Lane changing planning method and device, vehicle, equipment, storage medium and program product

Abstract

The invention provides a lane changing planning method and device, a vehicle, equipment, a storage medium and a program product, and relates to the technical field of intelligent driving. According to the lane changing planning method, a first initial feature embedding vector corresponding to each dynamic target and a second initial feature embedding vector corresponding to a target vehicle are obtained, then an enhanced embedding vector fused with gap role semantic information is generated for each dynamic target, and an enhanced embedding vector fused with gap role semantic information is generated for each dynamic target through an interactive game module. Global interaction relation modeling is carried out on the enhanced embedded vectors corresponding to the dynamic targets and the second initial feature embedded vectors, deep coupling of the enhanced embedded vectors corresponding to the dynamic targets and target vehicle lane changing trajectory planning is achieved, and finally the anthropomorphic (learning based on human driving interaction data) lane changing interaction game is achieved. The method can flexibly adapt to complex and changeable real traffic game scenes such as long-tail scenes (such as sudden lane change of aggressive vehicles), regional driving style differences and the like, and the generalization ability of lane change planning is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a lane change planning method, device, vehicle, equipment, storage medium, and program product. Background Technology

[0002] In intelligent driving systems such as Advanced Driver Assistance Systems (ADAS) and Autonomous Driving (AD), lane changing is a core and frequently used driving function. Especially in multi-lane, high-traffic scenarios such as urban expressways and highways, for autonomous vehicles to complete safe, comfortable, and efficient lane-changing maneuvers, they not only need to identify physically feasible gaps in the space, but also need to actively interact and engage in complex games with other moving object targets (MOTs) in the target lane through proactive longitudinal speed adjustments (acceleration or deceleration) to actively create or utilize lane-changing opportunities. This "gap-insertion lane changing" (referring to the autonomous vehicle actively adjusting its speed to insert itself into the spatial gap between dynamic targets in the target lane to complete the lane change) is a key operation reflecting the level of driving intelligence, directly impacting system safety, user comfort, and traffic flow efficiency.

[0003] In related technologies, lane change planning schemes typically rely on manually written rules or cost functions, which suffer from poor generalization ability. Summary of the Invention

[0004] This application provides a lane change planning method, apparatus, vehicle, equipment, storage medium, and program product to solve the problem that lane change planning methods in related technologies rely on manually written rules or cost functions, resulting in poor generalization ability.

[0005] In a first aspect, embodiments of this application provide a lane change planning method, comprising: acquiring current scene information of a target vehicle, the current scene information including at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle; generating an enhanced embedding vector corresponding to each dynamic target based on the first initial feature embedding vector corresponding to each dynamic target, the enhanced embedding vector fusing gap role semantic information; inputting the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into an interactive game module for global interactive modeling, obtaining a first fused embedding vector corresponding to each dynamic target and a second fused embedding vector corresponding to the target vehicle output by the interactive game module; and outputting at least one lane change trajectory corresponding to the target vehicle based on the first fused embedding vector and the second fused embedding vector corresponding to each dynamic target.

[0006] In one possible implementation, generating an enhanced embedding vector for each dynamic target based on a first initial feature embedding vector for each dynamic target includes: inputting the first initial feature embedding vector for each dynamic target into a gap role probability prediction module to perform gap role probability prediction, thereby obtaining a gap role probability distribution vector for each dynamic target output by the gap role probability prediction module; and generating an enhanced embedding vector for each dynamic target based on the gap role probability distribution vector for each dynamic target and the first initial feature embedding vector.

[0007] In one possible implementation, an enhanced embedding vector is generated for each dynamic target based on the gap role probability distribution vector and the first initial feature embedding vector. This includes: generating a gap role probability embedding vector for each dynamic target based on the gap role probability distribution vector; and fusing the gap role probability embedding vector and the first initial feature embedding vector to generate an enhanced embedding vector for each dynamic target.

[0008] In one possible implementation, the interactive game module includes a self-attention module and a cross-attention module. The enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into the interactive game module for global interactive modeling, resulting in a first fused embedding vector corresponding to each dynamic target and a second fused embedding vector corresponding to the target vehicle, output by the interactive game module. This includes: inputting the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the self-attention module for global feature association modeling, resulting in a first fused embedding vector corresponding to each dynamic target and a third fused embedding vector corresponding to the target vehicle, output by the self-attention module; and inputting the first fused embedding vector and the third fused embedding vector corresponding to each dynamic target into the cross-attention module for dynamic target-target vehicle game modeling, resulting in a second fused embedding vector output by the cross-attention module.

[0009] In one possible implementation, the interactive game module also includes a type embedding vector, which is used to distinguish between target vehicles and dynamic targets.

[0010] In one possible implementation, the type embedding vector includes a first type embedding vector and a second type embedding vector; the first type embedding vector is used to characterize the target vehicle; and the second type embedding vector is used to characterize the dynamic target.

[0011] In one possible implementation, the interactive game module includes a graph neural network. The enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into the interactive game module for global interactive modeling to obtain a first fused embedding vector corresponding to each dynamic target and a second fused embedding vector corresponding to the target vehicle output by the interactive game module. This includes: inputting the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the graph neural network for global interactive modeling to obtain a first fused embedding vector and a second fused embedding vector corresponding to each dynamic target output by the graph neural network.

[0012] In one possible implementation, based on the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle, at least one lane change trajectory corresponding to the target vehicle is output, including: inputting the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target into the trajectory prediction module to perform lane change trajectory prediction, and obtaining at least one lane change trajectory output by the trajectory prediction module.

[0013] In one possible implementation, each lane change trajectory in at least one lane change trajectory includes a confidence level. After outputting at least one lane change trajectory corresponding to the target vehicle, the implementation further includes: for each lane change trajectory in at least one lane change trajectory, determining the target lane change trajectory corresponding to the target vehicle based on the confidence level of the lane change trajectory.

[0014] In one possible implementation, the first initial feature embedding vector corresponding to each dynamic target is obtained by: acquiring the current environmental information of the target vehicle; inputting the current environmental information into the perception module for dynamic target perception, obtaining at least one dynamic target output by the perception module, and the first initial state feature corresponding to each dynamic target; encoding the first initial state feature corresponding to each dynamic target to obtain the first initial feature embedding vector corresponding to each dynamic target.

[0015] In one possible implementation, the second initial feature embedding vector is obtained by: acquiring the second initial state features of the target vehicle; encoding the second initial state features to obtain the second initial feature embedding vector.

[0016] In one possible implementation, the lane change planning method provided in this application is implemented through a lane change planning model, which includes a gap role probability prediction module and a lane change trajectory planning module. The lane change planning model is trained by acquiring multiple first training samples and multiple second training samples. Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a gap role ground truth probability distribution vector corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are based on the first probability in the multiple first dynamic targets. The gap role probability prediction module is trained based on multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. The gap role probability prediction module is used to predict the gap role probability of dynamic targets in lane change scenarios. The lane change trajectory planning module is trained based on multiple second training samples and multiple second gap role prediction probability distribution vectors to obtain a model for lane change planning. The lane change trajectory planning module is used to output the lane change trajectory of the target vehicle based on the gap role probability. The multiple second gap role prediction probability distribution vectors are obtained by selecting from multiple first gap role prediction probability distribution vectors based on the second probability.

[0017] In one possible implementation, the gap role probability prediction module is trained based on multiple first training samples, including: constructing a first loss function; and iteratively optimizing the gap role probability prediction module by minimizing the first loss function based on multiple first training samples until the gap role prediction probability distribution vector output by the gap role probability prediction module converges.

[0018] In one possible implementation, the lane change trajectory planning module is trained based on multiple second training samples and multiple second gap role prediction probability distribution vectors to obtain a lane change planning model, including: constructing a second loss function; iteratively optimizing the lane change trajectory planning module by minimizing the second loss function based on multiple second training samples and multiple second gap role prediction probability distribution vectors until the lane change trajectory output by the lane change trajectory planning module converges to obtain the lane change planning model.

[0019] In one possible implementation, the second loss function includes the goal achievement loss function, the velocity constraint loss function, and the trajectory error loss function.

[0020] Secondly, this application provides a lane change planning device, comprising:

[0021] The acquisition module is used to acquire the current scene information of the target vehicle. The current scene information includes at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle.

[0022] The generation module is used to generate an enhanced embedding vector corresponding to each dynamic target based on the first initial feature embedding vector corresponding to each dynamic target. The enhanced embedding vector integrates gap role semantic information.

[0023] The interactive modeling module is used to input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the interactive game module for global interactive modeling, so as to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module.

[0024] The output module is used to output at least one lane change trajectory corresponding to the target vehicle based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target.

[0025] In one possible implementation, the generation module is specifically used to: input the first initial feature embedding vector corresponding to each dynamic target into the gap role probability prediction module to perform gap role probability prediction, and obtain the gap role probability distribution vector corresponding to each dynamic target output by the gap role probability prediction module; and generate the enhanced embedding vector corresponding to each dynamic target based on the gap role probability distribution vector corresponding to each dynamic target and the first initial feature embedding vector.

[0026] In one possible implementation, the generation module is further configured to: generate a gap role probability embedding vector corresponding to each dynamic target based on the gap role probability distribution vector corresponding to each dynamic target; and fuse the gap role probability embedding vector corresponding to each dynamic target with the first initial feature embedding vector to generate an enhanced embedding vector corresponding to each dynamic target.

[0027] In one possible implementation, the interactive game module includes a self-attention module and a cross-attention module. The interactive modeling module is specifically used to: input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the self-attention module to perform global feature association modeling, thereby obtaining the first fusion embedding vector corresponding to each dynamic target and the third fusion embedding vector corresponding to the target vehicle output by the self-attention module; input the first fusion embedding vector and the third fusion embedding vector corresponding to each dynamic target into the cross-attention module to perform dynamic target and target vehicle game modeling, thereby obtaining the second fusion embedding vector output by the cross-attention module.

[0028] In one possible implementation, the interactive game module also includes a type embedding vector, which is used to distinguish between target vehicles and dynamic targets.

[0029] In one possible implementation, the type embedding vector includes a first type embedding vector and a second type embedding vector; the first type embedding vector is used to characterize the target vehicle; and the second type embedding vector is used to characterize the dynamic target.

[0030] In one possible implementation, the interactive game module includes a graph neural network, and the interactive modeling module is further used to: input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the graph neural network for global interactive modeling, so as to obtain the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target output by the graph neural network.

[0031] In one possible implementation, the output module is specifically used to: input the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target into the trajectory prediction module to perform lane change trajectory prediction, and obtain at least one lane change trajectory output by the trajectory prediction module.

[0032] In one possible implementation, each lane change trajectory in at least one lane change trajectory includes a confidence level. After outputting at least one lane change trajectory corresponding to the target vehicle, the lane change planning device further includes a determination module (not shown), which is used to: determine the target lane change trajectory corresponding to the target vehicle based on the confidence level corresponding to each lane change trajectory in at least one lane change trajectory.

[0033] In one possible implementation, the first initial feature embedding vector corresponding to each dynamic target is obtained by: acquiring the current environmental information of the target vehicle; inputting the current environmental information into the perception module for dynamic target perception, obtaining at least one dynamic target output by the perception module, and the first initial state feature corresponding to each dynamic target; encoding the first initial state feature corresponding to each dynamic target to obtain the first initial feature embedding vector corresponding to each dynamic target.

[0034] In one possible implementation, the second initial feature embedding vector is obtained by: acquiring the second initial state features of the target vehicle; encoding the second initial state features to obtain the second initial feature embedding vector.

[0035] In one possible implementation, the lane change planning method provided in this application is implemented through a lane change planning model, which includes a gap role probability prediction module and a lane change trajectory planning module. The lane change planning model is trained by acquiring multiple first training samples and multiple second training samples. Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a gap role ground truth probability distribution vector corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are based on the first probability in the multiple first dynamic targets. The gap role probability prediction module is trained based on multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. The gap role probability prediction module is used to predict the gap role probability of dynamic targets in lane change scenarios. The lane change trajectory planning module is trained based on multiple second training samples and multiple second gap role prediction probability distribution vectors to obtain a model for lane change planning. The lane change trajectory planning module is used to output the lane change trajectory of the target vehicle based on the gap role probability. The multiple second gap role prediction probability distribution vectors are obtained by selecting from multiple first gap role prediction probability distribution vectors based on the second probability.

[0036] In one possible implementation, the lane change planning device further includes a model training module (not shown), which is used to: construct a first loss function; and iteratively optimize the gap role probability prediction module by minimizing the first loss function based on multiple first training samples until the gap role prediction probability distribution vector output by the gap role probability prediction module converges.

[0037] In one possible implementation, the model training module is also used to: construct a second loss function; based on multiple second training samples and multiple second gap role prediction probability distribution vectors, iteratively optimize the lane change trajectory planning module by minimizing the second loss function until the lane change trajectory output by the lane change trajectory planning module converges, thus obtaining the lane change planning model.

[0038] In one possible implementation, the second loss function includes the goal achievement loss function, the velocity constraint loss function, and the trajectory error loss function.

[0039] Thirdly, this application provides a vehicle, including: a vehicle body and a lane change planning system, the lane change planning system being used to implement the lane change planning method provided in the first aspect.

[0040] Fourthly, embodiments of this application provide an electronic device, including:

[0041] At least one processor;

[0042] and memory that is communicatively connected to at least one processor;

[0043] The memory is used to store at least one processor-executable instruction, which is executed by at least one processor to enable at least one processor to execute the lane change planning method provided in the first aspect.

[0044] Fifthly, this application also provides a mobile platform that includes electronic devices. Exemplary examples include vehicles, robotic platforms (such as server robots, exploratory robots, scientific research robots, etc.), drones, and other devices, but are not limited thereto.

[0045] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions that are executed by a processor to implement the lane change planning method provided in the first aspect.

[0046] In a seventh aspect, embodiments of this application provide a program product comprising computer-executable instructions. When the computer-executable instructions are executed, they implement the lane change planning method provided in the first aspect.

[0047] The lane change planning method, apparatus, vehicle, device, storage medium, and program product provided in this application embodiment acquire current scene information of the target vehicle, including at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. Based on the first initial feature embedding vector corresponding to each dynamic target, an enhanced embedding vector for fusion gap role semantic information corresponding to each dynamic target is generated. Then, the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into an interactive game module for global interactive modeling to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module. Further, based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target, at least one lane change trajectory corresponding to the target vehicle is output. This application employs an end-to-end deep neural network learning model to directly learn the interactive game strategy between the target vehicle and each dynamic target in the current scene from data. Specifically, it obtains the first initial feature embedding vector corresponding to each dynamic target and the second initial feature embedding vector corresponding to the target vehicle. Then, it generates an enhanced embedding vector that integrates gap role semantic information for each dynamic target. Furthermore, through an interactive game module, it models the global interactive relationship between the enhanced embedding vectors and the second initial feature embedding vectors corresponding to each dynamic target, thereby achieving deep coupling between the enhanced embedding vectors corresponding to each dynamic target and the lane-changing trajectory planning of the target vehicle. Ultimately, it realizes anthropomorphic (learning based on human driving interaction data) lane-changing interactive game. Compared with the lane-changing planning methods in related technologies that rely on manually written rules or cost functions, the lane-changing planning method provided in this application can flexibly adapt to complex and ever-changing real traffic game scenarios, such as long-tail scenarios (e.g., aggressive vehicles suddenly changing lanes) and regional driving style differences, thereby improving the generalization ability of lane-changing planning. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 This is a schematic diagram of a gap-insertion lane change;

[0050] Figure 2 A flowchart illustrating the "pre-decision, post-planning" lane change planning approach;

[0051] Figure 3 A flowchart illustrating the "post-decision, pre-planning" lane change planning approach;

[0052] Figure 4 This is a schematic diagram of the lane change planning system provided in an embodiment of this application;

[0053] Figure 5 This is a flowchart illustrating the lane change planning method provided in the embodiments of this application. Figure 1 ;

[0054] Figure 6 A flowchart illustrating the training method of the lane change planning model provided in this application embodiment;

[0055] Figure 7 This is a schematic diagram of the lane change planning device provided in the embodiments of this application;

[0056] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0057] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0059] In ADAS and AD systems, the gap-insertion lane change is a key operation that reflects the level of driving intelligence and directly affects system safety, user ride comfort, and traffic flow efficiency. Figure 1 This is a schematic diagram of a gap-insertion lane change. (Example) Figure 1 As shown, there are vehicles traveling in front of and behind the lane change gap. In this scenario, the way the autonomous vehicle (autonomous vehicle) changes lanes from its current position to the lane change gap is called gap insertion lane change.

[0060] In related technologies, two lane change planning methods are commonly used to generate lane change trajectories: the "decoupled approach" or the "coupled approach".

[0061] Figure 2 This is a flowchart illustrating the "pre-decision, post-planning" lane change planning approach. (For example...) Figure 2 As shown, the implementation of the "pre-decision, post-planning" lane change planning method may include the following steps:

[0062] 1) Perception and environmental modeling: The system acquires information about the surrounding environment of the autonomous vehicle through sensors such as cameras and radar, and inputs the acquired information into the perception module. The perception module identifies and tracks all dynamic targets, including the target lane.

[0063] 2) Gap selection decision: The sensor module outputs all dynamic targets, including the target lane, as input to the decision module. The decision module then evaluates multiple lane change gaps on the target lane, such as... Figure 2 The diagram shows lane change gaps 1, 2, and N, and selects an optimal lane change gap, such as... Figure 2 The lane change clearance K shown above is usually defined by a pair of "front vehicle" and "rear vehicle". The decision rules of the decision module are based on a series of manually designed heuristic rules (such as safe distance, relative speed, time to collision (TTC), time headway (THW), etc.).

[0064] 3) Planning goal generation: The lane change gap K output by the decision module is converted into a specific planning goal, such as the position of the geometric center point of the lane change gap at a certain future time.

[0065] 4) Trajectory planning: Input the planning target into the trajectory planning module. Based on the planning target, the trajectory planning module generates a smooth trajectory that can guide the vehicle to the target point under the constraints of dynamics, comfort and safety.

[0066] Figure 3 This is a flowchart illustrating the "post-decision, pre-planning" lane change planning approach. Figure 3 As shown, the implementation of the "post-decision, pre-planning" lane change planning method may include the following steps:

[0067] 1) Perception and environmental modeling: The system acquires information about the surrounding environment of the autonomous vehicle through sensors such as cameras and radar, and inputs the acquired information into the perception module. The perception module identifies and tracks all dynamic targets, including the target lane.

[0068] 2) Candidate trajectory generation: The trajectory sampling module inputs all dynamic targets, including the target lane, from the perception module's output. Within the driving corridor surrounding the autonomous vehicle, the trajectory sampling module generates a large number of candidate trajectories covering different lane-changing strategies using sampling methods (such as sampling in the state space or control space). Figure 3 The candidate trajectories shown are 1, 2, 3, 4, and N. These candidate trajectories exhibit diversity in both the lateral and longitudinal directions, representing different lane change timings and speed changes.

[0069] 3) Trajectory evaluation and screening: Input the candidate trajectories generated in step 2) into the trajectory evaluation module. The trajectory evaluation module scores each candidate trajectory based on a cost function. The cost function is composed of multiple weighted sub-items. These sub-items are manually designed by engineers to quantify the trajectory's safety (such as the probability of collision with obstacles), comfort (such as acceleration), efficiency (such as lane change time), compliance (such as whether solid lines are crossed), etc.

[0070] 4) Optimal trajectory selection: The trajectory evaluation and screening module selects the candidate trajectory with the lowest total cost and the highest score as the final lane change trajectory to be executed.

[0071] However, the aforementioned "decoupled approach" and "coupled approach" lane-change planning methods have the following problems:

[0072] 1) It relies on manually designed heuristic rules or cost functions. However, real-world lane-changing games are full of uncertainty and non-cooperative behaviors (such as other vehicles suddenly accelerating to close the gap or decelerating). These complex and subtle interaction patterns are difficult to accurately capture with a limited, discrete set of rules or cost function weights. Therefore, when facing long-tail scenarios (corner cases) outside the rule base or traffic participants with different regional driving styles (such as aggressive and conservative), this approach often exhibits decision-making errors, rigid behavior, or excessive conservatism. It is unable to make anthropomorphic and flexible responses, has poor generalization ability, and is difficult to adapt to complex and ever-changing real-world traffic game scenarios.

[0073] 2) Regarding the "pre-decision planning and post-planning" lane change planning method, this method decouples the two tightly coupled tasks of "choosing which lane change gap" and "how to enter the lane change gap". In each frame, the decision module may reselect the "optimal" lane change gap based on minor environmental changes or perceived noise. This "frame-inconsistent" decision-making will cause the downstream trajectory planning module to frequently jump the planning target it is chasing (for example, deciding to accelerate to overtake car A one second and then deciding to decelerate to let car B the next second). This will cause unnecessary and repeated longitudinal acceleration and deceleration of the vehicle, resulting in a "jerking feeling" and seriously affecting the ride comfort. The ride comfort is poor. At the same time, once the decision module selects the lane change gap, the downstream planning module can only passively "follow" this planning target and cannot actively adjust its strategy to "create" or "maintain" a lane change gap according to the dynamic evolution of the interaction. This one-way information flow without feedback limits the vehicle's active game ability, resulting in a weak game ability.

[0074] 3) Regarding the "post-decision, pre-planning" lane change planning method, although it considers multiple possibilities, its merits are entirely dependent on the design and parameter tuning of a handwritten cost function. Designing a cost function that can accurately quantify the complex game-theoretic behaviors of "decisiveness," "yielding," and "anticipation" in human driving is extremely difficult. The weights of the cost function are unlikely to achieve an optimal balance in all scenarios, potentially leading to a selected trajectory that, while "safe" in some situations, feels very "mechanical" or "inappropriate," lacking the flexibility and social attributes of a human driver and exhibiting poor anthropomorphism. Furthermore, in dense traffic, to cover a sufficient number of lane change possibilities, a massive number of candidate trajectories need to be sampled and evaluated, resulting in a huge computational burden and posing a severe challenge to the real-time performance of the onboard computing platform, leading to high computational costs.

[0075] Based on the technical problems existing in related technologies, the embodiments of this application adopt an end-to-end deep neural network learning model to directly learn the interactive game strategy between the target vehicle and each dynamic target in the current scene from the data. That is, by obtaining the initial feature embedding vectors corresponding to each dynamic target and the initial feature embedding vector corresponding to the target vehicle, an enhanced embedding vector that integrates the semantic information of the gap role is generated for each dynamic target. Furthermore, through the interactive game module, the global interactive relationship model is performed on the enhanced embedding vectors corresponding to each dynamic target and the initial feature embedding vectors corresponding to the target vehicle, realizing the deep coupling between the enhanced embedding vectors corresponding to each dynamic target and the lane change trajectory planning of the target vehicle. Finally, an anthropomorphic (learning based on human driving interaction data) lane change interactive game is realized. Compared with the lane change planning method in related technologies that relies on manually writing rules or cost functions, the lane change planning method provided by this application can flexibly adapt to complex and ever-changing real traffic game scenarios such as long-tail scenarios (such as aggressive vehicles suddenly changing lanes) and regional driving style differences, thereby improving the generalization ability of lane change planning.

[0076] The application scenarios of the embodiments of this application will be described below first.

[0077] The lane change planning method provided in this application is applicable to lane change planning in intelligent driving systems such as ADAS and AD systems. Specifically, it can be applied to lane change planning for autonomous vehicles in multi-lane, high-traffic scenarios such as urban expressways and highways.

[0078] Figure 4 This is a schematic diagram of the lane change planning system provided in an embodiment of this application. Figure 4As shown, the lane change planning system includes a perception module, a first convolutional layer, a soft gap prompt module 41, and a trajectory prediction module. The soft gap prompt module includes a gap role probability prediction module, a second convolutional layer, multiple fusion modules, and an interactive game module.

[0079] The perception module is used to receive environmental information about the target vehicle’s surroundings collected by sensors such as millimeter-wave radar, lidar and cameras integrated in the target vehicle, in order to identify and track all dynamic targets, including the target lane, and obtain the dynamic target perception results output by the perception module, such as the kinematic features corresponding to each dynamic target. The dynamic target perception results can be determined based on the perception results corresponding to static obstacles and road structures.

[0080] The first convolutional layer receives the dynamic target perception results output by the perception module, encodes the dynamic perception results into features, and generates a first initial feature embedding vector corresponding to each dynamic target, such as... Figure 4 The first initial feature embedding vectors corresponding to dynamic target 1, dynamic target 2, and dynamic target N are shown in the figure.

[0081] For example, the first convolutional layer can be a multi-layer convolutional encoder.

[0082] The gap role probability prediction module receives the first initial feature embedding vector corresponding to each dynamic target output by the first convolutional layer, performs gap role probability prediction based on the first initial feature embedding vector corresponding to each dynamic target, and outputs the gap role probability distribution vector corresponding to each dynamic target, such as... Figure 4 The diagram shows the probability distribution vectors of the gap roles corresponding to dynamic target 1, dynamic target 2, and dynamic target N.

[0083] The second convolutional layer is used to map the probability distribution vector of the gap role corresponding to each dynamic target to a continuous, high-dimensional semantic space, generating the probability embedding vector of the gap role corresponding to each dynamic target.

[0084] Optionally, mapping the probability distribution vector of the gap role corresponding to each dynamic target to a continuous, high-dimensional semantic space can also be achieved through a multi-layer perceptron (MLP).

[0085] The fusion module is used to fuse the first initial feature embedding vector and the gap role probability embedding vector corresponding to each dynamic target to obtain the enhanced embedding vector corresponding to each dynamic target.

[0086] The interactive game module is used to perform global interactive modeling on the enhanced embedding vector corresponding to each dynamic target and the second initial feature embedding vector corresponding to the target vehicle, so as to obtain the first fused embedding vector corresponding to each dynamic target and the second fused embedding vector corresponding to the target vehicle.

[0087] The trajectory prediction module is used to predict lane change trajectories based on the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle, and output at least one lane change trajectory corresponding to the target vehicle.

[0088] from Figure 4 As can be seen from the above, the embodiments of this application adopt a "soft gap prompt" mechanism to deeply integrate lane change gap recognition (in probabilistic form) and trajectory planning (through an interactive game module) into the same model, which solves the problems of poor comfort and weak game ability caused by the decoupling of the two in traditional methods.

[0089] The following is based on the above. Figure 4 The lane change planning system shown is the execution entity. The implementation method of the lane change planning method provided in this application embodiment will be described in detail with reference to specific embodiments.

[0090] Figure 5 A flowchart illustrating the lane change planning method provided in this application embodiment. Figure 1 .like Figure 5 As shown, the implementation of this lane change planning method may include the following steps:

[0091] S501, Obtain the current scene information of the target vehicle, which includes at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle.

[0092] For example, the target vehicle can be a vehicle that integrates an ADAS system or an AD system.

[0093] For example, dynamic targets can be other motor vehicles, non-motor vehicles, and pedestrians in the current driving scenario besides the target vehicle.

[0094] In this step, the first initial feature embedding vector corresponding to each dynamic target can be obtained in the following way: First, the current surrounding environment information corresponding to the current driving scenario of the target vehicle is collected by sensors such as millimeter-wave radar, lidar, and cameras integrated in the target vehicle; then, the surrounding environment information is input as follows: Figure 4 The perception module shown identifies and tracks all dynamic targets, including the target lane, and obtains the dynamic target perception results output by the perception module. Finally, the dynamic target perception results are input as shown in the figure. Figure 4The first convolutional layer shown in the figure encodes features to generate the first initial feature embedding vector corresponding to each dynamic target.

[0095] For example, the dynamic target perception result includes multiple dynamic targets and the kinematic features corresponding to each dynamic target, such as kinematic information (position, speed, acceleration) relative to the target vehicle, geometric information (length, height, orientation, width), and category information (pedestrian, electric vehicle, truck, car), etc.

[0096] In this step, the second initial feature embedding vector corresponding to the target vehicle can be obtained in the following way: the initial state features of the target vehicle, such as the current speed, current acceleration, current position, current orientation, and historical speed of the target vehicle, are collected by sensors such as wheel speed sensors and inertial measurement units integrated in the target vehicle; the collected initial state features are extracted by high-dimensional semantic information through convolutional layers or MLP layers to obtain the second initial feature embedding vector corresponding to the target vehicle.

[0097] S502, based on the first initial feature embedding vector corresponding to each dynamic target, generate an enhanced embedding vector corresponding to each dynamic target, and fuse the gap role semantic information into the enhanced embedding vector.

[0098] For example, the gap role semantic information can be the role played by each dynamic target under the current lane change intention, such as a non-lane change gap vehicle, a vehicle in front of the lane change gap, and a vehicle behind the lane change gap. Among them, a non-lane change gap vehicle is a vehicle unrelated to the lane change gap, a vehicle in front of the lane change gap is a vehicle traveling in front of the lane change gap after the target vehicle changes lanes, and a vehicle behind the lane change gap is a vehicle traveling behind the lane change gap after the target vehicle changes lanes.

[0099] Optionally, this step may include the following steps:

[0100] S5021, the first initial feature embedded vector corresponding to each dynamic target is input into the gap role probability prediction module to perform gap role probability prediction, and the gap role probability distribution vector corresponding to each dynamic target is obtained from the output of the gap role probability prediction module.

[0101] The gap role probability prediction module, such as Figure 4 As shown, for example, the gap role probability prediction module can be an MLP layer. An MLP can contain multiple linear layers, and different linear layers are connected by activation functions (such as ReLU activation function or Sigmoid activation function). This application embodiment does not limit the number of linear layers contained in the MLP; it can be determined according to actual application requirements.

[0102] For example, the probability distribution vector of the gap role corresponding to each dynamic target can be a three-dimensional probability vector, and the probability distribution vector of the gap role corresponding to each dynamic target can be expressed as:

[0103]

[0104] Among them, P j This represents the probability distribution vector of the gap role corresponding to dynamic target j. This represents the probability that the dynamic target j corresponds to a vehicle that does not change lanes during a gap. This represents the probability that the dynamic target j corresponds to the vehicle in front during the lane change interval. This represents the probability of the vehicle following the lane change gap for the dynamic target j.

[0105] S5022, based on the gap role probability distribution vector and the first initial feature embedding vector corresponding to each dynamic target, generate the enhanced embedding vector corresponding to each dynamic target.

[0106] In this step, one possible implementation is as follows: input the probability distribution vector of the gap role corresponding to each dynamic target into an MLP layer, and map the discrete gap role probability information to a continuous, high-dimensional semantic space through the MLP layer to generate the gap role probability embedding vector corresponding to each dynamic target; fuse the gap role probability embedding vector corresponding to each dynamic target with the first initial feature embedding vector to generate the enhanced embedding vector corresponding to each dynamic target.

[0107] Another possible implementation is to input the gap role probability distribution vector corresponding to each dynamic target as follows: Figure 4 The second convolutional layer shown in the figure maps the discrete gap role probability information to a continuous, high-dimensional semantic space, generating a gap role probability embedding vector corresponding to each dynamic target; the gap role probability embedding vector corresponding to each dynamic target is fused with the first initial feature embedding vector to generate an enhanced embedding vector corresponding to each dynamic target.

[0108] Another possible implementation is to input the probability distribution vector of the gap role corresponding to each dynamic target into the self-attention module, and then use the self-attention module to map the discrete gap role probability information to a continuous, high-dimensional semantic space to generate the gap role probability embedding vector corresponding to each dynamic target; and then fuse the gap role probability embedding vector corresponding to each dynamic target with the first initial feature embedding vector to generate the enhanced embedding vector corresponding to each dynamic target.

[0109] Another possible implementation is to fuse the gap role probability distribution vector corresponding to each dynamic target with the first initial feature embedding vector to generate an enhanced embedding vector corresponding to each dynamic target.

[0110] For example, the fusion processing can be performed by addition, concatenation, gated fusion, or cross-attention fusion. This application does not limit the specific method, but can be determined according to the actual application requirements.

[0111] For example, when the fusion process involves splicing, it can be represented by the following formula:

[0112]

[0113] in, This represents the enhanced embedding vector corresponding to the dynamic target j. This represents the first initial feature embedding vector corresponding to the dynamic target j. This represents the gap role probability embedding vector corresponding to dynamic target j.

[0114] S503, input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the interactive game module for global interactive modeling, and obtain the first fused embedding vector corresponding to each dynamic target and the second fused embedding vector corresponding to the target vehicle output by the interactive game module.

[0115] In this step, one possible implementation is to construct a sequence from the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target, with the input as follows: Figure 4 The interactive game module shown in the figure models the complex interactive relationships between all traffic participants, i.e., all dynamic targets and target vehicles, and obtains the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module.

[0116] Understandably, by modeling the complex interactive relationships between all traffic participants—that is, all dynamic targets and target vehicles—through an interactive game theory module, the representation of target vehicles can "understand" the entire interactive scenario. For example, it can identify which dynamic target is a potential "car behind during a lane change gap" requiring acceleration to overtake; which dynamic target is a potential "car in front during a lane change gap" requiring deceleration to wait or coordinate; and which dynamic targets are irrelevant vehicles and can be ignored. This global, end-to-end interactive modeling enables anthropomorphic interactive game theory.

[0117] S504, based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target, output at least one lane change trajectory corresponding to the target vehicle.

[0118] In this step, one possible implementation is to decode the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target to generate at least one lane change trajectory corresponding to the target vehicle.

[0119] It is understandable that at least one lane-changing trajectory corresponding to the target vehicle implicitly contains anthropomorphic lane-changing game behavior.

[0120] This application employs an end-to-end deep neural network learning model to directly learn the interactive game strategy between the target vehicle and each dynamic target in the current scene from the data. Specifically, it obtains the first initial feature embedding vector corresponding to each dynamic target and the second initial feature embedding vector corresponding to the target vehicle. Then, it generates an enhanced embedding vector that integrates gap role semantic information for each dynamic target. Furthermore, through the interactive game module, it models the global interactive relationship between the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target, thereby achieving deep coupling between the enhanced embedding vector corresponding to each dynamic target and the lane change trajectory planning of the target vehicle. Ultimately, it realizes anthropomorphic (learning based on human driving interaction data) lane change interactive game. Compared with the lane change planning methods in related technologies that rely on manually writing rules or cost functions, the lane change planning method provided in this application can flexibly adapt to complex and ever-changing real traffic game scenarios such as long-tail scenarios (such as aggressive vehicles suddenly changing lanes) and regional driving style differences, thereby improving the generalization ability of lane change planning.

[0121] It is understood that in the lane change planning method provided in this application embodiment, in the lane change scenario, an end-to-end deep neural network learning method is used to replace handwritten rules and cost functions, realizing a paradigm shift from "instructional" to "learning". By coupling the gap role recognition of dynamic targets with the lane change trajectory planning of target vehicles, the core idea of ​​anthropomorphic game is realized, and the generalization ability of lane change planning is improved.

[0122] Optionally, one possible implementation of generating an enhanced embedding vector for each dynamic target based on the gap role probability distribution vector and the first initial feature embedding vector can be as follows: generate a gap role probability embedding vector for each dynamic target based on the gap role probability distribution vector; and fuse the gap role probability embedding vector and the first initial feature embedding vector to generate an enhanced embedding vector for each dynamic target.

[0123] The specific implementation method is similar to that described above, and will not be repeated here.

[0124] It is understood that, by predicting and embedding the gap role probability of dynamic targets, this application embodiment can protect the classification of dynamic targets in the scene into gap semantic roles such as "vehicle in front of lane change gap", "vehicle behind lane change gap" and "vehicle without lane change gap", and encode the classification probability as a kind of "soft hint" as a feature embedding to enhance the understanding of the interaction scene.

[0125] Optionally, the interactive game module includes a self-attention module and a cross-attention module.

[0126] Optionally, step S503, which inputs the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the interactive game module for global interactive modeling, to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module, can be implemented as follows: inputting the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the self-attention module for global feature association modeling, to obtain the first fusion embedding vector corresponding to each dynamic target and the third fusion embedding vector corresponding to the target vehicle output by the self-attention module; inputting the first fusion embedding vector and the third fusion embedding vector corresponding to each dynamic target into the cross-attention module for dynamic target and target vehicle game modeling, to obtain the second fusion embedding vector output by the cross-attention module.

[0127] In this implementation, a multi-head self-attention mechanism can be adopted. In the self-attention module, for each dynamic target, the pairwise attention weights between the dynamic target and other dynamic targets, as well as the target vehicle, are calculated. Using these attention weights, the enhanced embedding vector corresponding to the dynamic target is weighted and summed with the enhanced embedding vectors corresponding to other dynamic targets and the second initial feature embedding vector corresponding to the target vehicle to obtain the first fused embedding vector corresponding to each dynamic target. For the target vehicle, the attention weights between the target vehicle and all dynamic targets are calculated. Using these attention weights, the second initial feature embedding vector corresponding to the target vehicle is weighted and summed with the enhanced embedding vectors corresponding to all dynamic targets to obtain the third fused embedding vector corresponding to the target vehicle. In the cross-attention module, the third fused embedding vector corresponding to the target vehicle is used for game modeling with the first fused embedding vectors corresponding to all dynamic targets to obtain the second fused embedding vector corresponding to the target vehicle.

[0128] Understandably, the above calculation process is executed in parallel by multiple attention heads, and the outputs of all attention heads are further concatenated to form the final fused embedding vector in order to fully capture the correlation information in different dimensions.

[0129] Optionally, the interactive game module also includes a type embedding vector, which is used to distinguish between target vehicles and dynamic targets.

[0130] Understandably, by using this type of embedded vector, the interactive game module can accurately distinguish between dynamic targets and target vehicles when performing self-attention calculations.

[0131] Optionally, the type embedding vector includes a first type embedding vector and a second type embedding vector; the first type embedding vector is used to characterize the target vehicle; and the second type embedding vector is used to characterize the dynamic target.

[0132] For example, the first type of embedding vector is a learnable embedding vector, which can be represented as an Ego-type embedding.

[0133] For example, a second type of embedding vector can be represented as an MOT-type embedding.

[0134] Optionally, the interactive game module includes a graph neural network (GNN).

[0135] Optionally, step S503, which inputs the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the interactive game module for global interactive modeling, to obtain the first fused embedding vector corresponding to each dynamic target and the second fused embedding vector corresponding to the target vehicle output by the interactive game module, can be implemented as follows: inputting the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the graph neural network for global interactive modeling, to obtain the first fused embedding vector and the second fused embedding vector corresponding to each dynamic target output by the graph neural network.

[0136] In this implementation, all dynamic targets and target vehicles can be constructed into a graph, where nodes represent traffic participants (i.e., all dynamic targets and target vehicles), and edges represent the spatial relationships or interactions between traffic participants. GNN, through a message-passing mechanism, can effectively model complex interactive games, enabling global interactive modeling of the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target.

[0137] Optionally, step S504, which outputs at least one lane change trajectory corresponding to the target vehicle based on the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle, can be implemented by inputting the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target into the trajectory prediction module to predict the lane change trajectory, thereby obtaining at least one lane change trajectory output by the trajectory prediction module.

[0138] For example, the trajectory prediction module is as described above. Figure 4 As shown, it can be an MLP layer.

[0139] In this implementation, the first and second fusion embedding vectors corresponding to each dynamic target are reduced in dimensionality by the trajectory prediction module to predict the lane change trajectory.

[0140] For example, each lane change trajectory in at least one lane change trajectory may contain t time steps, and the dimension of the trajectory point corresponding to each time step may be 2, such as the horizontal spatial position dimension and the vertical spatial position dimension of the trajectory point.

[0141] Optionally, each lane change trajectory in at least one lane change trajectory includes a confidence level. The lane change planning method provided in this application embodiment further includes, after outputting at least one lane change trajectory corresponding to the target vehicle, determining the target lane change trajectory corresponding to the target vehicle based on the confidence level corresponding to each lane change trajectory in at least one lane change trajectory.

[0142] For example, the lane change trajectory with the highest confidence level is determined as the target lane change trajectory, and this target lane change trajectory is used as the lane change trajectory finally executed by the target vehicle.

[0143] Optionally, the first initial feature embedding vector corresponding to each dynamic target is obtained by: acquiring the current environmental information of the target vehicle; inputting the current environmental information into the perception module for dynamic target perception, obtaining at least one dynamic target output by the perception module, and the first initial state feature corresponding to each dynamic target; encoding the first initial state feature corresponding to each dynamic target to obtain the first initial feature embedding vector corresponding to each dynamic target.

[0144] For example, current environmental information can be obtained through sensors such as millimeter-wave radar, lidar, and cameras integrated on the target vehicle.

[0145] The perception module is as described above. Figure 4 As shown, the specific implementation of this embodiment is similar to that described above, and will not be repeated here.

[0146] Optionally, the second initial feature embedding vector corresponding to the target vehicle is obtained by: acquiring the second initial state features of the target vehicle; encoding the second initial state features to obtain the second initial feature embedding vector.

[0147] The specific implementation method is similar to that described above, and will not be repeated here.

[0148] It is understood that the lane change planning method provided in the above embodiments can be implemented using an end-to-end deep learning model, such as a lane change planning model, which includes a gap role probability prediction module and a lane change trajectory planning module. The gap role probability prediction module, for example... Figure 4 As shown, the lane change trajectory planning module includes, for example: Figure 4 The interactive game module and trajectory prediction module are shown in the figure.

[0149] The following is combined with Figure 6 The training method of the lane change planning model provided in the embodiments of this application will be described in detail.

[0150] Figure 6 This is a schematic flowchart illustrating the training method of the lane change planning model provided in an embodiment of this application. Figure 6 As shown, the training method for this lane change planning model includes the following steps:

[0151] S601, acquire multiple first training samples and multiple second training samples.

[0152] Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a gap role ground truth probability distribution vector corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are selected from the multiple first dynamic targets based on the first probability.

[0153] For example, multiple first training samples and multiple second training samples are all samples corresponding to successful lane changes in history.

[0154] The first initial feature embedding vector corresponding to each first dynamic target and the second initial feature embedding vector corresponding to the target vehicle are similar to those described above, and will not be repeated here.

[0155] In this step, for each second training sample, the gap where the target vehicle inserts at the moment the lane change is completed is located. The vehicle in front of and behind the gap is labeled as the ground truth (one-hot encoding) of "vehicle in front of lane change gap" and "vehicle behind lane change gap", respectively. All other vehicles are labeled as "vehicles not in lane change gap", thus obtaining the ground truth probability distribution vector of the gap role corresponding to each second dynamic target.

[0156] For example, the first probability can be 30%, meaning that when training the lane change trajectory planning module, a hybrid training strategy is used, such as using a 30% ground truth probability distribution vector of the gap role and a 70% predicted probability distribution vector of the gap role to train the lane change trajectory planning module. This application embodiment does not limit the magnitude of the first probability; it can be determined according to the actual application requirements.

[0157] In one possible implementation, multiple second dynamic targets are randomly selected from multiple first dynamic targets based on a first probability.

[0158] S602, the gap role probability prediction module is trained based on multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. The gap role probability prediction module is used to predict the gap role probability of dynamic targets in lane change scenarios.

[0159] Optionally, one possible implementation of this step is to: construct a first loss function; based on multiple first training samples, iteratively optimize the gap role probability prediction module by minimizing the first loss function until the gap role prediction probability distribution vector output by the gap role probability prediction module converges.

[0160] For example, the first loss function can be the cross-entropy loss function.

[0161] S603, based on multiple second training samples and multiple second gap role prediction probability distribution vectors, train the lane change trajectory planning module to obtain a model for lane change planning. This lane change trajectory planning module is used to output the lane change trajectory of the target vehicle based on the gap role probability.

[0162] Among them, the multiple second gap role prediction probability distribution vectors are obtained by filtering the multiple first gap role prediction probability distribution vectors based on the second probability.

[0163] For example, the second probability can be 70%. This application does not limit the magnitude of the second probability; it can be determined according to the actual application requirements.

[0164] In one possible implementation, the multiple second gap role prediction probability distribution vectors are obtained by randomly selecting from multiple first gap role prediction probability distribution vectors based on the second probability.

[0165] Optionally, one possible implementation of this step is to: construct a second loss function; based on multiple second training samples and multiple second gap role prediction probability distribution vectors, iteratively optimize the lane change trajectory planning module by minimizing the second loss function until the lane change trajectory output by the lane change trajectory planning module converges, thereby obtaining the lane change planning model.

[0166] It is understood that the hybrid training strategy adopted in the embodiments of this application for training the lane change trajectory planning module is to accelerate the convergence of the lane change trajectory planning module and improve the robustness of the lane change planning model. Specifically, 1) the first probability (30% probability) adopts the "Teach Forcing" mode, that is, it directly uses multiple second training samples containing the ground truth probability distribution vector of the gap role to generate the gap role probability embedding vector, forcing the lane change trajectory planning module to learn how to plan the trajectory when the correct lane change gap is known; 2) the second probability (70% probability) uses multiple second gap role prediction probability distribution vectors predicted by the gap role probability prediction module to generate the gap role probability embedding vector, forcing the lane change trajectory planning module to learn how to make robust decisions from uncertain predictions, thereby improving the generalization ability of the model.

[0167] Optionally, the second loss function includes the goal achievement loss function, the velocity constraint loss function, and the trajectory error loss function.

[0168] It should be noted that, in order to enable the lane change planning model to not only identify lane change gaps, but also to actively create and utilize lane change gaps through acceleration and deceleration, a customized target achievement loss function and speed constraint loss function are introduced when supervising at least one lane change trajectory output by the lane change planning model.

[0169] For example, the goal achievement loss function can be the longitudinal and lateral goal achievement loss (Goal Achievement Loss). This goal achievement loss function requires that the lane change trajectory planned by the lane change planning model, within the time T taken to complete the lane change in the ground truth data, should have a longitudinal travel distance that allows it to reach the middle position of the ground truth lane change gap, while the lateral displacement should be sufficient to complete the lane change.

[0170] For example, the goal achievement loss function can be expressed by the following formula:

[0171]

[0172] in, T represents the loss function for achieving the goal. gt S indicates the moment when the target vehicle completes the lane change. ego (T gt The ) represents the moment the lane change is completed, indicating the longitudinal position of the target vehicle at that moment, i.e., the longitudinal position of the target vehicle (compared to time 0). gap\_mid\_gt Indicates the longitudinal position of the lane change clearance (e.g., how many meters behind the vehicle in front when there is only one vehicle in front, or how many meters in front of the vehicle behind when there is only one vehicle behind), Y ego (T gtY represents the moment the lane change is completed, indicating the lateral position of the target vehicle at that moment, i.e., the lateral position of the target vehicle (compared to time 0). target\_lane Indicates the lateral position of the lane change clearance (e.g., if there is only one vehicle in front, how many meters behind the vehicle in front; if there is only one vehicle behind, how many meters in front of the vehicle behind).

[0173] For example, the speed constraint loss function could be the Average Speed ​​Loss, which requires that the average longitudinal speed of the planned trajectory within the time T taken to complete the lane change in the ground truth data should match the average speed required to travel from the initial position to the middle of the target lane change gap. This ensures that the planned acceleration and deceleration are dynamically feasible and reasonable, rather than instantaneous position jumps.

[0174] For example, the velocity constraint loss function can be expressed by the following formula:

[0175]

[0176] in, Let v represent the velocity constraint loss function. ego (t) represents the longitudinal velocity of the target vehicle at time t, S ego (0) indicates the longitudinal position of the target vehicle at time 0.

[0177] It is understood that the embodiments of this application explicitly supervise the model to learn proactive acceleration and deceleration game behavior through customized goal achievement loss function and speed constraint loss function, rather than passively searching for existing gaps, thereby improving the flexibility of lane change planning.

[0178] For example, the trajectory error loss function can be expressed by the following formula:

[0179]

[0180] Among them, L limitation This represents the trajectory error loss function, with dimensions [1, T, 2], where T... raj_gt T represents the human driving trajectory. raj_pred This represents the predicted lane change trajectory, with dimensions [1, T, 2].

[0181] For example, the second loss function can be expressed by the following formula:

[0182]

[0183] In this embodiment, multiple first training samples and multiple second training samples are obtained. Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a ground truth probability distribution vector of the gap role corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are selected from the multiple first dynamic targets based on a first probability. Then, the gap role probability prediction module is trained based on the multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. Based on the multiple second training samples and the multiple first gap role prediction probability distribution vectors, the lane change trajectory planning module is trained to obtain a model for lane change planning. By adopting a hybrid training strategy for model training, the generalization ability of the lane change planning model is improved.

[0184] In summary, the lane change planning method provided in this application has the following beneficial effects:

[0185] 1) The lane change planning method provided in this application is based on the core logic of obtaining the lane change trajectory through data-driven learning rather than manual coding. Therefore, it can better adapt to various long-tail scenarios and different driving styles in the training data. At the same time, the use of a hybrid training strategy can further enhance the tolerance of the lane change planning model to its own prediction error, so that the lane change planning model can show stronger robustness and adaptability when facing complex interaction scenarios that have not been accurately coded, reduce the risk of functional failure due to scenario changes, and reduce the customized development cost for driving habits in different regions.

[0186] 2) In the lane change planning method provided in this application embodiment, the interactive game module learns from massive data how human drivers interact with other vehicles in dense traffic flow through subtle acceleration and deceleration, rather than relying on fixed rules. This enables the lane change planning model to make more flexible, decisive, and predictive decisions. Compared to the traditional method, which may abandon lane change in congested scenarios due to the lack of a safe gap, the lane change planning method provided in this application embodiment can successfully "create" lane change opportunities through active longitudinal game, thereby significantly improving the success rate of lane change in complex traffic environments.

[0187] 3) In the lane change planning method provided in this application embodiment, by unifying lane change gap recognition and trajectory planning within an end-to-end framework, and using continuous gap role probability embedding vectors to replace discrete, frame-to-frame lane change gap decisions, the longitudinal acceleration and deceleration jitter caused by frequent switching of planning targets in the "pre-decision and post-planning" method of related technologies is avoided. This results in smoother and more continuous speed changes during lane changes, which is in line with human driving habits, greatly reduces the "jerking feeling", and improves the user's ride comfort and trust in the assisted driving system.

[0188] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0189] Figure 7 This is a schematic diagram of the lane change planning device provided in an embodiment of this application. Figure 7 As shown, the lane change planning device 70 includes: an acquisition module 710, a generation module 720, an interactive modeling module 730, and an output module 740.

[0190] The acquisition module 710 is used to acquire the current scene information of the target vehicle. The current scene information includes at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle.

[0191] The generation module 720 is used to generate an enhanced embedding vector corresponding to each dynamic target based on the first initial feature embedding vector corresponding to each dynamic target. The enhanced embedding vector integrates gap role semantic information.

[0192] The interactive modeling module 730 is used to input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the interactive game module for global interactive modeling, so as to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module.

[0193] The output module 740 is used to output at least one lane change trajectory corresponding to the target vehicle based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target.

[0194] In one possible implementation, the generation module 720 is specifically used to: input the first initial feature embedding vector corresponding to each dynamic target into the gap role probability prediction module to perform gap role probability prediction, and obtain the gap role probability distribution vector corresponding to each dynamic target output by the gap role probability prediction module; and generate the enhanced embedding vector corresponding to each dynamic target based on the gap role probability distribution vector corresponding to each dynamic target and the first initial feature embedding vector.

[0195] In one possible implementation, the generation module 720 is further configured to: generate a gap role probability embedding vector corresponding to each dynamic target based on the gap role probability distribution vector corresponding to each dynamic target; and fuse the gap role probability embedding vector corresponding to each dynamic target with the first initial feature embedding vector to generate an enhanced embedding vector corresponding to each dynamic target.

[0196] In one possible implementation, the interactive game module includes a self-attention module and a cross-attention module. The interactive modeling module 730 is specifically used to: input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the self-attention module to perform global feature association modeling, thereby obtaining the first fusion embedding vector corresponding to each dynamic target and the third fusion embedding vector corresponding to the target vehicle output by the self-attention module; input the first fusion embedding vector and the third fusion embedding vector corresponding to each dynamic target into the cross-attention module to perform dynamic target and target vehicle game modeling, thereby obtaining the second fusion embedding vector output by the cross-attention module.

[0197] In one possible implementation, the interactive game module also includes a type embedding vector, which is used to distinguish between target vehicles and dynamic targets.

[0198] In one possible implementation, the type embedding vector includes a first type embedding vector and a second type embedding vector; the first type embedding vector is used to characterize the target vehicle; and the second type embedding vector is used to characterize the dynamic target.

[0199] In one possible implementation, the interactive game module includes a graph neural network, and the interactive modeling module 730 is further configured to: input the enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target into the graph neural network for global interactive modeling, so as to obtain the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target output by the graph neural network.

[0200] In one possible implementation, the output module 740 is specifically used to: input the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target into the trajectory prediction module to perform lane change trajectory prediction, and obtain at least one lane change trajectory output by the trajectory prediction module.

[0201] In one possible implementation, each lane change trajectory in at least one lane change trajectory includes a confidence level. After outputting at least one lane change trajectory corresponding to the target vehicle, the lane change planning device further includes a determination module (not shown), which is used to: determine the target lane change trajectory corresponding to the target vehicle based on the confidence level corresponding to each lane change trajectory in at least one lane change trajectory.

[0202] In one possible implementation, the first initial feature embedding vector corresponding to each dynamic target is obtained by: acquiring the current environmental information of the target vehicle; inputting the current environmental information into the perception module for dynamic target perception, obtaining at least one dynamic target output by the perception module, and the first initial state feature corresponding to each dynamic target; encoding the first initial state feature corresponding to each dynamic target to obtain the first initial feature embedding vector corresponding to each dynamic target.

[0203] In one possible implementation, the second initial feature embedding vector is obtained by: acquiring the second initial state features of the target vehicle; encoding the second initial state features to obtain the second initial feature embedding vector.

[0204] In one possible implementation, the lane change planning method provided in this application is implemented through a lane change planning model, which includes a gap role probability prediction module and a lane change trajectory planning module. The lane change planning model is trained by acquiring multiple first training samples and multiple second training samples. Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a gap role ground truth probability distribution vector corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are based on the first probability in the multiple first dynamic targets. The gap role probability prediction module is trained based on multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. The gap role probability prediction module is used to predict the gap role probability of dynamic targets in lane change scenarios. The lane change trajectory planning module is trained based on multiple second training samples and multiple second gap role prediction probability distribution vectors to obtain a model for lane change planning. The lane change trajectory planning module is used to output the lane change trajectory of the target vehicle based on the gap role probability. The multiple second gap role prediction probability distribution vectors are obtained by selecting from multiple first gap role prediction probability distribution vectors based on the second probability.

[0205] In one possible implementation, the lane change planning device further includes a model training module (not shown), which is used to: construct a first loss function; and iteratively optimize the gap role probability prediction module by minimizing the first loss function based on multiple first training samples until the gap role prediction probability distribution vector output by the gap role probability prediction module converges.

[0206] In one possible implementation, the model training module is also used to: construct a second loss function; based on multiple second training samples and multiple second gap role prediction probability distribution vectors, iteratively optimize the lane change trajectory planning module by minimizing the second loss function until the lane change trajectory output by the lane change trajectory planning module converges, thus obtaining the lane change planning model.

[0207] In one possible implementation, the second loss function includes the goal achievement loss function, the velocity constraint loss function, and the trajectory error loss function.

[0208] The lane change planning device provided in this embodiment can be used to execute the method steps of the above method embodiment. The specific implementation and technical effects are similar, and will not be described again here.

[0209] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0210] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-On-a-Chip (SOC).

[0211] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Video Discs, DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0212] This application also provides a vehicle, including: a vehicle body and a lane change planning system, the lane change planning system being used to execute the lane change planning method provided in the above embodiments.

[0213] For example, the lane change planning system is as described above. Figure 4 As shown.

[0214] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 80 provided in this embodiment includes at least one processor 801 and a memory 802. Optionally, the electronic device 80 further includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus 804.

[0215] In a specific implementation, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to perform the above-described method.

[0216] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0217] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0218] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0219] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0220] This application also provides a mobile platform, which includes electronic devices. Exemplary examples include vehicles, robotic platforms (such as server robots, exploratory robots, scientific research robots, etc.), drones, and other devices, but are not limited thereto.

[0221] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0222] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0223] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0224] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0225] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0226] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0227] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0228] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0229] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0230] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A lane change planning method, characterized in that, include: Obtain the current scene information of the target vehicle, wherein the current scene information includes at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle; Based on the first initial feature embedding vector corresponding to each dynamic target, an enhanced embedding vector corresponding to each dynamic target is generated, and the enhanced embedding vector is fused with gap role semantic information; The enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into the interactive game module for global interactive modeling to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module. Based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target, at least one lane change trajectory corresponding to the target vehicle is output.

2. The lane change planning method according to claim 1, characterized in that, The step of generating an enhanced embedding vector corresponding to each dynamic target based on the first initial feature embedding vector corresponding to each dynamic target includes: The first initial feature embedding vector corresponding to each dynamic target is input into the gap role probability prediction module to perform gap role probability prediction, thereby obtaining the gap role probability distribution vector corresponding to each dynamic target output by the gap role probability prediction module. Based on the gap role probability distribution vector and the first initial feature embedding vector corresponding to each dynamic target, an enhanced embedding vector corresponding to each dynamic target is generated.

3. The lane change planning method according to claim 2, characterized in that, The step of generating an enhanced embedding vector corresponding to each dynamic target based on the gap role probability distribution vector and the first initial feature embedding vector for each dynamic target includes: Based on the probability distribution vector of the gap role corresponding to each dynamic target, generate the probability embedding vector of the gap role corresponding to each dynamic target; The gap role probability embedding vector and the first initial feature embedding vector corresponding to each dynamic target are fused to generate the enhanced embedding vector corresponding to each dynamic target.

4. The lane change planning method according to any one of claims 1 to 3, characterized in that, The interactive game module includes a self-attention module and a cross-attention module. The enhanced embedding vector corresponding to each dynamic target and the second initial feature embedding vector are input into the interactive game module for global interactive modeling, resulting in a first fused embedding vector corresponding to each dynamic target and a second fused embedding vector corresponding to the target vehicle output by the interactive game module. This includes: The enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into the self-attention module for global feature association modeling, so as to obtain the first fusion embedding vector corresponding to each dynamic target and the third fusion embedding vector corresponding to the target vehicle output by the self-attention module. The first fusion embedding vector and the third fusion embedding vector corresponding to each dynamic target are input into the cross-attention module to perform dynamic target and target vehicle game modeling, and the second fusion embedding vector output by the cross-attention module is obtained.

5. The lane change planning method according to claim 4, characterized in that, The interactive game module also includes a type embedding vector, which is used to distinguish between the target vehicle and the dynamic target.

6. The lane change planning method according to claim 5, characterized in that, The type embedding vector includes a first type embedding vector and a second type embedding vector; The first type of embedding vector is used to characterize the target vehicle; The second type of embedding vector is used to characterize the dynamic target.

7. The lane change planning method according to any one of claims 1 to 3, characterized in that, The interactive game module includes a graph neural network. The step of inputting the enhanced embedding vector corresponding to each dynamic target and the second initial feature embedding vector into the interactive game module for global interactive modeling, to obtain the first fused embedding vector corresponding to each dynamic target and the second fused embedding vector corresponding to the target vehicle output by the interactive game module, includes: The enhanced embedding vector and the second initial feature embedding vector corresponding to each dynamic target are input into the graph neural network for global interactive modeling, so as to obtain the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target output by the graph neural network.

8. The lane change planning method according to any one of claims 1 to 3, characterized in that, The step of outputting at least one lane-change trajectory corresponding to the target vehicle based on the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle includes: The first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target are input into the trajectory prediction module to perform lane change trajectory prediction, thereby obtaining the at least one lane change trajectory output by the trajectory prediction module.

9. The lane change planning method according to claim 8, characterized in that, Each lane change trajectory in the at least one lane change trajectory includes a confidence level, and after outputting the at least one lane change trajectory corresponding to the target vehicle, the following is also included: For each of the at least one lane change trajectory, the target lane change trajectory corresponding to the target vehicle is determined based on the confidence level corresponding to the lane change trajectory.

10. The lane change planning method according to any one of claims 1 to 3, characterized in that, The first initial feature embedding vector corresponding to each dynamic target is obtained in the following way: Obtain the current environmental information of the target vehicle; The current environment information is input into the perception module for dynamic target perception, and at least one dynamic target output by the perception module is obtained, as well as the first initial state feature corresponding to each dynamic target. The first initial state features corresponding to each dynamic target are encoded to obtain the first initial feature embedding vector corresponding to each dynamic target.

11. The lane change planning method according to any one of claims 1 to 3, characterized in that, The second initial feature embedding vector is obtained in the following way: Obtain the second initial state features of the target vehicle; The second initial state features are encoded to obtain the second initial feature embedding vector.

12. The lane change planning method according to any one of claims 1 to 3, characterized in that, The lane change planning method is implemented through a lane change planning model, which includes a gap role probability prediction module and a lane change trajectory planning module. The lane change planning model is trained in the following way: Multiple first training samples and multiple second training samples are obtained. Each first training sample includes multiple first dynamic targets and a first initial feature embedding vector corresponding to each first dynamic target. Each second training sample includes multiple second dynamic targets, a gap role ground value probability distribution vector corresponding to each second dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The multiple second dynamic targets are obtained by filtering from the multiple first dynamic targets based on a first probability. The gap role probability prediction module is trained based on the multiple first training samples to obtain multiple first gap role prediction probability distribution vectors output by the gap role probability prediction module. The gap role probability prediction module is used to predict the gap role probability of a dynamic target in a lane change scenario. The lane change trajectory planning module is trained based on the plurality of second training samples and the plurality of second gap role prediction probability distribution vectors to obtain a model for lane change planning. The lane change trajectory planning module is used to output the lane change trajectory of the target vehicle based on the gap role probability. The plurality of second gap role prediction probability distribution vectors are obtained by filtering the plurality of first gap role prediction probability distribution vectors based on the second probability.

13. The lane change planning method according to claim 12, characterized in that, The step of training the gap role probability prediction module based on the plurality of first training samples includes: Construct the first loss function; Based on the multiple first training samples, the gap role probability prediction module is iteratively optimized by minimizing the first loss function until the gap role prediction probability distribution vector output by the gap role probability prediction module converges.

14. The lane change planning method according to claim 12, characterized in that, The process of training the lane change trajectory planning module based on the plurality of second training samples and the plurality of second gap role prediction probability distribution vectors to obtain the lane change planning model includes: Construct a second loss function; Based on the multiple second training samples and the multiple second gap role prediction probability distribution vectors, the lane change trajectory planning module is iteratively optimized by minimizing the second loss function until the lane change trajectory output by the lane change trajectory planning module converges, thus obtaining the lane change planning model.

15. The lane change planning method according to claim 14, characterized in that, The second loss function includes the goal achievement loss function, the velocity constraint loss function, and the trajectory error loss function.

16. A lane change planning device, characterized in that, include: The acquisition module is used to acquire the current scene information of the target vehicle. The current scene information includes at least one dynamic target, a first initial feature embedding vector corresponding to each dynamic target, and a second initial feature embedding vector corresponding to the target vehicle. The generation module is used to generate an enhanced embedding vector corresponding to each dynamic target based on the first initial feature embedding vector corresponding to each dynamic target, wherein the enhanced embedding vector is fused with gap role semantic information. An interactive modeling module is used to input the enhanced embedding vector corresponding to each dynamic target and the second initial feature embedding vector into the interactive game module for global interactive modeling, so as to obtain the first fusion embedding vector corresponding to each dynamic target and the second fusion embedding vector corresponding to the target vehicle output by the interactive game module. The output module is used to output at least one lane change trajectory corresponding to the target vehicle based on the first fusion embedding vector and the second fusion embedding vector corresponding to each dynamic target.

17. A vehicle, characterized in that, include: The vehicle body and the lane change planning system, the lane change planning system being used to implement the lane change planning method as described in any one of claims 1 to 15.

18. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the lane change planning method as described in any one of claims 1 to 15.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the lane change planning method as described in any one of claims 1 to 15.

20. A computer program product, characterized in that, include: A computer program that, when executed by a processor, implements the lane change planning method as described in any one of claims 1 to 15.

Images