Vehicle trajectory prediction method and related equipment

Abstract

The invention relates to the technical field of vehicle trajectory prediction, and provides a vehicle trajectory prediction method and related equipment, and the method comprises the steps: training a driving intention recognition model and a trajectory prediction model through employing an improved clustering algorithm, and obtaining a trained driving intention recognition model and a trained trajectory prediction model; based on the adaptive graph and all the vehicle information, utilizing a trained driving intention recognition model to recognize the driving intention of the to-be-predicted vehicle, and utilizing a trained trajectory prediction model to predict and obtain an initial prediction trajectory of the to-be-predicted vehicle based on the driving intention of the to-be-predicted vehicle; and performing kinematics prediction on the to-be-predicted vehicle to obtain a motion prediction track, and superposing the motion prediction track and the initial prediction track of the vehicle to obtain a prediction track of the to-be-predicted vehicle. According to the method, the accuracy of trajectory prediction in different driving scenes and different prediction tasks can be improved, and the problem of prediction precision reduction caused by forgetting is solved.

Description

Technical Field

[0001] This application relates to the field of vehicle trajectory prediction technology, and in particular to a vehicle trajectory prediction method and related equipment. Background Technology

[0002] In today's rapidly developing autonomous driving technology, accurate trajectory prediction is widely recognized as a cornerstone technology for ensuring vehicle safety and system reliability. It forms the prerequisite for decision-making and planning modules, directly impacting the rationality of autonomous vehicle behavior in complex dynamic environments. However, current mainstream trajectory prediction technologies and their performance in practical applications are facing serious challenges.

[0003] Currently, most advanced trajectory prediction frameworks follow a similar development model: training, validating, and testing models on large-scale, well-labeled public benchmark datasets (such as Argoverse, nuScenes, and Waymo Open Dataset). These datasets capture rich interaction scenarios of traffic participants (vehicles, pedestrians, etc.) within specific regions and time periods. Based on this data, researchers have developed numerous complex model architectures aimed at inferring future movement intentions from historical trajectory sequences. These models can be broadly categorized as follows: 1. Deep learning-based methods: This is the current mainstream approach. They typically utilize encoder-decoder structures, where the encoder (such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), or Transformer) learns to extract features from the historical trajectories of traffic participants, and the decoder generates future trajectory points based on these features. 2. Graph neural network-based methods: To explicitly model the interaction relationships between traffic participants, graph neural networks (GNNs) have been introduced. Treating each participant in the scene as a node in a graph, and connecting potentially interacting entities with edges, graph neural networks are used for information transmission and aggregation, thereby learning complex social interaction representations. 3. Attention-based methods: The Transformer architecture and its attention mechanism have been widely applied in trajectory prediction. It can adaptively focus on other participants or environmental elements that have the greatest impact on the future trajectory of the target vehicle, thus better handling long-distance dependencies and a variable number of interacting objects.

[0004] On their respective test sets, these models demonstrated impressive prediction accuracy through metrics such as average displacement error and final displacement error. However, this static training paradigm based on a fixed dataset has the following drawbacks.

[0005] Pain Point 1: Catastrophic Forgetting – Knowledge Loss in Sequential Learning

[0006] Autonomous vehicles inevitably collect new data and encounter new scenarios throughout their lifecycle. An ideal system should be able to learn continuously from these new experiences, just like a human, and constantly improve itself. However, most current models face a fatal problem when confronted with the requirement of "continuous learning" or "lifelong learning"—catastrophic forgetting. Specifically, when we try to update a model trained on a previous dataset (Task A) with a dataset containing new scenarios and behaviors (Task B), the model will drastically adjust its parameters to adapt to the new data distribution. This process causes the model to quickly forget the knowledge and skills learned in Task A, resulting in a precipitous drop in its performance on that task. This means that learning a new task may come at the cost of sacrificing the performance of older tasks. For autonomous driving systems that need to maintain comprehensive capabilities throughout their entire operational lifecycle, this "learning new and forgetting old" characteristic is absolutely unacceptable. It prevents the model from accumulating and consolidating knowledge, leaving it perpetually a "short-sighted" student.

[0007] Pain Point 2: The unsustainability of computing and storage

[0008] Faced with catastrophic forgetting, an intuitive solution is "joint training," which involves merging all historical data with the new data and retraining the entire model whenever new data arrives. However, from a computational and storage perspective, this approach is completely unsustainable. As vehicle mileage increases, the amount of data collected will be massive and continuously growing. Storing all historical data requires enormous hardware costs. More importantly, each full retraining requires huge amounts of computing resources (such as GPU computing power) and a long time, which severely hinders the rapid iteration and deployment of models. Autonomous driving systems need an efficient, incremental learning solution, not one that becomes increasingly cumbersome and costly to train over time.

[0009] Pain Point 3: Difficulty in maintaining and reconstructing complex spatiotemporal dependencies

[0010] The essence of trajectory prediction is to understand and predict a dynamic system with complex spatiotemporal dependencies. These dependencies are not only reflected in the temporal continuity of the motion of individual objects, but also in the dynamic and non-stationary interactions between multiple objects as their spatial position and velocity change. Many existing lifelong learning methods (such as those directly from computer vision) often struggle to effectively maintain and reconstruct these finely structured spatiotemporal relationships when applied to trajectory prediction. They may fail to fully reproduce the complex interaction context from the initial training when replaying historical data, resulting in suboptimal or even erroneous interaction models that negatively impact prediction accuracy.

[0011] In summary, catastrophic forgetting makes continuous learning extremely difficult, and the high cost of full retraining is impractical in engineering. At the same time, existing methods face challenges in maintaining complex spatiotemporal interaction knowledge, resulting in poor model training performance and low accuracy in vehicle trajectory prediction. Summary of the Invention

[0012] This application provides a vehicle trajectory prediction method and related equipment, which can solve the problem of low accuracy in vehicle trajectory prediction.

[0013] In a first aspect, embodiments of this application provide a vehicle trajectory prediction method, which includes:

[0014] Obtain vehicle information of the vehicle to be predicted, as well as vehicle information of the vehicles surrounding the vehicle to be predicted, and construct an adaptive graph based on all vehicle information; multiple nodes of the adaptive graph correspond one-to-one with multiple vehicles, and the edges between nodes represent the proximity relationship between the corresponding two vehicles.

[0015] An improved clustering algorithm was used to train the driving intention recognition model and the trajectory prediction model, resulting in the trained driving intention recognition model and the trained trajectory prediction model.

[0016] Based on the adaptive graph and all vehicle information, the driving intention of the vehicle to be predicted is identified using the trained driving intention recognition model, and the initial predicted trajectory of the vehicle to be predicted is obtained based on the driving intention of the vehicle to be predicted using the trained trajectory prediction model.

[0017] Kinematic prediction is performed on the vehicle to be predicted to obtain the predicted trajectory. The predicted trajectory is then superimposed on the initial predicted trajectory of the vehicle to be predicted.

[0018] Optionally, based on the adaptive graph and all vehicle information, the driving intention of the vehicle to be predicted is identified using the trained driving intention recognition model, including:

[0019] Using the trained driving intention recognition model, spatial features are extracted from the adaptive map to obtain spatial intention features, and temporal features are extracted from each vehicle information to obtain temporal intention features for each vehicle information.

[0020] All temporal intent features are concatenated with spatial intent features to obtain concatenated intent features. These concatenated intent features are then decoded to obtain the driving intent of the vehicle to be predicted.

[0021] Optionally, the initial predicted trajectory of the vehicle to be predicted is obtained using the trained trajectory prediction model, including:

[0022] Using the trained trajectory prediction model, spatial features are extracted from the adaptive map to obtain spatial trajectory features, and temporal features are extracted from the vehicle information of the vehicle to be predicted to obtain temporal trajectory features of the vehicle to be predicted.

[0023] The temporal and spatial trajectory features of the vehicle to be predicted are concatenated to obtain the concatenated trajectory features. A fully connected neural network is selected based on the driving intention of the vehicle to be predicted, and the concatenated trajectory features are decoded using the fully connected neural network to obtain the initial predicted trajectory of the vehicle to be predicted.

[0024] Optionally, kinematic prediction is performed on the vehicle to be predicted to obtain the predicted motion trajectory, including:

[0025] Through the formula:

[0026]

[0027] Calculate the trajectory points in the predicted motion trajectory of the vehicle to be predicted. , ;

[0028] in, This represents the first trajectory point in the predicted trajectory of the vehicle to be predicted. The first element in the predicted trajectory of the vehicle to be predicted. A trajectory point, The first element in the predicted trajectory of the vehicle to be predicted. A trajectory point, This indicates the number of trajectory points in the predicted motion trajectory. Indicates displacement. Indicates speed, It represents acceleration.

[0029] Optionally, an improved clustering algorithm is used to train the driving intention recognition model and the trajectory prediction model, resulting in trained driving intention recognition models and trained trajectory prediction models, including:

[0030] An improved clustering algorithm is used to cluster vehicle trajectory data in an external historical memory storage device to obtain the centroids of multiple clusters; the centroids of the clusters are used as the centroids of the corresponding clusters of vehicle trajectory data.

[0031] The driving intention recognition model and trajectory prediction model are trained using all cluster centroids to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0032] Optionally, an improved clustering algorithm is used to cluster the vehicle trajectory data in the external historical memory storage to obtain the cluster centroids of multiple clusters, including:

[0033] Multiple initial centroids are selected from all vehicle trajectory data in the external historical memory storage;

[0034] Based on all initial centroids, the predicted trajectories of all vehicle trajectory data are clustered to obtain multiple clusters, and a new centroid is calculated for each cluster.

[0035] Calculate the average difference between all new centroids and all initial centroids;

[0036] Determine whether the average difference is less than a preset difference threshold;

[0037] If so, then the new centroid of each cluster is taken as the cluster centroid;

[0038] Otherwise, the new centroid is used as the initial centroid, and the steps of clustering all vehicle trajectory data based on all initial centroids are returned to obtain multiple clusters, and the new centroid of each cluster is calculated.

[0039] Optionally, multiple initial centroids can be selected from the predicted trajectories of all vehicles, including:

[0040] An initial centroid is randomly selected from all vehicle trajectory data;

[0041] Based on the selected initial centroid, the next initial centroid that meets the similarity condition is selected from all vehicle trajectory data;

[0042] Determine whether the number of selected initial centroids is less than the number threshold;

[0043] If so, the last selected initial centroid is taken as the current centroid, and the process of selecting the next initial centroid that meets the similarity condition from all vehicle trajectory data is returned.

[0044] Optional, similarity criteria are:

[0045] The similarity between the currently selected vehicle trajectory data and each selected initial centroid is less than the similarity threshold.

[0046] Secondly, embodiments of this application provide a vehicle trajectory prediction device, comprising:

[0047] The acquisition module is used to acquire vehicle information of the vehicle to be predicted, as well as vehicle information of the vehicles around the vehicle to be predicted, and to construct an adaptive graph based on all vehicle information; multiple nodes of the adaptive graph correspond one-to-one with multiple vehicles, and the edges between nodes represent the proximity relationship between the corresponding two vehicles.

[0048] The training module is used to train the driving intention recognition model and the trajectory prediction model using an improved clustering algorithm, so as to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0049] The recognition module is used to identify the driving intention of the vehicle to be predicted based on the adaptive map and all vehicle information, using a trained driving intention recognition model, and to predict the initial predicted trajectory of the vehicle to be predicted based on the driving intention of the vehicle to be predicted using a trained trajectory prediction model.

[0050] The kinematic prediction module is used to perform kinematic prediction on the vehicle to be predicted, obtain the motion prediction trajectory, and superimpose the motion prediction trajectory with the vehicle's initial prediction trajectory to obtain the predicted trajectory of the vehicle to be predicted.

[0051] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the vehicle prediction method described above.

[0052] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle prediction method described above.

[0053] The above-mentioned solution in this application has the following beneficial effects:

[0054] In the embodiments of this application, vehicle information of the vehicle to be predicted and vehicle information of the vehicles surrounding the vehicle to be predicted are obtained, and an adaptive graph is constructed based on all vehicle information. Then, an improved clustering algorithm is used to train the driving intention recognition model and the trajectory prediction model to obtain the trained driving intention recognition model and the trained trajectory prediction model. Based on the adaptive graph and all vehicle information, the driving intention of the vehicle to be predicted is identified using the trained driving intention recognition model, and the initial predicted trajectory of the vehicle to be predicted is predicted based on the driving intention of the vehicle to be predicted using the trained trajectory prediction model. Finally, kinematic prediction is performed on the vehicle to be predicted to obtain the motion prediction trajectory, and the motion prediction trajectory is superimposed with the initial predicted trajectory of the vehicle to obtain the predicted trajectory of the vehicle to be predicted. Among these methods, using a driving intention recognition model to identify the vehicle's driving intention and then predicting the vehicle's trajectory based on that intention can improve the richness of information considered in trajectory prediction. By incorporating the mining and analysis of vehicle behavior, the accuracy of trajectory prediction can be effectively improved. Using an improved clustering algorithm to train the driving intention recognition model and the trajectory prediction model can filter out representative training data, achieve knowledge review, mitigate catastrophic forgetting of the model, and improve model performance. Using a high-performance driving intention recognition model and trajectory prediction model for vehicle trajectory prediction can further improve the accuracy of vehicle trajectory prediction.

[0055] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart of a vehicle trajectory prediction method provided in an embodiment of this application;

[0058] Figure 2 This is a schematic diagram illustrating the construction of an adaptive graph provided in an embodiment of this application;

[0059] Figure 3 A schematic diagram illustrating the improved attention operation provided in an embodiment of this application;

[0060] Figure 4 This is a schematic diagram of the structure of a vehicle trajectory prediction device provided in an embodiment of this application;

[0061] Figure 5This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0062] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0063] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0064] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0065] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0066] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0067] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0068] To address the low accuracy of existing vehicle trajectory prediction methods, this application provides a vehicle trajectory prediction method. This method acquires vehicle information of the vehicle to be predicted, as well as vehicle information of surrounding vehicles. An adaptive graph is constructed based on all vehicle information. An improved clustering algorithm is then used to train a driving intention recognition model and a trajectory prediction model, resulting in trained driving intention recognition and trajectory prediction models. Based on the adaptive graph and all vehicle information, the trained driving intention recognition model identifies the driving intention of the vehicle to be predicted. The trained trajectory prediction model then predicts the initial trajectory of the vehicle based on its driving intention. Finally, kinematic prediction is performed on the vehicle to obtain a motion prediction trajectory, which is then superimposed on the initial prediction trajectory to obtain the final predicted trajectory of the vehicle. Among these methods, using a driving intention recognition model to identify the vehicle's driving intention and then predicting the vehicle's trajectory based on that intention can improve the richness of information considered in trajectory prediction. By incorporating the mining and analysis of vehicle behavior, the accuracy of trajectory prediction can be effectively improved. Using an improved clustering algorithm to train the driving intention recognition model and the trajectory prediction model can filter out representative training data, achieve knowledge review, mitigate catastrophic forgetting of the model, and improve model performance. Using a high-performance driving intention recognition model and trajectory prediction model for vehicle trajectory prediction can further improve the accuracy of vehicle trajectory prediction.

[0069] The vehicle trajectory prediction method provided in this application will be described exemplarily below.

[0070] like Figure 1 As shown, the vehicle trajectory prediction method provided in this application includes the following steps:

[0071] Step 11: Obtain vehicle information of the vehicle to be predicted, as well as vehicle information of the vehicles surrounding the vehicle to be predicted, and construct an adaptive graph based on all vehicle information.

[0072] In the adaptive graph described above, each node corresponds one-to-one with a vehicle (including the vehicle to be predicted and the vehicles surrounding it). Edges between nodes represent the proximity relationship between two corresponding vehicles; that is, there are no other vehicles between two vehicles. Two vehicles are considered adjacent and have a proximity relationship, with an edge between their corresponding nodes. The vehicle information includes the vehicle's position, speed, steering angle, acceleration, etc. The vehicles surrounding the vehicle to be predicted are those within a certain range centered on the vehicle to be predicted.

[0073] Specifically, a corresponding node is constructed for each vehicle. If two vehicles corresponding to two nodes are adjacent, an edge is generated between the two nodes.

[0074] It should be noted that after constructing the adaptive graph, for each vehicle, we can also analyze the regions of its neighboring vehicles relative to that vehicle (including eight regions: front, rear, left front, left side, left rear, right front, right side, and right rear). At the same time, we can analyze information such as the interaction distance between vehicles and use it as the attribute of the node in the adaptive graph.

[0075] For example, the construction of an adaptive graph is as follows: Figure 2 As shown, Figure 2 Figure 'a' is a diagram illustrating the actual driving of five vehicles. and Indicates the interaction distance between vehicles. This indicates the width of the target vehicle (i.e., the vehicle whose trajectory needs to be predicted), and the numbers represent the area number surrounding the target vehicle. Figure 2 b is based on Figure 2 Graph a is a schematic diagram of the adaptive graph constructed, where nodes correspond to vehicles, edges represent the proximity relationships between vehicles, and graph c is the topological graph of the vehicles. In graph c, circles correspond to vehicles in graph a, edges represent the proximity relationships between vehicles, and m represents the attribute of the corresponding edge. The expression for the interaction distance between vehicles is:

[0076]

[0077] in, Indicates speed.

[0078] As speed increases, the interaction distance in front of the target vehicle becomes greater than the interaction distance behind it, because vehicles in front have a more significant impact on future trajectories. If a vehicle exists within region i, the vehicle with the smallest relative Euclidean distance to the target vehicle, and not exceeding [a certain value], is selected. and The vehicle is used as the representative node of the region; if no vehicle is detected in region i, the corresponding node feature will be assigned a zero vector or a preset anchor point representing the farthest point of the region boundary, as shown below.

[0079]

[0080] in, This represents the two-dimensional position coordinates of vehicle v. Let i be the farthest anchor point in region i. This design ensures spatial integrity while effectively reflecting potential long-distance interaction possibilities when there are no vehicles in the adjacent area.

[0081] At any moment From node arrive The message is defined as As shown below, it represents the relative distance between vehicles.

[0082]

[0083] Then, the messages between nodes are combined into a message matrix, which can well represent spatial deviations.

[0084] Step 12: Using an improved clustering algorithm, train the driving intention recognition model and the trajectory prediction model to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0085] In some embodiments of this application, the steps of training the driving intention recognition model and the trajectory prediction model using the improved clustering algorithm to obtain the trained driving intention recognition model and the trained trajectory prediction model include:

[0086] The first step is to use an improved clustering algorithm to cluster the vehicle trajectory data in the external historical memory storage to obtain the cluster centroids of multiple clusters.

[0087] The aforementioned cluster centroids serve as the vehicle trajectory data for the corresponding clusters. The external historical memory storage contains a large amount of vehicle trajectory data, and new vehicle trajectory data is continuously added. This vehicle trajectory data consists of actual collected vehicle trajectory data.

[0088] Specifically, multiple initial centroids are selected from all vehicle trajectory data in the external historical memory storage; all vehicle trajectory data are clustered according to all initial centroids to obtain multiple clusters, and a new centroid is calculated for each cluster; the average difference between all new centroids and all initial centroids is calculated; it is determined whether the average difference is less than a preset difference threshold; if so, the new centroid of each cluster is used as the cluster centroid; otherwise, the new centroid is used as the initial centroid, and the process of clustering all vehicle trajectory data according to all initial centroids to obtain multiple clusters and calculating a new centroid for each cluster is returned.

[0089] It should be noted that the above steps for selecting multiple initial centroids from all vehicle trajectory data in the external historical memory storage include: randomly selecting an initial centroid from all vehicle trajectory data; based on the selected initial centroid, selecting the next initial centroid that meets the similarity condition from all vehicle trajectory data; determining whether the number of selected initial centroids is less than the number threshold; if so, taking the last selected initial centroid as the current centroid, and returning to the step of selecting the next initial centroid that meets the similarity condition from all vehicle trajectory data.

[0090] The similarity condition is that the similarity between the currently selected predicted trajectory and each selected initial centroid is less than the similarity threshold. The similarity calculation formula is:

[0091]

[0092] in, Indicates similarity. Represents the initial centroid With the initial centroid The distance between them.

[0093] For example, the above-mentioned process of clustering and calculating centroids can be implemented by K-means clustering. After obtaining the cluster centroids, the speed, acceleration of the vehicle corresponding to the cluster centroid and the position of the surrounding vehicles can be integrated to improve the information richness contained in the cluster centroids.

[0094] The second step is to train the driving intention recognition model and the trajectory prediction model using all cluster centroids to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0095] For example, all cluster centroids are used as training data, and the driving intention recognition model and trajectory prediction model are trained using methods such as gradient descent to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0096] It should be noted that, in order to improve the efficiency of trajectory prediction, a training cycle can be set for the model. If the model has already been trained, this step is not required before the training cycle is reached. When the training cycle is reached, this step is executed to retrain the model to ensure its performance.

[0097] Step 13: Based on the adaptive graph and all vehicle information, the driving intention of the vehicle to be predicted is identified using the driving intention recognition model, and the initial predicted trajectory of the vehicle to be predicted is predicted based on the driving intention of the vehicle to be predicted using the trajectory prediction model.

[0098] The aforementioned driving intention is used to describe the vehicle's state change in the next moment (such as changing lanes to the left, changing lanes to the right, keeping up with the car in front, etc.). The initial predicted trajectory is used to describe the vehicle's driving trajectory over a predicted time period under the influence of the driving intention and the current vehicle information. The predicted time period is a future time period corresponding to the vehicle's vehicle information.

[0099] In some embodiments of this application, the steps described above—identifying the driving intention of the vehicle to be predicted using a driving intention recognition model based on an adaptive graph and all vehicle information, and predicting the initial predicted trajectory of the vehicle to be predicted based on the driving intention using a trajectory prediction model—include:

[0100] The first step is to use the trained driving intention recognition model to extract spatial features from the adaptive map to obtain spatial intention features, and to extract temporal features from each vehicle information to obtain temporal intention features for each vehicle information.

[0101] For example, the driving intent recognition model described above can be a model based on CNN, LSTM, and fully connected layer (FC). For instance, CNN and FC can be used to encode the spatial interaction features of the adaptive graph in sequence to obtain spatial intent features, and LSTM and FC can be used to encode the temporal interaction features of the vehicle information in sequence to obtain the temporal intent features of the vehicle.

[0102] The second step is to concatenate all temporal intent features with spatial intent features to obtain concatenated intent features, and then decode the concatenated intent features to obtain the driving intent of the vehicle to be predicted.

[0103] For example, the feature concat operation can be used to concatenate the vehicle's temporal intent features with its spatial intent features to obtain concatenated intent features. Then, FC can be used to decode the concatenated intent features to obtain the vehicle's driving intent.

[0104] If the driving intention is a 1 A tensor of 1 is classified as 0 if its value is greater than -0.5 and less than 0.5, representing keeping straight. If its value is greater than or equal to 0.5, it is classified as 1, representing a left lane change. If its value is less than or equal to -0.5, it is classified as -1, representing a right lane change.

[0105] The third step involves using the trained trajectory prediction model to extract spatial features from the adaptive map, obtaining spatial trajectory features, and extracting temporal features from the vehicle information of the vehicle to be predicted, obtaining the temporal trajectory features of the vehicle to be predicted.

[0106] For example, the above trajectory prediction model is a model based on attention algorithm, FC, LSTM, etc. For example, attention module and FC can be used to encode the spatial interactive features of the adaptive map in sequence to obtain spatial trajectory features, and LSTM and FC can be used to encode the temporal interactive features of the vehicle information in sequence to obtain temporal trajectory features.

[0107] It should be noted that the above attention algorithm is an improved one. Before calculation, the absolute value of K is taken, so that the overall calculation result can better represent the spatial positional deviation. Since the adaptive image data contains both positive and negative values, the weight of negative values ​​is significantly reduced during the Softmax operation, which obviously affects the feature processing effect. Because Q is all positive, the absolute value of K needs to be taken. If the adaptive image data is not encoded by other neural networks and is directly transmitted to the improved attention mechanism, the calculation process of the improved attention is as follows: Figure 3 As shown, K and V are first encoded through a fully connected layer, then activated by the Tanh function, and finally subjected to dropout. Similarly, the query Q is encoded through a fully connected layer, then processed by Tanh activation and dropout. Following the design principles of Transformer, this module introduces residual connections to alleviate the vanishing or exploding gradient problem and accelerate convergence. The output of the attention mechanism is added to the dropout-processed Q, then normalized by the layers, and finally encoded and output through a fully connected layer.

[0108] The improved attention calculation process is shown below:

[0109]

[0110]

[0111] in, , , Represents a learnable matrix. This represents the query vector corresponding to a node in the adaptive graph. This represents the key vector corresponding to a node in the adaptive graph. This represents the value vector corresponding to a node in the adaptive graph. Representing dimension, This represents information from node 0 to node 0. Represents a node Information to node 0.

[0112] The fourth step involves concatenating the temporal and spatial trajectory features of the vehicle to be predicted to obtain concatenated trajectory features. Based on the driving intention of the vehicle to be predicted, a fully connected neural network is selected, and the concatenated trajectory features are decoded using the fully connected neural network to obtain the initial predicted trajectory of the vehicle to be predicted.

[0113] For example, the feature concat operation can be used to concatenate the temporal trajectory features and spatial trajectory features of a vehicle to obtain the concatenated trajectory features.

[0114] It should be noted that before proceeding to this step, multiple fully connected neural networks with different parameters can be set up. Each fully connected neural network is trained by sample vehicles with different driving intentions (which can be optimized using gradient descent, optimizers, etc.). In this step, the corresponding fully connected network is selected according to the vehicle's driving intention, and the fully connected neural network is used to decode the spliced ​​trajectory features to obtain the vehicle's initial predicted trajectory.

[0115] Step 14: Perform kinematic prediction on the vehicle to be predicted to obtain the predicted trajectory, and superimpose the predicted trajectory on the vehicle to obtain the predicted trajectory of the vehicle to be predicted.

[0116] The above predicted trajectory is the vehicle's trajectory within the predicted time period, considering only the vehicle's own information. The above predicted trajectory is the vehicle's trajectory within the predicted time period, considering both the vehicle itself and surrounding vehicles.

[0117] Specifically, through the formula:

[0118]

[0119] Calculate the trajectory points in the predicted motion trajectory of the vehicle to be predicted. , .

[0120] in, This represents the first trajectory point in the predicted trajectory of the vehicle to be predicted. The first element in the predicted trajectory of the vehicle to be predicted. A trajectory point, The first element in the predicted trajectory of the vehicle to be predicted. A trajectory point, This indicates the number of trajectory points in the predicted motion trajectory. Indicates displacement. Indicates speed, It represents acceleration.

[0121] It should be noted that noise can be introduced into the superposition operation when superimposing the motion prediction trajectory with the vehicle's initial prediction trajectory.

[0122] It is worth mentioning that using a driving intention recognition model to identify the vehicle's driving intention and then predicting the vehicle's trajectory based on that driving intention can improve the richness of information considered in trajectory prediction. By incorporating the mining and analysis of vehicle behavior, the accuracy of trajectory prediction can be effectively improved. Using an improved clustering algorithm to train the driving intention recognition model and the trajectory prediction model can select representative training data, realize knowledge review, alleviate the catastrophic forgetting of the model, and improve the model's performance. Using a high-performance driving intention recognition model and trajectory prediction model for vehicle trajectory prediction can further improve the accuracy of vehicle trajectory prediction.

[0123] Furthermore, this application has the following advantages:

[0124] 1. A dynamically adaptive interactive scene graph construction method: This method abandons the fixed grid-based scene representation approach and innovatively proposes a target vehicle-centric, dynamically adjustable interactive graph construction mechanism. The key to this mechanism lies in:

[0125] Semantic region segmentation divides the environment around the target vehicle into eight regions with clear semantics (front, rear, left front, left side, left rear, right front, right side, right rear), making the scene representation more interpretable and consistent with driving cognition.

[0126] The dynamic interaction distance is calculated and adjusted dynamically based on the real-time speed and physical size of the target vehicle, so that the perception range can adapt to different driving conditions (such as needing to perceive the risks ahead earlier at high speeds).

[0127] Robust node representation selects the nearest vehicle to the target vehicle within the interaction distance as a node for each region; if no vehicle is available, a preset "farthest anchor point" is used to fill the gap. This method ensures the spatial integrity of the scene graph and can effectively represent potential long-range interaction possibilities even in sparse vehicle environments, avoiding information loss.

[0128] 2. An improved attention mechanism incorporating spatial bias perception: The standard attention mechanism has been specifically improved to make it more suitable for spatial relationship modeling in trajectory prediction.

[0129] Graph-driven multi-head attention sets attention heads corresponding to the number of nodes in the interaction graph, enabling the model to pay attention to the interaction effects in different semantic directions in parallel.

[0130] Explicit modeling of spatial bias involves taking the absolute value of the key (K) vector before calculating the attention weights. This crucial operation addresses the issue of negative information being suppressed after the positive and negative position coordinates in the original data are processed by the Softmax function, ensuring that the attention weights accurately reflect the spatial positional bias (i.e., the feasible domain) of the target vehicle in all directions.

[0131] The enhanced modular structure introduces fully connected layers, Tanh activation, Dropout, and residual connections when encoding queries (Q), keys (K), and values ​​(V), which enhances the model's nonlinear fitting ability, training stability, and convergence speed.

[0132] 3. Dual-channel trajectory prediction model architecture: A collaborative prediction framework consisting of a driving intent recognition module and a trajectory prediction module was designed, specifically including:

[0133] Intent recognition first identifies the macro driving intent (straight ahead, left lane change, right lane change) of the target vehicle through a dual-channel network (spatial encoder + temporal encoder), transforming the continuous trajectory prediction problem into a serial process of "intent recognition - intent-specific prediction", thus reducing the prediction complexity.

[0134] Heterogeneous feature fusion and trajectory prediction also employ a dual-channel structure, encoding spatial interaction features (from surrounding vehicles) through an improved attention mechanism and temporal motion features (from the target vehicle's historical state) through LSTM. Finally, the two types of heterogeneous features are concatenated and decoded.

[0135] Multimodal output enhancement integrates neural network predictions, kinematic uniform velocity predictions, and small-amplitude random noise when decoding the final trajectory. This design considers the basic motion laws of the vehicle while increasing the diversity and realism of the predicted trajectory, making it closer to the subtle uncertainties present in actual driving.

[0136] 4. Lifelong learning mechanism based on improved clustering and fusion training: This is the core of achieving continuous model evolution, and includes two closely connected parts:

[0137] Improved K-means representativeness knowledge screening:

[0138] a. Centroid initialization optimization: When selecting the initial centroid, an iterative approach is used to ensure that there are sufficient differences between the centroids and to avoid getting trapped in local optima.

[0139] b. Real Trajectory Matching: After clustering, the calculated centroids are not used directly. Instead, they are matched with real trajectories in the original dataset, and the most similar original trajectory data (containing complete location, velocity, acceleration, and surrounding vehicle information) is selected as representative knowledge. This step is crucial, as it preserves all features and authenticity of the original data, avoiding the problem of missing key details in the clustering centroids.

[0140] Spatiotemporal knowledge playback and integration training:

[0141] a. Historical knowledge storage: Representative trajectories selected through improved clustering in each task are stored in an external historical memory storage device.

[0142] b. Cross-task fusion training: When learning a new task, each training batch mixes current task data with representative data from multiple old tasks randomly selected from historical memory. These mixed data are then subjected to uniform gradient descent optimization without zeroing out gradients, thereby achieving deep fusion and collaborative updating of new and old knowledge.

[0143] c. Incremental update of memory: After each new task training is completed, new representative samples are immediately clustered and stored in the history memory, so that the knowledge base can be continuously expanded.

[0144] The vehicle trajectory prediction device provided in this application is described below by way of example.

[0145] like Figure 4 As shown, this application embodiment provides a vehicle trajectory prediction device, the vehicle trajectory prediction device 400 including:

[0146] The acquisition module 401 is used to acquire vehicle information of the vehicle to be predicted, as well as vehicle information of the vehicles around the vehicle to be predicted, and to construct an adaptive graph based on all vehicle information; multiple nodes of the adaptive graph correspond one-to-one with multiple vehicles, and the edges between nodes represent the proximity relationship between the corresponding two vehicles.

[0147] Training module 402 is used to train the driving intention recognition model and trajectory prediction model using an improved clustering algorithm to obtain the trained driving intention recognition model and the trained trajectory prediction model.

[0148] The recognition module 403 is used to identify the driving intention of the vehicle to be predicted based on the adaptive map and all vehicle information, using the trained driving intention recognition model, and to predict the initial predicted trajectory of the vehicle to be predicted based on the driving intention of the vehicle to be predicted using the trained trajectory prediction model.

[0149] The kinematic prediction module 404 is used to perform kinematic prediction on the vehicle to be predicted, obtain the motion prediction trajectory, and superimpose the motion prediction trajectory with the initial prediction trajectory of the vehicle to obtain the prediction trajectory of the vehicle to be predicted.

[0150] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0151] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0152] like Figure 5 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 5 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0153] Specifically, when the processor D100 executes the computer program D102, it acquires vehicle information of the vehicle to be predicted and vehicle information of the vehicles surrounding the vehicle to be predicted, and constructs an adaptive graph based on all vehicle information. Then, it uses an improved clustering algorithm to train the driving intention recognition model and the trajectory prediction model to obtain the trained driving intention recognition model and the trained trajectory prediction model. Based on the adaptive graph and all vehicle information, it uses the trained driving intention recognition model to identify the driving intention of the vehicle to be predicted, and uses the trained trajectory prediction model to predict the initial predicted trajectory of the vehicle to be predicted based on the driving intention of the vehicle to be predicted. Finally, it performs kinematic prediction on the vehicle to be predicted to obtain the motion prediction trajectory, and superimposes the motion prediction trajectory with the initial predicted trajectory of the vehicle to obtain the predicted trajectory of the vehicle to be predicted. Among these methods, using a driving intention recognition model to identify the vehicle's driving intention and then predicting the vehicle's trajectory based on that intention can improve the richness of information considered in trajectory prediction. By incorporating the mining and analysis of vehicle behavior, the accuracy of trajectory prediction can be effectively improved. Using an improved clustering algorithm to train the driving intention recognition model and the trajectory prediction model can filter out representative training data, achieve knowledge review, mitigate catastrophic forgetting of the model, and improve model performance. Using a high-performance driving intention recognition model and trajectory prediction model for vehicle trajectory prediction can further improve the accuracy of vehicle trajectory prediction.

[0154] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0155] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0156] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0157] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0158] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to the vehicle trajectory prediction method device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, such as a USB flash drive, a portable hard drive, a magnetic disk, or an optical disk.

[0159] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0160] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0161] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A vehicle trajectory prediction method, characterized by, The method comprises the following steps: acquiring vehicle information of a to-be-predicted vehicle and vehicle information of vehicles around the to-be-predicted vehicle, and constructing an adaptive graph based on all the vehicle information; a plurality of nodes of the adaptive graph correspond to a plurality of vehicles one by one, and edges between the nodes are adjacent relationships between corresponding two vehicles; training the driving intention recognition model and the trajectory prediction model by using an improved clustering algorithm to obtain a trained driving intention recognition model and a trained trajectory prediction model; based on the adaptive graph and all the vehicle information, recognizing a driving intention of the to-be-predicted vehicle by using the trained driving intention recognition model, and predicting an initial prediction trajectory of the to-be-predicted vehicle based on the driving intention of the to-be-predicted vehicle by using the trained trajectory prediction model; kinematically predicting the to-be-predicted vehicle to obtain a motion prediction trajectory, and superimposing the motion prediction trajectory and the initial prediction trajectory of the vehicle to obtain a prediction trajectory of the to-be-predicted vehicle.

2. The vehicle trajectory prediction method of claim 1, wherein, The method comprises the following steps: extracting spatial features of the adaptive graph by using the trained driving intention recognition model to obtain spatial intention features, and extracting time features of each vehicle information to obtain time intention features of each vehicle information; splicing all the time intention features and the spatial intention features to obtain spliced intention features, and decoding the spliced intention features to obtain the driving intention of the to-be-predicted vehicle.

3. The vehicle trajectory prediction method of claim 2, wherein, The method comprises the following steps: extracting spatial features of the adaptive graph by using the trained trajectory prediction model to obtain spatial trajectory features, and extracting time features of the vehicle information of the to-be-predicted vehicle to obtain time trajectory features of the to-be-predicted vehicle; splicing the time trajectory features of the to-be-predicted vehicle and the spatial trajectory features to obtain spliced trajectory features, selecting a fully connected neural network based on the driving intention of the to-be-predicted vehicle, and decoding the spliced trajectory features by using the fully connected neural network to obtain the initial prediction trajectory of the to-be-predicted vehicle.

4. The vehicle trajectory prediction method of claim 1, wherein, The method comprises the following steps: by using the following formula: ； Computing a trajectory point in a motion prediction trajectory of a vehicle to be predicted , ; wherein represents a 1st trajectory point in a motion prediction trajectory of a vehicle to be predicted, represents a 2nd trajectory point in a motion prediction trajectory of a vehicle to be predicted, represents a 3rd trajectory point in a motion prediction trajectory of a vehicle to be predicted, represents a 4th trajectory point in a motion prediction trajectory of a vehicle to be predicted, represents a 5th trajectory point in a motion prediction trajectory of a vehicle to be predicted, represents a number of trajectory points in a motion prediction trajectory, represents a displacement, represents a velocity, represents an acceleration.

5. The vehicle trajectory prediction method of claim 1, wherein, The method comprises the following steps: clustering vehicle trajectory data in the external historical memory by using the improved clustering algorithm to obtain clustering centroids of a plurality of clustering clusters; the clustering centroids are vehicle trajectory data of corresponding clustering centroids; training the driving intention recognition model and the trajectory prediction model by using all the clustering centroids to obtain a trained driving intention recognition model and a trained trajectory prediction model.

6. The vehicle trajectory prediction method of claim 5, wherein, The improved clustering algorithm is used to cluster the vehicle trajectory data in the external historical memory to obtain clustering centroids of multiple clustering clusters, including: selecting multiple initial centroids from all vehicle trajectory data in the external historical memory; clustering all vehicle trajectory data according to all initial centroids to obtain multiple clustering clusters and calculate new centroids of each clustering cluster; calculating average difference between all new centroids and all initial centroids; determining whether the average difference is less than a preset difference threshold; if yes, taking the new centroid of each clustering cluster as a clustering centroid; otherwise, taking the new centroid as an initial centroid and returning to the step of clustering all vehicle trajectory data according to all initial centroids to obtain multiple clustering clusters and calculate new centroids of each clustering cluster.

7. The vehicle trajectory prediction method of claim 6, wherein, The selecting multiple initial centroids from all vehicle trajectory data in the external historical memory includes: randomly selecting an initial centroid from all vehicle trajectory data; selecting a next initial centroid satisfying a similarity condition from all vehicle trajectory data based on the selected initial centroid; determining whether the number of selected initial centroids is less than a number threshold; if yes, taking the last selected initial centroid as a current centroid and returning to the step of selecting a next initial centroid satisfying the similarity condition from all vehicle trajectory data based on the current initial centroid.

8. The vehicle trajectory prediction method of claim 7, wherein, The similarity condition is that: the similarity between the currently selected vehicle trajectory data and each selected initial centroid is less than a similarity threshold.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the vehicle trajectory prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the vehicle trajectory prediction method according to any one of claims 1 to 8.

Images