Trajectory prediction method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention discloses a trajectory prediction method and device, electronic equipment and a storage medium. The method comprises the steps that when a target vehicle is in a sparse interaction scene in an intelligent driving mode, a historical track set collected by the target vehicle is acquired; the historical track set is composed of historical motion tracks of traffic participants in a sparse interaction scene where the target vehicle is located; based on a preset sparse interaction feature extraction model and the historical trajectory set, determining sparse interaction features corresponding to the sparse interaction scene; and based on a preset initial noise trajectory, a preset time step, the sparse interaction features and a preset denoising diffusion probability model, determining a future motion trajectory corresponding to each traffic participant in the sparse interaction scene. Through the technical scheme of the embodiment of the invention, the prediction accuracy and efficiency of the future motion trail of each traffic participant in the sparse interaction scene of the intelligent driving vehicle are improved.

Description

A trajectory prediction method, apparatus, electronic device, and storage medium Technical Field

[0001] The present invention relates to the field of intelligent driving technology, and in particular to a trajectory prediction method, device, electronic device and storage medium. Background Technology

[0002] With the rapid development of technologies such as autonomous driving and intelligent robots, extremely high demands are placed on their ability to accurately predict the future trajectories of surrounding intelligent agents (such as pedestrians and vehicles) in dynamic environments. Accurate trajectory prediction is a crucial prerequisite for ensuring safe and efficient navigation decisions. Trajectory prediction refers to predicting the future trajectories of individuals based on their historical movement trajectories. In complex traffic environments, the factors influencing individual movement are diverse, and the movement process is abstract and difficult to model, posing challenges to trajectory prediction tasks. The problems and shortcomings of existing technologies are as follows:

[0003] Redundancy and inefficiency in interaction modeling: Existing interaction modules based on graph attention or Transformers typically operate under the assumption of dense connectivity. However, in sparse interaction scenarios, most interactions between agents are weak or meaningless. Existing methods cannot adaptively focus on the truly critical few interactions, leading to wasted computational resources and noise interference, making it difficult for models to capture interaction cues that have a decisive impact on predictions. Summary of the Invention

[0004] This invention provides a trajectory prediction method, device, electronic device, and storage medium to accurately and conveniently predict the future movement trajectory of each traffic participant in a sparse interaction scenario when an intelligent driving vehicle is in such a scenario, thereby improving the accuracy and efficiency of predicting the future movement trajectory of traffic participants.

[0005] In a first aspect, embodiments of the present invention provide a trajectory prediction method, comprising:

[0006] When the target vehicle is in a sparse interaction scenario in intelligent driving mode, the historical trajectory set collected by the target vehicle is obtained; the historical trajectory set is composed of the historical movement trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located.

[0007] Based on the preset sparse interaction feature extraction model and the historical trajectory set, the sparse interaction features corresponding to the sparse interaction scenario are determined.

[0008] Based on the preset initial noise trajectory, preset time step, the sparse interaction features, and preset denoising diffusion probability model, the future movement trajectory of each traffic participant in the sparse interaction scenario is determined.

[0009] Optionally, the preset sparse interaction feature extraction model includes: a location embedding module, a sparsification module, and a graph convolutional network; the method further includes: inputting the historical trajectory set into the location embedding module for vector transformation and location embedding to obtain a first trajectory feature; determining a sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario based on the first trajectory feature and the sparsification module; and performing graph convolution aggregation on the first trajectory feature based on the graph convolutional network and the sparse asymmetric interaction weight matrix to determine the sparse interaction feature corresponding to the sparse interaction scenario.

[0010] Optionally, the method further includes: inputting the first trajectory feature into the sparsification module, so that the interaction matrix corresponding to the first trajectory feature is determined in the sparsification module using an attention mechanism; determining the sparse interaction mask corresponding to the first trajectory feature in the sparsification module based on preset hyperparameters and preset activation functions; determining the sparse interaction matrix in the sparsification module based on the interaction matrix and the sparse interaction mask; and determining the sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario in the sparsification module based on a regularized adjacency matrix and a normalized sparse interaction matrix.

[0011] Optionally, the method further includes: performing graph convolution processing on the first trajectory features based on the graph convolutional network and the sparse asymmetric interaction weight matrix to obtain the second trajectory features; and determining the sparse interaction features corresponding to the sparse interaction scenario based on the first trajectory features and the second trajectory features.

[0012] Optionally, the method further includes: inputting a preset initial noise trajectory, a preset time step, and the sparse interaction features into a preset denoising diffusion probability model, so that the preset denoising diffusion probability model denoises the preset initial noise trajectory based on the sparse interaction features to obtain the processed current noise trajectory and the current time step; inputting the current noise trajectory, the current time step, and the sparse interaction features into the preset denoising diffusion probability model for iterative denoising to obtain the updated current noise trajectory and the current time step; when the current time step is updated to a preset step threshold, stopping the iterative denoising and obtaining the future movement trajectory corresponding to each traffic participant in the sparse interaction scenario.

[0013] Optionally, the method further includes: determining the future driving trajectory of the target vehicle based on the driving objective of the target vehicle, the traffic rules in the sparse interaction scenario, and the future motion trajectory of each traffic participant.

[0014] Secondly, embodiments of the present invention also provide a trajectory prediction device, the device comprising:

[0015] The historical trajectory set acquisition module is used to acquire the historical trajectory set collected by the target vehicle when the target vehicle is in a sparse interaction scenario in intelligent driving mode; the historical trajectory set is composed of the historical movement trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located.

[0016] The sparse interaction feature determination module is used to determine the sparse interaction features corresponding to the sparse interaction scenario based on the preset sparse interaction feature extraction model and the historical trajectory set.

[0017] The future motion trajectory determination module is used to determine the future motion trajectory of each traffic participant in the sparse interaction scenario based on a preset initial noise trajectory, a preset time step, the sparse interaction features, and a preset denoising diffusion probability model.

[0018] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0019] One or more processors;

[0020] Memory, used to store one or more programs;

[0021] When the one or more programs are executed by the one or more processors, the one or more processors implement the trajectory prediction method as provided in any embodiment of the present invention.

[0022] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the trajectory prediction method as provided in any embodiment of the present invention.

[0023] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the trajectory prediction method provided in any embodiment of the present invention.

[0024] The technical solution of this invention involves acquiring a set of historical trajectories of a target vehicle when it is in a sparse interaction scenario in intelligent driving mode. This set of historical trajectories consists of the historical movement trajectories of each traffic participant in the sparse interaction scenario. Based on a preset sparse interaction feature extraction model and the set of historical trajectories, sparse interaction features corresponding to the sparse interaction scenario are determined. This sparse perception interaction extraction mechanism adaptively identifies and focuses on key social interactions, effectively filtering unnecessary weakly related interactions, thereby significantly improving computational efficiency while maintaining prediction accuracy. Based on a preset initial noise trajectory, a preset time step, the sparse interaction features, and a preset denoising diffusion probability model, the future movement trajectory of each traffic participant in the sparse interaction scenario is determined. Through sparsification processing and structural optimization, while maintaining high prediction accuracy, computational resource consumption is significantly reduced. This allows the trajectory prediction model (i.e., the preset denoising diffusion probability model) to better meet the real-time requirements of practical applications and provide more accurate and reliable trajectory prediction results for the intelligent driving system, significantly improving the decision-making safety of the target vehicle in complex dynamic environments.

[0025] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 is a flowchart of a trajectory prediction method provided in Embodiment 1 of the present invention;

[0028] Figure 2 is a flowchart of a trajectory prediction method provided in Embodiment 2 of the present invention;

[0029] Figure 3 is a schematic diagram of a trajectory prediction device provided in Embodiment 3 of the present invention;

[0030] Figure 4 is a schematic diagram of the structure of an electronic device that implements the trajectory prediction method of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] Example 1

[0034] Figure 1 is a flowchart of a trajectory prediction method provided in Embodiment 1 of the present invention. This embodiment is applicable to common sparse interaction scenarios in the real world, such as open intersections and low-density crowd environments, and provides more accurate and reliable trajectory prediction results for intelligent driving systems. This method can be executed by a trajectory prediction device, which can be implemented in hardware and / or software and can be configured in an electronic device. As shown in Figure 1, the method includes:

[0035] S110. When the target vehicle is in a sparse interaction scenario in intelligent driving mode, acquire the historical trajectory set collected by the target vehicle; the historical trajectory set is composed of the historical movement trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located.

[0036] In this embodiment of the disclosure, the target vehicle can refer to a vehicle using driver assistance functions. Intelligent driving mode can refer to a driving mode with driver assistance functions. Sparse interaction scenarios can refer to scenarios where traffic participants (such as pedestrians, vehicles, etc.) are relatively far apart, have obvious and identifiable interaction relationships, but the connections are not densely complex. Compared to dense interaction scenarios (such as crowded sidewalks or congested traffic), interactions in sparse interaction scenarios are more decisive and critical, with less noise (meaningless interactions). For example, sparse interaction scenarios can be, but are not limited to, the following four types. In open city intersections: pedestrians stand on the roadside observing vehicles approaching from a distance. The pedestrian's decision (whether to cross the road) is primarily influenced by the speed and intentions of the nearest one or two vehicles, rather than all vehicles within a hundred meters. In parking lot lanes: a car is preparing to reverse into a parking space, and a pedestrian walks towards their car from a distance. Their paths potentially intersect, but there are no other people or vehicles nearby. In residential areas: a child is playing on the roadside, and a delivery truck slowly enters. Whether the child runs into the middle of the road is directly related to the distance and speed of the delivery truck. In low-density plazas: several people are scattered towards different exits, and their paths may involve slight mutual avoidance along the way. Historical movement trajectories can refer to the movement trajectories of traffic participants within a preset time period before the current moment (e.g., the previous 3-5 seconds). Historical movement trajectories can be discrete, represented by a sequence.

[0037] Specifically, the target vehicle's sensors (such as cameras and LiDAR) collect real-time location information of traffic participants around the vehicle. This real-time location information is used to determine the density of traffic participants around the target vehicle. If the density is below a preset density threshold, the target vehicle is considered to be in a sparse interaction scenario, and the system checks whether the vehicle is in intelligent driving mode. If so, the vehicle is determined to be in a sparse interaction scenario under intelligent driving mode. At this point, the historical movement trajectories of each traffic participant collected within the previous 5 seconds are acquired and compiled into a historical trajectory set.

[0038] S120. Based on the preset sparse interaction feature extraction model and historical trajectory set, determine the sparse interaction features corresponding to the sparse interaction scenario.

[0039] In this embodiment, the pre-set sparse interaction feature extraction model can refer to a pre-trained neural network encoder specifically designed to adaptively extract key social interaction relationships from historical trajectories of traffic scenarios. In this invention, it can specifically refer to a sparsely-aware interaction extraction network (SSA module), whose core structure includes a learnable sparse mask generator, a Zero-Softmax normalization layer, and a multi-head sparse graph convolutional network. The pre-set sparse interaction feature extraction model can be used to analyze the historical motion trajectories of multiple traffic participants to understand their motion relationships. It can also adaptively filter redundant weak interactions, identifying and retaining only a few key interaction connections (such as avoidance, following, and merging relationships) that have a decisive impact on prediction. Furthermore, the pre-set sparse interaction feature extraction model can encode the selected key interactions into a structured, compact mathematical representation (i.e., sparse interaction features), providing social rule constraints for subsequent trajectory generation.

[0040] In this embodiment, sparse interaction features refer to an asymmetric, graph-structured feature vector representation output by a pre-defined sparse interaction feature extraction model. Essentially, it is a mathematical encoding of key social dynamics in traffic scenarios. Sparse interaction features possess sparsity, criticality, structure, and conditionality. Sparsity means that sparse interaction features contain only a small number of interactions crucial to prediction, with the vast majority of elements in the matrix representation being zero. Criticality means that sparse interaction features adaptively focus on interactions with high predictive value (such as conflict points or cooperative behaviors), ignoring irrelevant or weak interactions. Structure means that sparse interaction features retain the graph structure and asymmetric direction of interactions (e.g., the influence strength of A on B is not equal to the influence strength of B on A). Conditionality means that sparse interaction features act as strong conditional signals, guiding subsequent trajectory generation models to generate trajectories that conform to the social rules of the current specific scenario.

[0041] Specifically, the historical trajectory set is input into a pre-defined sparse interaction feature extraction model. In this model, the historical location sequences (i.e., the historical trajectory set) of all traffic participants are transformed into initial temporal feature representations through linear transformation and location information embedding. The model then analyzes the potential relationships between participants using an attention mechanism and adaptively filters out a few truly important interactions using a learnable threshold mechanism, while masking a large number of weak or irrelevant connections, forming a sparse interaction relationship graph. Based on the sparse interaction relationship graph constructed in the previous step, a graph convolutional network is used to perform multiple rounds of information transmission and aggregation on the features of each participant. In each aggregation round, each participant only receives information from other participants with whom they have key interactions, thereby extracting high-level features rich in social context. Finally, the high-level features obtained after multiple aggregation rounds are fused with the initial features, ultimately outputting a compact vector that can represent the decisive social interaction rules in the current scenario—the sparse interaction feature.

[0042] S130. Based on the preset initial noise trajectory, preset time step, sparse interaction features and preset denoising diffusion probability model, determine the future movement trajectory of each traffic participant in the sparse interaction scenario.

[0043] In this embodiment, the preset initial noise trajectory can refer to a set of noise data randomly sampled from a standard multivariate Gaussian distribution during the model inference (prediction) phase. Its dimension is the same as the dimension of the future trajectory to be generated (e.g., [number of agents, future time steps, coordinate dimension]). The preset initial noise trajectory can represent the starting point of the generation process. The preset initial noise trajectory can be a source of diversity in the final predicted trajectory. Different random sampling will produce different initial noise, thereby generating multiple reasonable trajectories under the same social constraints (i.e., sparse interaction features).

[0044] In this embodiment, the preset time step can refer to a predefined, decreasing sequence of integers (e.g., K, K-1, ..., 1). The preset time step can be used to control the progress and granularity of the denoising diffusion generation process. Each time step corresponds to a specific stage of the denoising process, thus allowing the construction of an index using the time steps and the corresponding denoising process. The value of the time step is associated with the noise content in the current data (e.g., the larger the time step value, the more noise). The time step can be used to indicate to the denoising network how much denoising operation should be performed in the current step (e.g., early coarse-grained denoising, later fine-grained optimization).

[0045] In this embodiment, the preset denoising diffusion probability model can refer to a pre-trained generative neural network (such as LSDNet) conditioned on sparse interaction features. It can be built based on the Denoising Diffusion Probability Model (DDPM) framework. The preset denoising diffusion probability model can be used to perform an iterative denoising process from noise to a clear trajectory. In each denoising step, the current noisy trajectory, the current time step, and the sparse interaction features are received as inputs. Based on the social rules encoded by the sparse interaction features, unreasonable parts of the current noisy trajectory are predicted and removed, gradually shaping it into a future motion trajectory that conforms to social norms (i.e., sparse interaction features). Ultimately, this achieves the generation of diverse and reasonable future motion trajectories from random noise under strong social constraints.

[0046] In this embodiment of the disclosure, the future motion trajectory can refer to a series of predicted positions for each traffic participant over a future period of time. The future motion trajectory can include the two-dimensional or three-dimensional coordinates of the traffic participant at each predicted time point. Due to the randomness of the initial noise (i.e., the preset initial noise trajectory), the model can generate multiple possible future motion trajectories for each participant. Each generated trajectory is strongly constrained by sparse interaction features, conforming to physical laws (such as continuity) and social rules (such as avoiding collisions and maintaining polite distance).

[0047] Specifically, a preset initial noise trajectory, a preset time step, and sparse interaction features are input into a preset denoising and diffusion probability model. This model iteratively denoises the initial noise trajectory based on the sparse interaction features and then outputs the model's output to obtain the future movement trajectory of the traffic participant. The preset time step, sparse interaction features, and different preset initial noise trajectories can be further input into the preset denoising and diffusion probability model to obtain various future movement trajectories of the traffic participant based on the model's output.

[0048] As an optional implementation of this disclosure, determining the future movement trajectory of each traffic participant in a sparse interaction scenario based on a preset initial noise trajectory, a preset time step, sparse interaction features, and a preset denoising diffusion probability model may specifically include: inputting the preset initial noise trajectory, the preset time step, and the sparse interaction features into the preset denoising diffusion probability model, so that the preset denoising diffusion probability model denoises the preset initial noise trajectory according to the sparse interaction features to obtain the processed current noise trajectory and the current time step; inputting the current noise trajectory, the current time step, and the sparse interaction features into the preset denoising diffusion probability model for iterative denoising to obtain the updated current noise trajectory and the current time step; when the current time step is updated to a preset step threshold, stopping the iterative denoising and obtaining the future movement trajectory of each traffic participant in the sparse interaction scenario.

[0049] In this embodiment of the disclosure, the current noise trajectory can refer to the trajectory data state that is not yet fully clear before the start of each round (step) of denoising operation in the iterative generation process of the denoising diffusion model. The current noise trajectory can be the output result of the previous round of denoising operation and the input of the current round of denoising operation. As the iteration proceeds, the noise in the current noise trajectory is gradually removed, and its state gradually changes from pure random noise (initial state) to a clear and reasonable future motion trajectory (final state).

[0050] In this embodiment, the current time step refers to the index value corresponding to the currently executing denoising operation during the iterative generation process of the denoising diffusion model. The current time step can be an integer starting from a preset maximum value (such as K) and decreasing by 1 with each iteration until it reaches 0. The current time step indicates the stage of the generation process: the larger the value, the more noise there is, and the more the denoising operation focuses on correcting the macro trend; the smaller the value, the less noise there is, and the more the denoising operation focuses on fine-tuning and optimization of details.

[0051] In this embodiment, the preset step threshold can refer to the remaining number of iterations for the denoising diffusion model during the inference (prediction) phase. For example, iteration stops when the current time step decreases from K to 0. Alternatively, iteration can stop when the current time step reaches a preset step threshold (e.g., 3). More iterations result in a more refined denoising process and a smoother, more reasonable generated trajectory, but also a longer computation time.

[0052] Specifically, samples are first obtained from a standard Gaussian distribution. Then, LSTM (Long Short-Term Time Series Network) is used to predict the noise level. ,use and Relationship calculation The value of k is obtained by iterating step by step until k=0. The formulas involved in this step are as follows: ; ; ; .

[0053] in, This refers to the preset initial noise trajectory; Both refer to the current noise trajectory; It refers to the future trajectory of motion; and Predefined noise scheduling parameters related to time step k; It refers to the signal retention coefficient at step k; it is usually a value between 0 and 1, which determines the proportion of the original signal retained in the noise data of the current step. This refers to the noise variance added to the data at the k-th step; it is also a value between 0 and 1; it typically increases with the number of steps k. Z refers to sparse interaction features; Z refers to noise independent of model prediction. The model uses a random noise vector resampled from a standard normal distribution; Z is the newly sampled random noise, independent of model prediction, introducing additional randomness to each denoising step, ensuring that even starting from the same initial noise, the generated trajectory will differ depending on different samples of Z. Based on this, it eliminates the need for accurately predicting multiple different initial noise trajectories. By introducing scene-constrained noise priors and a hierarchical denoising mechanism, the generated trajectory can cover multiple possible future motion patterns while strictly adhering to environmental constraints and social rules, significantly reducing the probability of generating unreasonable trajectories.

[0054] It should be noted that, This is the algorithm implementation or forward propagation function of the pre-set sparse interaction feature extraction model in this embodiment. This can be understood as the historical trajectory set X being input into a pre-defined sparse interaction feature extraction model, running all the internal computational steps of the model, and finally outputting sparse interaction features. .

[0055] The technical solution of this invention involves acquiring a set of historical trajectories of the target vehicle when it is in a sparse interaction scenario in intelligent driving mode. This set of historical trajectories consists of the historical motion trajectories of each traffic participant in the sparse interaction scenario. Based on a preset sparse interaction feature extraction model and the set of historical trajectories, the sparse interaction features corresponding to the sparse interaction scenario are determined. This sparse perception interaction extraction mechanism adaptively identifies and focuses on key social interactions, effectively filtering unnecessary weakly related interactions, thereby significantly improving computational efficiency while maintaining prediction accuracy. Based on a preset initial noise trajectory, a preset time step, sparse interaction features, and a preset denoising diffusion probability model, the future motion trajectory of each traffic participant in the sparse interaction scenario is determined. Through sparsification processing and structural optimization, while maintaining high prediction accuracy, computational resource consumption is significantly reduced. This allows the trajectory prediction model (i.e., the preset denoising diffusion probability model) to better meet the real-time requirements of practical applications and provide more accurate and reliable trajectory prediction results for the intelligent driving system, significantly improving the decision-making safety of the target vehicle in complex dynamic environments.

[0056] As an optional implementation of this disclosure, the method further includes: determining the future driving trajectory of the target vehicle based on the driving objective of the target vehicle, the traffic rules in the sparse interaction scenario, and the future movement trajectory of each traffic participant.

[0057] In this embodiment, the driving objective can refer to the final or intermediate goal that the target vehicle expects to achieve in the current driving task, which is related to the path and state. The driving objective can integrate navigation instructions, driving tasks, and behavioral preferences. Driving objectives can include, but are not limited to, macro-path objectives, micro-behavioral objectives, and optimization preferences. For example, a macro-path objective could be to turn left at the next intersection, arrive at parking lot B, or drive straight for 3 kilometers along the current road. A micro-behavioral objective could be to complete the current lane change, overtake a slower vehicle ahead, or follow the vehicle in the current lane. Optimization preferences could be to arrive in the shortest time, pursue the smoothest ride experience, or achieve the most energy-efficient driving. Driving objectives are typically generated and issued by the vehicle's navigation system or a higher-level task planning module based on global route planning.

[0058] In this embodiment, traffic rules refer to a set of legal norms, traffic signs, and social conventions that all traffic participants in a public road environment must abide by. They set safe and legal boundaries for vehicle behavior. Traffic rules may include, but are not limited to, mandatory rules, structured environmental constraints, and implicit social norms. Mandatory rules include, but are not limited to, obeying traffic lights, stop signs, lane markings, speed limits, right-of-way regulations (such as yielding rules), and signs prohibiting turning or U-turns. Structured environmental constraints may include, but are not limited to, requiring vehicles to stay within road boundaries and not to drive on sidewalks or cross solid lines. Implicit social norms may refer to non-legal but widely accepted civilized driving habits, such as exchanging glances with pedestrians at intersections without traffic lights to yield, or not dangerously following other vehicles. Traffic rules can be used as both hard constraints and soft optimization targets when planning the trajectory of a target vehicle, ensuring the compliance and social acceptability of the generated trajectory.

[0059] In this embodiment, the future driving trajectory can refer to a specific, continuous, and executable sequence of states planned by the target vehicle based on a comprehensive understanding of the environment, to be executed within a certain period of time (e.g., 5-8 seconds in the future). The future driving trajectory may include, but is not limited to, spatiotemporal paths and motion states. The spatiotemporal path can refer to a series of future position points (x, y coordinates) arranged in chronological order, continuous in time and smooth in space; the motion state can refer to the state information at each future time point, usually accompanied by velocity, acceleration, heading angle, etc., constituting a complete motion plan. The future driving trajectory needs to possess safety, feasibility, compliance, and goal orientation. Safety can mean avoiding collisions with the predicted future motion trajectories of all other traffic participants, with sufficient safety margin; feasibility can mean that the trajectory must conform to the vehicle's dynamics (e.g., turning radius, acceleration limits) and the physical constraints of actuators (steering wheel, accelerator, brakes). Compliance can mean strictly adhering to the aforementioned traffic rules. Goal orientation can mean striving to achieve the preset driving goal efficiently and comfortably.

[0060] Specifically, based on the future trajectory of each traffic participant, a spatiotemporal decision-making environment is constructed, comprising the following elements: dynamic obstacle spatiotemporal occupancy zone, static constraint mapping, and target state definition. The dynamic obstacle spatiotemporal occupancy zone can refer to expanding each predicted trajectory of each participant into its spatiotemporal occupancy volume (considering the physical dimensions of vehicles / pedestrians and extending it in the time dimension). The static constraint mapping can refer to transforming traffic rules into spatiotemporal constraints, including: lane boundaries, curbs, medians, and other inviolable areas; traffic signal timing (red light prohibition periods, green light passage windows); and speed limit zones, stop lines, and yield sign locations. The target state definition can refer to defining the target state area that the target vehicle should reach at the end of the planned time period (such as the target lane center, target speed range, and position after turning at a specific intersection) based on the driving objective.

[0061] Specifically, based on the current state of the target vehicle and the driving objective, a set of candidate future driving trajectories is generated using parametric methods. Key sampling parameters can be discretized in the Frenet coordinate system (road coordinate system) or state space, including lateral offset and longitudinal velocity curves. Trajectory generation methods typically use polynomial (e.g., fifth-order polynomial) interpolation or spline curves to generate smooth candidate trajectories that conform to the basic constraints of vehicle kinematics. Covering the behavior space ensures that the candidate trajectories cover different driving behaviors, such as: maintaining the current lane and a safe distance from the vehicle in front when following another vehicle; changing lanes with a safe gap when overtaking; yielding to pedestrians or other vehicles with right-of-way; and emergency obstacle avoidance, which is a collision avoidance trajectory in extreme situations.

[0062] Specifically, each candidate trajectory undergoes a multi-dimensional evaluation, calculating a comprehensive cost function. This comprehensive cost function assesses the trajectory's safety costs, goal achievement costs, comfort costs, and social compliance costs. Safety costs (highest priority) include collision risk and rule violation penalties. Collision risk can be calculated by taking the minimum spatiotemporal distance between the candidate trajectory and all other participants' predicted trajectories, translating this into a collision probability. Rule violation penalties can involve checking for violations of traffic rules (such as crossing lines, running red lights, speeding) and imposing high penalties. Goal achievement costs include endpoint deviation and progress efficiency. Endpoint deviation measures the deviation between the trajectory's endpoint and the driving target's set endpoint. Progress efficiency assesses the time required to travel the same distance or the deviation from the desired speed. Comfort costs include jerk and lateral acceleration. Jerk measures the trajectory's acceleration (rate of change of acceleration); rapid acceleration and sharp turns result in high costs. Lateral acceleration assesses the centrifugal force during cornering, ensuring it remains within comfort thresholds. Social compliance costs include infringing on others' comfort zones and traffic etiquette. Violating another's comfort zone can refer to situations where, even without a collision, the trajectory forces other participants to brake suddenly or swerve sharply, incurring consequences. Traffic etiquette can refer to assessing adherence to non-mandatory but generally accepted etiquette (such as alternating passage when merging at ramps).

[0063] Specifically, optimal trajectory selection can be achieved by choosing the candidate trajectory with the lowest cost based on the overall cost, and using it as the base trajectory. Local fine-tuning can refer to using the selected trajectory as the initial solution and fine-tuning it using numerical optimization methods (such as Model Predictive Control, MPC) to solve for the optimal control sequence under more stringent dynamic models and constraints, generating the final smooth trajectory. Trajectory output can refer to outputting a dense sequence of trajectory points containing information such as timestamps, position coordinates, velocity, acceleration, and heading angle, as the future driving trajectory of the target vehicle, which is then delivered to the underlying control system for execution.

[0064] Example 2

[0065] Figure 2 is a flowchart of a trajectory prediction method provided in Embodiment 2 of the present invention. Based on the above embodiments, this embodiment describes in detail the process of determining the sparse interaction features corresponding to a sparse interaction scenario based on a preset sparse interaction feature extraction model and a historical trajectory set. Explanations of terms that are the same as or corresponding to those in the above embodiments are not repeated here. As shown in Figure 2, the method includes:

[0066] S210. When the target vehicle is in a sparse interaction scenario in intelligent driving mode, acquire the historical trajectory set collected by the target vehicle; the historical trajectory set is composed of the historical movement trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located.

[0067] S220. Input the historical trajectory set into the location embedding module for vector transformation and location embedding to obtain the first trajectory feature.

[0068] The preset sparse interaction feature extraction model includes: a location embedding module, a sparsification module, and a graph convolutional network.

[0069] In this embodiment, the first trajectory feature can refer to a high-dimensional feature representation that incorporates spatiotemporal information. The first trajectory feature can be the result of a historical trajectory after preliminary encoding and temporal information injection, serving as the input basis for subsequent sparse interaction extraction. The position embedding module can refer to a data processing sub-network in a pre-defined sparse interaction feature extraction model, typically consisting of a linear transformation layer and a positional encoding (PE) addition operation. The position embedding module can be used to map the input original trajectory coordinates (low-dimensional) to a higher-dimensional, more semantically rich feature space through a linear layer, and add a unique temporal encoding to the trajectory points at each time step, enabling the model to distinguish the positions at different times and understand the sequence of motion.

[0070] Specifically, given a set of input trajectories The trajectory data is transformed into learnable feature vectors through a linear layer. Then, position embedding (PE) is added to inject the positional information of the trajectory, resulting in the first trajectory feature vector. ,in, .

[0071] S230. Based on the first trajectory features and the sparsification module, determine the sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario.

[0072] In this embodiment, the sparsity module can refer to a core subnetwork responsible for identifying and filtering key connections from dense potential interactions. The sparsity module implements the sparsity of interaction relationships. It can receive first trajectory features, calculate initial interaction strengths through an attention mechanism, and generate a binary mask using a learnable thresholding mechanism to filter out weak or irrelevant interactions, retaining only the few connections crucial for prediction. The sparse asymmetric interaction weight matrix can refer to a graph-structured weight matrix. It defines the filtered, weighted, directional influence relationships between various traffic participants in a sparse interaction scenario. Matrix elements This indicates the strength of participant i's influence on participant j, and usually... (Asymmetric), the majority of elements in the matrix are zero (sparse).

[0073] Specifically, the first trajectory features are input into the sparsification module so that the key interaction relationships can be dynamically learned and selected from the first trajectory features through the sparsification module, forming a sparse asymmetric interaction weight matrix.

[0074] As an optional implementation of this disclosure, determining the sparse asymmetric interaction weight matrix between various traffic participants in a sparse interaction scenario based on the first trajectory features and the sparsification module may specifically include: inputting the first trajectory features into the sparsification module to determine the interaction matrix corresponding to the first trajectory features using an attention mechanism in the sparsification module; determining the sparse interaction mask corresponding to the first trajectory features in the sparsification module based on preset hyperparameters and preset activation functions; determining the sparse interaction matrix in the sparsification module based on the interaction matrix and the sparse interaction mask; and determining the sparse asymmetric interaction weight matrix between various traffic participants in the sparse interaction scenario based on a regularized adjacency matrix and a normalized sparse interaction matrix in the sparsification module.

[0075] In this embodiment, the interaction matrix can refer to a matrix initially calculated within the sparsity module using an attention mechanism (such as self-attention), representing the potential interaction strength between any two traffic participants. This matrix is ​​typically dense initially, meaning most elements are non-zero. Preset hyperparameters can refer to parameters manually set before model training and remaining unchanged during training, such as thresholds. This threshold is used to determine which interactions in the attention score are considered key interactions and retained (i.e., ...). Which are considered weak interactions and therefore blocked? The preset activation function can refer to a fixed function in a neural network used to introduce nonlinear transformations, such as the sigmoid activation function, which maps the attention score to the (0,1) interval for easy comparison with a threshold. The sparse interaction mask can refer to a binary matrix (elements are either 0 or 1), which is generated by the sparsification module based on preset hyperparameters (threshold). The interaction matrix is ​​generated by the activation function output. A value of 1 indicates that the corresponding interaction is considered critical and retained, while a value of 0 indicates that it is considered weak and masked. The sparse interaction matrix can be generated by element-wise multiplication of the dense interaction matrix R, the binary sparse interaction mask M, and the identity matrix I (e.g., ...). The resulting matrix Rs inherits the strength information of the interaction matrix, but its structure is controlled by a mask to become sparse. The adjacency matrix can be a predefined binary matrix based on the spatial proximity relationships (such as distance thresholds) of traffic participants in the scene. Elements =1 indicates that participants i and j are spatially adjacent and may interact; =0 indicates that they are not adjacent. It provides the physical basis for the interaction.

[0076] Specifically, the input feature vector is processed by an attention mechanism (multi-head) to learn an interaction matrix for each traffic participant relative to other traffic participants. The self-attention map has multiple heads M, representing the interaction between traffic participants, which may be influenced by various factors such as distance, speed, and direction. The process of determining the interaction matrix is ​​as follows: .

[0077] in, = Using hyperparameters exist The above generates a sparse interactive mask using element-level thresholding. ,Right now .

[0078] in, This is an index function; if the corresponding inequality is true, the output is 1, otherwise it is 0. It is a sigmoid activation function. To ensure that traffic participant nodes are self-connected, an identity matrix is ​​added to the interaction mask. Then, it is combined with the spacetime-dense interaction matrix through element-wise multiplication. By merging, a sparse interaction matrix is ​​obtained. ,Right now .

[0079] Specifically, matrix normalization is crucial for GCNs (Graph Convolutional Networks). Since the Softmax function outputs non-zero values, directly applying it for matrix normalization will cause the sparse matrix to revert to a dense matrix. Therefore, this embodiment of the invention employs a sparse algorithm to normalize the matrix. Normalization, i.e. .

[0080] in, express Total number of samples It is a negligible small constant to ensure numerical stability. Further, we can obtain: Sparse interactions Combined with the regularized adjacency matrix, a sparse asymmetric interaction weight matrix among traffic participants is obtained. ,Right now .

[0081] in, This represents the dot product between vectors, where m represents the current attention head. Sparse asymmetric interaction weight matrix It simultaneously encodes the spatial structure of the scene (such as who is near whom) and the semantic strength of the interaction (such as who has a greater influence on whom).

[0082] It should be noted that regularizing the adjacency matrix A yields a regularized adjacency matrix. ,Right now , where A is the adjacency matrix and I is the identity matrix.

[0083] It's important to note that the adjacency matrix provides hard constraints based on spatial distance, ensuring interactions occur within a reasonable physical range. The sparse asymmetric interaction weight matrix, learned through training, captures complex social semantics such as directionality and intentionality (e.g., the influence weight of a vehicle on a pedestrian is greater than that of the opposite direction). Combining the regularized adjacency matrix with the learned sparse interaction matrix further activates socially important connections on the physically possible interaction network. This results in an interaction graph that is both physically consistent and rich in social intelligence.

[0084] S240. Based on the graph convolutional network and the sparse asymmetric interaction weight matrix, the first trajectory features are aggregated by graph convolution to determine the sparse interaction features corresponding to the sparse interaction scenario.

[0085] In this embodiment of the disclosure, a graph convolutional network can refer to a neural network specifically designed for processing graph-structured data. A graph convolutional network is a graph neural network that uses a sparse asymmetric interaction weight matrix as the adjacency relation and first trajectory features as node features. Graph convolutional networks can be used to perform neighborhood information aggregation. Each participant (node)'s features are aggregated with information from its key neighbors according to the weights defined by the sparse interaction weight matrix, thereby updating its own features and making the final features contain rich social context.

[0086] Specifically, the sparse asymmetric interaction weight matrix and the first trajectory features are input into the graph convolutional network so that the graph convolutional network can exchange information among various traffic participants based on the sparse asymmetric interaction weight matrix, so as to extract features containing complex social relationships, namely sparse interaction features.

[0087] As an optional implementation of this disclosure, the method of performing graph convolution aggregation on the first trajectory features based on a graph convolutional network and a sparse asymmetric interaction weight matrix to determine the sparse interaction features corresponding to the sparse interaction scenario may specifically include: performing graph convolution processing on the first trajectory features based on a graph convolutional network and a sparse asymmetric interaction weight matrix to obtain the second trajectory features; and determining the sparse interaction features corresponding to the sparse interaction scenario based on the first trajectory features and the second trajectory features.

[0088] In this embodiment, the second trajectory feature can refer to a higher-level, more abstract feature representation obtained after the first trajectory feature has undergone multi-layer information propagation and aggregation through a graph convolutional network. It encodes the state of each traffic participant after considering the influence of sparse social interactions.

[0089] Specifically, the sparse asymmetric interaction weight matrix and the first trajectory features are input into the graph convolutional network. The graph convolutional network utilizes... As domain information for GC, the overall update rule for the joint representation is: .

[0090] in, It is the trajectory feature information input to the embedding layer. Indicates the first layer The network parameters for the head attention mechanism are used. Finally, the sparse interaction features corresponding to the sparse interaction scenario are obtained by performing a dot product between the first trajectory features output from the linear layer and the second trajectory features output from the graph convolutional network. .

[0091] S250: Based on the preset initial noise trajectory, preset time step, sparse interaction features and preset denoising diffusion probability model, determine the future motion trajectory of each traffic participant in the sparse interaction scenario.

[0092] The technical solution of this invention involves inputting a historical trajectory set into a location embedding module for vector transformation and location embedding to obtain a first trajectory feature. Based on the first trajectory feature and a sparsification module, a sparse asymmetric interaction weight matrix is ​​determined between various traffic participants in a sparse interaction scenario. Through a binary mask adaptively generated by the sparsification module, the model only performs calculations on key interaction connections, thus reducing the computational complexity of graph convolution operations from low to high due to the highly sparse matrix. sudden drop (where k is the average number of key connections), this approach is particularly suitable for handling scenarios with many agents but few interactions, such as open intersections, meeting the real-time requirements of autonomous driving and further improving trajectory prediction efficiency. Based on a graph convolutional network and a sparse asymmetric interaction weight matrix, the first trajectory features are aggregated using graph convolution to determine the sparse interaction features corresponding to the sparse interaction scenario. The sparse interaction features are not black-box vectors. They originate from an interpretable process: historical trajectory, key interaction filtering, and social information aggregation. This makes the sparse interaction features essentially a specification of the social rules of the current scenario, clarifying who influences whom and how, further improving the interpretability of the sparse interaction features and the predicted trajectory.

[0093] For example, the training process of the preset denoising diffusion probability model is as follows. First, from the sample dataset... Sampling Then, sampling is performed from the standard normal distribution. The future trajectory of the sample is determined according to a predefined schedule. By adding noise perturbation, we obtain the noisy future trajectory at any diffusion step k. Finally, the encoder's output features are used. (i.e., sparse interaction features), noisy future trajectories The method uses a given k value (i.e., a preset time step) to predict the magnitude of the noise added in that step. The mean square error between the predicted and actual noise values ​​is then used as the loss function and backpropagated. This process is repeated until convergence. This embodiment of the invention accelerates the model's convergence process and enhances training stability by optimizing the initial conditions and network structure of the denoising process, enabling the model to achieve ideal predictive performance within a shorter training period.

[0094] The following are embodiments of the trajectory prediction device provided in this invention. This device and the trajectory prediction methods in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the trajectory prediction device, please refer to the embodiments of the above trajectory prediction methods.

[0095] Example 3

[0096] Figure 3 is a schematic diagram of a trajectory prediction device provided in Embodiment 3 of the present invention. As shown in Figure 3, the device includes: a historical trajectory set acquisition module 310, a sparse interaction feature determination module 320, and a future motion trajectory determination module 330.

[0097] The historical trajectory set acquisition module 310 is used to acquire the historical trajectory set collected by the target vehicle when the target vehicle is in a sparse interaction scenario in intelligent driving mode; the historical trajectory set is composed of the historical motion trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located; the sparse interaction feature determination module 320 is used to determine the sparse interaction features corresponding to the sparse interaction scenario based on a preset sparse interaction feature extraction model and the historical trajectory set; the future motion trajectory determination module 330 is used to determine the future motion trajectory corresponding to each traffic participant in the sparse interaction scenario based on a preset initial noise trajectory, a preset time step, sparse interaction features and a preset denoising diffusion probability model.

[0098] The technical solution of this invention involves acquiring a set of historical trajectories of the target vehicle when it is in a sparse interaction scenario in intelligent driving mode. This set of historical trajectories consists of the historical motion trajectories of each traffic participant in the sparse interaction scenario. Based on a preset sparse interaction feature extraction model and the set of historical trajectories, the sparse interaction features corresponding to the sparse interaction scenario are determined. This sparse perception interaction extraction mechanism adaptively identifies and focuses on key social interactions, effectively filtering unnecessary weakly related interactions, thereby significantly improving computational efficiency while maintaining prediction accuracy. Based on a preset initial noise trajectory, a preset time step, sparse interaction features, and a preset denoising diffusion probability model, the future motion trajectory of each traffic participant in the sparse interaction scenario is determined. Through sparsification processing and structural optimization, while maintaining high prediction accuracy, computational resource consumption is significantly reduced. This allows the trajectory prediction model (i.e., the preset denoising diffusion probability model) to better meet the real-time requirements of practical applications and provide more accurate and reliable trajectory prediction results for the intelligent driving system, significantly improving the decision-making safety of the target vehicle in complex dynamic environments.

[0099] Based on the above technical solutions, the pre-defined sparse interaction feature extraction model includes: a location embedding module, a sparsification module, and a graph convolutional network;

[0100] The sparse interaction feature determination module 320 may include:

[0101] The first trajectory feature determination submodule is used to input the historical trajectory set into the position embedding module for vector transformation and position embedding to obtain the first trajectory features;

[0102] The sparse asymmetric interaction weight matrix determination submodule is used to determine the sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario based on the first trajectory features and the sparsification module.

[0103] The sparse interaction feature determination submodule is used to perform graph convolution aggregation on the first trajectory features based on graph convolutional networks and sparse asymmetric interaction weight matrices to determine the sparse interaction features corresponding to the sparse interaction scenario.

[0104] Based on the above technical solution, the sparse asymmetric interaction weight matrix determination submodule is specifically used for: inputting the first trajectory feature into the sparsification module, so that the interaction matrix corresponding to the first trajectory feature is determined in the sparsification module using an attention mechanism; determining the sparse interaction mask corresponding to the first trajectory feature in the sparsification module based on preset hyperparameters and preset activation functions; determining the sparse interaction matrix in the sparsification module based on the interaction matrix and the sparse interaction mask; and determining the sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario based on the regularized adjacency matrix and the normalized sparse interaction matrix in the sparsification module.

[0105] Based on the above technical solution, the sparse interaction feature determination submodule is specifically used to: perform graph convolution processing on the first trajectory feature based on graph convolution network and sparse asymmetric interaction weight matrix to obtain the second trajectory feature; and determine the sparse interaction feature corresponding to the sparse interaction scenario based on the first trajectory feature and the second trajectory feature.

[0106] Based on the above technical solution, the future motion trajectory determination module 330 is specifically used for: inputting a preset initial noise trajectory, a preset time step, and sparse interaction features into a preset denoising diffusion probability model, so that the preset denoising diffusion probability model denoises the preset initial noise trajectory according to the sparse interaction features to obtain the processed current noise trajectory and current time step; inputting the current noise trajectory, current time step, and sparse interaction features into the preset denoising diffusion probability model for iterative denoising to obtain the updated current noise trajectory and current time step; when the current time step is updated to a preset step threshold, stopping the iterative denoising and obtaining the future motion trajectory corresponding to each traffic participant in the sparse interaction scenario.

[0107] Based on the above technical solution, the device also includes:

[0108] The future driving trajectory determination module is used to determine the future driving trajectory of the target vehicle based on the target vehicle's driving objective, traffic rules in the sparse interaction scenario, and the future movement trajectory of each traffic participant.

[0109] The trajectory prediction device provided in the embodiments of the present invention can execute the trajectory prediction method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the trajectory prediction method.

[0110] It is worth noting that in the above-described trajectory prediction embodiments, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0111] Example 4

[0112] Figure 4 illustrates a schematic diagram of an electronic device 10 that can be used to implement embodiments of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0113] As shown in Figure 4, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0114] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0115] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as trajectory prediction methods.

[0116] In some embodiments, the trajectory prediction method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the trajectory prediction method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the trajectory prediction method by any other suitable means (e.g., by means of firmware).

[0117] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0118] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0119] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0120] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0121] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0122] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0123] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the trajectory prediction method provided in any embodiment of this application.

[0124] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). This program product belongs to the same inventive concept as the trajectory prediction methods disclosed in the embodiments of this application, and therefore will not be described further here.

[0125] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0126] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A trajectory prediction method, characterized in that, include: When the target vehicle is in a sparse interaction scenario in intelligent driving mode, the historical trajectory set collected by the target vehicle is obtained; the historical trajectory set is composed of the historical movement trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located; based on the preset sparse interaction feature extraction model and the historical trajectory set, the sparse interaction features corresponding to the sparse interaction scenario are determined. Based on the preset initial noise trajectory, preset time step, the sparse interaction features, and preset denoising diffusion probability model, the future movement trajectory of each traffic participant in the sparse interaction scenario is determined.

2. The method according to claim 1, characterized in that, The preset sparse interaction feature extraction model includes: a location embedding module, a sparsification module, and a graph convolutional network; the step of determining the sparse interaction features corresponding to the sparse interaction scenario based on the preset sparse interaction feature extraction model and the historical trajectory set includes: inputting the historical trajectory set into the location embedding module for vector transformation and location embedding to obtain a first trajectory feature; determining the sparse asymmetric interaction weight matrix between each traffic participant in the sparse interaction scenario based on the first trajectory feature and the sparsification module; and performing graph convolution aggregation on the first trajectory feature based on the graph convolutional network and the sparse asymmetric interaction weight matrix to determine the sparse interaction features corresponding to the sparse interaction scenario.

3. The method according to claim 2, characterized in that, The step of determining the sparse asymmetric interaction weight matrix among traffic participants in the sparse interaction scenario based on the first trajectory feature and the sparsification module includes: inputting the first trajectory feature into the sparsification module to determine the interaction matrix corresponding to the first trajectory feature using an attention mechanism in the sparsification module; determining the sparse interaction mask corresponding to the first trajectory feature in the sparsification module based on preset hyperparameters and preset activation functions; determining the sparse interaction matrix in the sparsification module based on the interaction matrix and the sparse interaction mask; and determining the sparse asymmetric interaction weight matrix among traffic participants in the sparse interaction scenario based on a regularized adjacency matrix and a normalized sparse interaction matrix in the sparsification module.

4. The method according to claim 2, characterized in that, The step of performing graph convolution aggregation on the first trajectory features based on the graph convolutional network and the sparse asymmetric interaction weight matrix to determine the sparse interaction features corresponding to the sparse interaction scenario includes: performing graph convolution processing on the first trajectory features based on the graph convolutional network and the sparse asymmetric interaction weight matrix to obtain the second trajectory features; and determining the sparse interaction features corresponding to the sparse interaction scenario based on the first trajectory features and the second trajectory features.

5. The method according to claim 1, characterized in that, The step of determining the future motion trajectory of each traffic participant in the sparse interaction scenario based on a preset initial noise trajectory, a preset time step, the sparse interaction features, and a preset denoising diffusion probability model includes: inputting the preset initial noise trajectory, the preset time step, and the sparse interaction features into the preset denoising diffusion probability model, so that the preset denoising diffusion probability model denoises the preset initial noise trajectory according to the sparse interaction features to obtain the processed current noise trajectory and current time step; inputting the current noise trajectory, the current time step, and the sparse interaction features into the preset denoising diffusion probability model for iterative denoising to obtain the updated current noise trajectory and current time step; and stopping iterative denoising and obtaining the future motion trajectory of each traffic participant in the sparse interaction scenario when the current time step is updated to a preset step threshold.

6. The method according to claim 1, characterized in that, The method further includes: determining the future driving trajectory of the target vehicle based on the driving objective of the target vehicle, the traffic rules in the sparse interaction scenario, and the future motion trajectory of each traffic participant.

7. A trajectory prediction device, characterized in that, The device includes: a historical trajectory set acquisition module, used to acquire a historical trajectory set collected by the target vehicle when the target vehicle is in a sparse interaction scenario in intelligent driving mode; the historical trajectory set is composed of the historical motion trajectories of each traffic participant in the sparse interaction scenario in which the target vehicle is located; a sparse interaction feature determination module, used to determine the sparse interaction features corresponding to the sparse interaction scenario based on a preset sparse interaction feature extraction model and the historical trajectory set; and a future motion trajectory determination module, used to determine the future motion trajectory corresponding to each traffic participant in the sparse interaction scenario based on a preset initial noise trajectory, a preset time step, the sparse interaction features, and a preset denoising diffusion probability model.

8. An electronic device, characterized in that, The electronic device includes: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the trajectory prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the trajectory prediction method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the trajectory prediction method as described in any one of claims 1-6.

Images