Prior-based normalized residual trajectory generation method, system, device, and medium

Abstract

The application provides a prior-based normalized residual trajectory generation method, system, device and medium, which comprises the following steps: obtaining current state information and current environment information of an automatic driving vehicle; determining inertia reference trajectory point coordinates based on the current state information; encoding and fusing the current environment information; and randomly sampling from a standard Gaussian random distribution to obtain initial noise normalized residuals; inputting the current environment features, the initial noise normalized residuals and a first time step vector into a residual prediction network model to obtain denoised normalized residuals; performing inverse normalization on the denoised normalized residuals; and adding the inverse normalized residuals and the inertia reference trajectory point coordinates element by element to obtain generated trajectory point coordinates. The application effectively suppresses the pseudo-correlation learning of the residual prediction network model, improves the robustness and interpretability of the decision logic, effectively solves the planning time domain dilemma, and solves the space-time imbalance problem in trajectory prediction.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a method, system, device and medium for generating normalized residual trajectories based on prior knowledge. Background Technology

[0002] Currently, end-to-end autonomous driving systems typically rely on deep learning models to directly generate vehicle trajectories from sensor data. However, this approach still faces two major challenges: (1) Overly heavy model learning tasks and spurious correlations. The model directly learns the mapping from high-dimensional perception to the future trajectory (usually an absolute coordinate sequence), making the task too complex. The model is prone to learning spurious correlations in the data (such as associating braking with the taillights of the vehicle in front, rather than understanding traffic lights), resulting in a lack of robustness in the decision-making logic; (2) The planning time-domain dilemma. Trajectory data has inherent spatiotemporal imbalances, and the uncertainty of the future trajectory increases significantly with the extension of the prediction time domain. This causes large errors at the far end to dominate the loss function during model training, while small adjustments at the near end that are crucial for safety are ignored.

[0003] Related technologies, such as DiffusionDrive, which aim to achieve vehicle trajectory planning by directly generating complete trajectories, have failed to fundamentally solve the problems of excessive model learning tasks, spurious correlations, and the temporal dilemma of planning. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, this application provides a priori-based normalized residual trajectory generation method, system, device and medium to solve the above technical problems.

[0005] According to one aspect of the embodiments of this application, a method for generating normalized residual trajectories based on prior knowledge is provided. The method includes: acquiring current state information and current environment information of an autonomous vehicle; determining the coordinates of inertial reference trajectory points of the autonomous vehicle based on the current state information; encoding and fusing the current environment information to obtain current environment features; and randomly sampling from a standard Gaussian random distribution to obtain an initial noisy normalized residual; inputting the current environment features, the initial noisy normalized residual, and a first time step vector into a residual prediction network model to obtain a denoised normalized residual; the residual prediction network model is used to generate a normalized residual based on prior knowledge. Given the current environmental characteristics, the initial noisy normalized residual is iteratively denoised and updated according to the first time step vector. The residual prediction network model is trained using historical environmental information and historical trajectory data. The first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The denoised normalized residual is inversely normalized to obtain the inversely normalized residual. The inversely normalized residual and the coordinates of the inertial reference trajectory point are added element-wise to obtain the coordinates of the generated trajectory point of the autonomous vehicle, so as to control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory point.

[0006] In one embodiment of this application, if the historical trajectory data includes: historical trajectory point coordinates and historical driving speed, then the process of training the pre-built network model with historical environmental information and historical trajectory data to obtain the residual prediction network model includes: determining the historical inertial reference trajectory point coordinates based on the historical trajectory point coordinates and the historical driving speed;

[0007] Based on the coordinates of the historical trajectory points and the coordinates of the historical inertial reference trajectory points, the true residual of the trajectory points is calculated; the true residual of the trajectory points is normalized to obtain the normalized true residual; noise is applied to the normalized true residual to obtain the noise residual; the historical environmental information is encoded and fused to obtain the historical environmental features; the noise residual, the historical environmental features, and the second time step vector are input into the pre-constructed network model to obtain the normalized prediction residual; the second time step vector is obtained by embedding and transforming the preset training time step of the diffusion process; a loss function is constructed according to the normalized true residual and the normalized prediction residual, and the parameters in the pre-constructed network model are updated with the goal of minimizing the loss function to obtain the residual prediction network model.

[0008] In one embodiment of this application, the calculation formula for the loss function includes: ,in, Represents the loss function. This represents the total number of trajectory points. Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis; the formula for calculating the noise residuals includes: ,in, Indicates the first The trajectory points are at noise residual on the shaft, Indicates the first Cumulative noise scheduling parameters at each time step Indicates the first The trajectory points are at Normalized truth residuals on the axis This represents standard Gaussian noise, which follows a Gaussian distribution. The axis is shaft or axis.

[0009] In one embodiment of this application, if the true residual of the trajectory points includes: different trajectory points in True residuals on the axis and different trajectory points at The process of normalizing the true residuals of the trajectory points on the axis to obtain the normalized true residuals includes: based on the true residuals of all trajectory points on the axis. The true residual on the axis is determined in Maximum and minimum true residuals on the axis; based on all trajectory points on The true residual on the axis is determined in The maximum and minimum true residuals on the axis; based on the The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized truth residuals on the axis; based on The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized truth residuals on the axis.

[0010] In one embodiment of this application, each trajectory point is in The formulas for calculating the normalized true residual on the axis include: ,in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis Represents a constant, used to prevent division by zero; each trajectory point is in The formulas for calculating the normalized true residual on the axis include: ,in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis This represents a constant and is used to prevent division by zero.

[0011] In one embodiment of this application, if the current environment information includes: surrounding multi-view images, surrounding radar point cloud data, and a preset precision map within a preset area, the process of encoding and fusing the current environment information to obtain current environment features includes: converting the surrounding multi-view images to the coordinate system of the autonomous vehicle using camera extrinsic parameters to obtain multi-view images in the vehicle coordinate system; converting the surrounding radar point cloud data to the coordinate system of the autonomous vehicle using radar extrinsic parameters to obtain radar point cloud data in the vehicle coordinate system; converting the preset precision map within the preset area to the coordinate system of the autonomous vehicle using the positioning information of the autonomous vehicle to obtain a preset precision map in the vehicle coordinate system; encoding the multi-view images in the vehicle coordinate system to obtain image bird's-eye view features; encoding the radar point cloud data in the vehicle coordinate system to obtain point cloud bird's-eye view features; encoding the preset precision map in the vehicle coordinate system to obtain map bird's-eye view features; and fusing the image bird's-eye view features, the point cloud bird's-eye view features, and the map bird's-eye view features at multiple scales to obtain the current environment features.

[0012] In one embodiment of this application, if the current state information includes: current position coordinates and current velocity, then the formula for calculating the coordinates of the inertial reference trajectory point includes: ,in, Indicates the first The coordinates of the inertial reference trajectory points at each time step. Indicates the current position coordinates. Indicates the current speed, in the... The coordinates of the inertial reference trajectory point, the current position coordinates, and the current velocity at each time step are all vectors.

[0013] According to another aspect of the embodiments of this application, a priori-based normalized residual trajectory generation system is also provided, comprising: an information acquisition module for acquiring current state information and current environment information of an autonomous vehicle; an information processing module for determining the coordinates of the inertial reference trajectory points of the autonomous vehicle based on the current state information; encoding and fusing the current environment information to obtain current environment features; and randomly sampling from a standard Gaussian random distribution to obtain an initial noisy normalized residual; and a residual prediction module for inputting the current environment features, the initial noisy normalized residual, and a first time step vector into a residual prediction network model to obtain a denoised normalized residual; the residual prediction network... The network model is used to iteratively denoise and update the initial noisy normalized residual based on the current environmental features and the first time step vector. The residual prediction network model is obtained by training a pre-constructed network model with historical environmental information and historical trajectory data. The first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The trajectory generation module is used to perform inverse normalization on the denoised normalized residual to obtain the inverse normalized residual. The inverse normalized residual and the coordinates of the inertial reference trajectory point are added element-wise to obtain the coordinates of the generated trajectory point of the autonomous vehicle, so as to control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory point.

[0014] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device implements the prior-based normalized residual trajectory generation method as described above.

[0015] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a computer processor, causes the computer to perform the prior-based normalized residual trajectory generation method as described above.

[0016] The beneficial effects of this application are as follows: This application obtains the current state information and current environment information of an autonomous vehicle. Based on the current state information, it determines the coordinates of the inertial reference trajectory points of the autonomous vehicle. It encodes and fuses the current environment information to obtain current environment features, and randomly samples from a standard Gaussian random distribution to obtain an initial noisy normalized residual. The current environment features, the initial noisy normalized residual, and the first time step vector are input into the residual prediction network model to obtain a denoised normalized residual. The residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual according to the first time step vector, using the current environment features as a condition. It then performs an inverse normalization operation on the denoised normalized residual to obtain an inverse normalized residual. Finally, it adds the inverse normalized residual and the coordinates of the inertial reference trajectory points element-wise to obtain the coordinates of the generated trajectory points of the autonomous vehicle, thereby controlling the vehicle to perform autonomous driving according to the generated trajectory point coordinates. The above process utilizes a constant velocity physical model. The output inertial reference trajectory point coordinates are used as strong priors and combined with the denoised normalized residuals output by the residual prediction network model. This ensures the rationality of the basic path and gives the residual prediction network model the flexibility to handle complex scenarios. Moreover, the residual prediction network model does not need to learn the complete kinematic laws, but focuses more on the deviation between the current inertial driving state and the optimal decision trajectory, which significantly reduces the learning complexity of the residual prediction network model. It effectively suppresses the pseudo-correlation learning of irrelevant features (such as taillight color) in historical environmental information and historical trajectory data, and improves the robustness and interpretability of the decision logic. The residual prediction network model generates denoised normalized residuals through iterative denoising, which balances the uncertainty of the vehicle trajectory at different time steps. This allows the residual prediction network model to focus on the macroscopic direction at the far end while not ignoring the subtle adjustments at the near end (safety-critical area), thus effectively solving the planning time domain dilemma and resolving the spatiotemporal imbalance problem in trajectory prediction.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0019] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application;

[0020] Figure 2 This is a flowchart illustrating a priori-based normalized residual trajectory generation method, as shown in an exemplary embodiment of this application.

[0021] Figure 3 This is a flowchart illustrating a priori-based normalized residual trajectory generation method, which is yet another exemplary embodiment of this application.

[0022] Figure 4 This is a flowchart illustrating residual calculation and normalization calculation in an exemplary embodiment of this application;

[0023] Figure 5 This is a schematic diagram illustrating the distribution of historical trajectory data, as shown in an exemplary embodiment of this application.

[0024] Figure 6 This is a schematic diagram illustrating the distribution of true residual data of historical trajectory points, as shown in an exemplary embodiment of this application.

[0025] Figure 7 This is an exemplary embodiment of the present application illustrating the distribution of true residual data of historical trajectory points after normalization by the PRNorm (Point-wise Residual Normalization) module;

[0026] Figure 8 This is a block diagram illustrating a priori-based normalized residual trajectory generation system, as shown in an exemplary embodiment of this application. Detailed Implementation

[0027] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0030] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application, such as... Figure 1 , refer to Figure 1 As shown, the system architecture may include a data acquisition device 101 and an electronic device 102. The electronic device 102 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, a neural network computer, etc. Technical personnel can use the electronic device 102 to acquire the current state information and current environment information of the autonomous vehicle. Based on the current state information, the coordinates of the inertial reference trajectory points of the autonomous vehicle are determined. The current environment information is encoded and fused to obtain the current environment features. Random sampling is performed from a standard Gaussian random distribution to obtain the initial noisy normalized residual. The current environment features, the initial noisy normalized residual, and the first time step vector are input into the residual prediction network model to obtain the denoised normalized residual. The residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual according to the first time step vector, using the current environment features as a condition. The denoised normalized residual is then inversely normalized to obtain the inverse normalized residual. The inverse normalized residual and the coordinates of the inertial reference trajectory points are added element-wise to obtain the coordinates of the generated trajectory points of the autonomous vehicle, so as to control the vehicle to drive autonomously according to the generated trajectory point coordinates. The acquisition device 101 is used to provide the current state information and current environment information of the autonomous vehicle to the electronic device 102 for processing after acquisition.

[0031] Schematic, after acquiring the current state information and current environment information of the autonomous vehicle from the acquisition device 101, the electronic device 102 determines the coordinates of the inertial reference trajectory points of the autonomous vehicle based on the current state information, encodes and fuses the current environment information to obtain the current environment features, and randomly samples from a standard Gaussian random distribution to obtain the initial noisy normalized residual. The current environment features, the initial noisy normalized residual, and the first time step vector are input into the residual prediction network model to obtain the denoised normalized residual. The residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual according to the first time step vector, based on the current environment features. The denoised normalized residual is then inversely normalized to obtain the inverse normalized residual. The inverse normalized residual and the coordinates of the inertial reference trajectory points are added element-wise to obtain the coordinates of the generated trajectory points of the autonomous vehicle, so as to control the vehicle to drive autonomously according to the generated trajectory point coordinates. The above process, through constant The coordinates of the inertial reference trajectory points output by the constant velocity physical model are used as strong priors and combined with the denoised normalized residuals output by the residual prediction network model. This ensures the rationality of the basic path and gives the residual prediction network model the flexibility to handle complex scenarios. Moreover, the residual prediction network model does not need to learn the complete kinematic laws, but focuses more on the deviation between the current inertial driving state and the optimal decision trajectory, which significantly reduces the learning complexity of the residual prediction network model. It effectively suppresses the pseudo-correlation learning of irrelevant features (such as taillight color) in historical environmental information and historical trajectory data, and improves the robustness and interpretability of the decision logic. The residual prediction network model generates denoised normalized residuals through iterative denoising, which balances the uncertainty of the vehicle trajectory at different time steps. This allows the residual prediction network model to pay attention to the macroscopic direction at the far end while not ignoring the subtle adjustments at the near end (safety-critical area), thus effectively solving the planning time domain dilemma and resolving the spatiotemporal imbalance problem in trajectory prediction.

[0032] It should be noted that the prior-based normalized residual trajectory generation method provided in this application embodiment is generally executed by electronic device 102, and correspondingly, the prior-based normalized residual trajectory generation system is generally set in electronic device 102.

[0033] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0034] Figure 2 This is a flowchart illustrating a priori-based normalized residual trajectory generation method, as shown in an exemplary embodiment of this application. (Refer to...) Figure 2 As shown, the prior-based normalized residual trajectory generation method includes at least steps S210 to S240, which are described in detail below:

[0035] In step S210, the current state information and current environment information of the autonomous vehicle are acquired. In one embodiment of this application, the current state information includes: current position coordinates and current speed. The current position coordinates are acquired through a Global Positioning System (GPS) module or an Inertial Navigation System (INS), and the current speed is acquired in real time through the vehicle's built-in speed sensor. The current environment information includes: surrounding multi-view images, surrounding radar point cloud data, and a preset precision map within a preset area. The surrounding multi-view images are acquired through multiple cameras installed around the vehicle, the surrounding radar point cloud data is acquired through a LiDAR device, and the preset precision map is pre-built and updated in real time by a high-precision map service provider. The precision of the high-precision map reaches the centimeter level (10-20 cm). In step S220, based on the current state information, the coordinates of the inertial reference trajectory points of the autonomous vehicle are determined; the current environment information is encoded and fused to obtain the current environment features; and random sampling is performed from a standard Gaussian random distribution to obtain the initial noise-normalized residual. In one embodiment of this application, if the current state information includes: current position coordinates and current speed, the calculation formula for the coordinates of the inertial reference trajectory points includes:

[0036] Equation (1)

[0037] in, Indicates the process Coordinates of the inertial reference trajectory point after time. Indicates the current position coordinates. Indicates the current speed, after which... The coordinates of the inertial reference trajectory point after a certain time, the current position coordinates, and the current velocity are all vectors. The coordinates of the inertial reference trajectory point include: the coordinates of the inertial reference trajectory point... Axis coordinates, inertial reference trajectory points Axis coordinates.

[0038] In step S230, the current environmental features, the initial noisy normalized residual, and the first time step vector are input into the residual prediction network model to obtain the denoised normalized residual. In one embodiment of this application, the residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual based on the current environmental features and the first time step vector; the residual prediction network model is trained on a pre-constructed network model using historical environmental information and historical trajectory data; the first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The embedding and transformation of the preset inference time step of the diffusion process is implemented through a sinusoidal position encoder, and the current environmental features are injected into the residual prediction network model through a cross-attention mechanism or feature concatenation to ensure that the predicted noise conforms to environmental constraints (such as avoiding obstacles).

[0039] In one embodiment of this application, the process of iteratively denoising and updating the initially noisy normalized residual based on the current environmental features and according to the first time step vector includes: inputting the current environmental features, the maximum time step in the first time step vector, and the initially noisy normalized residual into the residual prediction network model to obtain the first predicted noise; calculating the denoised residual of the maximum time step based on the first predicted noise and preset noise scheduling parameters, and updating the maximum time step; inputting the denoised residual of the maximum time step, the current environmental features, and the updated maximum time step (i.e., maximum time step - 1) into the residual prediction network model to obtain the second predicted noise; continuing to calculate the denoised residual of the updated maximum time step based on the second predicted noise and preset noise scheduling parameters, and continuing to update the time step; continuing to input the denoised residual of the updated maximum time step, the current environmental features, and the continuously updated maximum time step (i.e., maximum time step - 2) into the residual prediction network model until the maximum time step is updated to 1, and obtaining the denoised normalized residual.

[0040] In one embodiment of this application, the formula for calculating the denoised residual at the maximum time step is as follows:

[0041] Equation (2)

[0042] in, This represents the denoised residual at the maximum time step. This represents the normalized residual after initial noise reduction. Indicates the first Signal coefficients at each time step Indicates the first Cumulative noise scheduling parameters at each time step This represents the first prediction noise. Indicates the first The standard deviation of the time step, also known as the noise scaling factor, is a hyperparameter used to control the magnitude of randomness. This represents standard Gaussian noise.

[0043] In step S240, the denoised normalized residual is inversely normalized to obtain the inversely normalized residual. The inversely normalized residual is then added element-wise to the coordinates of the inertial reference trajectory points to obtain the coordinates of the generated trajectory points for the autonomous vehicle, thereby controlling the vehicle to perform autonomous driving according to these coordinates. In one embodiment of this application, the formula for calculating the inversely normalized residual is as follows:

[0044] Equation (3)

[0045] in, Indicates in The residuals after inverse normalization on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis This represents the normalized residual after denoising. This indicates the preset hyperparameter, with a value of 1.0. The axis is shaft or axis.

[0046] The formula for calculating the coordinates of the generated trajectory points is as follows:

[0047] Equation (4)

[0048] in, Indicates the coordinates of the generated trajectory points. Indicates the process Coordinates of the inertial reference trajectory point after time. The residual after inverse normalization is given by... The sum of the residuals after inverse normalization on the axis The composition of residuals after inverse normalization on the axis.

[0049] This application combines the coordinates of the inertial reference trajectory points output by the constant velocity physical model as a strong prior with the denoised normalized residuals output by the residual prediction network model. This ensures the rationality of the basic path while giving the residual prediction network model the flexibility to handle complex scenarios. Furthermore, the residual prediction network model does not need to learn complete kinematic laws, but focuses more on the deviation between the current inertial driving state and the optimal decision trajectory, significantly reducing the learning complexity of the residual prediction network model. This effectively suppresses the pseudo-correlation learning of irrelevant features (such as taillight color) in historical environmental information and historical trajectory data, improving the robustness and interpretability of the decision logic. The residual prediction network model generates denoised normalized residuals through iterative denoising, balancing the uncertainty of the vehicle trajectory at different time steps. This allows the residual prediction network model to focus on the macroscopic direction at the far end while not ignoring the subtle adjustments at the near end (safety-critical area), thus effectively overcoming the planning time domain dilemma and solving the spatiotemporal imbalance problem in trajectory prediction.

[0050] In one embodiment of this application, if the historical trajectory data includes: historical trajectory point coordinates and historical driving speed, then the process of training a pre-built network model using historical environmental information and historical trajectory data to obtain a residual prediction network model includes:

[0051] The coordinates of historical inertial reference points are determined based on historical trajectory point coordinates and historical driving speed. In one embodiment of this application, the coordinates of historical inertial reference points represent the default path that the autonomous vehicle will follow without any external control input (such as steering, acceleration, or braking), constituting a strong physical prior. The process of determining the coordinates of historical inertial reference points based on historical trajectory point coordinates and historical driving speed is the same as the process of determining the coordinates of the inertial reference points of the autonomous vehicle based on current state information.

[0052] The true residual of the trajectory points is calculated based on the coordinates of historical trajectory points and historical inertial reference trajectory points. In one embodiment of this application, the true residual of the trajectory points is the difference between the coordinates of historical trajectory points and the coordinates of historical inertial reference trajectory points, and there is a unique correspondence between historical trajectory points and historical inertial reference trajectory points. The number of historical trajectory points is equal to the number of historical inertial reference trajectory points. The true residual of the trajectory points quantifies the amount of correction applied to the default physical path by the driver (or expert strategy) in order to comply with traffic rules, avoid obstacles, or cope with road conditions.

[0053] The true residuals of the trajectory points are normalized to obtain normalized true residuals; noise is then applied to the normalized true residuals to obtain noisy residuals. In one embodiment of this application, the normalized true residual of each trajectory point is a vector, including the values ​​of each trajectory point in the vector vector. Normalized true residuals on the axis, for each trajectory point Normalized true residuals on the axis. The formula for calculating noise residuals is shown below:

[0054] Equation (5)

[0055] in, Indicates the first The trajectory points are at noise residual on the shaft, Indicates the first Cumulative noise scheduling parameters at each time step Indicates the first The trajectory points are at Normalized truth residuals on the axis This represents standard Gaussian noise, which follows a Gaussian distribution. The axis is shaft or axis.

[0056] In one embodiment of this application, normalizing the true residual of the trajectory point is beneficial to align the residual magnitudes under different time domains and different scenarios, and avoids large errors in the long time domain dominating gradient updates during training.

[0057] Historical environmental information is encoded and fused to obtain historical environmental characteristics. In one embodiment of this application, the process of encoding and fusing historical environmental information is the same as the process of encoding and fusing current environmental information.

[0058] The noise residual, historical environmental features, and second time step vector are input into a pre-built network model to obtain the normalized prediction residual. In one embodiment of this application, the second time step vector is obtained by embedding and transforming a preset training time step of the diffusion process; the embedding and transformation of the preset training time step of the diffusion process is implemented through a sinusoidal position encoder, and the historical environmental features are injected into the pre-built network model through a cross-attention mechanism or feature concatenation. The preset training time step can be the same as or different from the preset inference time step.

[0059] Based on the normalized true residual and the normalized predicted residual, a loss function is constructed. The parameters in the pre-built network model are updated with the goal of minimizing the loss function, resulting in a residual prediction network model. In one embodiment of this application, the residual prediction network model can serve as the implementation architecture of the residual predictor. The residual prediction network model can employ a generative network based on a diffusion model (such as Diffusion Transformer, DiT) or a generative network based on flow matching (such as Diffusion Transformer, DiT), which is more suitable for scenarios such as multimodal data prediction and uncertainty estimation. The formula for calculating the loss function includes:

[0060] Equation (6)

[0061] in, Represents the loss function. This represents the total number of trajectory points. Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis.

[0062] In one embodiment of this application, if the true residual of the trajectory points includes: different trajectory points in True residuals on the axis and different trajectory points at The true residuals on the axis are then normalized. The process of obtaining the normalized true residuals includes:

[0063] Based on all trajectory points The true residual on the axis is determined in The maximum and minimum true residuals on the axis. In one embodiment of this application, in The maximum true residual on the axis is the sum of the values ​​of all trajectory points on the axis. The maximum value of the true residual on the axis, in The minimum true residual on the axis is the sum of the values ​​of all trajectory points on the axis. The minimum value of the true residual on the axis.

[0064] Based on all trajectory points The true residual on the axis is determined in The maximum and minimum true residuals on the axis. In one embodiment of this application, in The maximum true residual on the axis is the sum of the values ​​of all trajectory points on the axis. The maximum value of the true residual on the axis, in The minimum true residual on the axis is the sum of the values ​​of all trajectory points on the axis. The minimum value of the true residual on the axis.

[0065] According to The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized true residuals on the axis. In one embodiment of this application, each trajectory point is on... The formulas for calculating the normalized true residual on the axis include:

[0066] Equation (7)

[0067] in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis Represents a constant, with values ​​of , used to prevent division by zero.

[0068] According to The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized true residuals on the axis. In one embodiment of this application, each trajectory point is on... The formulas for calculating the normalized true residual on the axis include:

[0069] Equation (8)

[0070] in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis Represents a constant, with values ​​of , used to prevent division by zero.

[0071] In one embodiment of this application, for each trajectory point in The purpose of normalizing the true residuals on the axis is to normalize the true residuals of each trajectory point on the axis. Transform the true residual on the axis to the symmetric interval [- , ], The preset hyperparameter is set to 1.0, and for each trajectory point... The purpose of normalizing the true residuals on the axis is to normalize the true residuals of each trajectory point. Transform the true residual on the axis to the symmetric interval [- , ].

[0072] In one embodiment of this application, if the current environmental information includes: surrounding multi-view images, surrounding radar point cloud data, and a preset precision map within a preset area, then the process of encoding and fusing the current environmental information to obtain the current environmental features includes:

[0073] By using camera extrinsic parameters, surrounding multi-view images are transformed into the coordinate system of the autonomous vehicle, resulting in multi-view images in the vehicle coordinate system. Similarly, by using radar extrinsic parameters, surrounding radar point cloud data is transformed into the coordinate system of the autonomous vehicle, resulting in radar point cloud data in the vehicle coordinate system. Furthermore, by using the autonomous vehicle's positioning information, a map of a preset precision within a preset area is transformed into the coordinate system of the autonomous vehicle, resulting in a map of a preset precision within the vehicle coordinate system. In one embodiment of this application, the process of transforming surrounding multi-view images into the coordinate system of the autonomous vehicle using camera extrinsic parameters to obtain multi-view images in the vehicle coordinate system can be implemented using conversion tools such as OpenCalib, CamOdoCal, and FAST-Calib. Similarly, the process of transforming surrounding radar point cloud data into the coordinate system of the autonomous vehicle using radar extrinsic parameters to obtain radar point cloud data in the vehicle coordinate system can be implemented using conversion tools such as OpenCalib, LAStools software suite, and rslidar_sdk. The process of converting a map with a preset accuracy within a preset area to the coordinate system of the autonomous vehicle using the positioning information of the autonomous vehicle, and obtaining a map with a preset accuracy in the vehicle coordinate system, can be achieved using conversion tools such as PROJ and CoordTransform.

[0074] Image encoding is performed on multi-view images in the vehicle coordinate system to obtain image bird's-eye view features; point cloud encoding is performed on radar point cloud data in the vehicle coordinate system to obtain point cloud bird's-eye view features; and map encoding is performed on a map with a preset precision in the vehicle coordinate system to obtain map bird's-eye view features. In one embodiment of this application, the process of image encoding of multi-view images in the vehicle coordinate system to obtain image bird's-eye view features can be implemented by an encoder, such as a BEVFormer encoder or a PETR (Position Embedding Transformation) encoder; the process of point cloud encoding of radar point cloud data in the vehicle coordinate system to obtain point cloud bird's-eye view features can be implemented by a point cloud voxel encoder, such as a Voxel-based encoder, a Pillar-based encoder, or a Dynamic Voxel encoder; and the process of map encoding of a map with a preset precision in the vehicle coordinate system to obtain map bird's-eye view features can be implemented by a map encoder, such as a Unified Vector Encoder (UVE) or a Transformer-based map encoder.

[0075] The current environment features are obtained by fusing image bird's-eye view features, point cloud bird's-eye view features, and map bird's-eye view features at multiple scales. In one embodiment of this application, the multi-scale fusion of image bird's-eye view features, point cloud bird's-eye view features, and map bird's-eye view features can be implemented by referring to the process of multi-scale fusion of image bird's-eye view features, point cloud bird's-eye view features, and map bird's-eye view features in related technologies.

[0076] Figure 3 This is a flowchart illustrating a priori-based normalized residual trajectory generation method, as shown in another exemplary embodiment of this application. Figure 3 In this method, the prior-based normalized residual trajectory generation method includes: (1) Sensor input: current state information, surrounding multi-view images, surrounding radar point cloud data, and a preset precision map within a preset area; (2) The perception encoder encodes the surrounding multi-view images, surrounding radar point cloud data, and the preset precision map within a preset area to obtain the current environmental features. The perception encoder can be R4Det (4D millimeter-wave radar-surround-view camera multimodal perception model) or CDM-Net (Cross-View Geo-Localization). Network); (3) The state encoder obtains the coordinates of the inertial reference trajectory point based on the current state information and the constant velocity kinematic model; (4) Input the current environmental features, the initial noise normalized residual and the first time step vector into the residual predictor to obtain the denoised normalized residual; (5) PRNorm module: Perform inverse normalization operation on the denoised normalized residual to obtain the inverse normalized residual; (6) Add the inverse normalized residual and the coordinates of the inertial reference trajectory point element by element to obtain the coordinates of the generated trajectory point of the autonomous vehicle; (7) Control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory point; (8) The residual predictor is obtained by training the pre-built network model through historical environmental information and historical trajectory data.

[0077] Figure 4 This is a flowchart illustrating residual calculation and normalization calculation as shown in an exemplary embodiment of this application, by... Figure 4 It can be seen that the residual calculation and normalization calculation process includes: (1) Inertial reference calculation: inputting historical vehicle state information into the constant speed model to obtain the coordinates of the historical inertial reference trajectory points of the autonomous vehicle; (2) Residual calculation: calculating the true residual of the trajectory points based on the historical trajectory point coordinates and the historical inertial reference trajectory point coordinates; (3) Statistical calculation: calculating the true residual of all trajectory points based on the historical trajectory point coordinates and the historical inertial reference trajectory point coordinates. The true residual on the axis is determined in The maximum and minimum true residuals on the axis are based on the true residuals of all trajectory points. The true residual on the axis is determined in Maximum and minimum truth residuals on the axis; (4) PRNorm normalization: based on The maximum and minimum true residuals on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized truth residuals on the axis; based on The maximum and minimum true residuals on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized true residuals on the axis; (5) Apply noise: for each trajectory point on Adding noise to the normalized true residuals on the axis, we obtain the trajectory point at each point. The noise residual on the axis, and for each trajectory point in Adding noise to the normalized true residuals on the axis, we obtain the trajectory point at each point. Noise residuals on the axis, for each trajectory point The noise residual on the axis and the noise residual at each trajectory point The noise residual on the axis is composed of the noise residual; (6) Model training: Input the noise residual, historical environmental features and the second time step vector into the pre-built network model to obtain the normalized prediction residual, and construct the loss function according to the normalized true residual and the normalized prediction residual, and update the parameters in the pre-built network model with the goal of minimizing the loss function to obtain the residual prediction network model; (7) Inverse normalization: After the residual prediction network model outputs the denoised normalized residual, the denoised normalized residual is inversely normalized to obtain the inversely normalized residual. The inversely normalized residual and the coordinates of the inertial reference trajectory point are added element by element to obtain the coordinates of the generated trajectory point of the autonomous vehicle.

[0078] Figure 5 This is a schematic diagram illustrating the distribution of historical trajectory data in an exemplary embodiment of this application. Figure 5 As can be seen, the historical trajectory data exhibits a distinct "trumpet-shaped" distribution, meaning it is densely distributed at the near end and dispersed at the far end.

[0079] Figure 6 This is a schematic diagram illustrating the distribution of true residual data of historical trajectory points in an exemplary embodiment of this application. Figure 6 As can be seen, although the true residual data of historical trajectory points are centralized, the magnitude differences are still very large.

[0080] Figure 7This is an exemplary embodiment of the present application illustrating the distribution of true residual data of historical trajectory points after normalization by the PRNorm (Point-wise Residual Normalization) module. Figure 7 As can be seen, the true residual data of historical trajectory points after PRNorm normalization is stably compressed within the interval [-1, 1], with a uniform distribution, effectively balancing the optimization weights of different time domains.

[0081] This application uses the coordinates of the inertial reference trajectory points output by the constant velocity physical model as a physical prior, and combines them with the PRNorm normalization method. The advantages compared with the prior art are shown in the table below:

[0082] Table 1 Advantages Comparison Table

[0083]

[0084] As shown in Table 1, this application significantly improves the training efficiency of the residual prediction network model by using the coordinates of the inertial reference trajectory points output by the constant velocity physical model as a strong physical prior and combining it with normalization processing. Experiments have verified that, because PRNorm normalizes the ground truth residuals of all trajectory points to a uniform range, the gradient updates are more stable, and the training convergence speed is improved by more than 30% compared to DiffusionDrive and ResAD. Trajectory quality is improved: the prediction accuracy of near-end trajectory points is improved, especially in scenarios requiring fine control (such as following other vehicles and parking). The robustness of the residual prediction network model is enhanced: the introduction of physical priors allows the residual prediction network model to maintain reasonable outputs even when facing out-of-distribution scenarios, reducing the incidence of abnormal behavior. Interpretability is improved: the visualization of residuals allows engineers to intuitively understand the model's decision logic, facilitating debugging and optimization.

[0085] This application combines a constant velocity physical model as a strong prior with a data-driven residual prediction network model, ensuring the rationality of the basic path while giving the model flexibility to handle complex scenarios. By normalizing point-state residuals, the differences in trajectory amplitude across different time domains and scenarios are eliminated, forcing the residual prediction network model to give equal importance to all prediction points, especially near-end safety critical points, thus significantly improving trajectory quality and training efficiency. The denoised normalized residuals obtained from the predictions are directly visualized as "correction actions." For example, a strong residual pointing to the left indicates "need to change lanes to the left," making the decision-making process of the residual prediction network model more transparent and interpretable. The modeling method of this application can be used as a plug-and-play module, seamlessly integrated into existing MLPs (Multilayer Perceptrons), Transformers (neural network architectures based on self-attention mechanisms), or diffusion model planners, bringing general performance improvements. By perturbing the initial velocity to generate multiple inertial references, a set of context-dependent and diverse trajectory assumptions can be naturally generated, avoiding the use of fixed, static trajectory libraries.

[0086] The following describes an apparatus embodiment of this application, which can be used to execute the prior-based normalized residual trajectory generation system in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the prior-based normalized residual trajectory generation method described above.

[0087] Figure 8 This is a block diagram illustrating a priori-based normalized residual trajectory generation system, as shown in an exemplary embodiment of this application.

[0088] like Figure 8 As shown, this exemplary prior-based normalized residual trajectory generation system 800 includes:

[0089] The information acquisition module 801 is used to acquire the current status information and current environment information of the autonomous vehicle.

[0090] The information processing module 802 is used to determine the coordinates of the inertial reference trajectory points of the autonomous vehicle based on the current state information; encode and fuse the current environmental information to obtain the current environmental features; and randomly sample from the standard Gaussian random distribution to obtain the initial noise-normalized residual.

[0091] The residual prediction module 803 is used to input the current environmental features, the initial noise-inducing normalized residuals, and the first time step vector into the residual prediction network model to obtain the denoised normalized residuals.

[0092] The trajectory generation module 804 is used to perform inverse normalization on the denoised normalized residual to obtain the inverse normalized residual; the inverse normalized residual and the coordinates of the inertial reference trajectory points are added element by element to obtain the coordinates of the generated trajectory points of the autonomous vehicle, so as to control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory points.

[0093] In one embodiment of this application, the current status information includes: current position coordinates and current speed. The current position coordinates are acquired through a Global Positioning System (GPS) module or an Inertial Navigation System (INS), and the current speed is acquired in real time through a speed sensor built into the vehicle. The current environmental information includes: surrounding multi-view images, surrounding radar point cloud data, and a preset precision map within a preset area. The surrounding multi-view images are acquired through multiple cameras installed around the vehicle, the surrounding radar point cloud data is acquired through a LiDAR device, and the preset precision map is pre-built and updated in real time by a high-precision map service provider, with the high-precision map achieving centimeter-level accuracy (10-20 cm).

[0094] In one embodiment of this application, if the current state information includes the current position coordinates and the current velocity, then the calculation formula for the coordinates of the inertial reference trajectory point is shown in equation (1).

[0095] In one embodiment of this application, the residual prediction network model is used to iteratively denoise and update the initially noisy normalized residuals based on the current environmental features and according to the first time step vector. The residual prediction network model is trained on a pre-constructed network model using historical environmental information and historical trajectory data. The first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The embedding and transformation of the preset inference time step of the diffusion process is implemented through a sinusoidal position encoder. The current environmental features are injected into the residual prediction network model through a cross-attention mechanism or feature concatenation to ensure that the predicted noise conforms to environmental constraints (such as avoiding obstacles).

[0096] In one embodiment of this application, the process of iteratively denoising and updating the normalized residual of the initial noise based on the current environmental features and according to the first time step vector includes: inputting the current environmental features, the maximum time step in the first time step vector, and the normalized residual of the initial noise into the residual prediction network model to obtain the first predicted noise; calculating the denoised residual of the maximum time step based on the first predicted noise and preset noise scheduling parameters, and updating the maximum time step; inputting the denoised residual of the maximum time step, the current environmental features, and the updated maximum time step (i.e., maximum time step - 1) into the residual prediction network model to obtain the second predicted noise; continuing to calculate the denoised residual of the updated maximum time step based on the second predicted noise and preset noise scheduling parameters, and continuing to update the time step; continuing to input the denoised residual of the updated maximum time step, the current environmental features, and the updated maximum time step (i.e., maximum time step - 2) into the residual prediction network model until the maximum time step is updated to 1, and obtaining the denoised normalized residual. The calculation formula for the denoised residual of the maximum time step is shown in equation (2).

[0097] In one embodiment of this application, the formula for calculating the inverse normalized residual is shown in Equation (3), and the formula for calculating the coordinates of the generated trajectory points is shown in Equation (4). This application combines the coordinates of the inertial reference trajectory points output by the constant velocity physical model as a strong prior with the denoised normalized residual output by the residual prediction network model. This ensures the rationality of the basic path and gives the residual prediction network model the flexibility to handle complex scenarios. Moreover, the residual prediction network model does not need to learn the complete kinematic laws and focuses more on the deviation between the current inertial driving state and the optimal decision trajectory, significantly reducing The learning complexity of the residual prediction network model is reduced, and the pseudo-correlation learning of irrelevant features (such as taillight color) in historical environmental information and historical trajectory data is effectively suppressed, thereby improving the robustness and interpretability of the decision logic. The residual prediction network model generates denoised normalized residuals through iterative denoising, which balances the uncertainty of vehicle trajectory at different time steps. This allows the residual prediction network model to focus on the macroscopic direction at the far end while not ignoring the subtle adjustments at the near end (safety-critical area), thus effectively solving the temporal domain dilemma of planning and addressing the spatiotemporal imbalance problem in trajectory prediction.

[0098] It should be noted that the prior-based normalized residual trajectory generation system and the prior-based normalized residual trajectory generation method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the prior-based normalized residual trajectory generation system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0099] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to implement the prior-based normalized residual trajectory generation method provided in the various embodiments above.

[0100] Embodiments of this application also provide a vehicle, which includes a priori-based normalized residual trajectory generation system or an electronic device as provided in the above embodiments.

[0101] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for generating normalized residual trajectories based on prior knowledge, characterized in that, The method includes: Obtain the current status information and current environmental information of the autonomous vehicle; Based on the current state information, the coordinates of the inertial reference trajectory points of the autonomous vehicle are determined; the current environmental information is encoded and fused to obtain the current environmental features; and random sampling is performed from a standard Gaussian random distribution to obtain the initial noise-normalized residual. The current environmental features, the initial noisy normalized residual, and the first time step vector are input into the residual prediction network model to obtain the denoised normalized residual. The residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual according to the first time step vector, based on the current environmental features. The residual prediction network model is obtained by training a pre-constructed network model using historical environmental information and historical trajectory data. The first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The denoised normalized residual is subjected to inverse normalization to obtain the inverse normalized residual; the inverse normalized residual and the coordinates of the inertial reference trajectory point are added element by element to obtain the coordinates of the generated trajectory point of the autonomous vehicle, so as to control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory point.

2. The prior-based normalized residual trajectory generation method according to claim 1, characterized in that, If the historical trajectory data includes: historical trajectory point coordinates and historical driving speed, then the process of training the pre-built network model using historical environmental information and historical trajectory data to obtain the residual prediction network model includes: Based on the historical trajectory point coordinates and the historical driving speed, determine the historical inertial reference trajectory point coordinates; Based on the coordinates of the historical trajectory points and the coordinates of the historical inertial reference trajectory points, calculate the true residual of the trajectory points; The true residual of the trajectory points is normalized to obtain the normalized true residual; and noise is applied to the normalized true residual to obtain the noise residual. The historical environmental information is encoded and fused to obtain historical environmental characteristics; The noise residual, the historical environmental features, and the second time step vector are input into the pre-constructed network model to obtain the normalized prediction residual; the second time step vector is obtained by embedding and transforming the preset training time step of the diffusion process. Based on the normalized true residual and the normalized prediction residual, a loss function is constructed, and the parameters in the pre-constructed network model are updated with the goal of minimizing the loss function to obtain the residual prediction network model.

3. The prior-based normalized residual trajectory generation method according to claim 2, characterized in that, The formula for calculating the loss function includes: ， in, Represents the loss function. This represents the total number of trajectory points. Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis Indicates the first The trajectory points are at Normalized truth residuals on the axis Indicates the first The trajectory points are at Normalized prediction residuals on the axis; The formula for calculating the noise residual includes: ， in, Indicates the first The trajectory points are at noise residual on the shaft, Indicates the first Cumulative noise scheduling parameters at each time step Indicates the first The trajectory points are at Normalized truth residuals on the axis This represents standard Gaussian noise, which follows a Gaussian distribution. The axis is shaft or axis.

4. The prior-based normalized residual trajectory generation method according to claim 2, characterized in that, If the true residual of the trajectory points includes: different trajectory points in True residuals on the axis and different trajectory points at The process of normalizing the true residuals on the axis to obtain the normalized true residuals includes: Based on all trajectory points The true residual on the axis is determined in The maximum and minimum true residuals on the axis; Based on all trajectory points The true residual on the axis is determined in The maximum and minimum true residuals on the axis; According to The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized truth residuals on the axis; According to The maximum true residual on the axis, in The minimum true residual on the axis, for each trajectory point in Normalize the true residuals on the axis to obtain the value of each trajectory point. Normalized truth residuals on the axis.

5. The prior-based normalized residual trajectory generation method according to claim 4, characterized in that, Each trajectory point is in Formula for calculating the normalized true residual on the axis include: ， in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis Represents a constant, used to prevent division by zero; Each trajectory point at The formulas for calculating the normalized true residual on the axis include: ， in, Indicates the first The trajectory points are at Normalized truth residuals on the axis This indicates the preset hyperparameters. Indicates the first The trajectory points are at True residuals on the axis Indicates in Minimum true residual on the axis, Indicates in Maximum true residual on the axis This represents a constant and is used to prevent division by zero.

6. The prior-based normalized residual trajectory generation method according to claim 1 or 2, characterized in that, If the current environmental information includes: surrounding multi-view images, surrounding radar point cloud data, and a map of preset precision within a preset area, then the process of encoding and fusing the current environmental information to obtain the current environmental features includes: Using camera extrinsic parameters, the surrounding multi-view images are converted to the coordinate system of the autonomous vehicle to obtain multi-view images in the vehicle coordinate system; using radar extrinsic parameters, the surrounding radar point cloud data is converted to the coordinate system of the autonomous vehicle to obtain radar point cloud data in the vehicle coordinate system; using the positioning information of the autonomous vehicle, the preset precision map in the preset area is converted to the coordinate system of the autonomous vehicle to obtain the preset precision map in the vehicle coordinate system. Image encoding is performed on the multi-view images in the vehicle coordinate system to obtain image bird's-eye view features; point cloud encoding is performed on the radar point cloud data in the vehicle coordinate system to obtain point cloud bird's-eye view features; and map encoding is performed on the preset precision map in the vehicle coordinate system to obtain map bird's-eye view features. The current environment features are obtained by multi-scale fusion of the image bird's-eye view features, the point cloud bird's-eye view features, and the map bird's-eye view features.

7. The prior-based normalized residual trajectory generation method according to claim 1 or 2, characterized in that, If the current state information includes: current position coordinates and current velocity, then the formula for calculating the coordinates of the inertial reference trajectory point includes: ， in, Indicates the first The coordinates of the inertial reference trajectory points at each time step. Indicates the current position coordinates. Indicates the current speed, in the... The coordinates of the inertial reference trajectory point, the current position coordinates, and the current velocity at each time step are all vectors.

8. A priori-based normalized residual trajectory generation system, characterized in that, include: The information acquisition module is used to obtain the current status information and current environmental information of the autonomous vehicle; The information processing module is used to determine the coordinates of the inertial reference trajectory points of the autonomous vehicle based on the current state information; and to encode and fuse the current environmental information to obtain the current environmental features. Then, random sampling is performed from the standard Gaussian random distribution to obtain the initial noise-enhanced normalized residuals; The residual prediction module is used to input the current environmental features, the initial noisy normalized residual, and the first time step vector into the residual prediction network model to obtain the denoised normalized residual. The residual prediction network model is used to iteratively denoise and update the initial noisy normalized residual according to the first time step vector, based on the current environmental features. The residual prediction network model is obtained by training a pre-constructed network model using historical environmental information and historical trajectory data. The first time step vector is obtained by embedding and transforming a preset inference time step of the diffusion process. The trajectory generation module is used to perform inverse normalization on the denoised normalized residual to obtain the inverse normalized residual; the inverse normalized residual and the coordinates of the inertial reference trajectory point are added element by element to obtain the coordinates of the generated trajectory point of the autonomous vehicle, so as to control the vehicle to perform autonomous driving according to the coordinates of the generated trajectory point.

9. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the prior-based normalized residual trajectory generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-readable instructions, which, when executed by the computer's processor, cause the computer to perform the prior-based normalized residual trajectory generation method as described in claims 1-7.

Images