A multi-agent trajectory prediction method based on a prompt-free large language model and a learnable catmull-rom spline

Abstract

The application discloses a multi-agent trajectory prediction method of a large language model without prompting and a learnable Catmull-Rom spline. The application combines non-prompt engineering time series coding and adaptive spline decoding to construct a lightweight and efficient trajectory prediction framework that is physically feasible. The application can efficiently model time series features without prompting engineering, while reducing computational overhead through low-rank adaptive fine-tuning, in order to solve the problems of single-scale extraction deficiency and limited long-range dependence capture of existing methods. The application can realize continuous and smooth trajectory output through a learnable Catmull-Rom spline to ensure continuous speed and acceleration, in order to solve the problems of non-smooth trajectory and insufficient physical rationality of traditional point-by-point regression trajectory. Through accurate time series modeling and physically reasonable decoding, the application achieves the purposes of improving trajectory prediction accuracy, smoothness and reasoning efficiency.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving environment perception and motion prediction technology, specifically to a multi-agent trajectory prediction method that combines a non-cue large language model with learnable Catmull-Rom splines. Background Technology

[0002] Accurate multi-agent trajectory prediction is crucial for path planning and safety decisions in autonomous driving systems. The prediction results determine the accuracy of autonomous vehicles' predictions of the behavior of dynamic traffic participants, thus impacting driving safety. Traditional physics-based multi-agent trajectory prediction models often rely on manually defined rules and simplified environmental assumptions, neglecting the dynamic nature of the traffic environment and the complex dependencies between agents in real-world autonomous driving scenarios. This leads to significant prediction errors and fails to meet actual driving requirements.

[0003] With the advancement of deep learning technology, multi-agent trajectory prediction technology has also made significant progress. Deep learning methods can effectively extract rich information from the historical trajectories of agents and the surrounding scene, capture complex interaction dynamics between agents, and significantly improve prediction accuracy, thus becoming the mainstream technology for multi-agent trajectory prediction. However, existing deep learning methods still have significant shortcomings and are difficult to adapt to complex autonomous driving scenarios. Among them, methods based on graph neural networks (GNNs) and graph attention networks (GATs), although capable of expressing the dependencies between agents, can lead to prediction bias and increase collision risk if adjacent nodes are not selected properly in practical applications. Fully connected graphs, on the other hand, result in excessively high computational costs, failing to meet the real-time decision-making needs of vehicles. Furthermore, their temporal modeling capabilities are weak, making it difficult to capture multi-scale temporal information, accurately predict sudden agent behaviors, and adapt to complex road conditions.

[0004] In recent years, large language models (LLMs) have provided a new approach to temporal modeling for multi-agent trajectory prediction. However, in practical applications, they rely on complex prompting engineering, which can easily introduce modeling biases. Furthermore, they are difficult to integrate effectively with existing prediction processes, increasing deployment difficulty and limiting their practical application value.

[0005] Furthermore, the trajectory decoding method directly affects the physical rationality of the predicted trajectory. Traditional point-by-point regression strategies are prone to generating jagged paths, resulting in unstable vehicle driving and potential safety hazards. Function approximation methods such as Bézier curves do not have interpolation characteristics and are prone to deviating from the actual movement trend of the intelligent agent. Neither of these can meet the requirements of actual autonomous driving for trajectory smoothness and physical rationality.

[0006] To improve the accuracy, real-time performance, and practicality of multi-agent trajectory prediction, and to address the issues of insufficient temporal modeling, reliance on complex prompting engineering in LLM applications, and unreasonable trajectory decoding in existing technologies, while ensuring the physical rationality of the predicted trajectory and adapting to the needs of actual autonomous driving scenarios.

[0007] Therefore, in order to improve the robustness and practicality of multi-agent trajectory prediction and adapt to the real-time decision-making and safe driving requirements of autonomous driving, a solution needs to be proposed to address the problems of insufficient temporal feature extraction, limited long-range dependency capture, poor trajectory physical rationality, and complex LLM application in existing technologies.

[0008] Currently, to address the key challenges of temporal modeling and trajectory decoding in multi-agent trajectory prediction, a combination of non-cue LLM with temporal compensation and low-rank adaptive fine-tuning techniques can be used to capture fine-grained temporal features in a lightweight manner and reduce computational costs. At the same time, a learnable curve decoding mechanism can be adopted to improve the physical rationality of the trajectory and the scene fit while ensuring the speed advantage of interpolation.

[0009] On the other hand, from the perspective of practical application and adaptation: the non-cue LLM application paradigm can avoid the modeling bias caused by complex prompting engineering. Combined with the learnable curve decoding mechanism, this collaborative approach can effectively overcome the limitations of existing methods such as insufficient temporal modeling and poor trajectory rationality. It can also ensure prediction performance while reducing model parameters, thus adapting to the actual deployment needs of autonomous driving. Summary of the Invention

[0010] This invention proposes a multi-agent trajectory prediction algorithm based on the collaboration of a non-cue large language model and learnable Catmull-Rom splines. It aims to solve the problems of insufficient temporal feature extraction, limited long-range dependency capture, complex LLM application process, and non-smooth trajectory decoding path and insufficient physical rationality in existing multi-agent trajectory prediction methods by combining the non-cue LLM application paradigm, temporal compensation, low-rank adaptive fine-tuning technology and learnable Catmull-Rom spline decoding mechanism.

[0011] This method uses a single time step of an agent as a token to construct an LLM temporal modeling paradigm that does not require explicit engineering prompts. It combines a local temporal compensation module to compensate for the shortcomings of fine-grained temporal feature modeling, and employs low-rank adaptive fine-tuning technology. This lightweight approach captures fine-grained temporal features of multiple agents and enhances the ability to capture long-range temporal correlations while reducing computational costs and model parameters, thus adapting to the real-time decision-making needs of autonomous driving. Simultaneously, it utilizes a learnable Catmull-Rom spline decoding mechanism, combined with adaptive tension prediction and segmented connection compensation design. Leveraging its local interpolation advantages, it ensures the smoothness, continuity, and physical rationality of the predicted trajectory, preventing the path from deviating from the agent's actual motion trend. This meets the practical requirements of trajectory usage in autonomous driving scenarios, ultimately improving the accuracy, robustness, and practicality of multi-agent trajectory prediction. This provides reliable trajectory prediction support for path planning and safety decision-making in autonomous driving systems, ensuring the safe and stable operation of autonomous vehicles and providing a superior technical solution for autonomous driving environmental perception and motion prediction.

[0012] To achieve the above objectives, this invention provides a multi-agent trajectory prediction method that combines a non-cue large language model with learnable Catmull-Rom splines, comprising the following steps: S1. Based on the agent's historical trajectory information, construct a non-cue large language model temporal coding module through local temporal enhancement and low-rank adaptive fine-tuning to obtain the agent's temporal features; S2. Based on the lane node sequence information, a lane map encoding module is constructed using PointNet and scene constraint enhancement to obtain lane scene features; S3. Based on the temporal characteristics of the agent and the characteristics of the lane scene, a spatial interaction modeling module is constructed by combining tokens and multi-head attention mechanism to obtain the agent fusion characteristics of the fused scene information; S4. Based on the fusion characteristics of the intelligent agent, a multimodal smooth trajectory decoding module is constructed through multimodal projection, adaptive tension prediction, historical trajectory alignment and segmented Catmull-Rom spline interpolation to obtain the multimodal predicted trajectory.

[0013] Preferably, S1 specifically includes: Based on the pre-processed trajectory features of the agent, local enhancement features are obtained by combining two layers of one-dimensional convolution with the GeLU activation function and then connecting the residuals. Based on the local enhancement features, global long-term temporal dependencies are captured through the self-attention mechanism of the pre-trained large language model, and a high-level truncation strategy is adopted to retain the last L-layer Transformer of the model as the core encoder to obtain LLM encoding features. Based on the LLM encoding features, low-rank decomposition is performed on the query matrix and key matrix of the attention layer of the large language model. The parameters are updated through the low-rank matrix, and global average pooling is performed along the time dimension to obtain the agent's temporal features of a fixed dimension.

[0014] Preferably, the step of obtaining local enhancement features includes: extracting fine-grained dynamic features based on the pre-processed agent trajectory features by using two layers of one-dimensional convolution combined with the GeLU activation function, and obtaining local enhancement features through residual connections.

[0015] Preferably, S2 specifically includes: Based on the ordered node sequence of the lane, an initial embedding transformation is performed through a multilayer perceptron to map the original coordinates, direction, and drivability information to a unified high-dimensional feature space and obtain the initial embedding features. Based on the initial embedded features, feature aggregation from local lane points to global road segments is achieved through two-level point feature aggregation and max pooling operations to obtain the aggregated lane features; Based on the aggregated lane features, a nonlinear transformation is performed through a scene constraint enhancement subnetwork to correct abnormal features that violate traffic physics rules and obtain lane scene features that fit the real road structure.

[0016] Preferably, S3 specifically includes: Based on the agent's temporal features and the lane scene features, the agent's temporal features and the lane's global representation are concatenated into a joint token sequence and injected with relative position encoding to obtain the joint token sequence; Construct an information tensor based on any pair of tokens in the joint token sequence, and all information tensors form an information matrix. Based on the information matrix, message passing and state updates are completed through multi-head attention computation, and agent features that separate and acquire fused scene information are obtained.

[0017] Preferably, based on the relative positional relationship of any pair of tokens and their own characteristics, an information tensor is constructed by fusing the relative positional encoding with the characteristics: in, This represents the relative positional characteristics of tokens i and j. This represents the inherent characteristics of token i. This represents the inherent characteristics of token j. The weight matrix represents the linear transformation. This represents the bias term of the linear transformation.

[0018] Preferably, S4 specifically includes: Based on the fusion features of the intelligent agent, K trajectory prediction modes are generated by multimodal feature projection and normalized by Softmax to obtain the modal probabilities; Based on the fusion features of the agents, the tension parameter τ corresponding to the trajectory mode prediction of each agent is obtained, and the trajectory smoothness is dynamically controlled by the tension value to obtain the adaptive tension parameter; Based on the fusion characteristics of the intelligent agent, the first control point of the predicted trajectory is aligned with the end point of the historical trajectory by learning the two-dimensional offset compensation amount, and the optimized trajectory control point is obtained by combining weighted prediction with residual correction. Based on the optimized trajectory control points and adaptive tension parameters, segmented predicted trajectories are generated through segmented Catmull-Rom interpolation. The velocity difference at the junction of adjacent segments is calculated and superimposed on the current segment trajectory as a compensation amount. After splicing and dynamic correction, a multimodal predicted trajectory is obtained.

[0019] Preferably, the step of generating segmented predicted trajectories through segmented Catmull-Rom interpolation includes: Based on the optimized trajectory control points and adaptive tension parameters, an adaptive tension coefficient matrix is ​​constructed based on the tension value τ, and a time power matrix is ​​constructed. The adaptive interpolation matrix is ​​then calculated through matrix multiplication. Based on the adaptive interpolation matrix and piecewise control points, a piecewise predicted trajectory is generated through matrix multiplication. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a lightweight, efficient, and physically feasible trajectory prediction framework by combining non-cue engineering temporal coding with adaptive spline decoding. It addresses the shortcomings of existing methods in single-scale extraction of temporal features and limited long-range dependency capture, achieving efficient temporal feature modeling without cue engineering, while reducing computational overhead through low-rank adaptive fine-tuning. Furthermore, it addresses the issues of uneven trajectory smoothness and insufficient physical plausibility in traditional point-by-point regression by using learnable Catmull-Rom splines to achieve continuous and smooth trajectory output, ensuring continuity of velocity and acceleration. Through precise temporal modeling and physically plausible decoding, this invention aims to improve trajectory prediction accuracy, smoothness, and inference efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the intelligent agent temporal semantic encoding process according to an embodiment of the present invention; Figure 2This is a flowchart of the lane map encoding process according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the spatial interaction modeling process in an embodiment of the present invention. Figure 4 This is a flowchart of the multimodal smooth decoding process according to an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 S1. Based on the agent's historical trajectory information, a non-cue large language model temporal coding module is constructed through local temporal enhancement and low-rank adaptive fine-tuning to obtain the agent's temporal features.

[0025] Based on local temporal augmentation and low-rank adaptive fine-tuning, a temporal encoding module for a cueless large language model is constructed. This module combines a one-dimensional convolutional local temporal augmentation unit with a pre-trained large language model to build a temporal modeling structure that requires no cueing engineering. Using a single time step as the processing unit, this module extracts fine-grained dynamic features through local temporal compensation and achieves lightweight parameter optimization through low-rank adaptive fine-tuning. While preserving the long-range dependency capture capability of the large model, it reduces computational overhead and avoids modeling biases introduced by cueing engineering, achieving efficient and accurate encoding of agent temporal features.

[0026] This step achieves temporal feature extraction for unannotated engineering, reusing the pre-training capabilities of large models and adapting them to trajectory prediction tasks. It is the core temporal coding step in this embodiment. Figure 1 As shown.

[0027] Trajectory Feature Mapping: The pre-processed agent trajectory features mapped to the LLM latent space dimension are denoted as... .in, The number of intelligent agents in the scene. For the hidden layer dimension of LLM, The length of the time step in the historical trajectory.

[0028] The local temporal enhancement submodule employs two layers of one-dimensional convolution combined with the GeLU activation function to capture fine-grained dynamic features such as instantaneous acceleration and subtle heading adjustments of the agent. These features are then processed through residual connections to obtain local enhancement features. The specific implementation is as follows: (1) (2) in, This is a one-dimensional convolution operation used to lightweight capture local short-range temporal dependencies; This represents enhanced temporal features that integrate the original temporal information with short-term local motion dependencies; This indicates the presence of local temporal characteristics; This represents the original agent trajectory features that have been mapped to the latent space dimension of the large model.

[0029] Unhinted LLM Temporal Encoding: A pre-trained large language model using Qwen2.5 is employed, without constructing any text prompts, directly using a single trajectory time step as a token input. A high-level truncation strategy is used, retaining only the last L layers of the Transformer as the core encoder, and a self-attention mechanism is used to capture global long-range temporal dependencies.

[0030] Efficient fine-tuning of LoRA parameters: for LLM attention layers and Matrix low-rank decomposition, low-rank dimension scaling factor The parameter update method is as follows: (3) in, To update the matrix; and It is a low-rank matrix decomposed from . This indicates that all elements are real numbers and the shape is r rows. The linear space to which the column matrix belongs corresponds to the tensor size of the low-rank matrix A.

[0031] Training only low-rank matrices and Freeze the backbone network parameters to obtain time-series-aware features: (4) in, This represents the processing module after the original pre-trained large model has undergone appropriate pruning and the addition of a low-rank adapter; The elements are all real numbers, and the dimension is the number of agents. Historical time series length Feature channels The linear space to which the three-dimensional feature tensor belongs.

[0032] Then, by performing global average pooling along the time dimension, a fixed-dimensional pool can be obtained. The temporal characteristics of the intelligent agent. The specific implementation is as follows: (5) in, t This represents the index of a single time step within a historical observation sequence, with values ​​ranging from t=1,2,… ; The elements are all real numbers, and the dimension is the number of agents. Aggregate time-series feature dimensions The two-dimensional feature matrix space.

[0033] S2. Based on the lane node sequence information, a lane map encoding module is constructed using PointNet and scene constraint enhancement to obtain lane scene features.

[0034] Based on PointNet and scene constraint enhancement, a lane map encoding module is constructed: combining the PointNet feature extraction structure with scene constraint correction to encode lane geometric and spatial information. This module achieves lane feature aggregation through multilayer perceptron and hierarchical pooling, and introduces a scene constraint enhancement subnetwork to correct abnormal features that violate traffic logic, making lane representations more closely resemble real road structures and providing reliable scene prior information for trajectory prediction.

[0035] This step transforms the lane geometry and spatial logic of the high-definition map into interactive scene features, with the coding architecture as follows: Figure 2 As shown.

[0036] First, execute the ordered sequence of lane nodes. The initial embedding transformation maps the original coordinates, orientation, drivability, and other information to a unified high-dimensional feature space, resulting in the initial embedding features. .in, The number of points for each lane; It is the dimension of the feature space of the mapping.

[0037] Next, a two-level point feature aggregation and max pooling operation are used to achieve feature aggregation from local lane points to the global road segment. The specific implementation is as follows: (6) (7) (8) (9) (10) in, For layer normalization, For tensor copy expansion, For max pooling along the point dimension, This represents the refined point-level features of each point in a single lane after undergoing a local linear transformation. This represents a joint representation where each point carries both its own local features and the global features of the entire lane. This represents the depth features of lane points with residual connections after hierarchical aggregation. This represents the final global feature representation of the j-th lane. This indicates a fixed dimension for the output feature vector. This represents the weight matrix of the first-level linear transformation. This represents the weight matrix of the second-level linear transformation. This represents the globally aggregated feature vector for the entire lane. This represents the bias of the first-level linear transformation. This represents the bias of the second-level linear transformation. Indicates the initial embedding of lane points. This represents a global one-dimensional max pooling operation. Indicates taking out the first k The complete feature vector corresponding to each lane point T This indicates transpose processing.

[0038] Finally, a scene constraint enhancement subnetwork is introduced, consisting of two layers. Constitutes nonlinear transformation This corrects abnormal features that violate traffic physics rules, making lane characteristics more closely resemble the smoothness and connectivity of real roads. The specific implementation is as follows: (11) in, This represents the modified lane point features after applying geometric prior constraints.

[0039] S3. Based on the temporal characteristics of the agent and the characteristics of the lane scene, a spatial interaction modeling module is constructed by combining tokens and multi-head attention mechanism to obtain the agent fusion characteristics of the fused scene information.

[0040] Based on a joint token and multi-head attention mechanism, a spatial interaction modeling module is constructed: the temporal features of the agent and the lane scene features are concatenated into a joint token sequence, and interaction modeling is completed through stacked spatiotemporal feature layers and multi-head attention. This module achieves bidirectional information fusion between agents and between agents and the environment through information matrix construction and attention weighting, accurately captures complex interaction dependencies, and outputs agent fusion features containing contextual information.

[0041] This step completes the two-way information exchange between agent features and lane features, and models the complex spatial dependencies between multiple agents and the traffic scene, such as... Figure 3 As shown.

[0042] First, the agent's temporal features and the lane's global representation are concatenated into a joint token sequence, and relative position encoding is injected to supplement prior spatial location information. The joint token sequence is as follows: (12) in, Indicates all within the scene The feature sequence obtained by temporal encoding of each agent Indicates all within the scene The global lane feature sequence obtained by the map encoder for each lane. This represents the tensor size of the spliced ​​mixed sequence H.

[0043] Next, construct an information tensor for any pair of tokens. ,in, The relative position encoded feature vector represents the spatial relative positional relationship between token i and token j. This represents the basic feature vector of the i-th hybrid token. This represents the basic feature vector of the j-th hybrid token. This represents the learnable weight matrix of the pairwise interactive feature projection layer. This represents the learnable bias term of the pairwise interactive feature projection layer. This information tensor fuses relative positional relationships with its own features. All information tensors together form an information matrix. .

[0044] Secondly, multi-head attention calculation is performed. The specific implementation is as follows: (13) (14) (15) (16) in, H represents the joint token sequence H obtained by concatenating the agent and lane features. This represents the tensor of all pairwise token interactions. A stacked three-dimensional global interactive information matrix. This indicates that the Query branch can learn a linear projection weight matrix. This indicates that the Key branch can learn a linear projective weight matrix. This indicates that the Value branch can learn a linear projective weight matrix. The query vector matrix, key vector matrix, and value vector matrix represent the multi-head attention mechanism. The feature channel dimension represents the single-head attention. This represents the total number of tokens in the scene. Hadamard product, which is the element-wise multiplication of matrices. This represents the scene interaction mask matrix.

[0045] Finally, by stacking multiple spatiotemporal feature layers to complete message passing and state updates, the intelligent agent features fused with scene information are obtained. .

[0046] S4. Based on the fusion characteristics of the intelligent agent, a multimodal smooth trajectory decoding module is constructed through multimodal projection, adaptive tension prediction, historical trajectory alignment and segmented Catmull-Rom spline interpolation to obtain the multimodal predicted trajectory.

[0047] Based on learnable Catmull-Rom splines and adaptive tension, a multimodal smooth trajectory decoding module is constructed. This module combines multimodal projection, adaptive tension prediction, historical trajectory alignment, and piecewise spline interpolation to build a learnable Catmull-Rom spline decoding structure. It generates multiple potential trajectories through multimodal projection, dynamically adjusts trajectory smoothness using adaptive tension, eliminates trajectory start-up jumps through historical alignment, and ensures continuous and smooth trajectory through piecewise interpolation and velocity compensation. Finally, it outputs a reasonable trajectory that conforms to kinematic constraints.

[0048] This step transforms high-level semantic features into smooth, kinematically feasible multimodal future trajectories, and is the core decoding step in this embodiment. Figure 4 .

[0049] Multimodal feature projection: through two layers Projecting the features of a single-modal agent into a multimodal space generates... A trajectory prediction model, and after Normalization yields the modal probabilities. The specific implementation is as follows: (17) (18) (19) in, Indicates the first The agent adopts the first The probability of a trajectory pattern, This represents the multimodal trajectory feature tensor obtained after multi-head projection transformation and dimensional rearrangement. The original unnormalized pattern score tensor output by the pattern classification subnetwork. This represents the index of the j-th candidate future trajectory pattern currently being calculated. The elements are all real numbers, and the dimension is the total number of trajectory patterns K and the number of scene agents. Single-modal feature channel The linear space to which the three-dimensional multimodal feature tensor belongs, This represents the trajectory pattern index used for summation and traversal.

[0050] Adaptive tension prediction: Predicting the tension parameters corresponding to the trajectory modes of each agent. Through tension value Dynamically control trajectory smoothness for smoother straight-line driving and better alignment with control points during lane changes or turns. This is achieved through the construction of an adaptive Catmull-Rom basis matrix. The format is as follows: (20) Historical trajectory alignment and control point refinement: By learning the two-dimensional offset compensation, the first control point of the predicted trajectory is aligned with the end point of the historical trajectory, eliminating trajectory jumps. Then, weighted prediction combined with residual correction is used to output the optimized trajectory control points. .

[0051] Piecewise Catmull-Rom interpolation: Transforming the future... Each time step is divided into Each segment shares two control points with adjacent segments to ensure seamless connection. For the nth segment, based on the tension value... Constructing an adaptive tension coefficient matrix And construct the time power matrix. The adaptive interpolation matrix is ​​calculated by multiplication. The specific implementation is as follows: (twenty one) (twenty two) (twenty three) in, For the first The number of prediction time steps for a segment; Normalized time values; Each row corresponds to a control point weight at a given time point.

[0052] Then, based on and segmented control points Segmented predicted trajectories are generated through matrix multiplication. .

[0053] Segmented connection speed consistency compensation: Calculate the speed difference at the connection point of adjacent segments, and use half of the difference as compensation to add to the current segment trajectory to ensure speed continuity; splice all segments and perform dynamic correction to output the final multimodal trajectory. .

[0054] In this embodiment, the implementation method of the temporal coding module of the unannotated large language model includes: constructing local temporal enhancement units using one-dimensional convolution and residual connections to extract short-term dynamic features; performing high-level pruning on the pre-trained large language model to retain effective temporal modeling capabilities; and employing low-rank adaptive fine-tuning technology to adapt to trajectory prediction tasks by optimizing only a small number of parameters. The implementation method of the multimodal smooth trajectory decoding module includes: generating multiple sets of trajectories and outputting probability distributions through multimodal projection; dynamically adjusting trajectory smoothness through adaptive tension prediction; ensuring trajectory continuity through historical alignment and control point optimization; and generating reasonable trajectories through segmented Catmull-Rom interpolation and velocity compensation.

[0055] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-agent trajectory prediction method that combines a non-cue large language model with learnable Catmull-Rom splines, characterized by the following steps: include: S1. Based on the agent's historical trajectory information, construct a non-cue large language model temporal coding module through local temporal enhancement and low-rank adaptive fine-tuning to obtain the agent's temporal features; S2. Based on the lane node sequence information, a lane map encoding module is constructed using PointNet and scene constraint enhancement to obtain lane scene features; S3. Based on the temporal characteristics of the agent and the characteristics of the lane scene, a spatial interaction modeling module is constructed by combining tokens and multi-head attention mechanism to obtain the agent fusion characteristics of the fused scene information; S4. Based on the fusion characteristics of the intelligent agent, a multimodal smooth trajectory decoding module is constructed through multimodal projection, adaptive tension prediction, historical trajectory alignment and segmented Catmull-Rom spline interpolation to obtain the multimodal predicted trajectory.

2. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines according to claim 1, characterized in that, S1 specifically includes: Based on the pre-processed trajectory features of the agent, local enhancement features are obtained by combining two layers of one-dimensional convolution with the GeLU activation function and then connecting the residuals. Based on the local enhancement features, global long-term temporal dependencies are captured through the self-attention mechanism of the pre-trained large language model, and a high-level truncation strategy is adopted to retain the last L-layer Transformer of the model as the core encoder to obtain LLM encoding features. Based on the LLM encoding features, low-rank decomposition is performed on the query matrix and key matrix of the attention layer of the large language model. The parameters are updated through the low-rank matrix, and global average pooling is performed along the time dimension to obtain the agent's temporal features of a fixed dimension.

3. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 2, is characterized in that... The steps for obtaining local enhancement features include: extracting fine-grained dynamic features based on the pre-processed agent trajectory features through two layers of one-dimensional convolution combined with the GeLU activation function, and obtaining local enhancement features through residual connections.

4. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 1, characterized in that... S2 specifically includes: Based on the ordered node sequence of the lane, an initial embedding transformation is performed through a multilayer perceptron to map the original coordinates, direction, and drivability information to a unified high-dimensional feature space to obtain the initial embedding features. Based on the initial embedded features, feature aggregation from local lane points to global road segments is achieved through two-level point feature aggregation and max pooling operations to obtain the aggregated lane features; Based on the aggregated lane features, a nonlinear transformation is performed through a scene constraint enhancement subnetwork to correct abnormal features that violate traffic physics rules and obtain lane scene features that fit the real road structure.

5. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 1, characterized in that... S3 specifically includes: Based on the agent's temporal features and the lane scene features, the agent's temporal features and the lane's global representation are concatenated into a joint token sequence and injected with relative position encoding to obtain the joint token sequence; Construct an information tensor based on any pair of tokens in the joint token sequence, and all information tensors form an information matrix. Based on the information matrix, message passing and state updates are completed through multi-head attention computation, and agent features that separate and acquire fused scene information are obtained.

6. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 5, is characterized in that... Based on the relative positional relationship and individual characteristics of any pair of tokens, an information tensor is constructed by fusing the relative positional encoding with the individual characteristics: in, This represents the relative positional characteristics of tokens i and j. This represents the inherent characteristics of token i. This represents the inherent characteristics of token j. The weight matrix represents the linear transformation. This represents the bias term of the linear transformation.

7. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 1, characterized in that, S4 specifically includes: Based on the fusion features of the intelligent agent, K trajectory prediction modes are generated by multimodal feature projection and normalized by Softmax to obtain the modal probabilities; Based on the fusion features of the agents, the tension parameter τ corresponding to the trajectory mode prediction of each agent is obtained, and the trajectory smoothness is dynamically controlled by the tension value to obtain the adaptive tension parameter; Based on the fusion characteristics of the intelligent agent, the first control point of the predicted trajectory is aligned with the end point of the historical trajectory by learning the two-dimensional offset compensation amount, and the optimized trajectory control point is obtained by combining weighted prediction with residual correction. Based on the optimized trajectory control points and adaptive tension parameters, segmented predicted trajectories are generated through segmented Catmull-Rom interpolation. The velocity difference at the junction of adjacent segments is calculated and superimposed on the current segment trajectory as a compensation amount. After splicing and dynamic correction, a multimodal predicted trajectory is obtained.

8. The multi-agent trajectory prediction method based on the collaboration between a non-cue large language model and learnable Catmull-Rom splines as described in claim 7, is characterized in that... The steps for generating piecewise predicted trajectories using piecewise Catmull-Rom interpolation include: Based on the optimized trajectory control points and adaptive tension parameters, an adaptive tension coefficient matrix is ​​constructed based on the tension value τ, and a time power matrix is ​​constructed. The adaptive interpolation matrix is ​​then calculated through matrix multiplication. Based on the adaptive interpolation matrix and segmented control points, segmented predicted trajectories are generated through matrix multiplication.

Images