Dynamic adaptive trajectory prediction method

Abstract

The invention belongs to the technical field of automatic driving, and particularly relates to a dynamic adaptive trajectory prediction method. Comprising the following steps: acquiring trajectory prediction data and performing preprocessing: performing coding fusion on current states of a vehicle and an environmental vehicle corresponding to position information subjected to coordinate conversion and map information to obtain map-state fusion features, and coding a historical trajectory of the environmental vehicle to obtain historical trajectory features; generating a priority feature vector of the environment vehicle; the map-state fusion features and the priority feature vectors are spliced, and priority-state fusion features are obtained; generating a first trajectory prediction at the current moment according to the priority-state fusion feature and the historical trajectory feature; and introducing a trajectory prediction result generated by the method at a historical moment, and interacting the first trajectory prediction with the historical trajectory prediction result based on an attention mechanism, thereby outputting a second trajectory prediction at the current moment.

Description

A Dynamically Adaptive Trajectory Prediction Method Technical Field

[0001] This application belongs to the field of autonomous driving technology, specifically relating to a dynamic adaptive trajectory prediction method. Background Technology

[0002] In the field of autonomous driving, accurately predicting the future trajectories of vehicles in the surrounding environment is crucial to ensuring the safe and efficient operation of autonomous vehicles. Early work relied heavily on simple physical models, such as constant speed or acceleration models, but these methods struggle to characterize the complex interactions between vehicles, limiting their practicality.

[0003] To address this challenge, existing technologies mostly employ deep learning-based methods for prediction. These methods typically integrate information such as the vehicle's status, historical vehicle trajectories in the environment, and high-precision maps to directly output future trajectory predictions. However, these methods still have significant drawbacks: a lack of dynamic risk assessment mechanisms. Existing methods struggle to differentiate threat levels based on factors such as the real-time relative position of the vehicle and its owner, and the time of collision, resulting in an inability to prioritize the prediction of high-risk targets and hindering the balance between safety and efficiency.

[0004] The existing models neglect the temporal consistency of prediction results. They make independent predictions at each time step, failing to effectively utilize historical prediction information as context. This can lead to unreasonable jumps in the generated continuous trajectories, reducing prediction reliability. Furthermore, the prediction strategy lacks adaptability, making it difficult to make differentiated and targeted predictions based on the real-time changing threat levels of vehicles in the environment, thus limiting further improvements in prediction performance.

[0005] Therefore, there is an urgent need in this field for an adaptive trajectory prediction method that can dynamically assess risks and ensure temporal consistency. Summary of the Invention

[0006] This application proposes a dynamic adaptive trajectory prediction method to address the core shortcomings of existing trajectory prediction technologies: insufficient prediction targeting due to the lack of dynamic risk assessment, poor trajectory temporal consistency due to ignoring historical prediction information, and weak adaptability of prediction strategies.

[0007] The first aspect of this application provides a dynamically adaptive trajectory prediction method, comprising: acquiring trajectory prediction data and preprocessing it, wherein the trajectory prediction data includes the current state of the vehicle and surrounding vehicles, the vehicle's planned path, the historical trajectories of the surrounding vehicles, and map information; the preprocessing includes: performing coordinate transformation on the location information; encoding and fusing the current state of the vehicle and surrounding vehicles corresponding to the coordinate-transformed location information with the map information to obtain map-state fusion features; and encoding the historical trajectories of the surrounding vehicles to obtain historical trajectory features; generating priority feature vectors for the surrounding vehicles, wherein: the vehicle's motion trend is determined based on the vehicle's planned path; the static priority of the surrounding vehicles is calculated based on the relative position of each surrounding vehicle with the motion trend; and the priority of each surrounding vehicle is calculated based on the relative position of the surrounding vehicles with the vehicle. The dynamic priority of the vehicles in the environment is calculated based on the estimated collision time, and the static priority is combined with the dynamic priority. Trajectory prediction is performed based on the map-state fusion feature, historical trajectory feature, and priority feature vector, including: concatenating the map-state fusion feature and the priority feature vector to obtain the priority-state fusion feature; generating a first trajectory prediction for the current moment based on the priority-state fusion feature and the historical trajectory feature; incorporating the trajectory prediction results from historical moments; and using an attention mechanism to enable the first trajectory prediction and the historical trajectory prediction results to interact, thereby outputting a second trajectory prediction for the current moment. In the first prediction, the historical trajectory prediction result is the initial state, default value, or real trajectory data; in subsequent predictions, the historical trajectory prediction result is generated by this method.

[0008] In some embodiments of this application, the coordinate transformation of the location information includes mapping the current position of the vehicle, the current position of the environmental vehicle, and the location information of the planned path of the vehicle to a Frenet coordinate system with the road centerline as the reference line, wherein the s-axis of the Frenet coordinate system represents the longitudinal displacement along the road centerline, and the d-axis represents the lateral offset relative to the road centerline.

[0009] In some embodiments of this application, the step of encoding and fusing the current state of the self-vehicle and surrounding vehicles corresponding to the coordinate-transformed position information with the map information to obtain map-state fusion features includes: encoding the current state and the map information respectively to obtain observation dynamic features and map encoding features, wherein the position information in the current state has undergone the coordinate transformation; fusing the observation dynamic features and the map encoding features through a spatiotemporal cross-attention mechanism to obtain the map-state fusion features. In some embodiments of this application, the step of calculating the static priority of the surrounding vehicles based on the relative position of each surrounding vehicle with the motion trend includes: defining three priority regions—high, medium, and low—in space through dot product operations based on the motion trend and the relative positional relationship between each surrounding vehicle and the self-vehicle; assigning static priority values ​​of 1.0, 0.7, and 0.4 to surrounding vehicles in the high, medium, and low priority regions, respectively.

[0010] In some embodiments of this application, calculating the dynamic priority of the environmental vehicles based on the estimated collision times between each environmental vehicle and the vehicle includes: calculating the estimated collision times between each environmental vehicle and the vehicle, using the following formula: ,in, This represents the estimated collision time between the vehicle and the driver in the j-th environment at time k. and Let be the current positions of the j-th environmental vehicle and the autonomous vehicle at time k, respectively. and Let be the speeds of the j-th environmental vehicle and the driver vehicle at time k, respectively. The calculated estimated collision times for each environmental vehicle are normalized to obtain the dynamic priority, calculated using the following formula: ,in, This represents the dynamic priority of the j-th environmental vehicle at time k. Let represent the estimated collision time between the vehicle and the driver in the j-th environment at time k. Let k represent the set of TTC values ​​for all vehicles in the environment at time k. and Let $k$ and $k$ represent the minimum and maximum TTC values ​​for all environmental vehicles at time $k$, respectively.

[0011] In some embodiments of this application, combining the static priority with the dynamic priority includes: for each environmental vehicle, multiplying its static priority value by its dynamic priority value to obtain the priority value of the environmental vehicle; and combining the priority values ​​of all environmental vehicles in sequence to form the priority feature vector of the environmental vehicles.

[0012] In some embodiments of this application, generating a first trajectory prediction for the current moment based on the priority-state fusion feature and the historical trajectory feature includes: concatenating the priority-state fusion feature and the historical trajectory feature, and inputting them into a recurrent neural network decoder for decoding to generate the first trajectory prediction.

[0013] In some embodiments of this application, the introduction of trajectory prediction results generated by the method at historical moments includes: for the current moment k, selecting environmental vehicle historical trajectory prediction results generated by the method for the previous t consecutive historical moments (kt, k-t+1, ..., k-1); performing feature extraction on the selected historical trajectory prediction result sequence to generate a historical trajectory embedding vector sequence; and using the historical trajectory embedding vector sequence as the historical trajectory prediction result for subsequent interaction with the first trajectory prediction at the current moment.

[0014] In some embodiments of this application, the attention mechanism is a multi-head attention mechanism.

[0015] In some embodiments of this application, the interaction between the first trajectory prediction and the historical trajectory prediction results based on the attention mechanism includes: using the first trajectory prediction as a query vector, using the historical trajectory embedded vector sequence as a key vector and a value vector, and inputting them into the multi-head attention mechanism to calculate the association weights and perform the interaction.

[0016] In summary, the dynamic adaptive trajectory prediction method provided in the embodiments of this application firstly, by introducing a priority feature vector based on the vehicle's motion trend and the expected collision time, and concatenating it with map-state fusion features, achieves a dynamic and refined assessment of potential vehicle risks in the environment, providing the prediction model with key information to distinguish the importance of vehicles; secondly, by constructing a historical prediction attention mechanism, the current prediction can interact with historical prediction results, effectively utilizing temporal context information and fundamentally ensuring the smoothness and inherent consistency of the generated trajectory; finally, the synergistic effect of the above technologies enables the entire system to more intelligently adapt to changes in complex dynamic environments, thereby significantly improving the accuracy and reliability of trajectory prediction overall. Attached Figure Description

[0017] The features and advantages of this application will be more clearly understood by referring to the accompanying drawings, which are illustrative and should not be construed as limiting the application in any way. In the drawings: Figure 1 shows the logical architecture for implementing the dynamic adaptive trajectory prediction method of this application; Figure 2 is a schematic diagram of a computer system to which this application applies; Figure 3 is a flowchart of a dynamic adaptive trajectory prediction method according to some embodiments of this application; Figure 4 is a flowchart of data acquisition and preprocessing in the input layer of this application; Figure 5 is a schematic diagram of the principle of static priority in this application; Figure 6 is a schematic diagram of the principle of dynamic priority in this application; Figure 7 is a schematic diagram of the historical prediction attention module of this application; Figure 8 is a network architecture diagram of the historical prediction attention module of this application. Detailed Implementation

[0018] In the following detailed description, numerous specific details of this application are illustrated by example to provide a thorough understanding of the relevant disclosure. However, it will be apparent to those skilled in the art that this application can be practiced without these details. It should be understood that the terms “system,” “apparatus,” “unit,” and / or “module” used in this application are one way of distinguishing different parts, elements, sections, or components at different levels in a sequential arrangement. However, these terms may be replaced with other expressions if other expressions can achieve the same purpose.

[0019] It should be understood that when a device, unit, or module is referred to as being "on," "connected to," or "coupled to" another device, unit, or module, it may be directly connected to or coupled to or communicate with other devices, units, or modules, or there may be intermediate devices, units, or modules present, unless the context explicitly indicates otherwise. For example, the term "and / or" as used herein includes any one and all combinations of one or more of the relevant listed items.

[0020] The terminology used in this application is for the purpose of describing specific embodiments only and is not intended to limit the scope of this application. As shown in the specification and claims of this application, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate that explicitly identified features, integrals, steps, operations, elements, and / or components are included, and such expressions do not constitute an exclusive list, and other features, integrals, steps, operations, elements, and / or components may also be included.

[0021] Referring to the following description and accompanying drawings, these and other features and characteristics, operating methods, functions of related structural elements, combinations of parts, and economics of manufacture of this application can be better understood, wherein the description and drawings form part of the specification. However, it is clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. It is understood that the drawings are not drawn to scale.

[0022] Various structural diagrams are used in this application to illustrate various variations of the embodiments according to this application. It should be understood that the preceding or following structures are not intended to limit this application. The scope of protection of this application is determined by the claims.

[0023] In real-world traffic scenarios, the behaviors of road users (such as vehicles, pedestrians, and bicycles) are complex and variable, exhibiting high uncertainty and interactivity. How to more efficiently and accurately predict their future trajectories is a major technical challenge facing current autonomous driving systems. Traditional methods based on physical models or rule-driven approaches, while possessing some interpretability, struggle to cope with diverse and dynamically changing real-world scenarios. In recent years, with the development of deep learning, data-driven trajectory prediction algorithms have received widespread attention. Novel methods, such as graph neural networks and Transformer networks, have gradually demonstrated superior performance on multiple publicly available autonomous driving datasets.

[0024] Nevertheless, existing methods still have significant shortcomings in modeling the dynamics and continuity of trajectories: current mainstream methods often lack an effective distinction between the current importance of different surrounding vehicles when dealing with multi-objective scenarios, making it difficult to reasonably allocate the focus of the autonomous driving system, thus affecting the overall accuracy and efficiency of the prediction; in addition, current prediction methods often adopt independent prediction strategies, without considering previous predicted trajectories when making each trajectory prediction, ignoring the temporal dependence of traffic participants' behavior and the inherent correlation between similar prediction stages, which makes the model prone to trajectory jumps or inconsistencies when facing continuous prediction tasks.

[0025] In view of this, this application proposes a dynamically adaptive trajectory prediction method. The core of this method lies in the introduction of two innovative modules: a priority judgment module and a historical prediction attention module.

[0026] Specifically, unlike existing technologies that predict surrounding vehicles "equally," this application's priority determination module achieves differentiated prediction through the following method: A coordinate system is established based on the vehicle's current observation state and planned path, and a coordinate transformation is performed to project surrounding vehicles onto this coordinate system. Then, a static priority is calculated based on the relative positional relationship between each vehicle and the vehicle's planned path, and a dynamic priority is calculated by combining the estimated time of collision (TTC) between the vehicle and each vehicle. Finally, the two are merged to obtain a priority vector for each vehicle in the environment. This process simulates the spatial attention allocation preferences of human drivers, assigning a "priority label" to each vehicle. In subsequent decoding and prediction, this priority vector is concatenated with observation features, enabling the prediction process to adjust attention based on priority features, thereby optimizing computational efficiency while maintaining accuracy.

[0027] Meanwhile, unlike existing technologies that rely solely on current observations for prediction, this application's historical prediction attention module concatenates historical predictions with features extracted from the input layer and feeds them into the decoding prediction network. This allows the current prediction to incorporate output information from previous time steps. This enhances the model's temporal correlation between similar predictions of the same vehicle, effectively improving the adaptability and coherence of the predicted trajectory.

[0028] Figure 1 illustrates the logical architecture of the trajectory prediction method for implementing the dynamic adaptive method of this application. The figure describes the main steps of the method and the interaction relationships between the functional modules.

[0029] Figure 2 is a schematic diagram of a computer system applicable to this application. The computer system shown in Figure 2 includes a trajectory dataset with data connectivity, a trajectory prediction calculation unit, and a control decision unit. The trajectory prediction calculation unit generates a predicted trajectory for the surrounding vehicles using trajectory prediction data loaded from the trajectory dataset; the control decision unit generates control commands for the vehicle based on this predicted trajectory. The trajectory dataset stores the data required for trajectory prediction, including historical state sequences of the vehicle and surrounding vehicles, the vehicle's planned path, historical trajectories of the surrounding vehicles, and corresponding high-precision map information.

[0030] The trajectory prediction calculation unit is the core processing unit for executing the method described in this application. This unit acquires and aggregates the current state of the vehicle and surrounding vehicles, the vehicle's planned path, the historical trajectories of the surrounding vehicles, and high-precision map information from the trajectory dataset through its internal data interface. It then performs the following core processing: transforming the location information to the Frenet coordinate system; encoding and fusing the transformed current state with the map information to obtain map-state fusion features; simultaneously encoding the historical trajectories of the surrounding vehicles to obtain historical trajectory features; subsequently, the unit calls its built-in priority judgment module to generate priority feature vectors for the surrounding vehicles, and based on the map-state fusion features, historical trajectory features, and priority feature vectors, completes trajectory interactive prediction through the historical prediction attention module, finally outputting the trajectory prediction for the surrounding vehicles at the current moment.

[0031] The control decision unit receives and uses the trajectory prediction results output by the trajectory prediction calculation unit to perform risk assessment and behavior decision-making, and finally generates vehicle control commands.

[0032] It should be understood that the above-mentioned units are functional sub-modules and can be deployed on the same or different computing devices, and this application does not impose any restrictions on this.

[0033] Figure 3 is a flowchart illustrating a dynamic adaptive trajectory prediction method according to some embodiments of this application. In some embodiments, the dynamic adaptive trajectory prediction method is executed by the trajectory prediction calculation unit shown in Figure 2. The method includes the following steps: S310, acquiring trajectory prediction data and preprocessing it. The trajectory prediction data includes the current state of the vehicle and surrounding vehicles, the planned path of the vehicle, the historical trajectory of the surrounding vehicles, and map information. The preprocessing includes: performing coordinate transformation on the location information; encoding and fusing the current state of the vehicle and surrounding vehicles corresponding to the coordinate-transformed location information with the map information to obtain map-state fusion features; and encoding the historical trajectory of the surrounding vehicles to obtain historical trajectory features.

[0034] This step is completed by the input layer shown in Figure 1, and its structure and flow are shown in Figure 4. Specifically, it includes: Data Acquisition: Load the following data from the trajectory dataset: Ego observation state: the state observed by the ego vehicle at the current moment, including the current state of the ego vehicle itself and the current state of the observed surrounding NPC vehicles.

[0035] Ego's path planning: Ego vehicles in The position after seconds and in this Within a second, the trajectory information of ego in the dataset is used to define the planned path of the ego vehicle at the current moment in this invention.

[0036] NPC Historical Tracks: All NPC Vehicles in the Past The trajectory information in seconds.

[0037] High-precision map information: Map information corresponding to the current trajectory data.

[0038] Preprocessing steps: Establishing a coordinate system and coordinate transformation First, establish a Frenet coordinate system, where the vertical axis is... The axis represents the longitudinal displacement of the vehicle, and the horizontal axis is... The axis represents the vehicle's offset relative to the road centerline. Then, the position information from the ego's observed state (including the ego's current position and the current positions of surrounding NPCs) and the position information of the ego's planned path are transformed into coordinates and mapped to this coordinate system for subsequent processing.

[0039] In one embodiment of this disclosure, the establishment of the coordinate system and the coordinate transformation specifically involve: firstly, extracting the vehicle state at the current time (denoted as k). (The position of the vehicle at time k) ,speed acceleration ), the planned position of the vehicle after t seconds and the current location of all NPC vehicles in the vicinity. Establish a Frenet coordinate system and transform the coordinates of all vehicles (including ego vehicles and NPC vehicles) at the current moment, projecting them into this coordinate system for subsequent priority judgment module.

[0040] Coding Fusion By using several different encoders and a strategy based on spatiotemporal cross-attention feature fusion, we can encode and fuse ego observation information, NPC historical trajectory information, and high-precision map information for subsequent trajectory prediction.

[0041] In one embodiment of this application, the encoding fusion specifically involves: using different encoders to encode the historical trajectory of the NPC vehicle, the current observation state of the ego vehicle, and the high-precision map, and then performing spatiotemporal cross-attention fusion of the observation encoding and the map encoding to obtain the fused features and high-dimensional historical trajectory vector features, which are used for subsequent decoding output.

[0042] S320, Generate priority feature vectors for environmental vehicles, wherein: the vehicle's motion trend is determined based on the vehicle's planned path; the static priority of the environmental vehicles is calculated based on the relative position of each environmental vehicle with the motion trend; the dynamic priority of the environmental vehicles is calculated based on the expected collision time between each environmental vehicle and the vehicle; and the static priority and dynamic priority are combined.

[0043] This step is completed by the priority determination module shown in Figure 1.

[0044] Specifically, the priority determination module determines the priority characteristics of each NPC vehicle. This rule-based module calculates a set of priority characteristics in real time based on the input information to guide the subsequent trajectory prediction process, thus achieving dynamic prediction. These priority characteristics consist of two parts: static priority... and dynamic priority Finally, by combining these two priority components, the priority characteristics for each NPC vehicle can be derived. Combined into a vector .

[0045] Static priority : Calculated based on the current position of the NPC vehicle relative to the ego vehicle.

[0046] The specific principle is shown in Figure 5. Specifically: based on the vehicle's current location... and position t seconds later To obtain the ego's movement trend (in and (These are unit vectors representing horizontal displacement and lateral deviation, respectively), and can be represented by two-dimensional points. express.

[0047] Meanwhile, the NPC vehicle's current position After coordinate transformation, it can also be represented by a set of two-dimensional points: The entire plane is divided into three parts—high priority, medium priority, and low priority—using dot product operations. The static priority of NPC vehicles in each of these three regions is then recorded. The values ​​are 1.0, 0.7, and 0.4.

[0048] Dynamic priority The TTC (Time to Collision) is calculated based on the threat level of the NPC vehicle relative to the ego vehicle at the current moment.

[0049] The specific principle is shown in Figure 6.

[0050] As shown in Figure 6, the potential threat posed by NPC vehicles in high-priority areas to ego vehicles may not be as great as that posed by vehicles in low-priority areas, and vehicles in the same priority area may also pose different threats to ego vehicles.

[0051] Therefore, after comprehensively considering factors such as relative distance and speed, TTC is adopted as the standard for measuring dynamic priority. , The dynamic priority of surrounding NPC vehicles is obtained by normalization.

[0052] Priority vector Finally, by combining static and dynamic priorities, the priority of surrounding NPC vehicles at the current moment is determined. And combine them into a feature vector Used to guide the generation of subsequent predicted trajectories of different lengths.

[0053] S330, trajectory prediction based on the map-state fusion feature, historical trajectory feature, and priority feature vector includes: concatenating the map-state fusion feature and the priority feature vector to obtain a priority-state fusion feature; generating a first trajectory prediction for the current moment based on the priority-state fusion feature and the historical trajectory feature; introducing the trajectory prediction result from the historical moment, and using an attention mechanism to enable the first trajectory prediction and the historical trajectory prediction result to interact, thereby outputting a second trajectory prediction for the current moment, wherein the historical trajectory prediction result is null, initial state, or default value during the first prediction, and the historical trajectory prediction result is generated by this method during subsequent predictions.

[0054] This step is completed by the historical prediction attention module shown in Figure 1. The core function of this module is to introduce historical prediction information to eliminate the isolation between prediction results at different times. It uses a novel attention mechanism (combining multi-head attention) to introduce the prediction results of previous predictions as context into the prediction at the current time, thereby ensuring that the output trajectory has temporal continuity.

[0055] The basic principle is shown in Figure 7. Existing trajectory prediction methods do not refer to the prediction results of the previous few seconds when making a prediction at the k-th second. However, it is easy to know that similar predictions should be extremely similar and have some kind of intrinsic correlation. Existing methods cannot reflect this correlation well. Therefore, I use a new attention mechanism to consider the predictions of the previous t seconds. Embedding is performed, and multi-head attention is used to improve the current prediction. It interacts with previous historical predictions, using information from those predictions to establish a temporal correlation between each prediction and historical predictions.

[0056] The network structure of the historical prediction attention module is shown in Figure 8. Encoded map information and observed dynamic information are fused through spatiotemporal cross-attention. The fused spatiotemporal information is then concatenated with priority features to obtain priority-included spatiotemporal information, which is then concatenated with historical trajectory features and input into the RNN decoder for decoding, generating a preliminary prediction. Finally, through a stacked historical prediction attention mechanism, the predictions from each step are embedded into the fused information for subsequent prediction generation.

[0057] The initialization process for trajectory prediction is detailed below through an embodiment: The trajectory dataset is cleaned, and time-aligned trajectory data from 0-5 seconds is loaded; the driver vehicle and NPC vehicles are selected, with the 5th second as the initial prediction time (simulation step size 10Hz). During the first prediction, the historical trajectory prediction results are set as the real trajectory data, specifically the actual trajectory information of the driver vehicle from 2-5 seconds; the NPC historical trajectory is set as the actual trajectory of the NPC vehicle from 2-5 seconds; and the driver vehicle's planned path is set as the actual path for the next 3 seconds. This ensures that the method operates based on real data from the initial stage. Subsequent predictions use the historical trajectory prediction results generated by this method. It is understood that the real trajectory data is a specific and preferred default value. In other embodiments, the default value can also be a zero tensor, a randomly initialized vector, or other methods well-known to those skilled in the art.

[0058] The historical prediction attention module can solve the problem of discontinuous predicted trajectories in existing methods, enhance the temporal consistency of predicted trajectories, and make full use of priority information to achieve adaptive trajectory prediction.

[0059] This application implements a novel dynamic adaptive trajectory prediction method through the collaborative interaction of a priority judgment module and a historical prediction attention module. It solves the problems of insufficient prediction targeting due to the lack of dynamic risk assessment, poor trajectory temporal consistency due to the neglect of historical prediction information, and weak adaptability of prediction strategies in existing works. In practical applications, it maintains high prediction accuracy and precision while exhibiting low computational requirements.

[0060] In summary, the dynamic adaptive trajectory prediction method provided in the embodiments of this application firstly, by introducing a priority feature vector based on the vehicle's motion trend and the expected collision time, and concatenating it with map-state fusion features, achieves a dynamic and refined assessment of potential vehicle risks in the environment, providing the prediction model with key information to distinguish the importance of vehicles; secondly, by constructing a historical prediction attention mechanism, the current prediction can interact with historical prediction results, effectively utilizing temporal context information and fundamentally ensuring the smoothness and inherent consistency of the generated trajectory; finally, the synergistic effect of the above technologies enables the entire system to more intelligently adapt to changes in complex dynamic environments, thereby significantly improving the accuracy and reliability of trajectory prediction overall.

[0061] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding descriptions in the foregoing device embodiments, and will not be repeated here.

[0062] Although the subject matter described herein is provided in the general context of execution on a computer system in conjunction with an operating system and applications, those skilled in the art will recognize that other implementations can also be executed in conjunction with other types of program modules. Generally, program modules include routines, programs, components, data structures, and other types of structures that perform specific tasks or implement specific abstract data types. Those skilled in the art will understand that the subject matter described herein can be practiced using other computer system configurations, including handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframes, etc., and can also be used in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules may reside on both local and remote memory storage devices.

[0063] Those skilled in the art will recognize that the units and method 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.

[0064] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A dynamic adaptive trajectory prediction method, characterized in that, include: The process involves acquiring and preprocessing trajectory prediction data. This data includes the current states of the vehicle and surrounding vehicles, the vehicle's planned path, historical trajectories of surrounding vehicles, and map information. The preprocessing includes: performing coordinate transformation on the location information; encoding and fusing the current states of the vehicle and surrounding vehicles corresponding to the transformed location information with the map information to obtain map-state fusion features; and encoding the historical trajectories of surrounding vehicles to obtain historical trajectory features. Finally, it generates priority feature vectors for surrounding vehicles, where: the vehicle's motion trend is determined based on its planned path; the static priority of each surrounding vehicle is calculated based on its relative position to the motion trend; and the dynamic priority of each surrounding vehicle is calculated based on the estimated collision time between each surrounding vehicle and the vehicle. The method combines static and dynamic priorities; it performs trajectory prediction based on the map-state fusion features, historical trajectory features, and priority feature vectors, including: concatenating the map-state fusion features and the priority feature vectors to obtain priority-state fusion features; generating a first trajectory prediction for the current moment based on the priority-state fusion features and the historical trajectory features; incorporating the trajectory prediction results from historical moments; and using an attention mechanism to enable the first trajectory prediction and the historical trajectory prediction results to interact, thereby outputting a second trajectory prediction for the current moment. In the first prediction, the historical trajectory prediction results are the initial state, default value, or actual trajectory data; in subsequent predictions, the historical trajectory prediction results are generated by this method.

2. The method according to claim 1, characterized in that, The coordinate transformation of the location information includes mapping the current position of the vehicle, the current position of the environmental vehicle, and the location information of the planned path of the vehicle to a Frenet coordinate system with the road centerline as the reference line. The s-axis of the Frenet coordinate system represents the longitudinal displacement along the road centerline, and the d-axis represents the lateral offset relative to the road centerline.

3. The method according to claim 2, characterized in that, The process of encoding and fusing the current state of the vehicle and surrounding vehicles corresponding to the coordinate-transformed position information with the map information to obtain map-state fusion features includes: encoding the current state and the map information respectively to obtain observation dynamic features and map encoding features, wherein the position information in the current state has undergone the coordinate transformation; and fusing the observation dynamic features and map encoding features through a spatiotemporal cross-attention mechanism to obtain the map-state fusion features.

4. The method according to claim 1, characterized in that, The step of calculating the static priority of the environmental vehicles based on their relative positions with the motion trend includes: defining three priority regions—high, medium, and low—in space based on the motion trend and the relative positional relationship between each environmental vehicle and the vehicle itself, using dot product operations; and assigning static priority values ​​of 1.0, 0.7, and 0.4 to environmental vehicles in the high, medium, and low priority regions, respectively.

5. The method according to claim 1, characterized in that, The calculation of the dynamic priority of the environmental vehicles based on the estimated collision time between each environmental vehicle and the vehicle itself includes: calculating the estimated collision time between each environmental vehicle and the vehicle itself, using the following formula: ,in, This represents the estimated collision time between the vehicle and the driver in the j-th environment at time k. and Let be the current positions of the j-th environmental vehicle and the autonomous vehicle at time k, respectively. and Let be the speeds of the j-th environmental vehicle and the driver vehicle at time k, respectively. The calculated estimated collision times for each environmental vehicle are normalized to obtain the dynamic priority, calculated using the following formula: ,in, This represents the dynamic priority of the j-th environmental vehicle at time k. Let represent the estimated collision time between the vehicle and the driver in the j-th environment at time k. Let k represent the set of TTC values ​​for all vehicles in the environment at time k. and Let $k$ and $k$ represent the minimum and maximum TTC values ​​for all environmental vehicles at time $k$, respectively.

6. The method according to claim 1, characterized in that, The step of combining the static priority with the dynamic priority includes: for each environmental vehicle, multiplying its static priority value by its dynamic priority value to obtain the priority value of the environmental vehicle; and combining the priority values ​​of all environmental vehicles in sequence to form the priority feature vector of the environmental vehicles.

7. As described in claim 1, characterized in that, The step of generating the first trajectory prediction for the current moment based on the priority-state fusion feature and the historical trajectory feature includes: concatenating the priority-state fusion feature and the historical trajectory feature, and inputting them into a recurrent neural network decoder for decoding to generate the first trajectory prediction.

8. The method according to claim 1, characterized in that, The introduction of trajectory prediction results generated by this method at historical moments includes: for the current moment k, selecting the environmental vehicle historical trajectory prediction results generated by this method for the previous t consecutive historical moments (kt, k-t+1, ..., k-1); performing feature extraction on the selected historical trajectory prediction result sequence to generate a historical trajectory embedding vector sequence; and using the historical trajectory embedding vector sequence as the historical trajectory prediction result for subsequent interaction with the first trajectory prediction at the current moment.

9. The method according to claim 8, characterized in that: The attention mechanism described is a multi-head attention mechanism.

10. The method according to claim 9, characterized in that, The interaction between the first trajectory prediction and the historical trajectory prediction results based on the attention mechanism includes: using the first trajectory prediction as a query vector, using the historical trajectory embedded vector sequence as a key vector and a value vector, and inputting them into the multi-head attention mechanism to calculate the association weights and perform the interaction.

Images