A method and device for predicting autonomous driving trajectory based on scene-guided gating

Abstract

The present invention discloses a method and device for autonomous driving trajectory prediction based on scene-guided gating. In the encoder module, a scene-guided gating unit is used to filter the agent code, and the filtered agent code is used to update the lane code. In the decoder module, the position of the target vehicle, the latent variables following the multivariate Gaussian distribution, and the updated target code are passed through a linear layer and then filtered through a scene-guided gating unit. The filtered features are passed through a linear layer to obtain a predicted trajectory. The present invention introduces a scene-guided gating unit to utilize the impact of scene heterogeneity on driving behavior, thereby facilitating the realization of more socially aware trajectory prediction. The present invention performs a diffusion convolution operation on the updated lane code and lane attention map, and uses a bidirectional random walk on the lane attention map to simulate the randomness of social interaction. Compared with the existing technical solutions, the present invention has better effectiveness and robustness.

Description

Technical Field

[0001] The present application belongs to the technical field of autonomous driving trajectory prediction, and in particular relates to an autonomous driving trajectory prediction method and device based on scene-guided gating. Background Art

[0002] Accurate trajectory prediction is crucial for improving the safety of autonomous driving. It predicts the future trajectory of a target vehicle by analyzing the historical trajectories of surrounding agents and complex lane maps. Accurate and reliable trajectory prediction supports downstream decision-making and planning tasks in autonomous driving to avoid potential hazards. However, the uncertain intentions of agents and the highly stochastic social interactions between them pose significant challenges to this task.

[0003] To address these challenges, some studies have proposed deep learning-based models for trajectory prediction. These models are typically based on an encoder-decoder structure, where the encoder converts historical trajectories and HD map information into vector representations, while the decoder generates multiple credible future trajectories. However, most existing studies have overlooked the impact of scene heterogeneity on driving behavior. Scene heterogeneity can affect driving behavior. In relatively simple scenarios, drivers may adopt more aggressive driving strategies, such as changing lanes and overtaking. Conversely, in relatively complex scenarios, such as when there are many vehicles or pedestrians around, drivers tend to adopt more cautious driving strategies. Existing models often fail to fully exploit the impact of scene heterogeneity on driving behavior, making it difficult for the models to achieve more socially aware trajectory predictions, which in turn affects the models' prediction performance in diverse scenarios.

[0004] Furthermore, the highly stochastic social interactions between agents have not been fully explored, reducing the reliability of predictions. The future trajectories of agents are largely influenced by their social interactions, which are inherently dynamic and uncertain and propagate randomly between agents. For example, social interactions between agents may change due to changes in the scene or the direction of the agents. Even in nearly identical scenes, these social interactions may vary due to different driving habits and intentions of the drivers. However, existing models primarily learn fixed or single-hop social interactions, which makes it difficult to fully describe the highly stochastic social interactions between agents. Summary of the Invention

[0005] The purpose of this application is to provide a method and device for autonomous driving trajectory prediction based on scene-guided gating to overcome the impact of scene heterogeneity on driving behavior and the impact of highly random social interactions between intelligent agents, thereby achieving more accurate trajectory prediction.

[0006] In order to achieve the above objectives, the technical solutions of this application are as follows:

[0007] A method for autonomous driving trajectory prediction based on scene-guided gating, comprising:

[0008] Use a gated recurrent unit to encode the target vehicle's past trajectory, lane map, and surrounding agent's past trajectory to obtain target encoding, lane encoding, and agent encoding;

[0009] The agent code is filtered using the scene-guided gating unit to obtain a filtered agent code, and the filtered agent code is used to update the lane code;

[0010] The lane map adjacency matrix and the updated lane encoding are used to generate a lane attention map through the self-attention mechanism;

[0011] Perform a dilated convolution operation on the updated lane code and lane attention map, and then add them to the updated lane code after random dropout operation to obtain lane node features;

[0012] Input the target encoding, lane node features, and the adjacency matrix of the lane graph into the strategy head to obtain the lane node set;

[0013] Get the lane node features corresponding to the lane node set, form a path code, and use the path code to update the target code;

[0014] The position of the target vehicle, the latent variables following the multivariate Gaussian distribution, and the updated target encoding are passed through a linear layer and then filtered by a scene-guided gating unit. The filtered features are passed through a linear layer to obtain the predicted trajectory.

[0015] Furthermore, the scene-guided gating unit performs the following operations:

[0016] Build the embedding vectors corresponding to the surrounding agents through the embedding layer;

[0017] Concatenate the embedding vectors corresponding to the surrounding agents and the position of the target vehicle to obtain heterogeneous scene information;

[0018] Input the heterogeneous scene information into the meta-learner to learn the weight parameters of the meta-linear layer;

[0019] The original code to be filtered is input into the meta-linear layer, and then after the activation operation, the Hadamard product is performed with the original code, and after the Hadamard product, the code is added to the original code to obtain the filtered code.

[0020] Furthermore, the updating of the lane code by using the filtered agent code includes:

[0021] Linearly map the filtered agent encoding to obtain the key vector and value vector;

[0022] Perform linear mapping on the lane encoding to obtain the query vector;

[0023] Perform scaled dot product attention operation to obtain lane encoding with agent context;

[0024] The lane coding with the agent context is then concatenated with the lane coding and linearly transformed to obtain the updated lane coding.

[0025] Furthermore, the lane attention map is generated by applying the adjacency matrix of the lane map and the updated lane encoding to a self-attention mechanism, including:

[0026] Perform linear mapping on the updated lane coding to obtain the query vector, key vector, and value vector;

[0027] Perform self-attention operation and then perform Hadamard product with the adjacency matrix of the lane map to obtain the lane attention map.

[0028] Furthermore, the updating of the target code using the path code includes:

[0029] Linearly map the path encoding to obtain key vector and value vector;

[0030] Linearly map the target encoding to obtain the query vector;

[0031] Perform scaled dot product attention operation to obtain target encoding with path context;

[0032] The target encoding with the path context is then concatenated with the target encoding to obtain the updated target encoding.

[0033] The present application also proposes an autonomous driving trajectory prediction device based on scene-guided gating, comprising a processor and a memory storing a plurality of computer instructions. When the computer instructions are executed by the processor, the steps of the above-mentioned autonomous driving trajectory prediction method based on scene-guided gating are performed.

[0034] This application proposes a method and device for autonomous driving trajectory prediction based on scene-guided gating. It proposes a new scene-guided gating mechanism to utilize the impact of scene heterogeneity on driving behavior. This mechanism selectively filters information by generating scene-specific non-shared parameters, thereby establishing a scene-guided information bottleneck, which is conducive to achieving more socially aware trajectory prediction. Secondly, a dynamic diffusion graph convolution module is proposed to deeply explore the highly random social interactions between intelligent agents. This module regards social interactions as a diffusion process on the lane map and uses bidirectional random walks on the lane attention map to simulate the randomness of social interactions. Compared with existing technical solutions, the technical solution of this application has better effectiveness and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the structural diagram of the dynamic diffusion graph convolutional network based on scene-guided gating for this application.

[0036] Figure 2 This is a flow chart of the autonomous driving trajectory prediction method based on scene-guided gating in this application.

[0037] Figure 3 This is a diagram of the structure of the gate control unit for this application scenario.

[0038] Figure 4 This is the structural diagram of the dynamic diffusion graph convolution module for this application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0040] For the trajectory prediction task, given the past trajectories of the target vehicle and surrounding agents (vehicles or pedestrians) and the lane map representation of the scene, the goal of trajectory prediction is to learn a mapping function f that can predict multiple possible trajectories of the target vehicle and the associated confidence scores. Each predicted trajectory includes the target vehicle from time step 1 to t f An ordered pair of two-dimensional coordinates of .

[0041] In each scenario, the past trajectory of agent i (vehicle or pedestrian) can be expressed as past t p time-step trajectory vector An ordered sequence of . Each in are the two-dimensional coordinate pair, velocity, acceleration, and yaw rate of agent i at time step t; I i Represents the type of agent i: vehicle (I i =0) or pedestrians (I i =1). In addition, i=0 represents the target vehicle and t=0 represents the current time step.

[0042] The lane graph is represented as a directed graph G = (V, E), where V is a set of nodes representing lane centerlines, and E is a set of edges (including successor edges and adjacent edges). Each lane centerline v is divided into fixed-length segments represented by N points. The nth point is represented by a vector c. n =[x n ,y n ,θ n ,I n ] means, where (x n ,y n ) and θ nare the two-dimensional coordinate pair and the yaw rate respectively; I n is a two-dimensional binary vector indicating whether the point is on a stop line or a crosswalk.

[0043] This application proposes a new deep learning model, namely the scene-guided gated dynamic diffusion graph convolutional network (SGG-DDGCN). The proposed SGG-DDGCN model is designed to learn the uncertain future relationships between intelligent agents and selectively filter information. Figure 1 As shown in Figure 1, it includes an encoder module, a policy head, and a decoder module. The encoder module encodes the target vehicle's past trajectory, the lane map, and the past trajectories of surrounding agents, and outputs a learned representation of the lane map that contains the context of the surrounding agents. This module contains a scene-guided gating unit (SGG) for selectively filtering information and a dynamic diffusion graph convolution module (DDGCM) for describing the uncertain future relationships between agents. Subsequently, the policy head outputs a probability distribution for each edge of the lane map to support path sampling on the lane map. Finally, based on the target vehicle's position, randomly generated latent variables, and the path traversed by the policy head, the decoder module implements multimodal trajectory prediction for multiple future time steps.

[0044] In one embodiment, Figure 2 As shown, a method for autonomous driving trajectory prediction based on scene-guided gating is provided, including:

[0045] Step S1: Use a gated recurrent unit to encode the target vehicle's past trajectory, lane map, and surrounding agent's past trajectory to obtain a target code, a lane code, and an agent code.

[0046] like Figure 1 As shown, the encoder module includes a gated recurrent unit (GRU), a scene-guided gating unit (SGG), and a dynamic diffusion graph convolution module (DDGCM).

[0047] This embodiment first uses a gated recurrent unit (GRU) to encode past trajectories, lane maps, and past trajectories of surrounding agents to obtain target encoding, lane encoding, and agent encoding.

[0048] Vehicle drivers tend to pay more attention to scene information in the adjacent time steps rather than the scene information in the distant time steps. Based on this assumption, the encoder module of this embodiment directly uses the gated recurrent unit (GRU) to encode the past trajectory of the target vehicle, the lane map, and the past trajectory of the surrounding agents, and outputs the final hidden state. Specifically, it outputs the target encoding Lane Coding and agent encoding Where D is the encoding feature dimension.

[0049] It should be noted that, given that the number of agents surrounding the target vehicle may vary in different scenarios, the GRU configured for each agent in this embodiment is independent, but its parameters are shared between the agents. Similarly, the GRU used in each lane also follows the same principle.

[0050] Step S2: Use the scene-guided gating unit to perform scene filtering on the agent code to obtain the filtered agent code, and use the filtered agent code to update the lane code.

[0051] This step uses the scene-guided gating unit (SGG) to filter the agent encoding The key information in the filter is obtained to obtain the filtered agent code. Figure 3 As shown in Figure 2, the proposed SGG extracts meta-knowledge from scene information, generates a set of scenario-specific, non-shared parameters, and uses these parameters to guide the information filtering process to fully leverage the impact of scene heterogeneity on driving behavior. SGG aims to assign similar model parameters to similar scenes, thereby constructing a scenario-guided information bottleneck, which in turn facilitates more socially aware trajectory prediction.

[0052] The scene guided gating unit (SGG) performs the following operations:

[0053] Step 2.1: Construct the embedding vector corresponding to the surrounding intelligent agents through the embedding layer.

[0054] In this embodiment, the embedding layer maps the number of surrounding agents into a vector space to obtain an embedding vector. This embodiment can generate an embedding vector based on the total number of surrounding agents, or it can generate embedding vectors for vehicles and pedestrians based on the number of vehicles and pedestrians in the surrounding agents.

[0055] Preferably, considering the different logics of pedestrian and vehicle impacts on the target vehicle, for example, surrounding vehicles tend to have a direct impact on the target vehicle, while pedestrians on the sidewalk have a smaller impact, while pedestrians at zebra crossings or intersections have a direct impact on the target vehicle. Therefore, a technical solution is adopted to generate embedding vectors for vehicles and pedestrians respectively based on the number of vehicles and pedestrians in the surrounding agents.

[0056] Specifically, in the embedding layer, the vector space is learnable based on the parameters and The number of surrounding vehicles and pedestrians in each scene (divided by 5 to facilitate model learning) can be matched to the corresponding vector representation and Where d (set to D / 2) is the vector dimension of the vector space. and They represent the maximum number of vehicles and pedestrians around the target vehicle respectively.

[0057] Step 2.2: Concatenate the embedding vectors corresponding to the surrounding intelligent agents and the position of the target vehicle to obtain heterogeneous scene information.

[0058] In this embodiment, E vehicle 、E pedestrian and the target vehicle's position Perform stitching to obtain heterogeneous scene information:

[0059]

[0060] in Represents heterogeneous scene information.

[0061] Step 2.3: Input the heterogeneous scene information into the meta-learner to learn the weight parameters of the meta-linear layer.

[0062] In this embodiment, HS is input into the meta-learner to extract meta-knowledge and obtain the parameters of the meta-linear layer. and

[0063] The meta-learner is responsible for generating the parameters of the meta-linear layer and A common practice is to use a single linear layer to map HS to W meta and b meta However, this will result in a large number of parameters that need to be learned by the model. For example, (D+5)×(D·D) parameters are required to map HS to W meta To address this issue, the meta-learner in this embodiment adopts a two-stage design: first, HS is mapped to a smaller dimension sd (set to 2), and then mapped to the dimension D·D. This two-stage design only requires (D+5)×sd+sd×(D·D) parameters to map HS to the D·D dimension, as shown in Equations (2) and (3).

[0064] W meta =σ(HSW w1 +b w1 )W w2 +b w2 (2)

[0065] b meta =σ(HSW b1 +b b1 )W b2 +b b2 (3)

[0066] in and are all learnable parameters; σ(.) represents the Sigmoid activation function.

[0067] Step 2.4: Input the agent code into the meta-linear layer, and then after the activation operation, perform the Hadamard product with the agent code, and then add it to the agent code after the Hadamard product to obtain the filtered agent code.

[0068] SGG uses scene information as a guide to filter the surrounding agent codes. For example, the process can be expressed as equation (4):

[0069]

[0070] in is the agent code after SGG filtering; ⊙ represents the Hadamard product.

[0071] The filtered agent encodings are then used to update the lane encodings. Social interactions between agents largely follow lane information. Based on this insight, the encoder module uses the SGG-filtered agent encodings to update lane encodings, enabling the DDGCM of this embodiment to model uncertain future relationships between agents.

[0072] In a specific embodiment, the lane code is updated using the filtered agent code, including:

[0073] Linearly map the filtered agent encoding to obtain the key vector and value vector;

[0074] Perform linear mapping on the lane encoding to obtain the query vector;

[0075] Perform scaled dot product attention operation to obtain lane encoding with agent context;

[0076] The lane coding with the agent context is then concatenated with the lane coding and linearly transformed to obtain the updated lane coding.

[0077] Formally, this process can be expressed as:

[0078]

[0079]

[0080] in, and are the key vector and value vector obtained by linearly mapping the filtered agent encoding; is the query vector obtained by linearly mapping the lane encoding; Attention(.) is the scaled dot product attention operation (agent-lane attention); is the lane encoding with agent context; and is a learnable parameter; || is a concatenation operation; a linear transformation is performed on the concatenated encoding to obtain the updated lane encoding, is the updated lane encoding. It should be noted that the scaled dot product attention operation only considers agents within a certain distance threshold to the lane node.

[0081] The future trajectory of an agent is inherently uncertain and depends on the scene information as well as the agent's intentions. Typically, a driver may make different decisions in different situations. For example, when there are more agents around (whether vehicles or pedestrians), the driver tends to adopt a more cautious driving strategy. Although such scene heterogeneity affects driving behavior, this has not been fully utilized. Therefore, this embodiment proposes SGG to selectively filter information based on scene-specific non-shared parameters. SGG will assign similar model parameters to similar scenes, thereby establishing a scene-guided information bottleneck, which in turn promotes the model to achieve more socially aware trajectory prediction.

[0082] Step S3: The adjacency matrix of the lane map and the updated lane encoding are used to generate a lane attention map through the self-attention mechanism.

[0083] Capturing social interactions between agents is crucial for achieving accurate and reliable trajectory prediction, and graph neural networks (GNNs) are well suited to this task. Currently, most GNN-based methods use graph convolutional networks and graph attention networks to capture social interactions between agents. However, these methods usually only focus on fixed or single-hop social interactions, and it is difficult to effectively model the randomly propagated social relationships between agents. Therefore, this embodiment proposes a dynamic diffusion graph convolution module DDGCM, which aims to model highly random social interactions between agents. Specifically, this module models the social interactions between agents as a diffusion process on the lane map, and constructs a lane attention map to dynamically calculate the diffusion ratio.

[0084] like Figure 4 As shown, DDGCM uses the adjacency matrix A of the lane graph lane and updated lane coding First, it uses the self-attention mechanism to generate the lane attention map, as shown in formula (7).

[0085]

[0086] in, are respectively the query vector, key vector, and value vector obtained by linearly mapping all updated lane encodings in the current scene; is the lane attention map under the lane map adjacency matrix mask; N V is the number of lane nodes, which varies in different scenarios; ⊙ denotes the Hadamard product; Attention(.) here is the self-attention operation.

[0087] Step S4: Perform a diffusion convolution operation on the updated lane code and lane attention map, and then add them to the updated lane code after random dropout operation to obtain the lane node feature.

[0088] The dynamic diffusion graph convolution module (DDGCM) proposed in this embodiment implements diffusion convolution to explicitly capture the dynamics and randomness of social interactions between agents. Formally, this process can be expressed as:

[0089]

[0090]

[0091] Among them, D K Indicates the number of diffusion steps (set to 2); A f =A / rowsum(A) and A b =A T / rowsum(A T ) represent the forward and backward transfer probability matrices of the lane attention map respectively; is the set of all updated lane codes, size N V ×D; and is a learnable parameter; is a lane encoding containing social interaction information; Dropout(.) represents a random dropout operation; are lane node features output by the proposed DDGCM.

[0092] It should be noted that, in this embodiment, steps S1 to S4 are a description of the entire encoder module, wherein steps 2.1 to 2.4 describe the operations performed by the scene-guided gating unit (SGG), and steps S3 and S4 are descriptions of the dynamic diffusion graph convolution module DDGCM. The dynamic diffusion graph convolution module DDGCM proposed in this embodiment is intended to comprehensively describe the highly random social interactions between agents. To achieve this goal, the module uses the dynamic diffusion process on the lane map to simulate the random social interactions between agents on the lane. Based on the bidirectional random walk on the lane attention map, the module explicitly captures the dynamics, uncertainty and random propagation of social interactions between agents.

[0093] Step S5: Input the target code, lane node features, and the adjacency matrix of the lane graph into the strategy header to obtain a lane node set.

[0094] In order to find the lane segment that is most relevant to the future trajectory of the target vehicle, the strategy head aims to predict the discrete probability distribution from each lane node to the next lane node. Based on this probability distribution, this embodiment can determine the most likely future driving path of the target vehicle through graph traversal. Specifically, the strategy head uses the target encoding E target , Lane node features and the adjacency matrix of the lane graph as input, and predict the probability distribution through multilinear mapping, as shown in formula (10):

[0095]

[0096] in, and yes A subset of ; (i, j) is the successor edge in the lane graph structure; φ[.] represents a three-layer multilayer perceptron; N E is the number of successor edges of lane node i; S i,j is the probability that the vehicle at lane node i will travel to lane node j. i,j , the strategy head can get a set of lane nodes [v1,v2,…,v M ], these nodes represent the most likely future driving paths of the target vehicle.

[0097] It should be noted that the strategy head predicts probability distribution through multiple linear mappings and is trained through behavioral cloning. This is a relatively mature technology in this field and will not be elaborated here.

[0098] Step S6: Obtain lane node features corresponding to the lane node set, form a path code, and use the path code to update the target code.

[0099] Based on the lane node set [v1, v2, ..., v M ], the decoder module can obtain the path encoding This embodiment uses E path To update the target encoding to align with the most likely future driving path.

[0100] In a specific embodiment, the updating of the target code using the path code includes:

[0101] Linearly map the path encoding to obtain key vector and value vector;

[0102] Linearly map the target encoding to obtain the query vector;

[0103] Perform scaled dot product attention operation to obtain target encoding with path context;

[0104] The target encoding with the path context is then concatenated with the target encoding to obtain the updated target encoding.

[0105] The specific process is shown in formulas (11) and (12).

[0106] E target ′=Attention(Q target ,K path ,V path ) (11)

[0107] H target =E target ||E target ′ (12)

[0108] Among them, K path and V path are the key vector and value vector obtained by linear mapping path encoding; Q target is the query vector obtained by linearly mapping the target encoding; is the target encoding with path context; is the updated target encoding, which contains information about the most likely future driving path. Here, Attention(.) is the scaled dot product attention operation (target vehicle-lane attention).

[0109] Step S7: The position of the target vehicle, the latent variables following the multivariate Gaussian distribution, and the updated target encoding are passed through a linear layer and then filtered by a scene-guided gating unit. The filtered features are passed through a linear layer to obtain a predicted trajectory.

[0110] In this embodiment, the decoder module aggregates information to generate the target vehicle's future trajectory. In the decoder module, the target vehicle's position, latent variables following a multivariate Gaussian distribution, and the updated target encoding are passed through a linear layer and then filtered by a scene-guided gating unit. The filtered features are then passed through a linear layer to produce the predicted trajectory.

[0111] The process can be expressed as the following formula:

[0112]

[0113] H hidden ′=f SGG (H hidden ) (14)

[0114]

[0115] in, is the position of the target vehicle; is a latent variable following a multivariate Gaussian distribution, used to enhance uncertainty modeling; and is a learnable parameter; D h is the dimension of the hidden feature; is a hidden feature; f SGG (.) is the operation set of the proposed SGG; is the hidden feature after filtering; is the predicted trajectory. LeakyReLU[.] is the activation function.

[0116] To achieve multimodal trajectory prediction, the proposed model samples various plausible future driving paths in the policy head. It obtains different path encodings and generates K predicted trajectories

[0117] It should be noted that the scene guide gating unit SGG in the decoder module needs to filter the original code Steps S6 and S7 describe the operations performed by the decoder module.

[0118] In one embodiment of the present application, a multi-task loss function is used to train the proposed SGG-DDGCN model:

[0119]

[0120] Among them, λ1 and λ2 are hyperparameters; E GT is the set of edges visited by the true trajectory; is the true trajectory at time step t.

[0121] Another embodiment of the present application further provides an autonomous driving trajectory prediction device based on scene-guided gating, comprising a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, perform the steps of the above-mentioned autonomous driving trajectory prediction method based on scene-guided gating.

[0122] The specific limitations of the scenario-guided gating-based autonomous driving trajectory prediction device can be found in the limitations of the scenario-guided gating-based autonomous driving trajectory prediction method described above and will not be repeated here. The scenario-guided gating-based autonomous driving trajectory prediction device described above can be implemented in whole or in part through software, hardware, or a combination thereof. It can be embedded in or independent of a processor in a computer device in hardware form, or it can be stored in a computer device's memory in software form, so that the processor can call and execute the corresponding operations described above.

[0123] The memory and processor are electrically connected, directly or indirectly, to enable data transmission or interaction. For example, these components may be electrically connected via one or more communication buses or signal lines. The memory stores a computer program executable on the processor, and the processor implements the methods of the embodiments of the present invention by executing the computer program stored in the memory.

[0124] The memory may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory is used to store a program, and the processor executes the program after receiving an execution instruction.

[0125] The processor may be an integrated circuit chip with data processing capabilities. The processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor may be a microprocessor or any conventional processor.

[0126] In order to evaluate the technical solution of this application, a challenging and widely used trajectory prediction dataset is used to evaluate the proposed SGG-DDGCN. The dataset provides 2 seconds of historical trajectories and 6 seconds of future trajectories with a data interval of 0.5 seconds (i.e., 2Hz), and also includes a high-definition map of the scene. The dataset is divided into three parts: training set, validation set, and test set, containing 32186, 8560, and 9041 samples respectively. Two widely used standard indicators are used for evaluation: (1) the minimum average displacement error (MinADE) of the top k predictions k ) and (2) miss rate (MissRate k,2 ), the former measures the minimum point-to-point L2 distance between the predicted trajectory and the true trajectory, while the latter indicates the proportion of the predicted results that are more than 2 meters away from the true trajectory.

[0127] The performance of the proposed SGG-DDGCN is compared with the state-of-the-art baseline models on the nuScenes benchmark, and the results are shown in Table 1. Overall, the traversal-based models (PGP, XHGP, and the proposed SGG-DDGCN) achieve more accurate trajectory predictions compared to other models, validating the effectiveness of the traversal-based prediction framework. In addition, driven by the proposed SGG and DDGCM, the proposed SGG-DDGCN outperforms the others in four metrics (MinADE5, MinADE 10 、MissRate 5,2 and MissRate 10,2 ) and achieved the best performance among traversal-based models.

[0128] Table 1

[0129]

[0130] To investigate the contributions of the key components of the proposed SGG-DDGCN, we conducted ablation experiments on the nuScenes benchmark. Specifically, we evaluated the learning rate schedule (LRS), the proposed SGG, and the DDGCM, and present the results in Table 2. Experimental results show that, compared to the baseline model (decay rate of 0.5 every 10 cycles) (row 1), the proposed LRS (decay rate of 0.97 per epoch) (row 2) better facilitates model training, thereby improving trajectory prediction accuracy. Furthermore, the addition of the proposed SGG (row 3) and DDGCM (row 6) improves all four model metrics. Notably, the addition of GCN (row 4) or GAT (row 5) produces ambiguous results: some metrics deteriorate while others improve. This may be due to the decoding module aggregating information about the agent's context along the traversal path, rendering fixed or single-hop social interactions redundant.

[0131] Table 2

[0132]

[0133] In order to analyze the computational performance of the proposed SGG-DDGCN, this application compares the running efficiency of different models on the nuScenes benchmark. Specifically, according to the source code and recommended hyperparameters provided by the authors of PGP and XHGP, the traversal-based models (PGP, XHGP and the proposed SGG-DDGCN) were tested on the same server. Table 3 shows the number of parameters of these models and the time required to complete a training or validation cycle. It can be observed that the proposed SGG-DDGCN does not significantly increase the number of parameters. In addition, the proposed SGG-DDGCN has higher training and inference efficiency while achieving excellent prediction performance.

[0134] Table 3

[0135]

[0136] Through extensive experiments on the nuScenes benchmark, this application verifies the effectiveness and robustness of the proposed SGG-DDGCN.

[0137] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for autonomous driving trajectory prediction based on scene-guided gating, characterized in that: The autonomous driving trajectory prediction method based on scene-guided gating includes: Use a gated recurrent unit to encode the target vehicle's past trajectory, the lane map, and the surrounding agent's past trajectory to obtain a target code, a lane code, and an agent code, where the surrounding agent is a surrounding vehicle or pedestrian; The agent code is filtered using the scene-guided gating unit to obtain a filtered agent code, and the filtered agent code is used to update the lane code; The lane map adjacency matrix and the updated lane encoding are used to generate a lane attention map through the self-attention mechanism; Perform a dilated convolution operation on the updated lane code and lane attention map, and then add them to the updated lane code after random dropout operation to obtain lane node features; Input the target encoding, lane node features, and the adjacency matrix of the lane graph into the strategy head to obtain the lane node set; Get the lane node features corresponding to the lane node set, form a path code, and use the path code to update the target code; The target vehicle’s position, latent variables following a multivariate Gaussian distribution, and the updated target encoding are passed through a linear layer and then filtered by a scene-guided gating unit. The filtered features are passed through a linear layer to obtain the predicted trajectory. The scene-guided gating unit performs the following operations: Build the embedding vectors corresponding to the surrounding agents through the embedding layer; Concatenate the embedding vectors corresponding to the surrounding agents and the position of the target vehicle to obtain heterogeneous scene information; Input the heterogeneous scene information into the meta-learner to learn the weight parameters of the meta-linear layer; The original code to be filtered is input into the meta-linear layer, and then after the activation operation, the Hadamard product is performed with the original code, and after the Hadamard product, the code is added to the original code to obtain the filtered code.

2. The method for autonomous driving trajectory prediction based on scene-guided gating according to claim 1, characterized in that: The method of using the filtered agent code to update the lane code includes: Linearly map the filtered agent encoding to obtain the key vector and value vector; Linearly map the lane encoding to obtain the query vector; Perform scaled dot product attention operation to obtain lane encoding with agent context; The lane coding with the agent context is then concatenated with the lane coding and linearly transformed to obtain the updated lane coding.

3. The method for autonomous driving trajectory prediction based on scene-guided gating according to claim 1, characterized in that: The lane attention map is generated by applying the adjacency matrix of the lane map and the updated lane encoding to the self-attention mechanism, including: Perform linear mapping on the updated lane coding to obtain the query vector, key vector, and value vector; Perform self-attention operation and then perform Hadamard product with the adjacency matrix of the lane map to obtain the lane attention map.

4. The method for autonomous driving trajectory prediction based on scene-guided gating according to claim 1, characterized in that: The method of updating the target code using the path code includes: Linearly map the path encoding to obtain key vector and value vector; Linearly map the target encoding to obtain the query vector; Perform scaled dot product attention operation to obtain target encoding with path context; The target encoding with the path context is then concatenated with the target encoding to obtain the updated target encoding.

5. An autonomous driving trajectory prediction device based on scene-guided gating, comprising a processor and a memory storing a plurality of computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images