. To address the challenges of acquiring, updating, and maintaining high-precision maps, this paper proposes an end-to-end trajectory prediction model that does not utilize high-precision map information. The algorithm includes the following aspects: For spatial feature extraction, a graph attention network based on a scaling dot product mechanism is introduced to aggregate features of neighboring vehicle states. A gated residual connection mechanism is also introduced to adaptively fuse the original features with the graph attention output, enhancing the model's feature robustness in dynamic interaction scenarios. For temporal feature extraction, a temporal attention network with a masking mechanism is constructed to model causal dependencies in trajectory sequences. To enhance the fusion and representation of spatiotemporal information, a hierarchical spatiotemporal feature fusion module is designed. Multi-level time interval sampling and a cross-attention mechanism are used to achieve a unified representation of spatiotemporal features at different scales. Finally, a segmented iterative decoder module is proposed, decomposing the overall trajectory prediction task into multiple local stages and gradually stitching them together to generate a complete future trajectory, thereby alleviating the error accumulation problem in long-term prediction.

[0075] 2. Based on the aforementioned end-to-end trajectory prediction model, a trajectory prediction method based on trajectory query distillation is proposed to achieve cross-model knowledge transfer. Under the teacher-student network architecture, a two-stage trajectory query distillation strategy is designed, which distills the trajectory query representation of the student network in the encoder and decoder stages, respectively. This allows the student network to approximately reconstruct the prior knowledge learned by the teacher network based on the high-precision map even in map-free scenarios during the testing phase. To further improve the distillation effect and model robustness, an adaptive distillation mechanism is introduced to dynamically adjust the distillation intensity and reduce the excessive dependence on the teacher network in the later stages of training. This strategy can improve the accuracy of student network trajectory prediction while maintaining the model's lightweight nature. Attached Figure Description

[0076] Figure 1This is an overall architecture diagram of a vehicle trajectory prediction method for scenarios without high-precision map coverage according to the present invention.

[0077] Figure 2 This is a diagram of the map encoder module architecture of the present invention;

[0078] Figure 3 This is an architecture diagram of the map-agent interaction fusion module of the present invention;

[0079] Figure 4 This is a schematic diagram of the distillation location for trajectory querying according to the present invention. Detailed Implementation

[0080] The technical solutions of 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.

[0081] Please see Figure 1-4 This invention provides a technical solution: a vehicle trajectory prediction method for scenarios without high-precision map coverage, comprising the following steps:

[0082] S1: High-precision map input map encoder, extracting topological and semantic features.

[0083] Figure 2 The diagram illustrates the specific architecture of the map encoder. The input to the map encoder consists of different elements of a high-precision map. In the Argoverse motion prediction dataset, the high-precision map mainly consists of lane centerlines and their geometric and semantic attributes. Each trajectory scene corresponds to a local map containing multiple adjacent lane segments. The feature matrix of the input high-precision map is defined as follows:

[0084] M = {m1, m2, ..., m} N},m i =(x i ,y i ,s i )

[0085] Where each m i This represents a lane segment whose feature vector consists of geometric and semantic information. (x i ,y i ) represents the coordinates of key points on the lane centerline, while semantic tags s iRepresenting lane types (such as main roads, ramps, or intersections), these points provide structural information, aiding the model in understanding the scene context. To unify the spatial reference frame and reduce coordinate scale differences, all map points are transformed to a relative coordinate system centered on the proxy vehicle. This process uses vehicle position and heading angle to rotate and translate map features, allowing the model to focus on local geometric relationships rather than global offsets. All input features are normalized to the [0,1] interval, with zeros padded if the number of centerlines is insufficient, ultimately forming a structurally consistent input tensor for subsequent feature encoding and trajectory query generation.

[0086] S2: Proxy historical trajectory input feature extraction module and hierarchical spatiotemporal feature fusion module.

[0087] S201. Constructing a dynamic graph structure:

[0088] Centered on the target vehicle, adjacent vehicle nodes are selected based on a Euclidean distance threshold of 100 meters, and directed edges are constructed:

[0089] τ i,j =(A j →A i )

[0090] Where A i For the target vehicle, A j For vehicles in the neighboring area;

[0091] Global coordinates of neighboring vehicles Convert to relative coordinates centered on the target vehicle:

[0092]

[0093] in, Used to characterize the spatial interaction relationships between vehicles;

[0094] Calculate each node A i The feature vector contains the position increment, instantaneous velocity, and heading angle:

[0095]

[0096] in The position increment of adjacent frames, Instantaneous velocity For heading angle;

[0097] Finally, the node features are normalized, mapping features such as coordinates and velocity to the [0,1] interval to eliminate the influence of dimensional differences on the model.

[0098] S202, Spatial feature extraction module processes historical trajectory information:

[0099] This model employs scaled dot product attention, fusing the attention results from the lower triangular mask and the relative temporal position embedding, as well as the nonlinear mapping of the feedforward network. The computation steps are as follows:

[0100] Calculate for node A j To A i Attention weight a j :

[0101]

[0102] Where D is the number of channels, Q i K j This represents the channel-wise dot product operation for Query and Key.

[0103] The normalized attention weights are used to calculate the weighted sum of the features of neighboring nodes, thereby fusing the information of neighboring nodes into the current node A. i attention characteristics

[0104]

[0105] In obtaining attention features Subsequently, a residual connection mechanism was introduced to learn residual characteristics. To adjust the original feature e i and attention characteristics The ratio allows the attention mechanism to adaptively adjust the degree of feature fusion in different scenarios. Ultimately, node A is obtained. i New features

[0106]

[0107] g i It is a gating coefficient obtained through neural network learning, used to control the fusion ratio of residual information and attention features.

[0108] S203, The time extraction module processes historical trajectory information:

[0109] Based on scaled dot product attention and combined with a lower triangular masking mechanism to ensure causality, temporal dependency information in historical trajectories is extracted, and the trajectory features of all time steps constitute a trajectory feature matrix:

[0110]

[0111] Define a mask for each time step

[0112]

[0113] The mask takes a value of 1 if and only if the historical positions at times t and t-1 are both valid. To ensure causality, the model introduces a lower triangular mask M∈R when calculating the attention score. T×T Its definition is as follows:

[0114]

[0115] After obtaining the attention output, operations such as feedforward network and layer normalization are added to further enhance feature representation and stabilize the training process. The calculation process is as follows:

[0116] f = LN(Q + Attention(Q,K,V))

[0117] I(K,Q,V)=LN(f+FFN(f))

[0118] The FFN(·) function typically consists of two fully connected layers and an activation function. It effectively integrates the input features through residual connections and layer normalization, ultimately yielding the temporal features.

[0119]

[0120] S204, the hierarchical spatiotemporal feature fusion module adopts a multi-level time interval sampling and cross-attention fusion method to gradually aggregate short-term maneuvering and long-term motion intentions, thereby generating a more discriminative trajectory query vector. The calculation process is as follows:

[0121] In the HA module, initialize H distinct hierarchical queries, where the h-th hierarchical query is used to aggregate at intervals of 2. h-1 The spatiotemporal characteristics of can be expressed by the following formula:

[0122]

[0123] Among them W h Let h be the learnable weight matrix. An indicator function is used to filter time steps that satisfy the sampling interval condition, when time step t satisfies t mod 2. h-1 The value is 1 when the value is 0, otherwise it is 0.

[0124] After concatenating all H hierarchical queries, input them into a multilayer perceptron for unified mapping to obtain a fused single trajectory query vector:

[0125] Q i =MLP(cat(q) i,1 ,q i,2 ,…,q i,H ))

[0126] In cross-attention calculation, the trajectory query Q is...i As a query vector, the fused spatiotemporal features As keys and values, an attention mechanism is used to progressively fuse key information from different times and spatial locations in the trajectory query. After this cross-attention update, each agent vehicle ultimately obtains K fused trajectory queries {Q}. i,1 Q i,2 ,…,Q i,K}, where each vector will serve as the initial query for the decoder, corresponding to a future predicted trajectory.

[0127] S3: Map-Agent Interaction Fusion Module, which aligns and interacts three types of features: temporal features, spatial features, and map features, to fully model the relationship between the vehicle and the environment. Its specific calculation steps are as follows:

[0128] Figure 3 The diagram illustrates the specific architecture of the map-agent interaction fusion module. First, a linear transformation is applied to the map features, projecting them into the same representation space as the agent's spatiotemporal features to reduce feature distribution differences. Let the query originate from the agent's spatiotemporal features f. st The keys and values ​​come from map features f m The fusion process can then be represented as:

[0129] f f =HA(MHA(f) st ,f m ,f m ))

[0130] In this context, MHA(·) represents multi-head attention operation, used to model the interaction between map information and agent features, while HA(·) is hierarchical aggregation, integrating the fused features. The calculation method for each attention head in the MHA mechanism is as follows:

[0131]

[0132] The outputs of all attention heads are concatenated and projected onto the final fused feature representation:

[0133] MHA(f st ,f m ,f m = Concat(head1,…,head) h W o

[0134] S4: Knowledge distillation of trajectories. In teacher-student network trajectory prediction architectures implemented based on various attention mechanisms (Q, K, V), the trajectory query representation Q is a key variable for the encoder and decoder in spatiotemporal modeling and prediction. Figure 4The following two stages are used to align and distill the trajectory query Q of the teacher-student network, illustrating the distillation process for trajectory queries.

[0135] S401, Encoder Stage Distillation:

[0136] Let the trajectory query vector generated by the teacher network at the encoder end be denoted as... The corresponding output of the student network is denoted as The former, by incorporating map-agent interaction fusion information, offers a more comprehensive representation in terms of global structure and spatiotemporal dependencies, while the latter can only rely on historical trajectories for interaction encoding. Distilling the two allows the student network to obtain scene understanding and global interaction representations similar to those of the teacher network at the encoder end.

[0137] S402, Encoder output trajectory query for aligned teacher and student networks and The loss function is set as follows:

[0138]

[0139] S403, Decoder stage distillation:

[0140] During the decoding phase, the teacher network uses This represents the trajectory query vector for the l-th decoding step; the corresponding vector in the student network is... Let represent the trajectory query vector of the l-th decoding step. The teacher network's decoding process combines the prediction of the previous moment with the map prior to generate a more reasonable motion state for the next moment. If the student network does not have sufficient scenario constraints, it is easy to accumulate deviations in multiple iterations. By gradually aligning the query vectors for different iteration steps, the student network can gradually approach the inference method of the teacher network in the dynamic prediction process, thereby improving its adaptability to complex road conditions.

[0141] S404, Aligning the trajectory query of each iteration step and Covering the keyframe iteration steps, the loss function is set as follows:

[0142]

[0143] S5: Adaptive distillation setting, setting the encoder distillation coefficient to α(t) = α0·e -γt The initial weight α0 = 1.0, and the attenuation rate γ = 0.01 are set; the decoder distillation coefficient is set as: β(t) = β0·e -γt Where the initial weight β0 = 0.5, and t is the number of training iterations; calculate the total distillation loss L. tqd =α(t)L enc +β(t)L decThe dynamic weighting balances teacher guidance with students' self-directed learning abilities.

[0144] Furthermore, Figure 1 This invention presents a framework based on an end-to-end trajectory prediction model that does not use high-precision map information. The teacher network is extended to include a map encoder and a map-agent interaction fusion module to effectively integrate map information with the trajectory prediction task. Specifically, the map encoder extracts structured road information from the high-precision map and converts it into high-dimensional features suitable for prediction. The extracted map features, along with the spatial and temporal features generated by the end-to-end trajectory prediction model, are input in parallel into the map-agent interaction fusion module. This module performs feature alignment to model the correlation between map information and the dynamic features of the agent vehicle. The fusion module employs a multi-head attention mechanism to calculate the impact of map features on agent features and combines a hierarchical aggregation module to enhance the spatiotemporal feature representation. The fused features ultimately enter a segmented iterative decoder. The teacher and student networks in this part maintain the same structure to ensure the effectiveness of the subsequent knowledge distillation process. The student network fully adopts the end-to-end trajectory prediction model structure and learns the trajectory query representation of the teacher network through distillation, compensating for the information gaps caused by the lack of high-precision maps.

[0145] Example 1: Evaluation based on the Argoverse motion prediction dataset. The model performance evaluation uses three standard metrics widely used in the field of trajectory prediction tasks: minimum average displacement error (minADE), minimum final displacement error (minFDE), and prediction failure rate (Miss Rate, MR), to comprehensively evaluate the performance of the proposed method in terms of trajectory prediction accuracy and robustness.

[0146] To further verify the performance improvement of the trajectory prediction model based on trajectory query distillation proposed in this invention under the condition of no high-precision map, this paper compares it with several mainstream methods on the Argoverse dataset, including LaneGCN-MF+FOKD, HiVT-64-MF+FOKD, MFTP, and HiVT-128-MF+FOKD. These methods have their own characteristics in terms of spatiotemporal feature encoding, vehicle interaction relationship modeling, and dependence on high-precision maps. To highlight the relationship between these models and high-precision map information and the effectiveness of the knowledge transfer mechanism, the following methods are briefly described:

[0147] HiVT: This model is a powerful baseline method that employs a hierarchical Transformer architecture with local and global encoders. HiVT achieves state-of-the-art results and is the best open-source method on the Argoverse.

[0148] LaneGCN: This model combines graph networks and 1D convolutions to extract vehicle interaction features while incorporating lane structure information. It performs well on Argoverse and has been widely used as a powerful baseline method.

[0149] FOKD: This model proposes a feature optimization knowledge distillation framework for trajectory prediction. By constructing a feature interaction mechanism between teacher and student networks, the road structure information learned by the teacher network is effectively transferred to the student network, thereby improving the trajectory prediction performance in scenarios without high-precision maps. In this experiment, the FOKD framework is combined with some baseline models to compare and analyze its effect on improving the accuracy of map-free prediction.

[0150] MFTP: This model designs a multi-level feature propagation mechanism based on the Transformer architecture and combines it with a distillation strategy to transmit fused information of spatiotemporal features, effectively improving the accuracy of trajectory prediction without high-precision map input.

[0151] TQDS-TS: This model is the complete teacher-student network model proposed in this invention. In the pre-training stage, the teacher network is based on high-precision map input, but in the subsequent training and testing stages, the student network relies entirely on historical trajectory data without map input. By using the high-precision map prior of the teacher network in the pre-training stage, the student network learns the latent representation of map information and completes prediction in a map-free environment through a distillation mechanism.

[0152] TQDS-T: This model is a variant of TQDS-TS and is a teacher network that relies on high-precision map input. In this model, the network explicitly utilizes high-precision map information and can fully utilize map prior knowledge such as lane geometry and topology during prediction. Therefore, TQDS-T is trained and tested using high-precision maps, with the goal of demonstrating the upper limit of the model's performance when map information is fully available.

[0153] Table I shows the performance of each model on the Argoverse test set.

[0154]

[0155] Specifically, in map-free trajectory prediction methods, compared with LaneGCN-MF+FOKD and HiVT-MF+FOKD, which employ traditional feature distillation strategies, TQDS-TS reduces the performance of minADE, minFDE, and MR by 15.63%, 19.02%, and 31.82%, and 11.96%, 16.98%, and 28.57%, respectively. This fully demonstrates the advantages of the proposed two-stage distillation mechanism in capturing dynamic trajectory evolution and structural priors. Compared with the MFTP model, which performs well in current map-free methods, TQDS-TS reduces the performance of the three metrics by 3.57%, 4.35%, and 6.25%, respectively, further validating the effectiveness of the trajectory query distillation strategy in terms of structural guidance and modeling ability. Under the premise of the same distillation structure, the comparison results with MGAT-MF also show that TQDS-TS can effectively alleviate the performance degradation problem caused by map absence and has significant advantages in improving model generalization ability and prediction accuracy.

[0156] Overall, TQDS-TS has demonstrated modeling capabilities that are close to or even surpass those of map-based methods without relying on high-precision map information, further validating the effectiveness of its structural design and distillation strategy synergistic optimization. As a map-free model, TQDS-TS possesses strong practicality and trajectory prediction accuracy.

[0157] 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 for scenarios without high-precision map coverage, characterized in that, Includes the following steps: S1. Constructing a dynamic graph structure and extracting spatial features: Construct a dynamic graph based on relative coordinates, with the target vehicle and neighboring vehicles as nodes. Filter adjacent nodes through Euclidean distance threshold, and use an improved graph attention network to fuse vehicle interaction features. The graph attention network uses a scaled dot product attention mechanism to calculate node weights, and adaptively fuses the original features and attention output through gated residual connections. S2. Temporal Feature Extraction and Temporal Dependency Modeling: A temporal attention module with a masking mechanism is adopted. Causality is ensured through a lower triangular mask, and the temporal modeling capability is enhanced by relative time position embedding to extract the temporal dependency features of historical trajectories. S3. Hierarchical spatiotemporal feature fusion: Through multi-level time interval sampling and cross-attention mechanism, a hierarchical spatiotemporal feature representation is constructed to generate a multimodal trajectory query vector. The multi-level time interval sampling realizes feature aggregation at different time scales through a hierarchical aggregation module. S4. Segmented Iterative Decoding and Trajectory Generation: A segmented iterative decoder with a fixed time window is used to gradually generate and stitch together future trajectories. Self-attention is used to optimize cross-modal trajectory information and alleviate the accumulation of errors in long-term prediction.

2. The vehicle trajectory prediction method for scenarios without high-precision map coverage according to claim 1, characterized in that, The dynamic graph structure and spatial feature extraction process in S1 includes: S101. Using the target vehicle as the center, filter adjacent vehicle nodes based on the Euclidean distance threshold of 100 meters, and construct a directed edge τ. i,j =(A j →A i ), where A i For the target vehicle, A j For vehicles in the neighboring area; S102, Obtain the global coordinates of neighboring vehicles. Convert to relative coordinates centered on the target vehicle: in, Used to characterize the spatial interaction relationships between vehicles; S103, Calculate A for each node. i The feature vector contains the position increment, instantaneous velocity, and heading angle: in The position increment of adjacent frames, Instantaneous velocity For heading angle; S104. Normalize the node features and map features such as coordinates and velocity to the [0,1] interval to eliminate the influence of dimensional differences on the model. In addition, scaled dot product attention is used, which integrates the attention results of the lower triangular mask and the relative temporal position embedding, as well as the nonlinear mapping of the feedforward network. The calculation steps are as follows: S105, Calculate the value for node A j To A i Attention weight a j : Where D is the number of channels, Q i K j This represents the channel-wise dot product operation between the Query and the Key; S106. The normalized attention weights are used to calculate the weighted sum of the features of neighboring nodes, thereby fusing the information of neighboring nodes into the current node A. i attention characteristics S107, Obtaining attention features Subsequently, a residual connection mechanism was introduced to learn residual characteristics. To adjust the original feature e i and attention characteristics The ratio allows the attention mechanism to adaptively adjust the degree of feature fusion in different scenarios. Ultimately, node A is obtained. i New features g i It is a gating coefficient obtained through neural network learning, used to control the fusion ratio of residual information and attention features.

3. The vehicle trajectory prediction method for scenarios without high-precision map coverage according to claim 1, characterized in that, The temporal attention module in S2 is based on scaled dot product attention and combines a lower triangular masking mechanism to ensure causality, thereby extracting temporal dependency information from historical trajectories. Its calculation steps include: S201. The trajectory features of all time steps constitute the trajectory feature matrix: S202, Define a mask for each time step The mask takes a value of 1 if and only if the historical positions at times t and t-1 are both valid. To ensure causality, the model introduces a lower triangular mask M∈R when calculating the attention score. T×T Its definition is as follows: S203. After obtaining the attention output, operations such as feedforward network and layer normalization are added to further enhance feature representation and stabilize the training process. The calculation process is as follows: f = LN(Q + Attention(Q,K,V)) I(K,Q,V)=LN(f+FFN(f)) FFN(·) typically consists of two fully connected layers and an activation function, which effectively integrates input features through residual connections and layer normalization. S204, Finally, the time characteristics are obtained:

4. The vehicle trajectory prediction method for scenarios without high-precision map coverage according to claim 1, characterized in that, The S3 hierarchical spatiotemporal feature fusion module employs multi-level time interval sampling and cross-attention fusion to gradually aggregate short-term maneuvering and long-term motion intentions, thereby generating a more discriminative trajectory query vector. The calculation process is as follows: S301, initialize H different hierarchical queries in the HA module, where the h-th hierarchical query is used to aggregate at a time interval of 2. h-1 The spatiotemporal characteristics of can be expressed by the following formula: Among them W h Let h be the learnable weight matrix. An indicator function is used to filter time steps that satisfy the sampling interval condition, when time step t satisfies t mod 2. h-1 The value is 1 when the value is 0, otherwise it is 0. S302. After concatenating all H hierarchical queries, input them into a multilayer perceptron for unified mapping to obtain a fused single trajectory query vector: Q i =MLP(cat(q i,1 ,q i,2 ,…,q i,H )) S303. In the cross-attention calculation, the trajectory query Q... i As a query vector, the fused spatiotemporal features As keys and values, the trajectory query is progressively fused from key information from different times and spatial locations through an attention mechanism. After this cross-attention update, each agent vehicle ultimately obtains K fused trajectory queries {Q}. i,1 Q i,2 ,…,Q i,K }, where each vector will serve as the initial query for the decoder, corresponding to a future predicted trajectory.

5. The vehicle trajectory prediction method for scenarios without high-precision map coverage according to claim 1, characterized in that, In S4, the segmented iterative decoder generates the trajectory segment by segment through a fixed time window. Each iteration performs self-attention optimization and generates absolute trajectory coordinates by accumulating displacement increments. The calculation process is as follows: S401. To optimize the multimodal trajectory distribution, self-attention calculation is performed on different trajectory queries for the same target vehicle in each iteration: S402. After completing the cross-modal optimization of trajectory query, a simple multilayer perceptron is used to predict future trajectories: in, This represents the predicted trajectory displacement increment (Δx, Δy) and its corresponding Gaussian distribution variance (σx, σy); S403. To obtain the final absolute trajectory point, the increment is accumulated. This calculation process is performed in L. T This process is executed in each iteration. Each iteration predicts a trajectory segment within a fixed time window and concatenates it with previous predictions. Finally, after L... T After the first iteration, for the target vehicle i, output K future trajectories and their confidence levels: Where, N L N represents the total prediction time step. A This represents the total number of vehicles handled by the agent.

6. The vehicle trajectory prediction method for scenarios without high-precision map coverage according to claim 1, characterized in that, The trajectory prediction model uses negative log-likelihood loss and classification loss to optimize the accuracy of the predicted trajectory and the ability to select multimodal trajectories.

7. A map-free vehicle trajectory prediction method based on trajectory query distillation, characterized in that: Constructing a teacher-student network knowledge distillation architecture, specifically including: S5. The teacher network includes a map encoder and a map-agent interaction fusion module, which are used to extract the topological and semantic features of high-precision maps and fuse them with vehicle spatiotemporal features. S6. The student network adopts the end-to-end dynamic graph model architecture described in claim 1. During the training phase, it learns the map prior knowledge of the teacher network through distillation, and no map input is required during the testing phase.

8. The map-free vehicle trajectory prediction method based on trajectory query distillation according to claim 7, characterized in that, The implementation steps of the map encoder and interactive fusion module are as follows: S501, the map encoder extracts features such as lane centerline coordinates, curvature, and semantic type from the high-precision map through scaling dot product attention, and outputs a high-dimensional map feature matrix f. m ; S502, Map-Agent Interaction Fusion Module uses agent spatiotemporal features f st For querying, map feature f m Using keys and values, the relationship between vehicles and the map is modeled through a multi-head attention mechanism: f f =HA(MHA(f st ,f m ,f m )) Among them, MHA is a multi-head attention operation, and HA is a hierarchical aggregation module used to fuse multi-scale spatiotemporal and map features.

9. A map-free vehicle trajectory prediction method based on trajectory query distillation according to claim 7, characterized in that, The implementation method of the two-stage trajectory query distillation strategy is as follows: S601, Encoder Stage Distillation: Aligning Encoder Output Trajectory Query for Teacher and Student Networks and The loss function is: S602, Decoder Stage Distillation: Aligning Trajectory Queries for Each Iteration Step and The keyframe iteration steps are covered, and the loss function is:

10. A map-free vehicle trajectory prediction method based on trajectory query distillation according to claim 7, characterized in that, The adaptive distillation mechanism dynamically adjusts the distillation intensity through exponential decay. The specific calculation steps are as follows: Set the encoder distillation coefficient to α(t) = α0·e -γt The initial weight α0 = 1.0, and the attenuation rate γ = 0.01 are set; the decoder distillation coefficient is set as: β(t) = β0·e -γt Where the initial weight β0 = 0.5, and t is the number of training iterations; calculate the total distillation loss L. tqd =α(t)L enc +β(t)L dec The dynamic weighting balances teacher guidance with students' self-directed learning abilities.

Images