Trajectory determination method and training method and device of trajectory determination model

Abstract

The invention discloses a trajectory determination method and a training method and device of a trajectory determination model. The trajectory determination method comprises the following steps: acquiring perception data of a surrounding environment of a vehicle; determining environmental characteristics based on the perception data; determining a transverse planning path of the vehicle based on the environmental characteristics; and based on the transverse planning path and the environment characteristics, determining a planning track of the vehicle. According to the technical scheme, the transverse planning path is determined firstly, and then the planning track of the vehicle is determined on the basis of the transverse planning path in combination with the environment characteristics; in this way, due to the fact that constraint dependence of transverse information exists in the process of determining the planned trajectory of the vehicle, the safety and stability of intelligent driving in a complex scene can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent driving technology, and in particular to a trajectory determination method, a training method and apparatus for a trajectory determination model. Background Technology

[0002] With the development of intelligent driving technology, end-to-end driving models have been extensively studied. These models learn and directly convert perceived environmental information into predicted vehicle trajectories for controlling steering and speed. In complex road and traffic environments, the predicted vehicle trajectory can be influenced by various factors, such as road shape, the movement of vehicles ahead, and traffic signals.

[0003] Two common approaches are used in related technologies to obtain the predicted trajectory of a vehicle for planning and control. The first approach involves two parallel branches that independently perform lateral path prediction and longitudinal trajectory prediction. Lateral path prediction controls the vehicle's direction, while longitudinal trajectory prediction generates the desired longitudinal velocity or displacement. Finally, the control module combines the lateral path prediction and longitudinal trajectory prediction results to further determine the executable complete vehicle control variables. The second approach directly outputs a sequence of complete trajectory points for the vehicle over a future period from a neural network, without distinguishing between lateral and longitudinal predictions.

[0004] However, the first approach suffers from unstable control results and repetitive path and trajectory encoding issues. The second approach faces challenges such as difficult model training and poor generalization ability. Summary of the Invention

[0005] To address the aforementioned technical problems, this disclosure provides a trajectory determination method, a trajectory determination model training method, and an apparatus.

[0006] A first aspect of this disclosure provides a trajectory determination method, comprising: acquiring perception data of the environment surrounding a vehicle; determining environmental features based on the perception data; determining a lateral planning path for the vehicle based on the environmental features; and determining the planned trajectory of the vehicle based on the lateral planning path and the environmental features.

[0007] A second aspect of this disclosure provides a training method for a trajectory determination model, comprising: determining an initial perception network and an initial planning network; determining perception sample data and corresponding label information; processing the perception sample data based on the initial perception network to obtain initial environmental features and initial object detection results; processing the initial environmental features based on the initial planning network to obtain an initial lateral planning path for the vehicle; processing the initial lateral planning path and initial environmental features based on the initial planning network to obtain an initial planned trajectory for the vehicle; iteratively training the initial perception network and the initial planning network based on the initial object detection results, the initial lateral planning path, the initial planned trajectory, and the label information to obtain a trained perception network and a trained planning network; and obtaining a trained trajectory determination model based on the trained perception network and the trained planning network.

[0008] A third aspect of this disclosure provides a trajectory determination device, comprising: a perception data acquisition module for acquiring perception data of the environment surrounding a vehicle; an environmental feature determination module for determining environmental features based on the perception data; a lateral path determination module for determining a lateral planned path for the vehicle based on the environmental features; and a planned trajectory determination module for determining a planned trajectory for the vehicle based on the lateral planned path and the environmental features.

[0009] A fourth aspect of this disclosure provides a training apparatus for a trajectory determination model, comprising: an initial network determination module for determining an initial perception network and an initial planning network; a sample data determination module for determining perception sample data and corresponding label information; a perception processing module for processing the perception sample data based on the initial perception network to obtain initial environmental features and initial object detection results; a lateral processing module for processing the initial environmental features based on the initial planning network to obtain an initial lateral planning path for the vehicle; a trajectory determination module for processing the initial lateral planning path and initial environmental features based on the initial planning network to obtain an initial planned trajectory for the vehicle; a network training module for iteratively training the initial perception network and the initial planning network based on the initial object detection results, the initial lateral planning path, the initial planned trajectory, and the label information to obtain a trained perception network and a trained planning network; and a model determination module for obtaining a trained trajectory determination model based on the trained perception network and the trained planning network.

[0010] A fifth aspect of this disclosure provides a computer-readable storage medium storing a computer program for performing the trajectory determination method provided in the first aspect embodiment, or the trajectory determination model training method provided in the second aspect embodiment.

[0011] A sixth aspect of this disclosure provides an electronic device comprising: a processor; a memory for storing processor-executable instructions; and a processor for reading executable instructions from the memory and executing the instructions to implement the trajectory determination method provided in the first aspect embodiment above, or the trajectory determination model training method provided in the second aspect embodiment above.

[0012] A seventh aspect of this disclosure provides a computer program product that, when instructions in the computer program product are executed by a processor, performs the trajectory determination method provided in the first aspect embodiment above, or the trajectory determination model training method provided in the second aspect embodiment above.

[0013] The trajectory determination method disclosed herein first determines the lateral planning path, and then determines the planned trajectory of the vehicle based on the lateral planning path and environmental characteristics. In this way, since there is a constraint dependency on lateral information in the process of determining the planned trajectory of the vehicle, it can avoid the situation where the lateral path and the longitudinal speed are inconsistent, thereby improving the safety and stability of intelligent driving in complex scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a vehicle in a driving scenario provided for an exemplary embodiment of this disclosure.

[0015] Figure 2 This is a schematic flowchart of a trajectory determination method provided for an exemplary embodiment of the present disclosure.

[0016] Figure 3 A schematic flowchart of a trajectory determination method provided for another exemplary embodiment of this disclosure.

[0017] Figure 4 A schematic flowchart of a trajectory determination method provided as yet another exemplary embodiment of this disclosure.

[0018] Figure 5 A schematic flowchart of a trajectory determination method provided as yet another exemplary embodiment of this disclosure.

[0019] Figure 6 A schematic flowchart of a trajectory determination method provided as yet another exemplary embodiment of this disclosure.

[0020] Figure 7 A schematic flowchart of a trajectory determination method provided as yet another exemplary embodiment of this disclosure.

[0021] Figure 8 A schematic flowchart of a trajectory determination method provided as yet another exemplary embodiment of this disclosure.

[0022] Figure 9AThis is a flowchart illustrating a training method for a trajectory determination model provided as an exemplary embodiment of the present disclosure.

[0023] Figure 9B A flowchart illustrating a training method for a trajectory determination model provided as another exemplary embodiment of this disclosure.

[0024] Figure 10A This is a schematic diagram of the structure of a trajectory determination model provided as an exemplary embodiment of the present disclosure.

[0025] Figure 10B A schematic diagram of the structure of an environmental feature determination submodule provided for an exemplary embodiment of this disclosure.

[0026] Figure 10C A schematic diagram of the structure of a lateral path planning submodule provided for an exemplary embodiment of this disclosure.

[0027] Figure 10D A schematic diagram of the structure of a longitudinal parameter planning submodule provided for an exemplary embodiment of this disclosure.

[0028] Figure 11 This is a schematic diagram of the structure of a trajectory determination device provided for an exemplary embodiment of the present disclosure.

[0029] Figure 12 A schematic diagram of the structure of a trajectory determination device provided for another exemplary embodiment of this disclosure.

[0030] Figure 13 This is a schematic diagram of the structure of a training apparatus for a trajectory determination model provided as an exemplary embodiment of the present disclosure.

[0031] Figure 14 A schematic diagram of the structure of a training apparatus for a trajectory determination model provided as another exemplary embodiment of this disclosure.

[0032] Figure 15 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this disclosure. Detailed Implementation

[0033] To explain this disclosure, exemplary embodiments of the disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the disclosure, and not all of them. It should be understood that the disclosure is not limited to exemplary embodiments.

[0034] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0035] Application Overview First, the application scenarios of this disclosure will be introduced. The trajectory determination method provided in the embodiments of this disclosure can be applied to, for example, autonomous driving scenarios, assisted driving scenarios, and any other feasible scenarios.

[0036] like Figure 1 As shown, the vehicle 10 is equipped with at least one visual sensor 11 and a radar sensor 12. Therefore, the visual sensor 11 and / or the radar sensor 12 can collect data on the vehicle's surrounding environment to obtain the perception data or perception sample data in this embodiment of the present disclosure.

[0037] It should be noted that, Figure 1 For illustrative purposes only, one vision sensor 11 or multiple vision sensors 11 may be installed on the vehicle 10. This disclosure embodiment... Figure 1 The number and location of the vision sensors 11 are not limited. In practical use, the vision sensors 11 can be the vehicle's forward-facing wide-angle / narrow-angle camera, the vehicle's side-view / surround-view / panoramic-view camera, or the vehicle's RGB-D (Red Green Blue-Depth) camera. The radar sensor 12 can be the vehicle's lidar.

[0038] Furthermore, the perception data collected by the vision sensor 11 and / or radar sensor 12 can be processed to obtain the vehicle's planned trajectory, and then the comprehensive planning and control command of the vehicle can be determined based on the planned trajectory to achieve vehicle driving control.

[0039] Two methods are typically used in related technologies to obtain the predicted driving trajectory of a vehicle, thereby enabling vehicle planning and control.

[0040] The first approach, which uses two parallel branches to independently predict the lateral path and longitudinal trajectory, is simple to implement. However, because these two branches are independent and lack information constraints, inconsistencies between the lateral path prediction and the longitudinal speed prediction can easily arise. For example, when a vehicle plans to change lanes to the left, the longitudinal branch might still output a higher speed, leading to unstable vehicle control during the lane change and affecting driving safety and comfort. Furthermore, this parallel approach typically predicts both the lateral path and longitudinal trajectory as two-dimensional pathpoints, requiring the model to encode both separately, resulting in redundant learning and computation. This not only increases training complexity but may also reduce the model's learning efficiency and generalization ability.

[0041] The second approach, which directly outputs the complete trajectory point sequence of the vehicle over a future period using a neural network, simplifies the planning process. However, the model needs to learn complex spatial and temporal relationships simultaneously, making training more difficult, and its generalization ability across different scenarios is limited. Furthermore, because the trajectory generation lacks structural constraints, the prediction results sometimes exhibit trajectory jitter or deviation from the road, affecting the model's controllability and interpretability.

[0042] To address the aforementioned issues, this disclosure provides a trajectory determination method. This method first determines a lateral planning path, and then determines the planned trajectory of the vehicle based on the lateral planning path and environmental features. Thus, since there is a constraint dependency on lateral information in the process of determining the vehicle's planned trajectory, it can avoid the situation where the lateral path and longitudinal speed are inconsistent, thereby improving the safety and stability of intelligent driving in complex scenarios.

[0043] Exemplary methods Figure 2 This is a flowchart illustrating a trajectory determination method provided in an exemplary embodiment of the present disclosure. The method of this embodiment can be applied to electronic devices, such as... Figure 2 As shown, the method includes steps S21-S24.

[0044] Step S21: Obtain perception data of the vehicle's surrounding environment.

[0045] For example, the perception data in this disclosure refers to raw and / or pre-processed information collected by the vehicle through various sensors for understanding its own state and surrounding environment. Figure 1 As shown, data collected by at least one visual sensor 11 on the vehicle 10 regarding the surrounding environment can be acquired and used as the perception data in step S21. For example, the vehicle's surround-view cameras (including front-view cameras, rear-view cameras, left-view cameras, and right-view cameras, etc.) collect image data of the surrounding environment, which can be used as perception data. Alternatively, image data collected by the vehicle's RGB-D cameras regarding the surrounding environment can also be used as perception data. Figure 1 As shown, data collected by the radar sensor 12 on the vehicle 10 from the surrounding environment can also be acquired and used as the perception data in step S21. For example, point cloud data collected by the lidar on the vehicle from the surrounding environment can be used as perception data. Of course, image data and point cloud data can also be used together as perception data, and this embodiment of the present disclosure does not limit this.

[0046] In some examples, image data collected by multiple visual sensors can be fused, and the fused image data can be used as the perception data in step S21. Alternatively, image data collected by visual sensors and point cloud data collected by radar sensors can be fused, and the fused data can be used as the perception data in step S21. That is, this embodiment does not limit the data type or the method of acquiring the perception data of the vehicle's surrounding environment.

[0047] Step S22: Determine environmental characteristics based on the perceived data.

[0048] For example, in this embodiment of the disclosure, environmental features refer to structured information with clear physical or semantic meaning extracted from perceived data. Environmental features can be used for driving decisions and motion planning of the vehicle in the future. These environmental features include dynamic environmental features and static environmental features. Dynamic environmental features can be structured information corresponding to dynamic objects (such as vehicles, pedestrians, etc.) in the environment surrounding the vehicle. Static environmental features can be structured information corresponding to static objects (such as roads, traffic lights, traffic signs, etc.) in the environment surrounding the vehicle. For example, dynamic environmental features can characterize the position, orientation, speed, etc., of dynamic objects, while static environmental features can characterize the position, state, etc., of static targets such as curbs, lane lines, traffic signs, and traffic lights. It should be noted that this embodiment of the disclosure does not limit the form of environmental features; the above are merely examples.

[0049] In some embodiments, the trained perception network can be used to process the perceived data to obtain corresponding environmental features. For example, the trained perception network may include a trained dynamic perception network and a trained static perception network. The trained dynamic perception network can output dynamic environmental features, and the trained static perception network can output static environmental features. Furthermore, by concatenating the dynamic and static environmental features, the environmental features corresponding to the perceived data can be obtained. For example, the input of the first deep residual network is the perceived data of the vehicle's surrounding environment, and the output is dynamic environmental features. The input of the second deep residual network is the perceived data of the vehicle's surrounding environment, and the output is static environmental features. Further, the dynamic and static environmental features can be concatenated to obtain the environmental features corresponding to the perceived data.

[0050] In some examples, feature extraction can be performed on the perceived data to obtain the corresponding perceived features. Simultaneously, object enhancement features can be determined. These object enhancement features are primarily used to fuse information from the perceived features to obtain information about dynamic and static objects in the vehicle's surrounding environment from the perceived data. Furthermore, the object enhancement features and perceived features can be fused to obtain the corresponding environmental features. It is understood that the specific method for determining the object enhancement features can be referred to the relevant description in step S223 below, and will not be repeated here.

[0051] Step S23: Determine the lateral planning path of the vehicle based on environmental characteristics.

[0052] For example, in this embodiment of the present disclosure, the lateral planning path of the vehicle refers to the lateral movement planning of the vehicle in a direction perpendicular to its forward direction (i.e., the horizontal direction). The lateral planning path is the lateral travel direction planning of the vehicle, which can be used to guide the directional control of the vehicle, and does not include information such as longitudinal planning speed and longitudinal planning displacement.

[0053] In some examples, lateral enhancement features can be identified, which are primarily used to fuse environmental features to determine the vehicle's lateral planning path in the current environment. Furthermore, the lateral enhancement features and environmental features can be fused to obtain the lateral planning path. It is understood that the specific method for determining the lateral enhancement features can be found in the relevant description in step S232 below, and will not be repeated here.

[0054] Step S24: Based on the lateral planning path and environmental characteristics, determine the planned trajectory of the vehicle.

[0055] For example, in this embodiment of the disclosure, the planned trajectory of the vehicle refers to the complete sequence of the vehicle's positions over a future period of time, including lateral information (such as lateral planned path) and longitudinal information (such as longitudinal planned speed). The planned trajectory of the vehicle can be used to guide the vehicle's directional control, and can also be used to generate control quantities for the vehicle's longitudinal speed or displacement.

[0056] In some examples, the longitudinal motion parameters of the vehicle along the lateral planning path can be determined first based on the lateral planning path and environmental characteristics. Then, the planned trajectory of the vehicle can be determined based on the lateral planning path and longitudinal motion parameters. The longitudinal motion parameters of the vehicle along the lateral planning path include, but are not limited to: the longitudinal velocity sequence, the longitudinal acceleration sequence, and the longitudinal displacement sequence of the vehicle along the lateral planning path.

[0057] The trajectory determination method provided in this embodiment first determines the lateral planning path, and then determines the planned trajectory of the vehicle based on the lateral planning path and environmental features. In this way, since there is a constraint dependency on lateral information in the process of determining the planned trajectory of the vehicle, the situation where the lateral path and the longitudinal speed are inconsistent can be avoided, thereby improving the safety and stability of intelligent driving in complex scenarios.

[0058] like Figure 3 As shown above, in the above Figure 2 Based on the illustrated embodiment, step S24 may include steps S241-S242.

[0059] Step S241: Based on the lateral planning path and environmental characteristics, determine the longitudinal motion parameters of the vehicle along the lateral planning path.

[0060] For example, in this embodiment of the disclosure, the lateral planning path of the vehicle can be determined first, and then the longitudinal motion parameters (such as the longitudinal speed sequence) of the vehicle can be predicted along the lateral planning path based on the current environmental characteristics of the vehicle, so as to achieve serial dependency between lateral and longitudinal planning. The lateral planning path is used for vehicle direction control, and the longitudinal motion parameters are used for vehicle longitudinal control. Furthermore, the lateral planning path and the longitudinal motion parameters can be combined to generate executable control quantities to achieve safe, stable, and controllable driving of the vehicle.

[0061] In some examples, longitudinal enhancement features can be determined based on the lateral planning path. These longitudinal enhancement features are mainly used to fuse environmental features and features corresponding to the lateral planning path, so as to determine the longitudinal motion parameters of the vehicle along the lateral planning path based on the environmental features. Furthermore, the longitudinal enhancement features and environmental features can be fused to obtain the longitudinal motion parameters of the vehicle along the lateral planning path. It is understood that the specific method for determining the longitudinal enhancement features can be referred to the relevant description in step S2412 below, and will not be repeated here.

[0062] Regarding the first parallel prediction scheme in related technologies, since both the path and trajectory are two-dimensional coordinates, the determination of these coordinates relies on map information. Therefore, both branches of this parallel prediction scheme require encoding map information during implementation, leading to repetitive encoding and complex model training. However, the trajectory determination method in this embodiment first determines the vehicle's lateral planning path, and then, based on the lateral planning path and the vehicle's environmental characteristics, determines the vehicle's longitudinal motion parameters (e.g., longitudinal velocity sequence) along the lateral planning path. That is, this embodiment predicts the longitudinal motion parameters based on the lateral planning path, eliminating the need for secondary encoding. Therefore, the trajectory determination method in this embodiment avoids the problem of repetitive encoding of the path and trajectory, thereby improving the efficiency and interpretability of trajectory planning. Simultaneously, it also improves the generalization ability of the trajectory determination model in complex traffic scenarios during training.

[0063] Step S242: Determine the planned trajectory of the vehicle based on the lateral planning path and longitudinal motion parameters.

[0064] For example, the lateral planning path of the vehicle refers to the lateral path planning perpendicular to the vehicle's forward direction during the vehicle's travel. The lateral planning path is mainly used to guide the vehicle's directional control. The longitudinal motion parameters typically refer to the time and motion state of the vehicle while traveling along the lateral planning path. Furthermore, the two can be combined to obtain the vehicle's planned trajectory. The vehicle's planned trajectory is not only a spatial curve, but also indicates the position information (with speed and time information) that the vehicle should reach on this curve at each time point. Further, the vehicle's control module can determine the control quantities of the vehicle's executable direction and longitudinal parameters (such as speed or displacement) based on the vehicle's planned trajectory. Of course, the vehicle's executable direction can also be determined directly based on the lateral planning path, while the control quantities of the vehicle's executable longitudinal parameters can be determined based on the longitudinal motion parameters. This disclosure does not limit this approach.

[0065] The trajectory determination method provided in this disclosure first determines the longitudinal motion parameters of the vehicle along the lateral planned path based on the lateral planning path and environmental features; then, based on the lateral planned path and longitudinal motion parameters, it determines the planned trajectory of the vehicle. Furthermore, the obtained lateral planned path and longitudinal motion parameters can be converted into control quantities of the vehicle's executable direction and longitudinal parameters for driving control. In this way, by conditionally coupling the longitudinal planning process with the lateral path planning results, the consistency between the lateral path and longitudinal planning parameters (such as longitudinal speed) can be improved, thereby reducing planning conflicts and enhancing the safety and stability of the intelligent driving system in complex scenarios.

[0066] like Figure 4 As shown above, in the above Figure 3Based on the illustrated embodiment, step S241 may include steps S2411-S2413.

[0067] Step S2411: Determine the initial longitudinal features and the preset reference speed.

[0068] For example, the initial longitudinal feature in this embodiment is obtained by training a trajectory determination model. The trajectory determination model processes perceived data of the vehicle's surrounding environment to obtain the vehicle's planned trajectory. In some embodiments, at least one set of longitudinal velocity representations can be initialized. Each set of longitudinal velocity representations includes an initial longitudinal feature and a preset reference speed, which can be a speed sequence. The preset reference speed is used to guide the prediction of subsequent longitudinal motion parameters. The initial value of the initial longitudinal feature can be generated randomly, and the value of the initial longitudinal feature is continuously updated as the model iterates through training. The initial longitudinal feature obtained after model training is the initial longitudinal feature in step S2411. Further, the longitudinal velocity representation obtained after training can be used to determine the enhanced longitudinal feature, thereby allowing it to interact with environmental features to obtain the corresponding longitudinal motion parameters. Furthermore, since the driving behavior of a vehicle is diverse, meaning there are multiple possible reference speeds, the longitudinal motion parameters determined using the multiple sets of longitudinal velocity representations are more accurate.

[0069] In some examples, the preset reference speed can be obtained through historical data and / or prior strategies, such as clustering based on expert driving data. Historical data can be real driving data of vehicles (such as ride-hailing vehicles, taxis, etc.) or autonomous driving test vehicles on specific roads, or vehicle driving data obtained through other means; this disclosure does not limit this. Prior strategies refer to pre-set rules, knowledge, or experience used to guide or constrain the behavior of the intelligent driving system, making it safer and more efficient in conforming to human driving logic and traffic environment requirements. Furthermore, in this disclosure, the reference speed included in each set of longitudinal speed representations can be a constant value or a speed sequence that varies over time; this disclosure does not limit this. For example, there are two sets of longitudinal speed representations: the first set includes a reference speed of v=3, and the second set includes a reference speed of v={3,5,1,7}. Of course, the reference speed v=3 included in the first set of longitudinal speed representations can be understood as a speed sequence with the same value.

[0070] Step S2412: Determine the longitudinal enhancement features based on the initial longitudinal features, reference velocity, and lateral planning path.

[0071] For example, in this embodiment of the disclosure, sampling can be performed on the lateral planning path based on a reference speed to obtain a reference path point sequence; then, based on the reference path point sequence and the initial longitudinal features, longitudinal enhancement features can be determined. It should be noted that the specific implementation of determining the longitudinal enhancement features based on the reference path point sequence and the initial longitudinal features can be referred to in step S23a below, and will not be repeated here.

[0072] For example, if there are five sets of trained longitudinal velocity representations, where the first set includes initial longitudinal feature 1 and reference velocity 1, the second set includes initial longitudinal feature 2 and reference velocity 2, ..., and so on, with the fifth set including initial longitudinal feature 5 and reference velocity 5, then sampling can be performed on the lateral planning path based on each reference velocity (reference velocity 1, reference velocity 2, ..., reference velocity 5) to obtain the corresponding reference path point sequence (reference path point sequence 1, ..., reference path point sequence 5). Furthermore, based on reference path point sequence 1 and initial longitudinal feature 1, longitudinal enhancement feature 1 can be determined; based on reference path point sequence 2 and initial longitudinal feature 2, longitudinal enhancement feature 2 can be determined; ..., and so on, based on reference path point sequence 5 and initial longitudinal feature 5, longitudinal enhancement feature 5 can be determined.

[0073] Step S2413: Determine the longitudinal motion parameters based on longitudinal enhancement features and environmental features.

[0074] For example, longitudinal enhancement features can be fused with environmental features to obtain the longitudinal motion parameters of the vehicle along the lateral planned path. Furthermore, if multiple longitudinal enhancement features exist, the longitudinal motion parameters corresponding to each longitudinal enhancement feature can be determined separately based on each feature and the environmental features. For instance, a cross-attention mechanism can be used to fuse longitudinal enhancement features and environmental features to obtain the longitudinal motion parameters.

[0075] For example, if there are five longitudinal enhancement features (Longitudinal enhancement feature 1, Longitudinal enhancement feature 2, Longitudinal enhancement feature 3, Longitudinal enhancement feature 4, and Longitudinal enhancement feature 5), then a cross-attention mechanism can be used to fuse Longitudinal enhancement feature 1 with environmental features to obtain longitudinal motion parameter 1. ... Similarly, a cross-attention mechanism can be used to fuse Longitudinal enhancement feature 5 with environmental features to obtain longitudinal motion parameter 5. Further, the scores for Longitudinal motion parameter 1, Longitudinal motion parameter 2, ..., and Longitudinal motion parameter 5 can be determined separately, and the final longitudinal motion parameter can be determined based on their respective scores. Of course, the above is merely an example; other methods can also be used to determine the final longitudinal motion parameters used by the vehicle from multiple longitudinal motion parameters, and this disclosure does not limit this approach.

[0076] The trajectory determination method provided in this disclosure determines longitudinal query features, initial longitudinal features, and a preset reference speed; based on the initial longitudinal features, reference speed, and lateral planning path, it determines longitudinal enhancement features; and based on the longitudinal enhancement features and environmental features, it determines longitudinal motion parameters. In this way, it can output longitudinal motion parameters along the time dimension based on the lateral planning path, thereby achieving conditional longitudinal control along the lateral path.

[0077] like Figure 5 As shown above, in the above Figure 4 Based on the illustrated embodiment, step S2412 may include steps S21a-S23a.

[0078] Step S21a: Sample along the lateral planning path based on the reference speed to obtain a reference path point sequence.

[0079] For example, the sampling time interval can be determined first, and then sampling can be performed on the lateral planning path based on the reference speed and the sampling time interval to obtain a reference path point sequence. For instance, sampling can be performed on the lateral planning path at a frequency of once every 0.5 seconds based on the reference speed.

[0080] Furthermore, in this embodiment of the present disclosure, based on the reference speed in each group of longitudinal speed representations, sampling is performed on the predicted lateral planning path at sampling time intervals to obtain a reference path point sequence corresponding to each group of longitudinal speed representations. The reference path point sequence is used to guide the prediction of longitudinal motion parameters, ensuring consistency between the longitudinal planning and the lateral path. Thus, this embodiment of the present disclosure can achieve effective coupling between the lateral planning path and longitudinal parameters (such as longitudinal speed) by sampling the reference path point sequence along the lateral planning path and performing longitudinal prediction based on the lateral planning path. This makes the vehicle's direction and speed more coordinated during lane changes, turns, and other operations, thereby reducing control conflicts and improving driving safety and comfort.

[0081] In some examples, the reference speed in each set of longitudinal velocity representations can be a constant speed, or it can vary with time. Alternatively, some longitudinal velocity representations can have a constant reference speed, while others can have a reference speed that varies with time; this disclosure does not impose any limitations on this. If the reference speed is constant, it can be used to perform equally spaced sampling on the lateral planning path at sampling time intervals to obtain a reference path point sequence. If the reference speed is non-constant, it can be used to perform non-equal spaced sampling on the lateral planning path at sampling time intervals to obtain a reference path point sequence.

[0082] Step S22a: Encode the reference path point sequence to obtain the reference path point features.

[0083] For example, the reference path point sequence can be feature-encoded using an MLP (Multi-Layer Perceptron) to obtain the reference path point features. The role of feature encoding is to encode features of different dimensions into the same feature space, thereby achieving the purpose of feature fusion (e.g., using a cross-attention mechanism to process features of different dimensions).

[0084] Step S23a: The reference path point features are fused with the initial longitudinal features to obtain the longitudinal enhanced features.

[0085] For example, in this embodiment of the disclosure, the reference path point features and the initial longitudinal features can be summed to obtain the longitudinal enhancement features. If there are multiple reference path point features, each reference path point feature is summed with its corresponding initial longitudinal features to obtain the longitudinal enhancement features corresponding to each reference path point feature.

[0086] The trajectory determination method provided in this disclosure obtains a reference path point sequence by sampling along the laterally planned path based on a reference speed; it then encodes the reference path point sequence to obtain reference path point features; and finally, it fuses the reference path point features with initial longitudinal features to obtain enhanced longitudinal features. This process yields a query feature vector for aggregating environmental information and the lateral reference path, thus laying the foundation for determining the longitudinal motion parameters of the vehicle along the laterally planned path.

[0087] like Figure 6 As shown above, in the above Figure 4 Based on the illustrated embodiment, step S2413 may include steps S21b-S22b.

[0088] Step S21b: Fuse the longitudinal enhancement features with the environmental features to obtain the longitudinal parameter features.

[0089] For example, in this embodiment of the disclosure, an attention mechanism can be used to fuse longitudinal enhancement features with environmental features, so that the longitudinal enhancement features can be updated according to the environmental features to obtain longitudinal parameter features that are more consistent with the information of the vehicle's surrounding environment.

[0090] For example, a cross-attention mechanism can be used to treat the vertical enhancement feature as a query vector and interact with the environmental features as key and value vectors to obtain the association weight between the vertical enhancement feature and the environmental features; and the value features are dynamically weighted and fused according to the association weight to generate a new feature representation that incorporates the contextual information related to the environmental features, which is the vertical parameter feature in step S21b.

[0091] Step S22b: Decode the longitudinal parameter features to obtain the longitudinal motion parameters.

[0092] For example, longitudinal parameter features can be decoded using a multilayer perceptron (MLP) to obtain longitudinal motion parameters for vehicle longitudinal control. The function of feature decoding is to map the longitudinal parameter features to the required longitudinal motion parameters. Of course, other feature decoding methods can also be used to process the longitudinal parameter features to obtain the longitudinal motion parameters, and this disclosure does not limit this approach.

[0093] The trajectory determination method provided in this disclosure fuses longitudinal enhancement features with environmental features to obtain longitudinal parameter features; then, it decodes these longitudinal parameter features to obtain longitudinal motion parameters. In this way, longitudinal motion parameters can be obtained by associating environmental features through feature fusion.

[0094] like Figure 7 As shown above, in the above Figure 2 Based on the illustrated embodiment, step S22 may include steps S221-S224.

[0095] Step S221: Determine the initial features of the object and the preset spatial reference position.

[0096] For example, in this embodiment of the present disclosure, the initial object feature is obtained by training a trajectory determination model. The trajectory determination model processes the perceived data of the vehicle's surrounding environment to obtain the vehicle's planned trajectory. In some embodiments, since there are typically multiple traffic participants in the vehicle's surrounding environment, multiple sets of perception instance representations can be initialized. Each set of perception instance representations includes an initial object feature and a preset spatial reference position, used to acquire dynamic and static environmental information. The initial value of the initial object feature can be generated randomly, and the value of the initial object feature is continuously updated as the model iterates through training. The initial object feature obtained after model training is the initial object feature in step S221. Further, the perception instance representation obtained after training can be used to determine enhanced object features, thereby enabling them to interact with the perception features to obtain corresponding environmental features.

[0097] In some examples, the spatial reference location refers to a set of three-dimensional coordinates, which are prior knowledge obtained by clustering the dataset. The dataset can be obtained by collecting real data from vehicles on specific roads, or by other means (using general datasets in the driving domain), and this disclosure does not impose any limitations on this. Prior knowledge refers to existing information and constraints determined before the vehicle travels, through methods such as manual design, rule injection, historical data learning, or physical laws.

[0098] Step S222: Extract features from the perceived data to obtain perceived features.

[0099] For example, if the perceived data is multi-view images captured by a surround-view camera, feature extraction can be performed on the multi-view images to obtain image features corresponding to the multi-view images, and these image features can be used as the perceived features in step S222. If the perceived data is point cloud data collected by a LiDAR scanner of the surrounding environment, feature extraction can be performed on the point cloud data to obtain point cloud features, and these point cloud features can be used as the perceived features in step S222. If the perceived data is an image captured by an RGB-D camera, feature extraction can be performed on the image, and the extracted image features can be used as the perceived features in step S222. For example, after acquiring multiple images captured by a surround-view camera, the multiple images can be respectively put into a deep residual network (ResNet) to obtain features corresponding to the multiple images.

[0100] In some examples, if the perceived data is a fusion of image data and point cloud data, feature extraction can be performed on the fused data to obtain the perceived features. Alternatively, if the perceived data includes both image data and point cloud data, feature extraction can be performed separately on the image data and point cloud data to obtain the corresponding image features and point cloud features, and then the image features and point cloud features can be fused to obtain the perceived features.

[0101] It should be noted that the implementation method of feature extraction in this embodiment is not limited; the use of a deep residual network is merely an example. Those skilled in the art can select an appropriate method to extract features from the perceived data based on the type of perceived data and actual usage, thereby obtaining perceived features. Furthermore, the execution order of steps S221 and S222 in this embodiment is not limited. Step S221 can be executed first, followed by step S222; or step S222 can be executed first, followed by step S221; or steps S221 and S222 can be executed in parallel.

[0102] Step S223: Determine the enhanced features of the object based on the initial features of the object and the spatial reference position.

[0103] For example, in this embodiment of the disclosure, the spatial reference position can first be feature-encoded to obtain reference position features. Then, the initial features of the object and the reference position features are fused to obtain enhanced object features.

[0104] For example, if there are four sets of trained perceptual instance representations: the first set includes initial object feature 1 and spatial reference position 1; the second set includes initial object feature 2 and spatial reference position 2; the third set includes initial object feature 3 and spatial reference position 3; and the fourth set includes initial object feature 4 and spatial reference position 4, then feature encoding can be performed on each spatial reference position (spatial reference position 1, ..., spatial reference position 4) to obtain the corresponding reference position features (reference position feature 1, ..., reference position feature 4). Further, initial object feature 1 and reference position feature 1 are summed to obtain object enhancement feature 1, ..., and so on. Similarly, initial object feature 5 and reference position feature 5 are summed to obtain object enhancement feature 5.

[0105] Step S224: Based on the spatial reference position, the enhanced features of the object and the perceived features are fused to obtain the environmental features.

[0106] For example, in this embodiment of the disclosure, an attention mechanism can be used to fuse object enhancement features with perceptual features to obtain environmental features.

[0107] For example, a cross-attention mechanism can be used to treat the object enhancement feature as a query vector and interact with the perceptual features as key and value vectors to obtain the association weight between the object enhancement feature and the perceptual features; and the value features are dynamically weighted and fused according to the association weight to generate a new feature representation that incorporates the contextual information related to the perceptual features, which is the environmental feature in step S224.

[0108] In some examples, since object augmentation features contain three-dimensional spatial coordinates, they can interact with perception features using these coordinates. During the interaction, a cross-attention mechanism is used to achieve feature fusion between object augmentation features and perception features, enabling object augmentation features to effectively acquire perception information related to their own position. In some embodiments, if multiple object augmentation features exist, environmental sub-features corresponding to each object augmentation feature can be determined based on each object augmentation feature and perception feature, and then environmental features of the vehicle's surrounding environment can be determined based on multiple environmental sub-features.

[0109] For example, if there are five object enhancement features (object enhancement feature 1, ..., and so on, with initial object feature 5), then a cross-attention mechanism can be used to fuse object enhancement feature 1 with the perceptual features to obtain environmental sub-feature 1. ... and so on, a cross-attention mechanism can be used to fuse object enhancement feature 5 with the perceptual features to obtain environmental sub-feature 5. Further, environmental sub-feature 1, environmental sub-feature 2, environmental sub-feature 3, environmental sub-feature 4, and environmental sub-feature 5 constitute the environmental features corresponding to the perceptual data.

[0110] The trajectory determination method provided in this disclosure involves determining initial object features and a preset spatial reference position; extracting features from the perceived data to obtain perceived features; determining enhanced object features based on the initial object features and the spatial reference position; and fusing the enhanced object features with the perceived features based on the spatial reference position to obtain environmental features. This allows for the acquisition of dynamic and static environmental features containing image information, thus laying the foundation for subsequent lateral path planning and longitudinal parameter planning using these environmental features.

[0111] In some embodiments, step S223, determining object enhancement features based on initial object features and spatial reference position, includes: performing feature encoding on the spatial reference position to obtain reference position features; and fusing the initial object features with the reference position features to obtain object enhancement features.

[0112] For example, a multilayer perceptron (MLP) can be used to encode features of a spatial reference location to obtain the reference location features. The role of feature encoding is to encode features of different dimensions into the same feature space, thereby achieving the purpose of feature fusion of features of different dimensions (such as using a cross-attention mechanism to process features of different dimensions).

[0113] In some examples, the initial features of the object can be summed with the features at the reference positions to obtain the enhanced features of the object. If multiple reference position features exist, each reference position feature is summed with its corresponding initial object feature to obtain the enhanced object feature for each reference position feature.

[0114] like Figure 8 As shown above, in the above Figure 2 Based on the illustrated embodiment, step S23 may include steps S231-S233.

[0115] Step S231: Determine the initial lateral features and the preset lateral reference path.

[0116] For example, in this embodiment, the initial lateral feature is obtained by training a trajectory determination model. The trajectory determination model processes perceived data of the vehicle's surrounding environment to obtain the vehicle's planned trajectory. In some embodiments, at least one set of lateral path representations can be initialized, each set including an initial lateral feature and a preset lateral reference path. The initial value of the initial lateral feature can be generated randomly, and its value is continuously updated as the model iterates through training. The initial lateral feature obtained after model training is the same as the initial lateral feature in step S231. Further, the lateral path representation obtained after training can be used to determine a lateral enhancement vector, which interacts with environmental features to obtain the corresponding lateral planning path. Moreover, since the driving behavior of a vehicle is diverse, meaning there are multiple possible lateral reference paths, using multiple sets of lateral path representations results in a more accurate lateral planning path.

[0117] In some examples, the lateral reference path is a typical driving path of the vehicle, which can be obtained based on expert driving data clustering. Furthermore, one lateral reference path can correspond to one driving intention; for example, one lateral reference path corresponds to turning right to overtake, one corresponds to turning left to overtake, one corresponds to following another vehicle, and one corresponds to turning, etc. It should be noted that this disclosure does not limit the type or number of preset lateral reference paths, and those skilled in the art can set them according to actual usage.

[0118] Step S232: Determine the lateral enhancement features based on the initial lateral features and the lateral reference path.

[0119] For example, in this embodiment of the disclosure, the lateral reference path can first be feature-encoded to obtain lateral reference path features; then, the initial lateral features and the lateral reference path features can be fused to obtain lateral enhanced features. For example, if there are six sets of trained lateral path representations, the first set of lateral path representations includes initial lateral feature 1 and lateral reference path 1, the second set of lateral path representations includes initial lateral feature 2 and lateral reference path 2, ..., and so on, and the sixth set of lateral path representations includes initial lateral feature 6 and lateral reference path 6, then each lateral reference path (lateral reference path 1, ..., lateral reference path 6) can be feature-encoded to obtain the corresponding lateral reference path features (lateral reference path feature 1, ..., lateral reference path feature 6). Further, lateral reference path feature 1 and initial lateral feature 1 are summed to obtain lateral enhanced features 1, ..., and so on, lateral reference path feature 6 and initial lateral feature 6 are summed to obtain lateral enhanced features 6.

[0120] Step S233: Determine the lateral planning path based on lateral enhancement features and environmental features.

[0121] For example, lateral enhancement features can be first fused with environmental features to obtain lateral path features; then, the lateral path features can be decoded to obtain the lateral planning path. Furthermore, if multiple lateral enhancement features exist, the lateral planning path corresponding to each lateral enhancement feature can be determined separately based on each lateral enhancement feature and the environmental features. In some examples, a cross-attention mechanism can be used to fuse lateral enhancement features and environmental features to obtain the lateral planning path.

[0122] For example, if there are five lateral enhancement features (lateral enhancement feature 1, lateral enhancement feature 2, ..., lateral enhancement feature 5), then a cross-attention mechanism can be used to fuse lateral enhancement feature 1 with environmental features to obtain lateral planning path 1. ... and so on, a cross-attention mechanism can be used to fuse lateral enhancement feature 5 with environmental features to obtain lateral planning path 5. Furthermore, the longitudinal motion parameters of the vehicle along each lateral planning path can be determined based on each lateral planning path.

[0123] The trajectory determination method provided in this disclosure determines lateral initial features and a preset lateral reference path; based on the lateral initial features and the lateral reference path, it determines lateral enhancement features; and based on the lateral enhancement features and environmental features, it determines a lateral planning path. In this way, the lateral planning path of the vehicle in the current environment can be obtained, thus laying the foundation for determining the longitudinal motion parameters of the vehicle along the lateral planning path.

[0124] In some embodiments, step S232, determining the lateral enhancement features based on the lateral initial features and the lateral reference path, includes: performing feature encoding on the lateral reference path to obtain lateral reference path features; and performing feature fusion on the lateral initial features and the lateral reference path features to obtain lateral enhancement features.

[0125] For example, the lateral reference path can be feature-encoded using a multilayer perceptron (MLP) to obtain the lateral reference path features. The role of feature encoding is to encode features of different dimensions into the same feature space, thereby achieving the purpose of feature fusion (e.g., using a cross-attention mechanism to process features of different dimensions).

[0126] In some examples, the lateral initial features can be summed with the lateral reference path features to obtain the lateral enhancement features. If multiple lateral reference path features exist, each lateral reference path feature is summed with its corresponding lateral initial feature to obtain the lateral enhancement features corresponding to each lateral reference path feature.

[0127] In some embodiments, step S233, determining the lateral planning path based on lateral enhancement features and environmental features, includes: fusing the lateral enhancement features and environmental features to obtain lateral path features; and decoding the lateral path features to obtain the lateral planning path.

[0128] For example, in this embodiment of the disclosure, an attention mechanism can be used to fuse lateral enhancement features with environmental features, so that the lateral enhancement features can be updated according to the environmental features to obtain lateral path features that are more consistent with the information of the vehicle's surrounding environment.

[0129] For example, a cross-attention mechanism can be used to treat the lateral enhancement features as query vectors and perform feature interaction calculations with the environmental features as key and value vectors to obtain the association weight between the lateral enhancement features and the environmental features; and the value features are dynamically weighted and fused according to the association weight to generate a new feature representation that incorporates the contextual information related to the environmental features, which is the lateral path feature in step S233.

[0130] In some examples, a multilayer perceptron (MLP) can be used to decode the lateral path features to obtain the lateral planning path. The role of feature decoding is to map the lateral path features to the desired lateral planning path. Of course, other feature decoding methods can also be used to process the lateral path features to obtain the lateral planning path, and this disclosure does not limit this approach.

[0131] Figure 9A This is a flowchart illustrating a training method for a trajectory determination model provided as an exemplary embodiment of the present disclosure. The method of this embodiment can be applied to electronic devices, such as… Figure 9A As shown, the method includes steps S901-S907.

[0132] Step S901: Determine the initial sensing network and the initial planning network.

[0133] For example, the trajectory determination model provided in this embodiment includes at least a perception network and a planning network. In this embodiment, the initial perception network is first pre-trained to obtain a pre-trained perception network. Then, the pre-trained perception network and the initial planning network are trained together to obtain the trained trajectory determination model.

[0134] Step S902: Determine the perceived sample data and the corresponding label information of the perceived sample data.

[0135] For example, image sample data collected by at least one visual sensor 11 on the vehicle 10 can be acquired and used as perception sample data. Point cloud sample data collected by a radar sensor 12 on the vehicle 10 can also be acquired and used as perception sample data. Alternatively, image sample data and point cloud sample data can be fused, and the fused sample data can be used as perception sample data. That is, this disclosure does not limit the type or method of determining perception sample data.

[0136] In some examples, the perceived sample data contains corresponding label information, including but not limited to: detection labels, path labels, and trajectory labels. Detection labels can include static detection labels and dynamic detection labels. Path labels can be the actual lateral path of the vehicle in the perceived sample data, and trajectory labels can be the actual trajectory of the vehicle in the perceived sample data.

[0137] It should be noted that the execution order of steps S901 and S902 is not limited in this embodiment. Step S901 can be executed first, followed by step S902; or step S902 can be executed first, followed by step S901; or steps S901 and S902 can be executed in parallel.

[0138] Step S903: Process the perceived sample data based on the initial perception network to obtain the initial environmental features and initial object detection results.

[0139] It is understandable that the specific implementation method for processing the perceived sample data based on the initial perception network to obtain the initial environmental features can refer to the implementation method of the aforementioned step S22, and will not be repeated here.

[0140] For example, the initial object perception results include, but are not limited to, the initial detection results of static objects in the perception sample data and the initial detection results of dynamic objects. It should be noted that, in this embodiment of the disclosure, the input to the initial perception network is the perception sample data, and the output of the initial perception network includes not only initial environmental features but also initial object detection results. Furthermore, the initial object detection results can be used to train the initial perception network subsequently.

[0141] Step S904: Process the initial environmental features based on the initial planning network to obtain the initial lateral planning path of the vehicle.

[0142] It is understandable that the specific implementation of processing the initial environmental features based on the initial planning network to obtain the initial lateral planning path of the vehicle can refer to the implementation of step S23 mentioned above, and will not be repeated here.

[0143] Step S905: Process the initial lateral planning path and initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle.

[0144] It is understandable that the specific implementation method of processing the initial lateral planning path and initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle can refer to the implementation method of the aforementioned step S24, and will not be repeated here.

[0145] Step S906: Based on the initial object detection results, initial lateral planning path, initial planning trajectory and label information, iteratively train the initial perception network and the initial planning network to obtain the trained perception network and the trained planning network.

[0146] For example, the initial perception network can be iteratively trained based on the initial object detection results and the detection labels in the label information to obtain the pre-trained perception network; then, the initial planning network and the pre-trained perception network can be iteratively trained based on the initial lateral planning path, the initial planning trajectory and the label information to obtain the trained perception network and the trained planning network.

[0147] Step S907: Based on the trained perception network and the trained planning network, the trained trajectory determination model is obtained.

[0148] For example, the trajectory determination model in this embodiment includes a perception network and a planning network, so a trained trajectory determination model can be obtained based on the trained perception network and the trained planning network.

[0149] The trajectory determination model training method provided in this disclosure involves: determining an initial perception network and an initial planning network; determining perception sample data and corresponding label information; processing the perception sample data based on the initial perception network to obtain initial environmental features and initial object detection results; processing the initial environmental features based on the initial planning network to obtain the initial lateral planning path of the vehicle; processing the initial lateral planning path and initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle; iteratively training the initial perception network and initial planning network based on the initial object detection results, initial lateral planning path, initial planned trajectory, and label information to obtain a trained perception network and a trained planning network; and finally, obtaining a trained trajectory determination model based on the trained perception network and trained planning network. Thus, because there is a constraint dependency on lateral information when determining the initial planned trajectory of the vehicle during training, the trained trajectory determination model can avoid inconsistencies between the lateral path and longitudinal speed, thereby improving the safety and stability of intelligent driving in complex scenarios.

[0150] like Figure 9B As shown above, in the above Figure 9A Based on the illustrated embodiment, step S906 may include steps S9061-S9062.

[0151] Step S9061: Based on the initial object detection results and the detection labels in the label information, iteratively train the initial perception network to obtain the pre-trained perception network.

[0152] For example, a first loss value can be determined based on the initial object detection results and the detection labels in the label information; then, based on the first loss value, the initial perception network can be iteratively trained to obtain the pre-trained perception network.

[0153] In some examples, the Euclidean distance between the initial object detection result and the detection label can be determined as the first loss value. For example, if the perceptual network is a dynamic perceptual network, the initial object detection result can be the bounding box of the dynamic object, whose attributes include, but are not limited to, size, category, position, orientation angle, and velocity. Then, the Euclidean distances between the position, orientation angle, and velocity and their corresponding ground truth values ​​in the detection label can be calculated, and the first loss value can be determined accordingly. The size, category, position, and orientation angle can be determined from the single-frame detection result of the dynamic object, and the velocity can be determined from the detection results of two adjacent frames of the dynamic object. The specific content of the attributes and the corresponding determination methods are not limited in this embodiment; this is merely an example.

[0154] Step S9062: Based on the initial lateral planning path, initial planning trajectory and label information, iteratively train the initial planning network and the pre-trained perception network to obtain the trained perception network and the trained planning network.

[0155] For example, a second loss value can be determined first based on the path labels in the initial lateral planning path and label information; then, a third loss value can be determined based on the trajectory labels in the initial planned trajectory and label information; further, the initial planning network and the pre-trained perception network can be iteratively trained based on the first, second, and third loss values ​​to obtain the trained perception network and the trained planning network. For instance, the Euclidean distance between the initial lateral planning path and the path labels can be determined as the second loss value. The Euclidean distance between the initial planned trajectory and the trajectory labels can be determined as the third loss value.

[0156] The trajectory determination model training method provided in this disclosure first pre-trains an initial perception network, and then trains an initial planning network and the pre-trained perception network to obtain a trained perception network and a trained planning network. In this way, a trajectory determination model with high prediction accuracy can be obtained through two stages of training (the first stage of pre-training and the second stage of training).

[0157] In some embodiments, step S9061, iteratively training an initial perception network based on the initial object detection results and the detection labels in the label information to obtain a pre-trained perception network, includes: determining a first loss value based on the initial object detection results and the detection labels in the label information; and iteratively training the initial perception network based on the first loss value to obtain a pre-trained perception network.

[0158] For example, a first loss value can be determined based on the initial object detection results and the detection labels in the label information. This first loss value is then used to update the network parameters in the initial perceptual network until the updated perceptual network meets the convergence condition. If the updated perceptual network meets the convergence condition, the pre-trained perceptual network is obtained. The updated perceptual network meeting the convergence condition can be achieved in three ways: first, the loss output by the perceptual network is less than a preset value; second, the change in network parameters between two iterations is less than a preset value, where network parameters include learnable parameters such as weights and biases; and third, the number of iterations reaches a preset number.

[0159] In some embodiments, step S9062, iteratively training the initial planning network and the pre-trained perception network based on the initial lateral planning path, the initial planning trajectory, and the label information to obtain the trained perception network and the trained planning network, includes: determining a second loss value based on the path labels in the initial lateral planning path and the label information; determining a third loss value based on the trajectory labels in the initial planning trajectory and the label information; and iteratively training the initial planning network and the pre-trained perception network based on the first loss value, the second loss value, and the third loss value to obtain the trained perception network and the trained planning network.

[0160] For example, a second loss value can be determined first based on the path labels in the initial lateral planning path and label information, and a third loss value can be determined based on the trajectory labels in the initial planned trajectory and label information. Then, the network parameters in the pre-trained perceptual network and the initial planning network can be updated using the first, second, and third loss values ​​until the updated network (including the pre-trained perceptual network and the initial planning network) satisfies the convergence condition. If the updated network satisfies the convergence condition, the trained perceptual network and the trained planning network are obtained. The updated network satisfying the convergence condition can include: the initial planning network satisfying the convergence condition; or, both the pre-trained perceptual network and the initial planning network satisfying the convergence condition.

[0161] In some examples, the updated network satisfies the convergence condition, which can be achieved in the following three ways: the first is that the loss of the network output is less than a certain preset value; the second is that the change of the network parameters between two iterations is less than a certain preset value, where the network parameters include learnable parameters such as weights and biases; and the third is that the number of iterations reaches a preset number.

[0162] In some embodiments, step S905, processing the initial lateral planning path and initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle, includes: determining the initial longitudinal motion parameters of the vehicle along the initial lateral planning path based on the initial lateral planning path and initial environmental features; and determining the initial planned trajectory of the vehicle based on the initial lateral planning path and initial longitudinal motion parameters.

[0163] Furthermore, if the initial longitudinal motion parameters are an initial longitudinal velocity sequence, then sampling can be performed on the initial lateral planning path based on a preset time interval and the initial longitudinal velocity sequence to obtain an initial displacement sequence. Then, based on the initial displacement sequence and the displacement labels in the trajectory labels, a third loss value is determined.

[0164] In some embodiments, determining the initial longitudinal motion parameters of the vehicle along the initial lateral planning path based on the initial lateral planning path and initial environmental features includes: determining initial longitudinal initial features and a preset reference speed; sampling on the initial lateral planning path based on the reference speed to obtain an initial reference path point sequence; performing feature encoding on the initial reference path point sequence to obtain initial reference path point features; fusing the initial reference path point features with the initial longitudinal initial features to obtain initial longitudinal enhancement features; and determining the initial longitudinal motion parameters based on the initial longitudinal enhancement features and initial environmental features.

[0165] In some embodiments, determining the initial longitudinal motion parameters based on the initial longitudinal enhancement features and the initial environmental features includes: fusing the initial longitudinal enhancement features and the initial environmental features to obtain the initial longitudinal parameter features; and decoding the initial longitudinal parameter features to obtain the initial longitudinal motion parameters.

[0166] Thus, since the embodiments of this disclosure can separate the lateral planning path from the longitudinal planning trajectory, and enhance the model through the interaction of the reference path point sequence after feature encoding and environmental features, the model does not have to learn complex global spatial and temporal relationships at the same time, thereby reducing the difficulty of training, enhancing the stability of prediction, avoiding trajectory jitter or deviation from the road, and improving the executability and interpretability of the model.

[0167] Figure 10A A schematic diagram of the structure of a trajectory determination model provided in an exemplary embodiment of this disclosure, as shown below. Figure 10AAs shown, the trajectory determination model 10 includes at least: a perception feature extraction module 11, an environmental feature determination model 12, a lateral path planning module 13, a longitudinal parameter planning module 14, and a lateral and longitudinal planning combined module 15.

[0168] The perceptual feature extraction module 11 is mainly used to extract features from the perceptual data to obtain the corresponding perceptual features.

[0169] The environmental feature determination model 12 is mainly used to process perceived features to obtain the corresponding environmental features.

[0170] The lateral path planning module 13 is mainly used to process environmental features to obtain the lateral planning path of the vehicle.

[0171] The longitudinal parameter planning module 14 is mainly used to process the lateral planning path and environmental features to obtain the longitudinal motion parameters of the vehicle.

[0172] The horizontal and vertical planning combined module 15 is mainly used to process the horizontal planning path and the vertical motion parameters to obtain the planned trajectory of the vehicle.

[0173] In some embodiments, the environmental feature determination module 12 may include one environmental feature determination submodule or multiple environmental feature determination submodules connected in series. The lateral path planning module 13 may include one lateral path planning submodule or multiple lateral path planning submodules connected in series. Similarly, the longitudinal parameter planning module 14 may include one longitudinal parameter planning submodule or multiple longitudinal parameter planning submodules connected in series.

[0174] It should be noted that the number of environmental feature determination submodules, horizontal path planning submodules, and vertical parameter planning submodules in this embodiment are not limited, and those skilled in the art can set them according to actual application conditions.

[0175] Figure 10B A schematic diagram of the structure of the environmental feature determination submodule provided in an exemplary embodiment of this disclosure is shown below. Figure 10B As shown, the environmental feature determination submodule 120 includes at least a feature encoding module 121, a feature summing module 122, and a feature fusion module 123.

[0176] The feature encoding module 121 is mainly used to determine the preset spatial reference position and to perform feature encoding on the spatial reference position to obtain the reference position feature.

[0177] The feature summing module 122 is mainly used to sum the initial features of the object with the features of the reference position to obtain environment-related query features (such as object enhancement features).

[0178] The feature fusion module 123 is mainly used to fuse environment-related query features and perception features using a cross-attention mechanism to obtain environment features.

[0179] Figure 10C A schematic diagram of the structure of a lateral path planning submodule provided in an exemplary embodiment of this disclosure is shown below. Figure 10C As shown, the lateral path planning submodule 130 includes at least a feature encoding module 131, a feature summing module 132, a feature fusion module 133, and a feature decoding module 134.

[0180] The feature encoding module 131 is mainly used to determine the preset lateral reference path and to perform feature encoding on the lateral reference path to obtain the lateral reference path features.

[0181] The feature summing module 132 is mainly used to sum the initial horizontal features and the horizontal reference path features to obtain the query features related to the horizontal path (such as the horizontal enhanced features).

[0182] The feature fusion module 133 is mainly used to fuse the query features related to the horizontal path with the environmental features using the cross attention mechanism to obtain the horizontal path features.

[0183] The feature decoding module 134 is mainly used to decode the lateral path features to obtain the lateral planning path of the vehicle.

[0184] Figure 10D A schematic diagram of the structure of the longitudinal parameter planning submodule provided in an exemplary embodiment of this disclosure is shown below. Figure 10D As shown, the longitudinal parameter planning submodule 140 includes at least a path sampling module 141, a feature summing module 142, a feature fusion module 143, and a feature decoding module 144.

[0185] The path sampling module 141 is mainly used to determine a preset reference speed, and to sample on the horizontally planned path based on the reference speed to obtain a reference path point sequence, and to perform feature encoding on the reference path point sequence to obtain reference path point features.

[0186] The feature summing module 142 is mainly used to sum the initial vertical features with the reference path point features to obtain query features related to the vertical parameters (such as vertical enhancement features).

[0187] The feature fusion module 143 is mainly used to fuse query features related to vertical parameters with environmental features using a cross-attention mechanism to obtain vertical parameter features.

[0188] The feature decoding module 144 is mainly used to perform feature decoding on the longitudinal parameter features to obtain the longitudinal motion parameters of the vehicle along the lateral planning path.

[0189] It should be noted that the trajectory determination model (including the perception network and the planning network) used in the trajectory determination method in this disclosure embodiment can be obtained based on any embodiment corresponding to the above-described trajectory determination model training method.

[0190] Table 1 shows performance examples of different trajectory determination methods. As shown in Table 1, in the Bench2Drive benchmark test, task success requires the vehicle controlled by the model to correctly complete road driving and avoid collisions with other vehicles during the driving process. The Bench2Drive scoring system includes indicators such as driving score, task success rate score, and driving efficiency score, where the driving score includes road completion rate and traffic rule compliance.

[0191]

[0192] Table 1. Performance examples of different trajectory determination methods As can be seen from Table 1, the trajectory determination method provided in this embodiment achieves higher driving scores, task success rates, and driving efficiency scores than other related technologies in the Bench2Drive benchmark. Therefore, the trajectory determination method provided in this embodiment has significant effects on ensuring vehicle driving safety, improving path consistency, and enhancing overall driving performance.

[0193] Exemplary device Figure 11 A trajectory determination device provided in the embodiments of this disclosure, such as Figure 11 As shown, the trajectory determination device 110 includes a perception data acquisition module 111, an environmental feature determination module 112, a lateral path determination module 113, and a planned trajectory determination module 114.

[0194] The perception data acquisition module 111 is used to acquire perception data of the environment around the vehicle. The environmental feature determination module 112 is used to determine environmental features based on perceived data; The lateral path determination module 113 is used to determine the lateral planning path of the vehicle based on environmental characteristics; The trajectory determination module 114 is used to determine the planned trajectory of the vehicle based on the lateral planning path and environmental characteristics.

[0195] In some embodiments, such as Figure 12 As shown, the trajectory determination module 114 includes a longitudinal parameter determination unit 1141 and a trajectory determination unit 1142.

[0196] The longitudinal parameter determination unit 1141 is used to determine the longitudinal motion parameters of the vehicle along the lateral planning path based on the lateral planning path and environmental characteristics. The trajectory determination unit 1142 is used to determine the planned trajectory of the vehicle based on the lateral planning path and longitudinal motion parameters.

[0197] In some embodiments, the longitudinal parameter determination unit 1141 is used to determine the longitudinal initial features and the preset reference speed; determine the longitudinal enhancement features based on the longitudinal initial features, the reference speed and the lateral planning path; and determine the longitudinal motion parameters based on the longitudinal enhancement features and environmental features.

[0198] In some embodiments, the longitudinal parameter determination unit 1141 is specifically used to determine the longitudinal initial features and the preset reference speed; to sample on the lateral planning path based on the reference speed to obtain a reference path point sequence; to perform feature encoding on the reference path point sequence to obtain reference path point features; to perform feature fusion with the longitudinal initial features to obtain longitudinal enhancement features; and to determine the longitudinal motion parameters based on the longitudinal enhancement features and environmental features.

[0199] In some embodiments, the longitudinal parameter determination unit 1141 is specifically used to determine the longitudinal initial features and the preset reference speed; determine the longitudinal enhancement features based on the longitudinal initial features, the reference speed and the lateral planning path; fuse the longitudinal enhancement features with the environmental features to obtain the longitudinal parameter features; and decode the longitudinal parameter features to obtain the longitudinal motion parameters.

[0200] In some embodiments, the environmental feature determination module 112 is used to determine the initial features of the object and a preset spatial reference position; extract features from the perceived data to obtain perceived features; determine the enhanced features of the object based on the initial features of the object and the spatial reference position; and fuse the enhanced features of the object with the perceived features based on the spatial reference position to obtain environmental features.

[0201] In some embodiments, the environmental feature determination module 112 is specifically used to determine the initial features of the object and the preset spatial reference position; extract features from the perceived data to obtain perceived features; encode the spatial reference position to obtain reference position features; fuse the initial features of the object and the reference position features to obtain enhanced features of the object; and fuse the enhanced features of the object and the perceived features based on the spatial reference position to obtain environmental features.

[0202] In some embodiments, the lateral path determination module 113 is used to determine lateral initial features and a preset lateral reference path; determine lateral enhancement features based on the lateral initial features and the lateral reference path; and determine a lateral planning path based on the lateral enhancement features and environmental features.

[0203] In some embodiments, the lateral path determination module 113 is specifically used to determine lateral initial features and a preset lateral reference path; perform feature encoding on the lateral reference path to obtain lateral reference path features; perform feature fusion on the lateral initial features and the lateral reference path features to obtain lateral enhancement features; and determine the lateral planning path based on the lateral enhancement features and environmental features.

[0204] In some embodiments, the lateral path determination module 113 is specifically used to determine lateral initial features and a preset lateral reference path; perform feature encoding on the lateral reference path to obtain lateral reference path features; perform feature fusion on the lateral initial features and the lateral reference path features to obtain lateral enhanced features; perform feature fusion on the lateral enhanced features and environmental features to obtain lateral path features; and perform feature decoding on the lateral path features to obtain a lateral planned path.

[0205] The beneficial technical effects corresponding to the exemplary embodiment of the trajectory determination device 110 described above can be found in the corresponding beneficial technical effects of the exemplary method section described above, and will not be repeated here.

[0206] Figure 13 A training apparatus for a trajectory determination model provided in this disclosure embodiment, such as Figure 13 As shown, the training device 130 includes an initial network determination module 131, a sample data determination module 132, a perception processing module 133, a lateral processing module 134, a trajectory determination module 135, a network training module 136, and a model determination module 137.

[0207] The initial network determination module 131 is used to determine the initial sensing network and the initial planning network; The sample data determination module 132 is used to determine the perceived sample data and the label information corresponding to the perceived sample data; The perception processing module 133 is used to process the perception sample data based on the initial perception network to obtain the initial environmental features and the initial object detection results; The lateral processing module 134 is used to process the initial environmental features based on the initial planning network to obtain the initial lateral planning path of the vehicle. The trajectory determination module 135 is used to process the initial lateral planning path and initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle. The network training module 136 is used to iteratively train the initial perception network and the initial planning network based on the initial object detection results, the initial lateral planning path, the initial planning trajectory and the label information, so as to obtain the trained perception network and the trained planning network. The model determination module 137 is used to obtain a trained trajectory determination model based on the trained perception network and the trained planning network.

[0208] In some embodiments, such as Figure 14 As shown, the network training module 136 includes a perception training unit 1361 and an overall training unit 1362.

[0209] The perception training unit 1361 is used to iteratively train the initial perception network based on the initial object detection results and the detection labels in the label information to obtain the pre-trained perception network. The overall training unit 1362 is used to iteratively train the initial planning network and the pre-trained perception network based on the initial lateral planning path, initial planning trajectory and label information, to obtain the trained perception network and the trained planning network.

[0210] In some embodiments, the perception training unit 1361 is specifically used to determine a first loss value based on the initial object detection results and the detection labels in the label information; and to iteratively train the initial perception network based on the first loss value to obtain the pre-trained perception network.

[0211] In some embodiments, the overall training unit 1362 is specifically used to determine a second loss value based on the path label in the initial lateral planning path and label information; determine a third loss value based on the trajectory label in the initial planning trajectory and label information; and iteratively train the initial planning network and the pre-trained perception network based on the first loss value, the second loss value and the third loss value to obtain the trained perception network and the trained planning network.

[0212] The beneficial technical effects of the training device 130 for the trajectory determination model described above can be found in the corresponding beneficial technical effects in the exemplary method section above, and will not be repeated here.

[0213] Exemplary electronic devices Figure 15 A structural diagram of an electronic device 150 provided in this disclosure embodiment includes at least one processor 151 and a memory 152.

[0214] The processor 151 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 150 to perform desired functions.

[0215] The memory 152 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 151 may execute one or more computer program instructions to implement the trajectory determination method and / or trajectory determination model training method and / or other desired functions of the various embodiments of this disclosure described above.

[0216] In one example, the electronic device 150 may also include an input device 153 and an output device 154, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0217] The input device 153 may also include, for example, a keyboard, a mouse, etc.

[0218] The output device 154 can output various information to the outside, including, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0219] Of course, for the sake of simplicity, Figure 15 Only some of the components of the electronic device 150 relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 150 may include any other suitable components depending on the specific application.

[0220] Exemplary computer program products and computer-readable storage media In addition to the methods and apparatus described above, embodiments of this disclosure may also provide a computer program product, including computer program instructions, which, when executed by a processor, cause the processor to perform steps in the trajectory determination method and / or the training method for the trajectory determination model as described in the various embodiments of this disclosure in the "Exemplary Methods" section above.

[0221] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of embodiments of this disclosure. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0222] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform steps in the trajectory determination method and / or the training method for the trajectory determination model in the various embodiments of this disclosure described in the "Exemplary Methods" section above.

[0223] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, but is not limited to, systems, apparatuses, or devices that are electrical, magnetic, optical, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0224] The basic principles of this disclosure have been described above with reference to specific embodiments. However, the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0225] Various modifications and variations can be made to this disclosure without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A trajectory determination method, comprising: Acquire perception data of the vehicle's surrounding environment; Based on the perceived data, environmental characteristics are determined; Based on the environmental characteristics, the lateral planning path of the vehicle is determined; Based on the lateral planning path and the environmental features, the planned trajectory of the vehicle is determined.

2. The method according to claim 1, wherein, Determining the planned trajectory of the vehicle based on the lateral planning path and the environmental features includes: Based on the lateral planning path and the environmental features, the longitudinal motion parameters of the vehicle along the lateral planning path are determined; Based on the lateral planning path and the longitudinal motion parameters, the planned trajectory of the vehicle is determined.

3. The method according to claim 2, wherein, The step of determining the longitudinal motion parameters of the vehicle along the lateral planning path based on the lateral planning path and the environmental features includes: Determine the initial longitudinal features and the preset reference velocity; Based on the initial longitudinal features, the reference velocity, and the lateral planning path, determine the longitudinal enhancement features; The longitudinal motion parameters are determined based on the longitudinal enhancement features and the environmental features.

4. The method according to claim 3, wherein, The determination of longitudinal enhancement features based on the initial longitudinal features, the reference velocity, and the lateral planning path includes: Based on the reference speed, samples are taken on the lateral planning path to obtain a reference path point sequence; The reference path point sequence is feature-encoded to obtain reference path point features; The reference path point features are fused with the initial longitudinal features to obtain the enhanced longitudinal features.

5. The method according to claim 3, wherein, The determination of the longitudinal motion parameters based on the longitudinal enhancement features and the environmental features includes: The longitudinal enhancement features are fused with the environmental features to obtain longitudinal parameter features; The longitudinal motion parameters are obtained by feature decoding of the longitudinal parameter features.

6. The method according to any one of claims 1-5, wherein, The determination of environmental characteristics based on the perceived data includes: Determine the initial features of the object and the preset spatial reference position; Feature extraction is performed on the perceived data to obtain perceived features; Based on the initial features of the object and the spatial reference position, the enhanced features of the object are determined; Based on the spatial reference position, the enhanced features of the object and the perceived features are fused to obtain the environmental features.

7. The method according to claim 6, wherein, The process of determining the enhanced features of the object based on the initial features of the object and the spatial reference position includes: The spatial reference position is feature-encoded to obtain the reference position features; The initial features of the object are fused with the features of the reference position to obtain the enhanced features of the object.

8. The method according to any one of claims 1-5, wherein, Determining the lateral planning path of the vehicle based on the environmental characteristics includes: Determine the initial lateral features and the preset lateral reference path; Based on the initial lateral features and the lateral reference path, determine the lateral enhancement features; The lateral planning path is determined based on the lateral enhancement features and the environmental features.

9. The method according to claim 8, wherein, The determination of lateral enhancement features based on the lateral initial features and the lateral reference path includes: The lateral reference path is feature-encoded to obtain the lateral reference path features; The lateral initial features and the lateral reference path features are fused to obtain the lateral enhanced features.

10. The method according to claim 8, wherein, The process of determining the lateral planning path based on the lateral enhancement features and the environmental features includes: The lateral enhancement features are fused with the environmental features to obtain lateral path features; The lateral path features are decoded to obtain the lateral planning path.

11. A training method for a trajectory determination model, comprising: Determine the initial sensing network and the initial planning network; Determine the perceived sample data and the corresponding label information; The initial sensing network is used to process the sensing sample data to obtain initial environmental features and initial object detection results. The initial environmental features are processed based on the initial planning network to obtain the initial lateral planning path of the vehicle; The initial lateral planning path and the initial environmental features are processed based on the initial planning network to obtain the initial planned trajectory of the vehicle. Based on the initial object detection results, the initial lateral planning path, the initial planning trajectory, and the label information, the initial perception network and the initial planning network are iteratively trained to obtain the trained perception network and the trained planning network. Based on the trained perception network and the trained planning network, a trained trajectory determination model is obtained.

12. The method according to claim 11, wherein, The process of iteratively training the initial perception network and the initial planning network based on the initial object detection results, the initial lateral planning path, the initial planning trajectory, and the label information to obtain the trained perception network and the trained planning network includes: Based on the initial object detection results and the detection labels in the label information, the initial perception network is iteratively trained to obtain the pre-trained perception network. Based on the initial lateral planning path, the initial planning trajectory, and the label information, the initial planning network and the pre-trained perception network are iteratively trained to obtain the trained perception network and the trained planning network.

13. The method according to claim 12, wherein, The step of iteratively training the initial perception network based on the initial object detection results and the detection labels in the label information to obtain the pre-trained perception network includes: Based on the initial object detection results and the detection labels in the label information, a first loss value is determined; Based on the first loss value, the initial perception network is iteratively trained to obtain the pre-trained perception network.

14. The method according to claim 13, wherein, The step of iteratively training the initial planning network and the pre-trained perception network based on the initial lateral planning path, the initial planning trajectory, and the label information to obtain the trained perception network and the trained planning network includes: Based on the initial lateral planning path and the path labels in the label information, a second loss value is determined; Based on the initial planned trajectory and the trajectory labels in the label information, a third loss value is determined; Based on the first loss value, the second loss value, and the third loss value, the initial planning network and the pre-trained perception network are iteratively trained to obtain the trained perception network and the trained planning network.

15. A trajectory determining device, comprising: The perception data acquisition module is used to acquire perception data of the vehicle's surrounding environment; An environmental feature determination module is used to determine environmental features based on the perceived data; A lateral path determination module is used to determine the lateral planning path of the vehicle based on the environmental characteristics. The trajectory determination module is used to determine the planned trajectory of the vehicle based on the lateral planning path and the environmental features.

16. A training device for a trajectory determination model, comprising: The initial network determination module is used to determine the initial sensing network and the initial planning network; The sample data determination module is used to determine the perceived sample data and the label information corresponding to the perceived sample data; The perception processing module is used to process the perception sample data based on the initial perception network to obtain initial environmental features and initial object detection results; The lateral processing module is used to process the initial environmental features based on the initial planning network to obtain the initial lateral planning path of the vehicle. The trajectory determination module is used to process the initial lateral planning path and the initial environmental features based on the initial planning network to obtain the initial planned trajectory of the vehicle. The network training module is used to iteratively train the initial perception network and the initial planning network based on the initial object detection results, the initial lateral planning path, the initial planning trajectory and the label information, to obtain the trained perception network and the trained planning network. The model determination module is used to obtain a trained trajectory determination model based on the trained perception network and the trained planning network.

17. A computer-readable storage medium storing a computer program for performing the trajectory determination method according to any one of claims 1-10, or the training method for the trajectory determination model according to any one of claims 11-14.

18. An electronic device, the electronic device comprising: processor; Memory used to store the processor's executable instructions; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the trajectory determination method according to any one of claims 1-10, or the training method of the trajectory determination model according to any one of claims 11-14.

Images