A vehicle driving track planning method and device, a vehicle and a storage medium

Abstract

The embodiments of this application relate to the field of autonomous driving technology, specifically to a vehicle driving trajectory planning method, device, vehicle, and storage medium. The main steps of the aforementioned method include: acquiring environmental perception data, motion state data, and pose data of the vehicle at multiple historical moments; extracting scene features of the current driving scenario based on the environmental perception data; generating multiple candidate trajectories based on the scene features; determining the target planning trajectory based on the cost scores of the multiple candidate trajectories; when generating each candidate trajectory: if the confidence level of the previously generated trajectory point is not lower than a preset threshold, then generating the next trajectory point based on the scene features; if the confidence level of the previous trajectory point is lower than the preset threshold, then using a preset environmental prediction model to generate environmental prediction features for the next trajectory point, fusing the environmental prediction features with the scene features, and generating the next trajectory point based on the fused features. This method achieves on-demand allocation and efficient utilization of computing resources.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a vehicle driving trajectory planning method, apparatus, vehicle, and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, end-to-end autonomous driving planning methods have demonstrated high planning accuracy and computational efficiency in conventional driving scenarios. The mainstream approach directly maps the observation data collected by sensors to the vehicle trajectory at multiple future moments. However, methods that rely solely on current observation information are prone to long-term trajectory drift in complex interaction scenarios, turning sections, and long-term planning nodes due to the lack of projection of future environmental evolution, which affects planning accuracy.

[0003] Related technologies incorporate world models to predict future environmental states and use the prediction results to assist in trajectory generation. These schemes typically calculate and fuse the world model output at each future time step, employing a dense future prediction and fusion strategy. However, in practical applications, the repeated calculation of high-dimensional future states at each trajectory point easily leads to significant computational redundancy and inference latency, increasing the deployment pressure on the vehicle. Furthermore, during recursive rolling prediction, early errors accumulate along the time chain, causing the far-end prediction results to deviate from reality. Summary of the Invention

[0004] Therefore, it is necessary to provide a vehicle driving trajectory planning method, device, vehicle, and storage medium to address at least one of the aforementioned technical problems.

[0005] In a first aspect, embodiments of this application provide a vehicle driving trajectory planning method, the method comprising: Acquire environmental perception data, motion state data, and pose data of the vehicle at multiple historical moments; Extract scene features of the current driving scenario based on environmental perception data; Based on scene features, multiple candidate trajectories are generated; each candidate trajectory includes a preset number of trajectory points arranged in chronological order. The target planning trajectory is determined based on the cost scores of multiple candidate trajectories; When generating each candidate trajectory: If the confidence level of the previously generated trajectory point is not lower than the preset threshold, then the next trajectory point is generated based on the scene features. If the confidence level of the previous trajectory point is lower than the preset threshold, a preset environment prediction model is used. Based on scene features, motion state data, pose data, and the generated driving instruction sequence, the environment prediction features for the next trajectory point at the next moment are generated. The environment prediction features are then fused with the scene features, and the next trajectory point is generated based on the fused features.

[0006] In some embodiments, the vehicle driving trajectory planning method further includes: Based on the network hidden state of the current trajectory point and the planned pose data corresponding to the current trajectory point, a pre-set confidence estimation network is used to determine the prediction bias scale of the current trajectory point. The confidence level of the current trajectory point is determined based on the prediction bias scale; where the prediction bias scale is negatively correlated with the confidence level.

[0007] In some embodiments, the vehicle driving trajectory planning method further includes: Obtain the reference trajectory points at the time corresponding to the simulated trajectory points in the reference trajectory; the reference trajectory is the standard trajectory generated by driving according to the standard driving route; A confidence loss is constructed based on the error between the reference trajectory point and the simulated trajectory point, and the initial prediction bias scale corresponding to the simulated trajectory point; wherein, the initial prediction bias scale is obtained through an initial confidence estimation network; the larger the initial prediction bias scale, the smaller its contribution to the confidence loss; The parameters of the initial confidence network are adjusted based on the confidence loss to determine the pre-set confidence estimation network.

[0008] In some embodiments, a preset environment prediction model is used to generate future time-at-time environment prediction features for the next trajectory point based on scene features, motion state data, pose data, and the currently generated driving command sequence, including the following steps: Arrange scene features, motion state data, pose data, and driving command sequences into a unified sequence; A causal attention mechanism is used to encode a unified sequence; the causal attention mechanism is used to limit the attention of any element in the unified sequence to elements preceding the current element during encoding; Environmental prediction features are generated based on the coding results.

[0009] In some embodiments, fusing environmental prediction features with scene features includes the following steps: Use scene features as the query matrix and environmental prediction features as the key matrix and value matrix; The attention weights of the query matrix and the key matrix are calculated using a cross-attention mechanism. The value matrix is ​​weighted and summed according to the attention weights to obtain the fused features.

[0010] In some embodiments, the vehicle driving trajectory planning method further includes: Obtain the sample dataset; the sample dataset includes: environmental perception data, sample motion state data, sample pose data, and sample driving command sequence; An initial environment prediction model is used to determine the sample environment prediction characteristics at the target time based on the sample dataset. The first loss value is determined based on the predicted features of the sample environment and the actual environmental features corresponding to the target time. The parameters of the initial environmental prediction model are updated based on the first loss value to reduce the difference between the predicted environmental features and the actual environmental features.

[0011] In some embodiments, the vehicle driving trajectory planning method further includes: An initial environment prediction model is used to determine the first environment prediction features at a specified time based on the first sample driving instruction sequence from the first start time to the specified time. Using an initial environment prediction model, the second environment prediction features at the specified time are determined based on the second sample driving instruction sequence from the second start time to the specified time. The second loss value is determined based on the first environmental prediction characteristics and the second environmental prediction characteristics; The parameters of the initial environmental prediction model are updated based on the second loss value to reduce the difference between the first and second environmental prediction features.

[0012] In some embodiments, the vehicle driving trajectory planning method further includes: Using an initial environment prediction model, the first sample environment prediction features of the target at future times are determined based on the first sample dataset corresponding to the first driving scenario. Based on the second sample dataset corresponding to at least one second driving scenario in the same training batch, determine the second sample environment prediction features of the target at future time. The third loss value is determined based on the environmental prediction characteristics of the first sample and the environmental prediction characteristics of the second sample. The parameters of the initial environment prediction model are updated based on the third loss value to reduce the similarity between the first sample environment prediction features and the second sample environment prediction features.

[0013] In a second aspect, embodiments of this application provide a vehicle driving trajectory planning device, the device comprising: The data acquisition module is used to acquire environmental perception data, motion state data, and pose data of the vehicle at multiple historical moments; The scene feature extraction module is used to extract scene features of the current driving scene based on environmental perception data; The candidate trajectory generation module is used to generate multiple candidate trajectories based on scene features; each candidate trajectory includes a preset number of trajectory points arranged in chronological order. The target trajectory determination module is used to determine the target planned trajectory based on the cost scores of multiple candidate trajectories. The candidate trajectory module further includes a first generation unit and a second generation unit. When generating each candidate trajectory, the candidate trajectory module: The first generation unit is used to generate the next trajectory point based on scene features if the confidence level of the previously generated trajectory point is not lower than a preset threshold. The second generation unit is used to generate environmental prediction features for the next trajectory point if the confidence level of the previous trajectory point is lower than a preset threshold. It uses a preset environment prediction model to generate environmental prediction features for the next trajectory point at the next moment based on scene features, motion state data, pose data, and the generated driving instruction sequence. It then fuses the environmental prediction features with the scene features and generates the next trajectory point based on the fused features.

[0014] In a third aspect, embodiments of this application provide an autonomous vehicle, the vehicle comprising: Sensor components are used to collect environmental perception data, motion state data, and pose data of the vehicle. The vehicle controller includes one or more processors and a memory storing a computer program, which, when executed by the processor, implements the steps of the vehicle driving trajectory planning method provided in any embodiment of the first aspect of this application.

[0015] In a fourth aspect, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the vehicle driving trajectory planning method provided in any embodiment of the first aspect of this application.

[0016] The aforementioned vehicle driving trajectory planning method, device, vehicle, and storage medium assign confidence scores to each trajectory point in the candidate trajectories and determine whether to invoke the environmental prediction model based on the confidence scores. This makes the invocation of the environmental prediction model no longer fixed and intensive, but rather triggered on demand based on the reliability of the trajectory points themselves. This avoids the computational redundancy problem caused by repeatedly calculating high-dimensional environmental prediction features at each trajectory point, reducing additional computational overhead and inference latency in the trajectory planning process. Furthermore, the computational resources of the environmental prediction model can be accurately allocated to the critical moments when future information is truly needed, achieving on-demand allocation and efficient utilization of computational resources. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating the application environment of the vehicle driving trajectory planning method in some embodiments; Figure 2 This is a flowchart illustrating the vehicle driving trajectory planning method in some embodiments; Figure 3 This is a flowchart illustrating the confidence step in some embodiments; Figure 4This is a flowchart illustrating the steps involved in training a pre-configured confidence estimation network in some embodiments; Figure 5 This is a flowchart illustrating the steps involved in generating environmental prediction features in some embodiments; Figure 6 This is a flowchart illustrating the steps involved in generating fusion features in some embodiments; Figure 7 This is a flowchart illustrating the step of generating a first loss value in some embodiments; Figure 8 This is a flowchart illustrating the step of generating a second loss value in some embodiments; Figure 9 This is a flowchart illustrating the step of generating a third loss value in some embodiments; Figure 10 This is a structural block diagram of the vehicle driving trajectory planning device in some embodiments; Figure 11 Here are some structural block diagrams of autonomous vehicles in some embodiments; Figure 12 This is a diagram showing the internal structure of the vehicle controller in some embodiments. Detailed Implementation

[0018] To make the technical solutions and advantages of this application clearer, the embodiments and related technical content of this application will be further described in detail below with reference to the accompanying drawings and text description. It should be understood that the embodiments described below are only used to explain the technical solutions of the embodiments of this application and are not intended to limit more possible implementations of this application.

[0019] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0020] In the multiple formulas appearing in this application, the meaning of the variables or symbols contained in each formula shall be subject to the explanation below the corresponding formula. The meaning of the same variables or symbols may be different in different formulas.

[0021] In a first aspect, embodiments of this application provide a vehicle driving trajectory planning method. This method can be applied to, for example... Figure 1In the application environment shown, the processor 110 can communicate with the sensor component 120 via a network. The sensor component 120 is used to collect perception data around the vehicle and the vehicle's own motion state and pose data, and transmits them to the processor 110. The processor 110 is used to execute a trajectory planning method based on the received data to generate a target planned trajectory and output it to the vehicle's drive-by-wire system.

[0022] The processor 110 can be an in-vehicle domain controller, a high-performance in-vehicle computing unit, or a distributed in-vehicle computing cluster consisting of multiple computing nodes.

[0023] The processor 110 may be implemented in at least one of the following hardware forms: programmable logic array (PLA), field-programmable gate array (FPGA), digital signal processor (DSP), application-specific integrated circuit (ASIC), general-purpose processor, or other programmable logic device.

[0024] The sensor assembly 120 may include, but is not limited to, a forward-looking camera, a surround-view camera, a millimeter-wave radar, a lidar, an inertial measurement unit, a wheel speed sensor, and a global positioning system.

[0025] Of course, the vehicle driving trajectory planning method provided in this application embodiment can also be applied to more scenarios not shown in the figure.

[0026] Applying vehicle driving trajectory planning methods to Figure 1 Taking the processor 110 as an example, in some embodiments, such as Figure 2 As shown, the vehicle driving trajectory planning method includes steps S210 to S240 that can be executed by the processor 110.

[0027] Step S210: Acquire environmental perception data, motion state data and pose data of the vehicle at multiple historical moments.

[0028] Environmental perception data refers to raw sensor data used to characterize information about the vehicle's surrounding environment. In this embodiment, environmental perception data specifically refers to the surrounding environment data collected by the vehicle at multiple historical moments. Its specific form can be selected according to the actual sensor configuration, such as a sequence of continuous video frames from a forward-facing camera.

[0029] Motion state data is a set of parameters used to characterize the current kinematic state of a vehicle, reflecting its motion trend and dynamic characteristics at the current moment. Motion state data includes, but is not limited to, longitudinal velocity, longitudinal acceleration, lateral velocity, lateral acceleration, and yaw rate.

[0030] Pose data is a set of parameters used to characterize a vehicle's position and orientation in space. Position information can be obtained through methods such as GPS, inertial navigation systems, or wheel speed odometers; attitude information includes parameters such as heading angle, pitch angle, and roll angle.

[0031] Step S220: Extract scene features of the current driving scenario based on environmental perception data.

[0032] Scene features are semantic feature representations obtained by encoding and extracting environmental perception data. Deep neural networks are typically used as encoders to map the raw data into a low-dimensional semantic feature space, thereby retaining key information related to driving decisions (such as lane line positions, drivable areas, etc.) while removing redundant information unrelated to driving decisions (such as texture details, lighting changes, etc.).

[0033] Step S230: Generate multiple candidate trajectories based on scene features.

[0034] Each candidate trajectory includes a preset number of trajectory points arranged in chronological order.

[0035] In the field of trajectory planning, due to the uncertainty of the future environment and the possibility of multiple local optima for the planning objective, a planner generates multiple alternative driving paths as candidate trajectories. Each candidate trajectory consists of a series of trajectory points arranged in chronological order. Each trajectory point typically contains information such as the vehicle's expected position, heading angle, and other information at a future moment. The entire sequence of trajectory points constitutes the complete driving path of the vehicle from the current moment to a future moment.

[0036] In this embodiment, multiple candidate trajectories are generated in parallel based on scene features, and the multiple candidate trajectories may have different driving styles or spatial orientations.

[0037] For example, the total number of candidate trajectories can be 32, each candidate trajectory can include 8 trajectory points, the time interval between each trajectory point is 0.5 seconds, and the total planning time is 4 seconds. The above parameters are for illustrative purposes only and can be adjusted according to the computing resources and control cycle of a specific vehicle model.

[0038] Step S240: Determine the target planning trajectory based on the cost scores of multiple candidate trajectories.

[0039] In this embodiment of the application, the trajectory points in each candidate trajectory are assigned a confidence level during the generation process, and the generation of trajectory points has a temporal dependency relationship, that is, the generation of the next trajectory point depends on the state and confidence level determination result of the previous trajectory point.

[0040] Confidence level is used to characterize the reliability or accuracy of trajectory points. In the field of vehicle driving trajectory planning, due to noise in environmental perception and uncertainty in prediction, not all trajectory points output by the planner are equally reliable. Confidence level is related to the prediction error of the trajectory point; the larger the prediction error, the lower the confidence level, and vice versa. When the confidence level of the previous trajectory point is not lower than a preset threshold, it indicates that the reliability of the current trajectory point is high enough, and no additional information is needed; when the confidence level is lower than the preset threshold, it indicates that the reliability of the current trajectory point is insufficient, and it is necessary to use an environmental prediction model to obtain future information for correction.

[0041] Specifically, the generation of each candidate trajectory includes steps S231 and S232.

[0042] Step S231: If the confidence level of the previously generated trajectory point is not lower than the preset threshold, then generate the next trajectory point based on the scene features.

[0043] The previously generated trajectory point refers to the closest trajectory point that has been generated in chronological order before the current trajectory point to be generated during the point-by-point autoregressive generation process of the candidate trajectory.

[0044] When the confidence level is not lower than a preset threshold, the current trajectory point is considered to be in a reliable state. Since the planner is trained based on a large amount of current observation data, its output already has high accuracy in simple scenarios or near-field trajectory points. Injecting additional future features may actually create noise interference. Therefore, when the trajectory point output by the planner already has sufficient confidence, there is no need to introduce additional future information for correction; the next trajectory point can be generated solely based on the scene features of the current driving scenario.

[0045] While ensuring accuracy, the trajectory point generation method in step S231 has lower computational latency and less resource consumption because it does not involve the reasoning process of the environmental prediction model.

[0046] It should be noted that when the trajectory point to be generated is the first trajectory point in the candidate trajectories, there is no previous trajectory point. Therefore, the first trajectory point should also be generated directly based on the scene features.

[0047] Step S232: If the confidence level of the previous trajectory point is lower than the preset threshold, a preset environment prediction model is used to generate the environment prediction features for the next trajectory point at the future time based on scene features, motion state data, pose data, and the generated driving instruction sequence. The environment prediction features are then fused with the scene features, and the next trajectory point is generated based on the fused features.

[0048] Environmental prediction features refer to the feature representation obtained by predicting the environmental state at a certain future moment through an environmental prediction model.

[0049] An environmental prediction model is a machine learning model that outputs a feature representation (i.e., environmental prediction feature) of the environmental state at a future moment based on input environmental perception data, motion state data, pose data, and driving command sequence.

[0050] A driving instruction sequence refers to a set of driving control instructions that the vehicle has executed or planned from the start time of the candidate trajectory to the last trajectory point generated so far. Driving control instructions include, but are not limited to, low-level control parameters such as steering wheel angle, throttle opening, and braking pressure, and may also include motion control quantities such as lateral acceleration and longitudinal acceleration.

[0051] By using the generated sequence of driving instructions, the environmental prediction model can determine how the vehicle travels from the starting point to the current trajectory point. Based on this historical control information and the target time, it can infer the environmental state at future times while maintaining the trend of the driving instructions.

[0052] In related technologies, models used to predict future environmental states typically employ a recursive rolling prediction approach. This involves first predicting the state at the next time step, then using that state as input to predict the state at the next time step after that, and so on, until the target time step. The drawback of this recursive approach is that early prediction errors accumulate gradually along the time chain, leading to significant deviations in far-end prediction results, a phenomenon known as far-end illusion.

[0053] In this application, the environmental prediction model adopts a direct prediction method. After receiving the target time information, it directly outputs the environmental prediction features corresponding to the target time, without having to generate the environmental prediction features of each intermediate time from the current time to the target time in sequence, thereby avoiding the cumulative propagation of errors along the time chain.

[0054] For example, the mapping relationship of the environmental prediction model can be: ; in, Let C represent the environmental prediction features at time k, C represent the scene features of the current driving scenario, and e represent the vehicle's motion state data. This represents the prefix conditional feature obtained after causal encoding based on the generated driving instruction sequence. This represents the time embedding information at time k. This represents the mapping function of the environmental prediction model. The time embedding information can be a high-dimensional vector obtained by mapping the absolute time value of the target time through a position encoding function. By introducing time embedding information, the environmental prediction model can respond to prediction requests at any target time.

[0055] The mapping result of this mapping relationship does not depend on the prediction result at time k-1, so that the prediction error only comes from a single mapping from the current time to the target time, rather than the cumulative error generated by multiple recursion steps.

[0056] In some embodiments, the generation process of the above two candidate trajectories can be described by the following expression: ; in, Let T represent the candidate trajectory and T represent the number of planned steps. This represents the trajectory point at time t+k. This represents the trajectory generation function. Let C represent the generated trajectory points before time t+k, and let C represent the scene features of the current driving scenario. This represents the environmental prediction features at time t+k. Represents a gated variable. This indicates a feature fusion operation. When the gating variable m=0, the trajectory points are generated based on scene features; when m=1, the trajectory points are generated based on the fusion result of scene features and environmental prediction features.

[0057] Gating variables can be determined in the following ways: ; in, This represents the gate variable for the k-th trajectory point in the p-th candidate trajectory. This represents the confidence level of the (k-1)th trajectory point. This is a preset threshold. This indicates an indicator function that takes the first value (i.e., m=1) when the condition is met, and the second value (i.e., m=0) when the condition is not met.

[0058] This means that for the first trajectory point of each candidate trajectory, since there is no previous trajectory point, the gating variable is fixed to the value that does not trigger the environment prediction model, that is, the first trajectory point is directly generated based on the scene features.

[0059] In this application, a confidence level is assigned to each trajectory point in the candidate trajectory, and the decision to call the environmental prediction model is based on the confidence level gating. This makes the calling of the environmental prediction model no longer fixed and intensive, but triggered on demand according to the reliability of the trajectory point itself. This avoids the computational redundancy problem caused by the repeated calculation of high-dimensional environmental prediction features at each trajectory point, and reduces the additional computational overhead and inference latency in the trajectory planning process. Furthermore, the computational resources of the environmental prediction model can be accurately allocated to the critical moments when future information is truly needed, realizing on-demand allocation and efficient utilization of computational resources.

[0060] Optionally, the preset threshold can be a fixed value or dynamically adjusted according to the current driving scenario. For example, in a highway scenario, the preset threshold can be appropriately increased to increase the frequency of calling the environmental prediction model and improve the prediction accuracy of the long-term trajectory; in a low-speed parking scenario, the preset threshold can be appropriately decreased to reduce computational overhead. In practical applications, the specific value of the preset threshold can be selected within the range of 0.2 to 0.6.

[0061] In some embodiments, such as Figure 3 As shown, the vehicle driving trajectory planning method may also include steps S310 and S320.

[0062] Step S310: Based on the network hidden state of the current trajectory point and the planned pose data corresponding to the current trajectory point, a pre-set confidence estimation network is used to determine the prediction bias scale of the current trajectory point.

[0063] The hidden state of a network is the internal state at each time step during the candidate trajectory generation process. Essentially, it is a compressed encoding of the input information by models such as recurrent neural networks and Transformer decoders. It integrates all historical information from the start time to the current trajectory point, including the history of the vehicle's motion trajectory, the generated sequence of driving instructions, and the constraints of scene features on the current decision.

[0064] Specifically, when generating the k-th trajectory point, the hidden state of the trajectory generation network... It can be obtained through the following recursive relation: ; Where p is the candidate trajectory number. Here, C represents the vehicle's pose data at the previous trajectory point, C represents scene features, and e represents motion state data. To embed information at all times, Candidate trajectory generation function.

[0065] Planned pose data refers to the expected position and heading of the vehicle, calculated by the trajectory generation network, corresponding to the current trajectory point.

[0066] Prediction bias scale is an uncertainty parameter used to quantify the degree of deviation between the planned pose data of the current trajectory point and its true value. Prediction bias scale can be implemented in the form of logarithmic scaling parameters, variance parameters, or standard deviation parameters.

[0067] Specifically, the formula for calculating the prediction bias scale is as follows: ; Among them, the confidence estimation network It receives two inputs: the network hidden state. and planning pose data These two inputs are concatenated and fed into the network, outputting a scalar. That is, the logarithmic scaling parameter. The larger the value, the greater the prediction deviation of the current trajectory point and the higher the uncertainty.

[0068] Step S320: Determine the confidence level of the current trajectory point based on the prediction deviation scale.

[0069] Among them, the prediction bias scale is negatively correlated with the confidence level.

[0070] Specifically, the confidence level can be obtained through the following mapping relationship: ; in, For a function that maps real numbers to (0,1), when When it increases, Decrease, confidence level It then decreases.

[0071] By determining the confidence level point by point, the system can provide an independent reliability assessment for each trajectory point, avoiding the problem of individual difficult trajectory points being masked by the high average confidence level of the entire trajectory.

[0072] In some embodiments, such as Figure 4 As shown, the vehicle driving trajectory planning method may also include steps S410 to S430.

[0073] Step S410: Obtain the reference trajectory point at the time corresponding to the simulated trajectory point in the reference trajectory.

[0074] The reference trajectory is a standard trajectory generated by driving along a prescribed route, used as a supervisory signal during the training phase of the autonomous driving model. The reference trajectory also consists of a series of reference trajectory points arranged in chronological order, with each reference trajectory point corresponding to the vehicle's position and heading at a given time step.

[0075] Step S420: Construct the confidence loss based on the error between the reference trajectory point and the simulated trajectory point, and the initial prediction deviation scale corresponding to the simulated trajectory point.

[0076] Among them, simulated trajectory points refer to the trajectory point prediction results generated by the trajectory generation network based on the input conditions during the model training phase.

[0077] The initial prediction bias scale is obtained through the initial confidence estimation network; the larger the initial prediction bias scale, the smaller its contribution to the confidence loss.

[0078] Confidence loss is a loss function used to supervise and optimize the parameters of the confidence estimation network. Its expression is: ; in, To simulate the vehicle's pose data at trajectory points, The vehicle's pose data is used as a reference trajectory point. The initial prediction bias scale corresponding to the simulated trajectory points. This represents the confidence loss value.

[0079] The loss function described above matches the initial prediction bias scale with the actual prediction error. Trajectory points with larger errors are assigned a larger prediction bias scale, corresponding to lower confidence levels; trajectory points with smaller errors are assigned a smaller prediction bias scale, corresponding to higher confidence levels.

[0080] Step S430: Adjust the parameters of the initial confidence network according to the confidence loss to determine the pre-set confidence estimation network.

[0081] Specifically, in each iteration, the weight and bias parameters of the initial confidence network are updated using the backpropagation algorithm based on the gradient information calculated from the confidence loss, so that the confidence loss gradually decreases. As the number of iterations increases, the parameters of the initial confidence network gradually converge to a state that satisfies the convergence condition. At this point, the network parameters are fixed, and the pre-set confidence estimation network is determined.

[0082] The convergence condition can be that the number of training rounds reaches a preset value, or the confidence loss drops below a preset threshold, etc.

[0083] By iteratively correcting the parameters of the initial confidence network through confidence loss, the confidence estimation network can automatically learn the mapping relationship between the network hidden state and the planned pose data and the prediction bias scale during the training process. This avoids the subjectivity and limitations of manually setting confidence mapping rules, making the confidence calibration more objective and accurate.

[0084] In some embodiments, such as Figure 5 The method shown uses a preset environment prediction model to generate the environment prediction features for the next trajectory point at the future time based on scene features, motion state data, pose data, and the currently generated driving command sequence, including steps S510 to S530.

[0085] Step S510: Arrange the scene features, motion state data, pose data, and driving command sequence into a unified sequence.

[0086] A unified sequence refers to a single input sequence formed by arranging the various input information required by an environmental prediction model in a preset order.

[0087] Specifically, a unified sequence arranges scene features, motion state data, pose data, and driving command sequences into different elements within the same sequence. For example, a unified sequence can be represented as: ; Where C represents the scene characteristics of the current driving scenario, and e represents the vehicle motion state data. For historical pose data, to This is the sequence of generated driving instructions from the start time to the current time.

[0088] Step S520: Encode the uniform sequence using a causal attention mechanism.

[0089] Causal attention mechanisms are used to restrict any element in a uniform sequence to focus only on elements preceding the current element during encoding.

[0090] Step S530: Generate environmental prediction features based on the encoding results.

[0091] Specifically, through the causal attention mask matrix The attention weights are constrained, and their encoding expression is as follows: ; Among them, the mask matrix The mechanism works as follows: when the i-th element and the j-th element in the unified sequence satisfy i < j (i.e., the j-th element is after the i-th element), the attention weight between the two is reset to negative infinity, so that the weight approaches zero after normalization.

[0092] The encoding result refers to the prefix conditional features obtained after processing a uniform sequence through a causal attention mechanism. The encoding result integrates all information from scene features, motion state data, historical pose data, and the driving command sequence, while preserving the temporal causal order between the driving commands. Each element in the encoding result corresponds to a position in a unified sequence, and the features at that position only contain information from that position and preceding it. The environment prediction model further generates environment prediction features for the target time based on this encoding result.

[0093] The causal attention mechanism ensures that any element in the unified sequence can only access the element preceding it during encoding, so that the encoding of each driving instruction in the driving instruction sequence strictly follows the temporal causal order. Furthermore, the complete causal encoding enables the environmental prediction model to capture the continuous changing trend of driving operations, thereby making the predicted future environmental features highly consistent with the actual movement trend of the vehicle.

[0094] In some embodiments, such as Figure 6As shown, the environmental prediction features are fused with the scene features, including steps S610 to S630.

[0095] Step S610: Use the scene features as the query matrix and the environment prediction features as the key matrix and value matrix.

[0096] In attention mechanisms, the query matrix is ​​a matrix obtained by linearly transforming the feature representation for which information needs to be obtained, and is used to initiate information retrieval from other feature representations. Specifically, the query matrix is ​​obtained by linearly mapping the scene features of the current driving scenario.

[0097] The current scene features are the baseline information for candidate trajectory generation. Their semantic meaning is the current environmental state of the vehicle. Using scene features as a query means taking the current observation as the main factor and retrieving future information that is relevant to the current scene and useful for trajectory correction from the environmental prediction features.

[0098] In the attention mechanism, the key matrix and value matrix are matrices obtained by performing two different linear transformations on the retrieved feature representation. The key matrix is ​​used to calculate the similarity with the query matrix, and the value matrix is ​​used for weighted summation based on the similarity. Specifically, both the key matrix and value matrix are obtained by linearly mapping the environment prediction features. The environment prediction features characterize the environmental state at future moments, containing future information that may affect the current trajectory point. The key matrix measures which locations in the future environment are most relevant to the current scene, and the value matrix is ​​used to actually extract this relevant information.

[0099] Step S620: Calculate the attention weights of the query matrix and the key matrix through a cross-attention mechanism.

[0100] Step S630: Perform a weighted summation of the value matrix according to the attention weights to obtain the fused features.

[0101] Attention weights are weight coefficients obtained by normalizing the similarity between the query matrix and the key matrix. Specifically, attention weights are calculated as follows: ; Where Q is the query matrix, K is the key matrix, V is the value matrix, and d is the feature dimension.

[0102] Each element in the attention weight matrix represents the degree of attention a certain location in the current scene features pays to a certain location in the environment prediction features. When a certain attention weight is large, it indicates that the features at that location in the current scene are strongly correlated with the information at the corresponding location in the environment prediction features, and the information at that location will be prioritized for preservation during fusion.

[0103] Through cross-attention computation, each location in the scene features can selectively extract relevant information from the environmental prediction features based on its correlation with locations in the future environment, achieving targeted fusion of the two features and avoiding interference from irrelevant future information on the current scene features.

[0104] Alternatively, in addition to cross-attention mechanisms, the fusion of environmental prediction features and scene features can also be achieved through direct addition, multilayer perceptrons, or self-attention mechanisms, and this application does not impose any restrictions on this.

[0105] In some embodiments, such as Figure 7 As shown, the vehicle driving trajectory planning method also includes steps S710 to S740.

[0106] Step S710: Obtain the sample dataset.

[0107] A sample dataset refers to the set of data used to train model parameters during the training phase of an environmental prediction model. The sample dataset includes: environmental perception data, sample motion state data, sample pose data, and sample driving command sequences.

[0108] Step S720: Using the initial environment prediction model, based on the sample dataset, determine the sample environment prediction characteristics at the target time.

[0109] The initial environment prediction model refers to the environment prediction model that has not yet completed parameter optimization during the training process. Its inputs include scene features, sample motion state data, sample pose data, and sample driving command sequences from the sample dataset. The output is the sample environment prediction features at the target time. The target time refers to the future point in time that needs to be predicted; it is a prediction target manually specified during the training process.

[0110] In the initial stage of model training, the parameters of the initial environment prediction model are randomly initialized or pre-trained values, and its predictive ability has not yet reached its optimal level.

[0111] Step S730: Determine the first loss value based on the predicted features of the sample environment and the real environmental features corresponding to the target time.

[0112] Step S740: Update the parameters of the initial environment prediction model based on the first loss value to reduce the difference between the predicted features of the sample environment and the real environment features.

[0113] Real-world environmental features refer to the feature representations that characterize the actual environmental state at a given time, extracted from real-world observation data at the target time. During the training phase, the real-world observation data at the target time is known because the training samples are derived from complete driving records, and the actual environmental state at the target time has already been collected and stored by sensors.

[0114] The first loss value is a measure of the difference between the predicted features of the sample environment and the actual features of the environment. For example, the first loss value can be calculated using the mean squared error, and its expression is as follows: ; in, Predicting features for the sample environment. To reflect real-world environmental characteristics, The first loss value is used. The smaller the first loss value, the closer the predicted environmental features of the sample are to the actual environmental features, and the higher the prediction accuracy of the environmental prediction model. Updating the parameters of the initial environmental prediction model based on the first loss value to reduce the difference between the predicted environmental features of the sample and the actual environmental features is one of the basic goals of training the environmental prediction model.

[0115] By using real-world environmental features at the target time as supervisory signals, the environmental prediction model can learn an accurate mapping from the current scene and historical driving operations to the future environmental state. Simultaneously, the use of mean squared error to calculate the first loss value provides a strong penalty for large errors, allowing the environmental prediction model to prioritize correcting predictions with significant biases during training, thereby improving overall prediction accuracy.

[0116] In some embodiments, such as Figure 8 As shown, the vehicle driving trajectory planning method also includes steps S810 to S840.

[0117] Step S810: Using the initial environment prediction model, determine the first environment prediction features at the specified time based on the first sample driving instruction sequence from the first start time to the specified time.

[0118] Step S820: Using the initial environment prediction model, determine the second environment prediction features at the specified time based on the second sample driving instruction sequence from the second start time to the specified time.

[0119] The first and second starting times refer to two different historical time points selected during the training of the environmental prediction model to construct path consistency constraints. The first starting time corresponds to the first sample driving instruction sequence, and the second starting time corresponds to the second sample driving instruction sequence.

[0120] The designated time point refers to the moment that serves as the prediction target during the training process. Located after the first and second start times, the designated time point is the future moment corresponding to both the first and second environment prediction features, and is also a prediction target manually designated during the training process.

[0121] In sequence prediction tasks, the model predicts the output at the target time based on the input sequence from a certain starting point to the target time. If the model obtains different results when predicting the same target time from different starting points, it indicates that the model has failed to learn the predictive ability to truly reflect the environmental state at the target time, and instead encodes the path information between the starting point and the destination into the prediction result.

[0122] The first environmental prediction feature refers to the environmental prediction feature at a specified time, predicted by the initial environmental prediction model based on the first sample driving instruction sequence from the first start time to the specified time. Similarly, the second environmental prediction feature refers to the environmental prediction feature at a specified time, predicted by the initial environmental prediction model based on the second sample driving instruction sequence from the second start time to the specified time. Since the result times of the two prediction tasks are the same but the start times are different, if the model is path-independent, the first environmental prediction feature and the second environmental prediction feature should tend to be consistent.

[0123] Step S830: Determine the second loss value based on the first environmental prediction characteristics and the second environmental prediction characteristics.

[0124] The second loss value refers to a measure of the difference between the first environmental prediction feature and the second environmental prediction feature. In this application, the second loss value can be calculated using the mean squared error form, and its expression is as follows: ; in, Indicates from the intermediate time anchor point Departure time, combined with the corresponding driving instruction sub-segment prediction time The obtained environmental prediction features (i.e., the first environmental prediction features) This indicates predicting the same time starting from the initial time. The obtained environmental prediction features (i.e., the second environmental prediction features). The smaller the second loss value, the lower the dependence of the model prediction results on the initial time and the stronger the independence of the inference path.

[0125] Step S840: Update the parameters of the initial environmental prediction model based on the second loss value to reduce the difference between the first environmental prediction features and the second environmental prediction features.

[0126] In the early stages of training, because the model parameters are not yet optimized, there may be significant differences between the first and second environment prediction features, resulting in a high second loss value. This second loss value, after backpropagation, is used to update the parameters of the initial environment prediction model, enabling the model to learn a mapping rule that the prediction result should be consistent regardless of the starting time, as long as the target time is the same. After multiple rounds of training, the second loss value gradually decreases, the model's sensitivity to the starting time decreases, and path independence increases.

[0127] By predicting the environmental features for the same specified time from two different starting times and calculating a second loss value between the two predictions to constrain the model parameters, the environmental prediction model learns path-independent prediction capabilities. After training, regardless of which historical time the prediction begins, the model outputs consistent environmental prediction features for the same target time.

[0128] In some embodiments, such as Figure 9 As shown, the vehicle driving trajectory planning method also includes steps S910 to S940.

[0129] Step S910: Using the initial environment prediction model, determine the first sample environment prediction features of the target at future times based on the first sample dataset corresponding to the first driving scenario.

[0130] Step S920: Based on the second sample dataset corresponding to at least one second driving scenario in the same training batch, determine the second sample environment prediction features for the target's future time.

[0131] The first driving scenario is the driving scenario corresponding to a sample in the current training batch during the training process of the environment prediction model. Its corresponding first sample dataset includes environmental perception data, sample motion state data, sample pose data and sample driving command sequence in that scenario.

[0132] The second driving scenario refers to another driving scenario within the same training batch that differs from the first driving scenario. The second sample dataset corresponding to the second driving scenario also includes various types of sample data within that scenario. Differences between the first and second driving scenarios may manifest in multiple aspects, such as road structure (e.g., straight roads versus curves), distribution of traffic participants (e.g., open roads versus congested roads), and traffic signal status (e.g., green light versus red light). By selecting multiple different driving scenarios within the same training batch, the model can learn discriminative features between different scenarios during the comparison process.

[0133] The first sample environment prediction feature and the second sample environment prediction feature refer to the environmental prediction features of the target at future times predicted by the initial environment prediction model based on the first sample dataset corresponding to the first driving scenario and the second sample dataset corresponding to the second driving scenario, respectively.

[0134] Step S930: Determine the third loss value based on the first sample environment prediction characteristics and the second sample environment prediction characteristics.

[0135] The third loss value refers to the loss value used to measure the similarity between the predicted environmental features of the first sample and the predicted environmental features of the second sample. For example, the third loss value can be calculated using the InfoNCE loss form in contrastive learning, and its expression is: ; in, The target is the real environmental characteristics at a future moment. As a positive sample, This refers to either the environment prediction feature for the first sample or the environment prediction feature for the second sample. When the first sample environment prediction feature is used, the loss function increases the similarity between the first sample environment prediction feature and the positive sample, while decreasing the similarity between the first sample environment prediction feature and the second sample environment prediction feature. When the second sample environment prediction feature is used, the loss function increases the similarity between the second sample environment prediction feature and the positive sample, while decreasing the similarity between the second sample environment prediction feature and the first sample environment prediction feature.

[0136] In other words, in a specific loss calculation, The anchor point can only be one of the first sample environment prediction feature or the second sample environment prediction feature. When the first sample environment prediction feature is used as the anchor point, the second sample environment prediction feature acts as a negative sample; and vice versa.

[0137] Step S940: Update the parameters of the initial environment prediction model based on the third loss value to reduce the similarity between the first sample environment prediction features and the second sample environment prediction features.

[0138] By calculating the similarity between the first and second sample environment prediction features and constructing a third loss value accordingly, the environment prediction model is incentivized during training to differentiate the prediction features of different driving scenarios in the feature space. After training, the environment prediction model exhibits significant discriminative power in outputting environment prediction features for different driving scenarios, avoiding the loss of scene recognition ability due to representation shrinkage.

[0139] It should be noted that the calculation of the first, second, and third loss values ​​can share the same input data. During training, the three loss values ​​can be calculated simultaneously and backpropagated together, making the training process simple and efficient.

[0140] For example, the total loss function for the three loss values ​​can be expressed as: ; in, The first loss value, This is the second loss value. This is the third loss value. The comparison loss weight for the third loss value, This represents the total loss value.

[0141] In some embodiments, the step S240 of determining the target planning trajectory based on the cost scores of multiple candidate trajectories may further include the following steps: determining the cost score of each candidate trajectory based on any one or more of the following: the distance between each trajectory point and the obstacle, the distance between each candidate trajectory and the boundary of the drivable area, the rate of curvature change of each candidate trajectory, and the lateral acceleration amplitude of each candidate trajectory; and determining the candidate trajectory whose cost score meets the preset conditions as the target planning trajectory.

[0142] Obstacle distance refers to the distance between each trajectory point in the candidate trajectory and surrounding obstacles. In vehicle driving trajectory planning, obstacles include other vehicles, pedestrians, non-motorized vehicles, roadblocks, and other dynamic or static objects that may affect the safe driving of the vehicle. The closer the trajectory point is to the obstacle, the higher the collision risk and the greater the safety cost of the trajectory. Specifically, for each trajectory point in the candidate trajectory, the minimum distance between the vehicle position corresponding to that trajectory point and each perceived obstacle can be calculated, and this distance can be statistically analyzed over the entire trajectory range (e.g., taking the minimum, average, or weighted sum) as the cost of the candidate trajectory in the obstacle distance dimension.

[0143] The distance to the drivable area boundary refers to the distance between each trajectory point in the candidate trajectory and the boundary of the drivable area. The drivable area is the area within which vehicles can safely travel, determined based on road structure, lane lines, traffic markings, and other information. Its boundary typically corresponds to road edges, lane lines, medians, etc. If a candidate trajectory point is too close to the boundary of the drivable area, it means that the vehicle is at risk of crossing the boundary, potentially leading to lane departure or exiting the road. Therefore, the smaller the distance between the trajectory point and the boundary of the drivable area, the greater the potential cost.

[0144] The rate of change of curvature refers to the degree to which the curvature of a candidate trajectory changes along its length. Curvature is a geometric quantity describing the degree of bending of a curve, while the rate of change of curvature reflects the smoothness of the transition from a straight section to a curved section or vice versa. An excessively large rate of change of curvature indicates a sharp turn or abrupt change in direction, which can lead to significant lateral acceleration and steering impact on the vehicle, affecting ride comfort and driving stability. Therefore, the larger the absolute value of the rate of change of curvature, the greater the cost.

[0145] Lateral acceleration amplitude refers to the magnitude of the vehicle's lateral acceleration at each trajectory point in the candidate trajectory. Lateral acceleration is the lateral acceleration generated by a vehicle when turning or changing lanes, and its amplitude directly reflects the magnitude of the lateral force felt by passengers inside the vehicle. Excessive lateral acceleration amplitude not only affects ride comfort but can also lead to vehicle instability, especially dangerous on slippery roads or at high speeds. Therefore, the larger the absolute value of the lateral acceleration amplitude, the greater the cost.

[0146] The cost score is a numerical value obtained by quantitatively evaluating candidate trajectories across multiple dimensions. For example, the cost score can be a weighted sum of the cost values ​​calculated for four dimensions: obstacle distance, drivable area boundary distance, rate of curvature change, and lateral acceleration amplitude. A lower cost score indicates better overall performance of the candidate trajectory in terms of safety, compliance, and comfort.

[0147] Preset conditions refer to the filtering criteria used to determine the target planning trajectory from multiple candidate trajectories. Specifically, the candidate trajectory with the lowest cost score can be selected as the target planning trajectory. In other embodiments, preset conditions may also include a cost score below a certain preset upper limit, or a cost score ranking among the top few in the sorting, etc.

[0148] By calculating the cost score of candidate trajectories across four dimensions—obstacle distance, drivable area boundary distance, rate of curvature change, and lateral acceleration amplitude—the selection of the target trajectory simultaneously considers safety, compliance, and comfort. Compared to selection methods that only consider a single dimension, multi-dimensional comprehensive scoring can more comprehensively evaluate the quality of candidate trajectories, avoiding unsatisfactory selection results due to good performance in one dimension but deficiencies in others.

[0149] It should be understood that, although Figures 2 to 9 The steps in the flowchart are shown sequentially according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Figures 2 to 9 Unless otherwise expressly stated herein, the steps illustrated and other steps involved in the embodiments are not subject to strict order restrictions and may be performed in other orders. Furthermore, at least some steps in the foregoing embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but may be performed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0150] In the second aspect, such as Figure 10 As shown, an embodiment of this application provides a vehicle driving trajectory planning device 1000, which includes: The data acquisition module 1010 is used to acquire environmental perception data, motion state data and pose data of the vehicle at multiple historical moments; The scene feature extraction module 1020 is used to extract scene features of the current driving scene based on environmental perception data; The candidate trajectory generation module 1030 is used to generate multiple candidate trajectories based on scene features; each candidate trajectory includes a preset number of trajectory points arranged in chronological order. The target trajectory determination module 1040 is used to determine the target planned trajectory based on the cost scores of multiple candidate trajectories. The candidate trajectory module further includes a first generation unit 1031 and a second generation unit 1032, which generate each candidate trajectory as follows: The first generation unit 1031 is used to generate the next trajectory point based on scene features if the confidence level of the previously generated trajectory point is not lower than a preset threshold. The second generation unit 1032 is used to generate environmental prediction features for the next trajectory point at future time if the confidence level of the previous trajectory point is lower than a preset threshold. The second generation unit 1032 is used to generate environmental prediction features for the next trajectory point at future time based on scene features, motion state data, pose data, and the generated driving instruction sequence, and to fuse the environmental prediction features with scene features and generate the next trajectory point based on the fused features.

[0151] In some embodiments, the vehicle driving trajectory planning device 1000 further includes: The deviation scale prediction unit (not shown) is used to determine the prediction deviation scale of the current trajectory point based on the network hidden state of the current trajectory point and the planning pose data corresponding to the current trajectory point, using a pre-set confidence estimation network. The confidence determination unit (not shown) is used to determine the confidence level of the current trajectory point based on the prediction deviation scale; wherein the prediction deviation scale is negatively correlated with the confidence level.

[0152] In some embodiments, the vehicle driving trajectory planning device 1000 further includes: The reference trajectory determination unit (not shown) is used to obtain the reference trajectory points at the time corresponding to the simulated trajectory points in the reference trajectory; the reference trajectory is a standard trajectory generated by driving according to the standard driving route; The confidence loss determination unit (not shown) is used to construct the confidence loss based on the error between the reference trajectory point and the simulated trajectory point, as well as the initial prediction deviation scale corresponding to the simulated trajectory point; wherein, the initial prediction deviation scale is obtained through the initial confidence estimation network; the larger the initial prediction deviation scale, the smaller its contribution to the confidence loss; The confidence estimation network correction unit (not shown) is used to correct the parameters of the initial confidence network based on the confidence loss, and to determine the preset confidence estimation network.

[0153] In some embodiments, the second generation unit 1032 further includes: The sequence generation subunit (not shown) is used to arrange scene features, motion state data, pose data, and driving command sequences into a unified sequence. The causal coding subunit (not shown) is used to encode a uniform sequence using a causal attention mechanism; the causal attention mechanism is used to limit any element in the uniform sequence to only the elements preceding the current element during encoding; The environmental feature prediction subunit (not shown) is used to generate environmental prediction features based on the coding results.

[0154] In some embodiments, the second generation unit 1032 further includes: The matrix allocation sub-unit (not shown) is used to use scene features as the query matrix and environment prediction features as the key matrix and value matrix; The weight calculation subunit (not shown) is used to calculate the attention weights of the query matrix and the key matrix through a cross-attention mechanism; The fusion subunit (not shown) is used to perform a weighted summation of the value matrix based on the attention weights to obtain the fused features.

[0155] In some embodiments, the vehicle driving trajectory planning device 1000 further includes: The sample acquisition module (not shown) is used to acquire the sample dataset; the sample dataset includes: environmental perception data, sample motion state data, sample pose data, and sample driving command sequence; The sample environment feature prediction module (not shown) is used to determine the sample environment prediction features at the target time based on the sample dataset using the initial environment prediction model. The first loss value determination module (not shown) is used to determine the first loss value based on the predicted features of the sample environment and the real environmental features corresponding to the target time. The first parameter update module (not shown) is used to update the parameters of the initial environment prediction model based on the first loss value, so as to reduce the difference between the predicted features of the sample environment and the actual environment features.

[0156] In some embodiments, the vehicle driving trajectory planning device 1000 further includes: The first off-path environment feature prediction module (not shown) is used to determine the first environment prediction feature at the specified time based on the first sample driving instruction sequence from the first start time to the specified time using the initial environment prediction model. The second off-path environment feature prediction module (not shown) is used to determine the second environment prediction features at the specified time based on the second sample driving instruction sequence from the second start time to the specified time using the initial environment prediction model. The second loss value determination module (not shown) is used to determine the second loss value based on the first environmental prediction characteristics and the second environmental prediction characteristics. The second parameter update module (not shown) is used to update the parameters of the initial environmental prediction model based on the second loss value, so as to reduce the difference between the first environmental prediction feature and the second environmental prediction feature.

[0157] In some embodiments, the vehicle driving trajectory planning device 1000 further includes: The first batch of internal contrast environment feature prediction modules (not shown) is used to determine the first sample environment prediction features of the target at future times based on the first sample dataset corresponding to the first driving scenario using the initial environment prediction model. The first batch of internal contrast environment feature prediction modules (not shown) is used to determine the second sample environment prediction features of the target at future time based on the second sample dataset corresponding to at least one second driving scenario in the same training batch. The third loss value determination module (not shown) is used to determine the third loss value based on the environmental prediction features of the first sample and the environmental prediction features of the second sample. The third parameter update module (not shown) is used to update the parameters of the initial environment prediction model based on the third loss value, in order to reduce the similarity between the first sample environment prediction features and the second sample environment prediction features.

[0158] In the third aspect, such as Figure 11 As shown, an embodiment of this application provides an autonomous driving vehicle 1100, which includes: Sensor component 1110 is used to collect environmental perception data, motion state data and pose data of the vehicle; The vehicle controller 1120 includes one or more processors and a memory storing a computer program. When the processor executes the computer program, it implements the steps of the vehicle driving trajectory planning method provided in any embodiment of the first aspect of this application.

[0159] in, Figure 11 Sensor component 1110 and Figure 1 The sensor component 120 in the above is the same sensor component.

[0160] Sensor components may include, but are not limited to: at least one of a front-view camera, a surround-view camera, a side-view camera, and a rear-view camera, used to collect environmental perception data around the vehicle; at least one of a wheel speed sensor, an inertial measurement unit, an acceleration sensor, and a yaw rate sensor, used to collect vehicle motion state data; and at least one of a global positioning system, an inertial navigation system, a wheel speed odometer, and a heading angle sensor, used to collect vehicle pose data. Furthermore, sensor components may also include active ranging sensors such as lidar, millimeter-wave radar, and ultrasonic radar to supplement or replace the environmental perception capabilities of cameras. The specific configuration of the above sensors can be selected according to the vehicle model, autonomous driving level, and cost requirements; this application does not impose any limitations on this.

[0161] In some embodiments, the internal structure diagram of the vehicle controller 1120 can be as follows: Figure 12 As shown. The vehicle controller 1120 includes a processor, memory, network interface, and database connected via a system bus. The processor of the vehicle controller 1120 provides computing and control capabilities. The memory of the vehicle controller 1120 includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The vehicle controller 1120 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the vehicle controller 1120 stores trajectory planning-related data. The network interface of the vehicle controller 1120 is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the vehicle driving trajectory planning method in any embodiment of this document.

[0162] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the embodiments of this application and does not constitute a limitation on the vehicle controller to which the embodiments of this application are applied. A specific vehicle controller may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0163] In a fourth aspect, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the vehicle driving trajectory planning method provided in any embodiment of the first aspect of this application.

[0164] The computer-readable storage medium may be Figure 11 The non-volatile storage medium in the vehicle controller shown.

[0165] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The aforementioned computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments of this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0166] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0167] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A vehicle driving trajectory planning method, characterized in that, The method includes: Acquire environmental perception data, motion state data, and pose data of the vehicle at multiple historical moments; The scene features of the current driving scenario are extracted based on the environmental perception data; Based on the scene features, multiple candidate trajectories are generated; each candidate trajectory includes a preset number of trajectory points arranged in chronological order. The target planning trajectory is determined based on the cost scores of the multiple candidate trajectories; When generating each candidate trajectory: If the confidence level of the previously generated trajectory point is not lower than a preset threshold, then the next trajectory point is generated based on the scene features. If the confidence level of the previous trajectory point is lower than the preset threshold, a preset environment prediction model is used to generate the environment prediction features for the future time corresponding to the next trajectory point based on the scene features, the motion state data, the pose data, and the generated driving instruction sequence. The environment prediction features are then fused with the scene features, and the next trajectory point is generated based on the fused features.

2. The method according to claim 1, characterized in that, The method further includes: Based on the network hidden state of the current trajectory point and the planned pose data corresponding to the current trajectory point, a pre-set confidence estimation network is used to determine the prediction bias scale of the current trajectory point. The confidence level of the current trajectory point is determined based on the prediction deviation scale; wherein the prediction deviation scale is negatively correlated with the confidence level.

3. The method according to claim 2, characterized in that, The method further includes: Obtain the reference trajectory point at the time corresponding to the simulated trajectory point in the reference trajectory; the reference trajectory is a standard trajectory generated by driving according to a standard driving route; A confidence loss is constructed based on the error between the reference trajectory point and the simulated trajectory point, and the initial prediction deviation scale corresponding to the simulated trajectory point; wherein, the initial prediction deviation scale is obtained through an initial confidence estimation network; the larger the initial prediction deviation scale, the smaller its contribution to the confidence loss; The parameters of the initial confidence network are corrected based on the confidence loss, and the pre-set confidence estimation network is determined.

4. The method according to claim 1, characterized in that, The method employs a preset environment prediction model to generate future time-at-time environment prediction features for the next trajectory point based on the scene features, motion state data, pose data, and the currently generated driving command sequence, including: The scene features, motion state data, pose data, and driving command sequence are arranged into a unified sequence; The unified sequence is encoded using a causal attention mechanism; the causal attention mechanism is used to limit any element in the unified sequence to only focus on elements preceding the current element during encoding; The environmental prediction features are generated based on the encoding results.

5. The method according to claim 1, characterized in that, The process of fusing the environmental prediction features with the scene features includes: The scene features are used as the query matrix, and the environment prediction features are used as the key matrix and value matrix; The attention weights between the query matrix and the key matrix are calculated using a cross-attention mechanism; The value matrix is ​​weighted and summed according to the attention weights to obtain the fused features.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Obtain a sample dataset; the sample dataset includes: environmental perception data, sample motion state data, sample pose data, and sample driving command sequence; Using an initial environment prediction model, based on the sample dataset, the sample environment prediction characteristics at the target time are determined; The first loss value is determined based on the predicted features of the sample environment and the real environment features corresponding to the target time. The parameters of the initial environment prediction model are updated based on the first loss value to reduce the difference between the sample environment prediction features and the real environment features.

7. The method according to claim 6, characterized in that, The method further includes: Using the initial environment prediction model, the first environment prediction feature at the specified time is determined based on the first sample driving instruction sequence from the first start time to the specified time. Using the initial environment prediction model, the second environment prediction features at the specified time are determined based on the second sample driving instruction sequence from the second start time to the specified time. The second loss value is determined based on the first environmental prediction feature and the second environmental prediction feature; The parameters of the initial environmental prediction model are updated based on the second loss value to reduce the difference between the first environmental prediction feature and the second environmental prediction feature.

8. The method according to claim 6, characterized in that, The method further includes: Using the initial environment prediction model, the first sample environment prediction features of the target at future time are determined based on the first sample dataset corresponding to the first driving scenario. Based on the second sample dataset corresponding to at least one second driving scenario in the same training batch, determine the second sample environment prediction features of the target at future time. Based on the first sample environment prediction features and the second sample environment prediction features, a third loss value is determined; The parameters of the initial environment prediction model are updated based on the third loss value to reduce the similarity between the first sample environment prediction features and the second sample environment prediction features.

9. A vehicle driving trajectory planning device, characterized in that, The device includes: The data acquisition module is used to acquire environmental perception data, motion state data, and pose data of the vehicle at multiple historical moments; The scene feature extraction module is used to extract scene features of the current driving scene based on the environmental perception data; The candidate trajectory generation module is used to generate multiple candidate trajectories based on the scene features; each candidate trajectory includes a preset number of trajectory points arranged in chronological order. The target trajectory determination module is used to determine the target planned trajectory based on the cost scores of the multiple candidate trajectories; The candidate trajectory module further includes a first generation unit and a second generation unit. When generating each candidate trajectory, the candidate trajectory module: The first generation unit is used to generate the next trajectory point based on the scene features if the confidence level of the previously generated trajectory point is not lower than a preset threshold. The second generation unit is used to, if the confidence level of the previous trajectory point is lower than the preset threshold, adopt a preset environment prediction model, generate the environment prediction features for the future time corresponding to the next trajectory point based on the scene features, the motion state data, the pose data, and the generated driving instruction sequence, fuse the environment prediction features with the scene features, and generate the next trajectory point based on the fused features.

10. An autonomous vehicle, characterized in that, The vehicles include: Sensor components are used to collect environmental perception data, motion state data, and pose data of the vehicle. An onboard controller includes one or more processors and a memory storing a computer program, which, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images