A method and system for trajectory prediction of a shuttle vehicle

Abstract

The application relates to a kind of week vehicle trajectory prediction methods, it is characterized in that, it includes: obtaining environment perception data;Based on environment perception data, establish self-vehicle local coordinate system and determine current driving scene category;Based on current driving scene category, from the target anchor point set in the pre-set anchor point library is screened, and environment perception data is mapped to the local space-time trajectory data obtained under self-vehicle local coordinate system;Determine the interactive relationship of target week vehicle and surrounding traffic participants based on local space-time trajectory data, generate sparse attention mask for shielding irrelevant interaction information;The target anchor point set is encoded into initial query vector, and combined with local space-time trajectory data and sparse attention mask, residual offset relative to target anchor point set is predicted using trajectory prediction model;The target anchor point set and residual offset are superimposed, and the multi-modal prediction trajectory of target week vehicle is obtained.

Description

Technical Field

[0001] This application relates to the field of motion prediction for autonomous vehicles, and in particular to a method and system for predicting vehicle trajectories. Background Technology

[0002] With the rapid development of autonomous driving technology, understanding the motion intentions and predicting the trajectories of surrounding traffic participants (hereinafter referred to as "surrounding vehicles") has become a core module in autonomous driving decision-making and planning systems. Accurate trajectory prediction can help autonomous vehicles identify potential collision risks in advance, thereby making smooth and safe driving decisions.

[0003] Currently, data-driven deep learning methods are the mainstream technology in trajectory prediction. These methods typically utilize recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or Transformer architectures to extract the historical motion state of vehicles and semantic features from high-precision maps, thereby regressing future trajectory distributions. However, existing trajectory prediction technologies still suffer from the following significant technical shortcomings in practical vehicle-side deployments: First, existing methods mostly use a global coordinate system as the benchmark for model input and output. In the global coordinate system, the trajectory data of the surrounding vehicles includes the motion components of the vehicle itself, causing the model to need to additionally learn the motion compensation relationship of the vehicle, increasing the learning difficulty. More importantly, the downstream planning and control module usually performs risk assessment and path generation based on the vehicle's local coordinate system. The prediction results in the global coordinate system need to undergo a secondary transformation, which not only increases the computation time but also easily introduces errors during the transformation process, making it difficult to intuitively reflect the collision risk level of the surrounding vehicles relative to the vehicle itself.

[0004] Secondly, existing query vector-based prediction models typically use fixed learnable parameters or random noise as their initial query, lacking guidance from prior physical knowledge. This "blind guessing" initialization method results in poor model adaptability to different driving scenarios. For example, in a straight-ahead scenario, the model may generate a trajectory with significant lateral jitter; while in an intersection scenario, it may fail to effectively cover multiple modalities such as left turns and right turns. Due to the lack of scenario-based trajectory prior constraints, the model needs to learn the physical distribution of the trajectory from scratch, leading to slow training convergence and the predicted trajectory easily violating vehicle kinematic constraints.

[0005] Furthermore, existing methods typically employ a global attention mechanism when handling multi-vehicle interactions, calculating the interaction weights between the target vehicle and all other vehicles and obstacles within its perception range. However, in real-world road scenarios, most distant vehicles or those without interaction intent have negligible impact on the target vehicle's trajectory. This indiscriminate, full-scale computation introduces a significant amount of redundant calculations, resulting in high model inference latency and failing to meet the stringent real-time requirements of automotive embedded chips.

[0006] Finally, existing models mostly employ direct regression of absolute trajectory coordinates. Due to the high uncertainty of future trajectories, directly predicting absolute coordinate values ​​requires extremely strong model fitting capabilities, and errors tend to accumulate in long-term predictions. In contrast, utilizing historical statistical patterns as a benchmark and predicting only the correction relative to that benchmark would significantly reduce prediction difficulty and improve the physical plausibility of the trajectory. However, current technologies lack a solution that effectively combines scenario-based prior anchor points with residual prediction mechanisms. Summary of the Invention

[0007] This application provides a method and system for predicting vehicle trajectories to solve the technical problems in related technologies, such as poor adaptability to different scenarios, high computational redundancy, and prediction results that do not conform to the laws of physical motion.

[0008] Firstly, a method for predicting vehicle trajectories is provided, which includes: Acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants; A local coordinate system for the vehicle is established based on the environmental perception data, and the current driving scenario category is determined. Based on the current driving scenario category, a set of target anchor points is selected from a preset anchor point library, and the environmental perception data is mapped to the vehicle's local coordinate system to obtain local spatiotemporal trajectory data; Based on the local spatiotemporal trajectory data, the interaction relationship between the target vehicle and the surrounding traffic participants is determined, and a sparse attention mask is generated to shield irrelevant interaction information. The target anchor point set is encoded into an initial query vector, and combined with the local spatiotemporal trajectory data and the sparse attention mask, the residual offset relative to the target anchor point set is predicted using a trajectory prediction model. The target anchor point set is superimposed with the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

[0009] In conjunction with the first aspect, in one implementation, establishing a vehicle local coordinate system based on the environmental perception data and determining the current driving scenario category includes: With the center of the rear axle of the vehicle as the origin, the direction of the vehicle's front as the positive X-axis, and the direction perpendicular to the direction of the front to the left as the positive Y-axis, a local rectangular coordinate system for the vehicle is constructed. The vehicle perception data is input into the scene classifier, which outputs the current driving scene category. The vehicle perception data includes at least vehicle status information, target surrounding vehicle trajectory features, and road topology features. The current driving scene category includes at least straight driving scene, lane changing scene, and intersection passage scene.

[0010] In conjunction with the first aspect, in one embodiment, the target surrounding vehicle perception data includes position coordinates and velocity vectors in the original perception coordinate system, and mapping the environmental perception data to the vehicle's local coordinate system includes: The position coordinates and velocity vectors in the original sensing coordinate system are transformed to the vehicle's local coordinate system, and the vehicle's kinematic model is used to compensate for the vehicle's motion displacement within the data acquisition time window.

[0011] In conjunction with the first aspect, in one implementation, obtaining a preset anchor point library includes: Collect weekly vehicle trajectory fragments from historical driving data; The vehicle trajectory segments are subjected to statistical analysis or clustering to generate an offset sequence representing typical driving intentions. The offset sequence is categorized and stored according to the corresponding driving scenario category to form the preset anchor point library.

[0012] In conjunction with the first aspect, in one implementation, the step of selecting a set of target anchor points from a preset anchor point library based on the current driving scenario category includes: Based on the current driving scenario category, retrieve the anchor point sub-library that matches the current driving scenario category from the preset anchor point library; A preset number of offset sequences are selected from the anchor point sub-library as the target anchor point set.

[0013] In conjunction with the first aspect, in one implementation, generating a sparse attention mask for masking irrelevant interaction information includes: Calculate the interaction strength between the target vehicle and surrounding traffic participants, the interaction strength being determined based on at least one of relative distance, relative speed, or collision time; The interaction strength is compared with a preset threshold: if the interaction strength is lower than the preset threshold, the corresponding surrounding traffic participants are marked as ignored; if the interaction strength is higher than or equal to the preset threshold, the corresponding surrounding traffic participants are marked as active. Construct an attention mask matrix; The mask value corresponding to the ignored state is set as a mask value to block attention calculation, and the mask value corresponding to the valid state is set as a reserved value to allow attention calculation, thus obtaining the sparse attention mask.

[0014] In conjunction with the first aspect, in one implementation, predicting the residual offset relative to the target anchor point set using a trajectory prediction model includes: The local spatiotemporal trajectory data is input into the encoder of the trajectory prediction model to obtain environmental semantic features and interaction features; The initial query vector, the environmental semantic features, and the interaction features are input into the decoder of the trajectory prediction model; The residual offset is output through the cross-attention mechanism of the decoder.

[0015] In conjunction with the first aspect, in one implementation, predicting the residual offset relative to the target anchor point set using the trajectory prediction model further includes: Output the intent probability corresponding to each trajectory in the multimodal predicted trajectory; The multimodal predicted trajectories are weighted and sorted according to the intent probability, and a preset number of trajectories at the top of the sort are output as the final prediction result.

[0016] In conjunction with the first aspect, in one implementation, the vehicle perception data includes vehicle motion state data provided by the vehicle CAN bus; The target surrounding vehicle perception data and surrounding traffic participant perception data include at least one of the following: point cloud data collected by vehicle-mounted radar, image data collected by vehicle-mounted cameras, or lane line information provided by high-precision maps.

[0017] In conjunction with the first aspect, in one implementation, the step of superimposing the target anchor point set with the residual offset includes: In the local coordinate system of the vehicle, the coordinates of each time point of each anchor point sequence in the target anchor point set are added point by point to the offset coordinates of the corresponding time point in the residual offset to obtain the absolute trajectory coordinate sequence.

[0018] In conjunction with the first aspect, in one implementation, a vehicle trajectory prediction method further includes: The multimodal predicted trajectory is sent to the autonomous driving planning and control module; The autonomous driving planning and control module performs collision risk assessment based on the multimodal predicted trajectory and generates an obstacle avoidance planning path for the vehicle.

[0019] Secondly, a vehicle trajectory prediction system is provided, including: The data acquisition module is used to acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants. The data processing module is used to establish a local coordinate system of the vehicle based on the environmental perception data and determine the current driving scenario category, filter a set of target anchor points from a preset anchor point library based on the current driving scenario category, and map the environmental perception data to the local coordinate system of the vehicle to obtain local spatiotemporal trajectory data. A mask generation module is used to determine the interaction relationship between the target vehicle and the surrounding traffic participants based on the local spatiotemporal trajectory data, and generate a sparse attention mask to mask irrelevant interaction information. And a trajectory prediction module, which is used to encode the target anchor point set into an initial query vector, and combine the local spatiotemporal trajectory data and the sparse attention mask to predict the residual offset relative to the target anchor point set using a trajectory prediction model, and to superimpose the target anchor point set and the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

[0020] The beneficial effects of the technical solution provided in this application include: using a local coordinate system of the vehicle as a spatiotemporal reference, and dynamically selecting a set of target anchor points from a statistical offset anchor point library in conjunction with the current driving scenario category, using a sparse attention mask to shield invalid interactive connections, and superimposing the residual offset relative to the set of target anchor points predicted by a trajectory prediction model to obtain the final trajectory. Therefore, this application can effectively eliminate vehicle motion compensation errors, significantly improve the scenario adaptability and physical rationality of trajectory prediction through scenario-based anchor point priors, and reduce computational redundancy by using a sparse attention mechanism to meet the real-time requirements of vehicle deployment, thereby providing a more accurate and efficient risk assessment basis for the autonomous driving planning and control module. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a vehicle trajectory prediction method provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] This application provides a method and system for predicting vehicle trajectories, which can solve the technical problems of poor adaptability to different scenarios, high computational redundancy, and prediction results that do not conform to the laws of physical motion in related technologies.

[0025] Reference Figure 1 This application discloses a method and system for predicting vehicle trajectories, wherein the method for predicting vehicle trajectories includes the following steps: S1: Acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants. In this step, the autonomous vehicle collects information about its surrounding environment through its onboard sensor network. The vehicle's perception data includes motion status data provided by the vehicle's CAN bus, such as vehicle speed, acceleration, yaw rate, and steering wheel angle.

[0026] The target vehicle perception data and surrounding traffic participant perception data include at least one of the following: point cloud data collected by vehicle-mounted radar (such as lidar or millimeter-wave radar), image data collected by vehicle-mounted cameras, or lane line information provided by high-precision maps. The target vehicle refers to the specific vehicle whose future trajectory needs to be predicted, and surrounding traffic participants refer to other vehicles, pedestrians, or non-motorized vehicles that may interact with the target vehicle.

[0027] S2: Establish a local coordinate system for the vehicle based on the environmental perception data and determine the current driving scenario category; S201: Construct a local rectangular coordinate system for the vehicle with the center of the rear axle as the origin, the direction of the vehicle's front as the positive X-axis, and the direction perpendicular to the direction of the front to the left as the positive Y-axis; To eliminate reference frame errors caused by vehicle motion and make relative position representation more intuitive, this step defines a unified coordinate reference system. Using the vehicle's local coordinate system instead of a global coordinate system allows all perceived data to be uniformly converted to the vehicle's perspective.

[0028] Using the vehicle's center at time t as the origin, the vehicle's front direction as the positive x-axis, and the left side of the vehicle as the positive y-axis, a local predicted coordinate system for the vehicle's center is established. This coordinate system is used to unify the reference frame of all sensing data and eliminate vehicle motion errors in the global coordinate system. In other implementations, the origin of the coordinate system can also be selected from the vehicle's center of gravity; the coordinate system type can also be the Frenet coordinate system based on the lane centerline, as long as it can achieve a standardized expression of relative positions.

[0029] S202: Input the vehicle perception data into the scene classifier and output the current driving scene category. The vehicle perception data includes at least vehicle status information, target surrounding vehicle trajectory features and road topology features. The current driving scene category includes at least straight driving scene, lane changing scene and intersection passage scene.

[0030] This step identifies the driving scenario by concatenating the historical trajectory features of the vehicle to be predicted, the vehicle's motion state features, the current lane topology features, the navigation path features, and the intersection distance features in a high-dimensional array to construct a scene feature vector. This scene feature vector is then input into a classifier, which outputs the scene classification probability and dynamically determines whether the scenario is a straight-ahead driving situation, a lane-changing / merging situation, or a turning / intersection situation. The scene classifier can be a multilayer perceptron (MLP), decision tree, hidden Markov model (HMM), or temporal convolutional network (TCN); a rule engine can also be used, such as classifying a turn if the lateral speed is greater than a threshold and there is an intersection ahead. As a specific example, the calculation formula is as follows:

[0031]

[0032]

[0033] S3: Based on the current driving scenario category, select a set of target anchor points from the preset anchor point library, and map the environmental perception data to the local coordinate system of the vehicle to obtain local spatiotemporal trajectory data; S301: Obtain the preset anchor point library; Specifically, the preset anchor point library is a set of typical trajectory offset sequences that are pre-clustered and statistically analyzed from historical real data. The real weekly vehicle trajectory data is pre-mined, normalized, and clustered to extract multiple statistical offset sequences covering different potential driving intentions.

[0034] For example, the straight-ahead anchor point library contains a statistical offset sequence that is primarily longitudinally advanced and has limited or near-zero lateral offset amplitude. This is used to suppress lateral jitter that does not conform to road constraints in straight-ahead scenarios. Specifically, it includes the following steps: S301(a): Collect weekly vehicle trajectory segments from historical driving data; S301(b): Perform statistical analysis or clustering processing (such as K-Means clustering or Gaussian mixture model) on the said vehicle trajectory segment to generate an offset sequence representing typical driving intentions; S301(c): The offset sequence is classified and stored according to the corresponding driving scenario category to form a preset anchor point library.

[0035] Unsupervised clustering or quantile statistics are performed on the relative coordinate changes in large-scale real-world vehicle trajectory data to extract representative trajectories representing different driving intentions. These trajectories are defined as a set of statistical offset sequences for the next T time steps. Each anchor point is a two-dimensional offset sequence for the next T predicted times, rather than a single endpoint. Based on scene semantics, straight-ahead anchor point libraries, lane-change / merging anchor point libraries, and turning / intersection anchor point libraries are constructed respectively.

[0036]

[0037]

[0038] The straight-ahead anchor point library contains statistical offset sequences that primarily advance longitudinally with limited or near-zero lateral offset amplitude. The lane-change / merging anchor point library and the turning / intersection anchor point library contain statistical offset sequences covering different curvatures, deflection angles, and endpoint areas for left / right lane changes, merging, left turns, right turns, and U-turns. Vehicle motion characteristics vary significantly across different scenarios; straight-ahead scenarios exhibit minimal lateral variation, while intersection scenarios involve multiple modal intentions. Scenario classification provides a basis for subsequently calling different anchor point libraries.

[0039] The lane change / merge anchor point library and the turn / intersection anchor point library contain statistical offset sequences covering different curvatures, deflection angles, and endpoint regions for left / right lane changes, merging, left turns, right turns, and U-turns, used to improve multimodal coverage in complex scenarios. This scenario-based anchor point mechanism avoids unnecessary lateral swaying of general anchor points when going straight, or missing intentions such as left turns / U-turns at intersections. Furthermore, the anchor point generation method is not limited to statistical clustering; it can also be replaced with typical trajectories generated by generative adversarial networks (GANs), the central parameters of Gaussian mixture models (GMMs), or polynomial curves preset by human experience.

[0040] S302: Transform the position coordinates and velocity vectors in the original sensing coordinate system to the vehicle's local coordinate system, and use the vehicle's kinematic model to compensate for the vehicle's motion displacement within the data acquisition time window.

[0041] The position transformation matrix is ​​calculated using the vehicle's speed and yaw rate. This transforms the historical observations of the surrounding vehicle into the vehicle's local coordinate system, eliminating motion compensation errors and obtaining a standardized historical trajectory sequence. Specifically, the vehicle's speed v, yaw rate r, heading angle yaw, and sampling time interval are obtained at time t and within the historical time window. t, calculated based on the vehicle kinematics model from historical time t The rotation matrix R from time i to the current time t t i>t With translation vector T t i>t The historical time window can be from 1 second to 5 seconds in the past, preferably from 2 seconds in the past, and the sampling frequency can be from 5 Hz to 20 Hz, preferably 10 Hz.

[0042] Using the position transformation matrix R and translation vector T, the global observation coordinates of the vehicle under the historical timestamp are uniformly mapped and aligned to the local prediction coordinate system of the vehicle center at time t, thereby eliminating the compensation error caused by the vehicle's motion and obtaining a standardized historical trajectory sequence of the vehicle.

[0043] S303: Based on the current driving scenario category, retrieve the anchor point sub-library that matches the current driving scenario category from the preset anchor point library; Based on the scene classification probability output in step S202 A dynamic gating mechanism is used to route to the corresponding anchor point library. If the current environment is determined to be a straight-ahead condition, the statistical offset sequence of group K1 in A_straight is extracted; if it is determined to be a lane-changing / merging condition, the statistical offset sequence of group K2 in A_lane is extracted; if it is determined to be a turning / intersection condition, the statistical offset sequence of group K3 in A_turn is extracted. Alternatively, it can be configured according to... Multiple anchor point libraries are weighted and fused to obtain a set of anchor points for mixed scenarios.

[0044] S304: Select a preset number of offset sequences from the anchor point sub-library as the target anchor point set.

[0045] The extracted discrete statistical offset sequence is updimensionalized using a multilayer perceptron or linear mapping layer, and then fused with learnable location encoding, time encoding, or anchor category encoding to transform it into a high-dimensional intent query vector required by the decoder. These queries carry the multimodal offset patterns of real traffic flow before entering the decoder.

[0046] S4: Based on the local spatiotemporal trajectory data, determine the interaction relationship between the target vehicle and the surrounding traffic participants, and generate a sparse attention mask to shield irrelevant interaction information; To reduce computational complexity and minimize interference from irrelevant vehicles in the prediction results, this invention introduces a sparse attention mechanism. Existing technologies typically perform fully connected interaction modeling for all perceived vehicles, resulting in significant computational redundancy. This invention uses physical thresholding to filter neighboring vehicles that have a substantial impact on the vehicle being predicted, based on longitudinal distance, lateral distance, lane width, relative speed, and collision time. This generates a sparse attention mask for subsequent attention calculations, which masks or reduces the attention weights of irrelevant vehicles, allowing only neighboring vehicles with potential interactions with the target vehicle to participate in the calculation.

[0047] S401: Calculate the strength of the interaction relationship between the target vehicle and surrounding traffic participants, the strength of the interaction relationship being determined based on at least one of relative distance, relative speed, or collision time; Calculate the longitudinal and lateral relative distances between the vehicle to be predicted and the j-th surrounding vehicle within the perception range. Based on the speed of the vehicle to be predicted, relative speed, system response time, and lane width, dynamically set the longitudinal and lateral interaction thresholds.

[0048]

[0049]

[0050] in, Based on the vertical interaction distance; This is a preset system response time constant; The set number of lateral radial lanes; This is the standard lane width for the current road section; The speed of the train to be predicted; Let be the relative speed between the j-th car and its neighboring car.

[0051] S402: Compare the strength of the interaction relationship with a preset threshold: If the strength of the interaction relationship is lower than the preset threshold, the corresponding surrounding traffic participants will be marked as ignored. If the strength of the interaction relationship is higher than or equal to the preset threshold, the corresponding surrounding traffic participants will be marked as valid.

[0052] S403: Construct the attention mask matrix; Based on the adaptive interaction threshold, identify neighboring vehicles that are actually interacting and generate an attention mask matrix for spatial interaction calculation; for vehicles that meet the longitudinal and lateral interaction thresholds, allow them to participate in attention interaction, otherwise mask or weaken their attention weight.

[0053]

[0054] S404: Set the mask value corresponding to the ignored state to a mask value used to block attention calculation, and set the mask value corresponding to the valid state to a reserved value used to allow attention calculation, to obtain the sparse attention mask.

[0055] S5: Encode the target anchor point set into an initial query vector, and combine the local spatiotemporal trajectory data and the sparse attention mask to predict the residual offset relative to the target anchor point set using a trajectory prediction model; This model employs an encoder-decoder prediction architecture, preferably using a Transformer, but can also be replaced by a graph neural network, recurrent neural network, attention network, or a hybrid of these networks. The model uses scene-specific statistical offset anchor points as intent priors, and the decoder predicts the residual offsets and intent probabilities relative to each anchor point, thus mapping from "anchor point candidate trajectories" to "final multimodal predicted trajectories." The environmental semantic joint encoder extracts static road semantics such as lane centerlines, lane boundaries, road curvature, intersection topology, and navigation directions, as well as dynamic traffic information such as historical trajectories of surrounding vehicles, the state of vehicles to be predicted, and the states of interacting neighboring vehicles. These features are encoded and fused to obtain the environmental semantic features used for subsequent decoding.

[0056] Anchor Point Query Cross-Decoder: The statistical offset Query generated in step S4 is input into the decoder. The decoder uses a cross-attention mechanism to query the Query carrying prior offset patterns and aggregate environmental semantic features and interactive vehicle features to obtain decoded features adapted to the current scene, interaction relationships, and road structure.

[0057] Anchor point residual trajectory and probability output: After feature aggregation, the decoder outputs the intent probability and residual offset corresponding to each candidate anchor point in parallel through the prediction head; the statistical offset anchor point, residual offset and the position of the vehicle to be predicted at the current time are superimposed to obtain the multimodal vehicle prediction trajectory.

[0058] Among them, the statistical offset anchor points are used to define the prior trajectory distribution under different scenarios, while the residual offset is used to refine and correct the anchor point trajectory based on the current interaction relationship and environmental semantics. In the straight-ahead scenario, the lateral jitter of the final predicted trajectory is suppressed because the selected anchor point library has a small lateral offset; in the intersection / turning scenario, the final predicted trajectory can cover a variety of potential driving intentions because the selected anchor point library covers different curvatures and endpoint areas.

[0059] S501: Input the local spatiotemporal trajectory data into the encoder of the trajectory prediction model to obtain environmental semantic features and interaction features; Joint encoding of environmental semantics: Features such as lane centerline, lane boundary, road curvature, intersection topology, navigation direction, surrounding vehicle historical trajectory, predicted surrounding vehicle state, and interacting neighbor vehicle state are extracted, vectorized, and input into a multi-layer encoder. Combined with the attention mask matrix generated in step S4, features of interacting vehicles and environmental semantics are fused to output high-dimensional environmental semantic features. .

[0060] S502: Input the initial query vector, the environmental semantic features, and the interaction features into the decoder of the trajectory prediction model; The intent query vector Q generated in step S3 is input into the decoder, and the cross-attention mechanism is used to dynamically query and aggregate the environmental semantic features to obtain the decoded features. ;

[0061]

[0062] S503: Output the residual offset through the cross-attention mechanism of the decoder.

[0063] After feature aggregation, the decoder outputs the probability distribution of each multimodal intent of the vehicle in parallel through the prediction head, as well as the corresponding prediction residual offset.

[0064] S504: Output the intent probability corresponding to each trajectory in the multimodal predicted trajectory; S505: The multimodal predicted trajectories are weighted and sorted according to the intent probability, and the preset number of trajectories at the top of the sort are output as the final prediction result.

[0065] S6: The target anchor point set is superimposed with the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

[0066] S601: In the local coordinate system of the vehicle, the coordinates of each time point of each anchor point sequence in the target anchor point set are added point by point to the offset coordinates of the corresponding time point in the residual offset to obtain the absolute trajectory coordinate sequence.

[0067] By superimposing the statistical anchor point, the predicted residual offset, and the initial coordinates of the vehicle at the current moment, the absolute trajectory coordinates of each mode within the future prediction window are obtained.

[0068]

[0069]

[0070]

[0071] Among them, the statistical offset anchor points are used to define the prior trajectory distribution under different scenarios, while the residual offset is used to refine and correct the anchor point trajectory based on the current interaction relationship and environmental semantics. In the straight-ahead scenario, the lateral jitter of the final predicted trajectory is suppressed because the selected anchor point library has a small lateral offset; in the intersection / turning scenario, the final predicted trajectory can cover a variety of potential driving intentions because the selected anchor point library covers different curvatures and endpoint areas.

[0072] S7: Send the multimodal predicted trajectory to the autonomous driving planning and control module; The trained model can be exported as an inference engine format. It receives real-time perception data, vehicle status data, and map / navigation data from the vehicle system to achieve online multimodal vehicle trajectory prediction. In a preferred embodiment, the system uses the historical vehicle trajectory of the past 2 seconds as input, with a sampling frequency of 10Hz, to predict the vehicle trajectory of the next 3 to 5 seconds, with a prediction frequency of 10Hz. For each vehicle to be predicted, N candidate trajectories and their corresponding probabilities are output, where N can be 6, 8, 12, or 16.

[0073] S8: The autonomous driving planning and control module performs a collision risk assessment based on the multimodal predicted trajectory and generates an obstacle avoidance planning path for the vehicle.

[0074] A joint loss function is constructed to perform end-to-end training on the trajectory prediction model. The joint loss function includes trajectory regression loss, final endpoint error loss, false negative rate boundary optimization loss, and anchor point intent classification loss. The training can employ minimum distance matching between the real trajectory and candidate anchor point trajectories, optimal modality matching, or soft assignment to determine the supervision anchor points.

[0075]

[0076] Specifically, the trajectory regression loss constrains the overall similarity between the complete predicted trajectory and the actual trajectory; the final endpoint error loss enhances the convergence ability at the end of the trajectory under a long prediction window; the missed detection rate boundary optimization loss improves the coverage of low-probability, high-risk intentions; and the anchor point intent classification loss constrains the model to select the anchor point modality that best matches the actual trajectory. The expected results are: reduced lateral jitter in straight-ahead scenarios, reduced missed detection rate in intersection scenarios, improved multimodal intent coverage, and online inference time meeting the requirements for real-vehicle deployment.

[0077] Taking a specific embodiment of this application as an example, the system uses the historical trajectory of the target vehicle in the past 2 seconds as input, with a sampling frequency of 10Hz, to predict the target vehicle trajectory in the next 3 to 5 seconds, with a prediction frequency of 10Hz. For each target vehicle to be predicted, N candidate trajectories and corresponding intent probabilities are output, where N can be 6, 8, 12, or 16, ultimately forming a multimodal predicted trajectory.

[0078] When constructing the preset anchor point library, the real trajectory is first uniformly mapped to the vehicle's local coordinate system or the lane's Frenet coordinate system, and then the offset sequence at the next T time points is extracted. For straight-ahead scenarios, trajectories with lateral offset absolute values ​​less than a preset threshold can be selected for clustering; for intersection or lane-changing scenarios, they can be grouped according to turning direction, trajectory curvature, endpoint area, or heading angle change, and then clustered separately to form anchor point sub-libraries corresponding to different current driving scenario categories.

[0079] During online inference, if the straight-ahead scenario has the highest probability among the classification probabilities of the current driving scenario category, the anchor point sub-library corresponding to the straight-ahead scenario is called first to reduce the lateral sway of the predicted trajectory; if the intersection passage scenario or lane change scenario has the highest probability, the corresponding anchor point sub-library is called to cover candidate intentions such as left turn, right turn, and U-turn; if the probabilities of multiple current driving scenario categories are similar, the offset sequence from multiple anchor point sub-libraries is selected according to probability weighting to form a mixed target anchor point set.

[0080] In a comparative test, this invention can be compared with a random initial query vector model without anchor points, a fixed anchor point model, and a model without a scene classification anchor point library. Evaluation metrics include average displacement error, final endpoint error, false negative rate, lateral error, intersection scene modal coverage, and single-frame inference time. Expected results include: reduced lateral jitter in straight-ahead scenes, reduced false negative rate in intersection scenes, improved multimodal intent coverage, and online inference time meeting the requirements for real-vehicle deployment.

[0081] Based on the vehicle trajectory prediction method disclosed in this application, the present invention also provides a vehicle trajectory prediction system, comprising: The data acquisition module is used to acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants. The data processing module is used to establish a local coordinate system of the vehicle based on the environmental perception data and determine the current driving scenario category, filter a set of target anchor points from a preset anchor point library based on the current driving scenario category, and map the environmental perception data to the local coordinate system of the vehicle to obtain local spatiotemporal trajectory data. A mask generation module is used to determine the interaction relationship between the target vehicle and the surrounding traffic participants based on the local spatiotemporal trajectory data, and generate a sparse attention mask to mask irrelevant interaction information. And a trajectory prediction module, which is used to encode the target anchor point set into an initial query vector, and combine the local spatiotemporal trajectory data and the sparse attention mask to predict the residual offset relative to the target anchor point set using a trajectory prediction model, and to superimpose the target anchor point set and the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

[0082] In the description of this application, it should be noted that the terms "upper," "lower," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0083] It should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0084] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting the trajectory of a vehicle, characterized in that, It includes: Acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants; A local coordinate system for the vehicle is established based on the environmental perception data, and the current driving scenario category is determined. Based on the current driving scenario category, a set of target anchor points is selected from a preset anchor point library, and the environmental perception data is mapped to the vehicle's local coordinate system to obtain local spatiotemporal trajectory data; Based on the local spatiotemporal trajectory data, the interaction relationship between the target vehicle and the surrounding traffic participants is determined, and a sparse attention mask is generated to shield irrelevant interaction information. The target anchor point set is encoded into an initial query vector, and combined with the local spatiotemporal trajectory data and the sparse attention mask, the residual offset relative to the target anchor point set is predicted using a trajectory prediction model. The target anchor point set is superimposed with the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

2. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The process of establishing a local coordinate system for the vehicle based on the environmental perception data and determining the current driving scenario category includes: With the center of the rear axle of the vehicle as the origin, the direction of the vehicle's front as the positive X-axis, and the direction perpendicular to the direction of the front to the left as the positive Y-axis, a local rectangular coordinate system for the vehicle is constructed. The vehicle perception data is input into the scene classifier, and the current driving scene category is output. The vehicle perception data includes at least vehicle status information, target surrounding vehicle trajectory features and road topology features. The current driving scene category includes at least straight driving scene, lane changing scene and intersection passing scene.

3. The method for predicting a vehicle trajectory as described in claim 2, characterized in that: The target surrounding vehicle perception data includes position coordinates and velocity vectors in the original perception coordinate system. Mapping the environmental perception data to the vehicle's local coordinate system includes: The position coordinates and velocity vectors in the original sensing coordinate system are transformed to the local coordinate system of the vehicle, and the vehicle's kinematic model is used to compensate for the vehicle's motion displacement within the data acquisition time window.

4. The method for predicting a vehicle trajectory as described in claim 1, characterized in that, It also includes: Obtain a preset anchor point library, which includes: Collect weekly vehicle trajectory fragments from historical driving data; The vehicle trajectory segments are subjected to statistical analysis or clustering to generate an offset sequence representing typical driving intentions. The offset sequence is categorized and stored according to the corresponding driving scenario category to form the preset anchor point library.

5. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The step of selecting a set of target anchor points from a preset anchor point library based on the current driving scenario category includes: Based on the current driving scenario category, retrieve the anchor point sub-library that matches the current driving scenario category from the preset anchor point library; A preset number of offset sequences are selected from the anchor point sub-library as the target anchor point set.

6. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The generation of the sparse attention mask for masking irrelevant interaction information includes: Calculate the interaction strength between the target vehicle and surrounding traffic participants, the interaction strength being determined based on at least one of relative distance, relative speed, or collision time; The interaction strength is compared with a preset threshold: if the interaction strength is lower than the preset threshold, the corresponding surrounding traffic participants are marked as ignored; if the interaction strength is higher than or equal to the preset threshold, the corresponding surrounding traffic participants are marked as active. Construct an attention mask matrix; The mask value corresponding to the ignored state is set as a mask value to block attention calculation, and the mask value corresponding to the valid state is set as a reserved value to allow attention calculation, thus obtaining the sparse attention mask.

7. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The step of predicting the residual offset relative to the target anchor point set using a trajectory prediction model includes: The local spatiotemporal trajectory data is input into the encoder of the trajectory prediction model to obtain environmental semantic features and interaction features; The initial query vector, the environmental semantic features, and the interaction features are input into the decoder of the trajectory prediction model; The residual offset is output through the cross-attention mechanism of the decoder.

8. The method for predicting a vehicle trajectory as described in claim 7, characterized in that: The step of predicting the residual offset relative to the target anchor point set using the trajectory prediction model further includes: Output the intent probability corresponding to each trajectory in the multimodal predicted trajectory; The multimodal predicted trajectories are weighted and sorted according to the intent probability, and a preset number of trajectories at the top of the sort are output as the final prediction result.

9. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The vehicle perception data includes vehicle motion status data provided by the vehicle's CAN bus; The target surrounding vehicle perception data and surrounding traffic participant perception data include at least one of the following: point cloud data collected by vehicle-mounted radar, image data collected by vehicle-mounted cameras, or lane line information provided by high-precision maps.

10. The method for predicting a vehicle trajectory as described in claim 1, characterized in that: The step of superimposing the target anchor point set with the residual offset includes: In the local coordinate system of the vehicle, the coordinates of each time point of each anchor point sequence in the target anchor point set are added point by point to the offset coordinates of the corresponding time point in the residual offset to obtain the absolute trajectory coordinate sequence.

11. A method for predicting a vehicle trajectory as described in any one of claims 1 to 10, characterized in that, It also includes: The multimodal predicted trajectory is sent to the autonomous driving planning and control module; The autonomous driving planning and control module performs collision risk assessment based on the multimodal predicted trajectory and generates an obstacle avoidance planning path for the vehicle.

12. A vehicle trajectory prediction system, characterized in that, include: The data acquisition module is used to acquire environmental perception data, which includes vehicle perception data, target surrounding vehicle perception data, and perception data of surrounding traffic participants. The data processing module is used to establish a local coordinate system of the vehicle based on the environmental perception data and determine the current driving scenario category, filter a set of target anchor points from a preset anchor point library based on the current driving scenario category, and map the environmental perception data to the local coordinate system of the vehicle to obtain local spatiotemporal trajectory data. A mask generation module is used to determine the interaction relationship between the target vehicle and the surrounding traffic participants based on the local spatiotemporal trajectory data, and generate a sparse attention mask to mask irrelevant interaction information. And a trajectory prediction module, which is used to encode the target anchor point set into an initial query vector, and combine the local spatiotemporal trajectory data and the sparse attention mask to predict the residual offset relative to the target anchor point set using a trajectory prediction model, and to superimpose the target anchor point set and the residual offset to obtain the multimodal predicted trajectory of the target vehicle.

Images