A driving trajectory planning method, device and terminal equipment

Abstract

The application provides a driving trajectory planning method and device and a terminal equipment, which are suitable for the technical field of data processing. The method comprises the following steps: generating driving trajectory planning donor knowledge information and driving trajectory planning receptor knowledge information according to historical driving scene image information, a driving trajectory planning knowledge donor model and a driving trajectory planning knowledge receptor model; obtaining driving trajectory planning donor-receptor alignment information according to the driving trajectory planning donor knowledge information, the driving trajectory planning receptor knowledge information and a driving trajectory planning donor-receptor alignment model; and training the driving trajectory planning knowledge receptor model according to the driving trajectory planning donor-receptor alignment information to obtain a driving trajectory planning model. The application effectively avoids the problems of feature extraction failure caused by noise interference of a traditional model and trajectory drift caused by blind inheritance of incorrect knowledge, and greatly improves the trajectory planning capability of an end-to-end automatic driving system in a complex unstructured scene.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to driving trajectory planning methods, devices and terminal equipment. Background Technology

[0002] Currently, autonomous driving technology is evolving from assisted driving to advanced autonomous driving. The limitations of traditional modular technical architectures are becoming increasingly apparent. End-to-end autonomous driving architectures, capable of constructing a direct mapping from raw sensor inputs to vehicle control signal outputs, have become an important direction for technological development. The introduction of large-scale visual language models provides a new path to solve the "cognitive long tail" problem in autonomous driving. However, high-performance multimodal large-scale models suffer from large parameter scales, high computational overhead, and high memory consumption, which significantly conflict with the computing power, power consumption limitations, and real-time requirements of in-vehicle edge computing platforms. Knowledge distillation, as a mainstream model lightweighting technique, is widely used in research on transferring the reasoning capabilities of large models to lightweight models, becoming a key exploration direction for the implementation of end-to-end autonomous driving.

[0003] In existing technologies, end-to-end driving planning schemes based on knowledge distillation mainly adopt a combination of feature distillation and logical distillation. In the feature distillation stage, the mean squared error is used as a metric to rigidly align the intermediate layer features of the lightweight model and the large model in Euclidean space. In the logical distillation stage, the KL divergence constraint is used to approximate the output distribution of the lightweight model to approximate the output of the large model. At the same time, some schemes introduce variational information bottleneck theory for feature compression, using fixed Lagrange multipliers to control the compression intensity, attempting to achieve noise filtering in the feature extraction stage. Overall, this approach completes the transfer of driving planning knowledge from the large model to the lightweight model to adapt to the deployment requirements of the vehicle platform.

[0004] Existing technologies face numerous technical challenges in practical applications. Feature distillation lacks a noise filtering mechanism, resulting in insufficient model robustness. Logical distillation lacks awareness of uncertainties in the output of large models, posing a risk of erroneous knowledge transmission. Variational information bottleneck methods suffer from rigid parameter adjustment mechanisms, making it difficult to adapt to dynamically changing driving scenarios. Lightweight models, in inheriting the reasoning capabilities of large models, are prone to overfitting environmental noise, limiting their ability to extract key driving semantics. In long-tail driving scenarios, lightweight models may blindly inherit erroneous planning logic from large models, leading to potential safety hazards in autonomous driving. Fixed-ratio feature compression methods are prone to losing key details or leaving noise residue, failing to achieve a balance between semantic preservation and noise removal. Summary of the Invention

[0005] In view of this, embodiments of this application provide a driving trajectory planning method, device, and terminal equipment, aiming to solve the problems existing in the prior art, such as models being susceptible to interference from environmental noise such as rain, snow, and light, lack of uncertainty perception leading to lightweight models blindly inheriting the "illusion" error of large models, inability to adapt to dynamic driving scenarios, high computational load of large models being incompatible with limited resources of vehicle edge platforms, lightweight models being prone to overfitting environmental noise, weak key semantic extraction capabilities, and easy inheritance of incorrect planning logic causing safety hazards.

[0006] The first aspect of this application provides a driving trajectory planning method, including:

[0007] Obtain image information of the current driving scene;

[0008] Based on the current driving scene image information and the driving trajectory planning model, current driving trajectory planning information is generated.

[0009] One aspect of this application provides a driving trajectory planning model generation step, including:

[0010] Acquire historical driving scene image information, driving trajectory planning knowledge donor model, and driving trajectory planning knowledge recipient model;

[0011] Based on the historical driving scene image information, the driving trajectory planning knowledge donor model, and the driving trajectory planning knowledge recipient model, driving trajectory planning donor knowledge information and driving trajectory planning recipient knowledge information are generated.

[0012] Based on the driving trajectory planning donor knowledge information, driving trajectory planning recipient knowledge information, and driving trajectory planning donor-recipient alignment model, driving trajectory planning donor-recipient alignment information is obtained.

[0013] Based on the alignment information of the driving trajectory planning donor and recipient, the driving trajectory planning knowledge recipient model is trained to obtain the driving trajectory planning model.

[0014] A second aspect of this application provides a driving trajectory planning device, comprising:

[0015] The current driving scene image information acquisition module is used to acquire current driving scene image information;

[0016] The current driving trajectory planning information generation module is used to generate current driving trajectory planning information based on the current driving scene image information and the driving trajectory planning model.

[0017] A third aspect of this application provides a terminal device, the terminal device including a memory and a processor, the memory storing a computer program executable on the processor, the processor executing the computer program to implement the steps of the driving trajectory planning method as described in the first aspect above.

[0018] A fourth aspect of this application provides a computer-readable storage medium, comprising: storing a computer program, wherein when executed by a processor, the computer program implements the steps of the driving trajectory planning method as described in the first aspect above.

[0019] Compared with the prior art, the beneficial effects of this application are: this application can effectively avoid the problems of feature extraction failure caused by noise interference and trajectory drift caused by blindly inheriting wrong knowledge in traditional models. At the same time, through the lightweight model architecture, it adapts to the computing power, power consumption and real-time requirements of the vehicle edge computing platform, ensuring that the generation of current driving trajectory planning information is both accurate and efficient, and greatly improving the trajectory planning capability of the end-to-end autonomous driving system in complex unstructured scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 1 of this application;

[0022] Figure 2 This is a schematic diagram illustrating the implementation process of the driving trajectory planning model generation steps provided in Embodiment 1 of this application;

[0023] Figure 3 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 2 of this application;

[0024] Figure 4 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 3 of this application;

[0025] Figure 5 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 4 of this application;

[0026] Figure 6 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 5 of this application;

[0027] Figure 7This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment Six of this application;

[0028] Figure 8 This is a schematic diagram illustrating the implementation process of the driving trajectory planning method provided in Embodiment 7 of this application;

[0029] Figure 9 This is a schematic diagram of the driving trajectory planning device provided in the embodiments of this application;

[0030] Figure 10 This is a schematic diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0031] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0032] To illustrate the technical solution described in this application, specific embodiments are provided below.

[0033] Figure 1 A flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment 1 of this application is shown, and is described in detail below:

[0034] Step S101: Obtain the current driving scene image information.

[0035] In this embodiment, the current driving scene image information can refer to the image information collected in real time by the vehicle-mounted vision sensor during the vehicle's driving process, which can cover lane lines, traffic signals, road obstacles, surrounding vehicles, and environmental features such as rain, snow, and changes in lighting. It can be obtained by continuously capturing various road scenes during driving using the vehicle-mounted vision acquisition device, and thus serve as the current driving scene image information containing complete road environmental features.

[0036] Step S102: Generate current driving trajectory planning information based on the current driving scene image information and the driving trajectory planning model.

[0037] In this embodiment, the current driving scene image information can be input into the driving trajectory planning model, which then completes a series of processing steps, including adaptive variational compression of visual features and planning logic. Finally, the driving trajectory planning model outputs the current driving trajectory planning information, which includes the coordinates of the vehicle's future waypoints and driving action decisions, thereby providing accurate trajectory decision-making basis for the autonomous driving control of the vehicle.

[0038] The driving trajectory planning method provided in this application ensures that the generation of current driving trajectory planning information is both accurate and efficient, greatly improving the trajectory planning capability of end-to-end autonomous driving systems in complex unstructured scenarios, and providing high signal-to-noise ratio and high safety technical support for the lightweight deployment of high-level autonomous driving in vehicles.

[0039] Figure 2 The flowchart illustrating the steps for generating the driving trajectory planning model according to Embodiment 1 of this application is shown below in detail:

[0040] Step S1101: Obtain historical driving scene image information, driving trajectory planning knowledge donor model, and driving trajectory planning knowledge recipient model.

[0041] In this embodiment, historical driving scene image information can refer to image information obtained from historical driving datasets collected by vehicle-mounted vision sensors, covering various conventional and complex unstructured road driving scenarios. The driving trajectory planning knowledge donor model is a multimodal large model with strong semantic understanding, causal logical reasoning and long-tail scene generalization capabilities. The driving trajectory planning knowledge recipient model is a lightweight end-to-end model adapted to the vehicle edge computing platform, which can obtain corresponding information and models from the autonomous driving model training data resource library and the model development platform, respectively.

[0042] Step S1102: Based on the historical driving scene image information, the driving trajectory planning knowledge donor model, and the driving trajectory planning knowledge recipient model, generate driving trajectory planning donor knowledge information and driving trajectory planning recipient knowledge information.

[0043] In this embodiment, historical driving scene image information can be simultaneously input into the driving trajectory planning knowledge donor model and the driving trajectory planning knowledge recipient model. Then, the two models perform multimodal feature extraction and planning logic reasoning on the historical driving scene image information, respectively. The driving trajectory planning knowledge donor model generates driving trajectory planning donor knowledge information containing high-confidence visual features and planning logic, and the driving trajectory planning knowledge recipient model generates driving trajectory planning recipient knowledge information containing lightweight visual features and planning logic, thereby obtaining driving trajectory planning donor knowledge information and driving trajectory planning recipient knowledge information.

[0044] Step S1103: Based on the driving trajectory planning donor knowledge information, driving trajectory planning recipient knowledge information, and driving trajectory planning donor-recipient alignment model, obtain driving trajectory planning donor-recipient alignment information.

[0045] In this embodiment, the driving trajectory planning donor-recipient alignment model includes an adaptive variational information bottleneck module and a dynamic entropy weight filtering module. The model can input driving trajectory planning donor and recipient knowledge information. The adaptive variational information bottleneck module then dynamically adjusts the compression coefficient based on the similarity of the two features, adaptively filtering out environmental noise while aligning semantics. The dynamic entropy weight filtering module then quantifies the prediction uncertainty of the driving trajectory planning knowledge donor model, adaptively generating distillation weights and blocking the transmission of low-confidence or erroneous planning logic. Finally, the two types of knowledge information, after feature alignment and logic filtering, are fused to generate driving trajectory planning donor-recipient alignment information.

[0046] Step S1104: Based on the alignment information of the driving trajectory planning donor and recipient, train the driving trajectory planning knowledge recipient model to obtain the driving trajectory planning model.

[0047] In this embodiment, a total loss function can be constructed by combining task supervision loss, adaptive variational information bottleneck loss, and dynamic entropy weight distillation loss. The various loss weights in this total loss function can be preset by humans. Then, the alignment information of the driving trajectory planning knowledge recipient is used as the training basis, and the driving trajectory planning knowledge recipient model is jointly optimized and trained end-to-end according to the total loss function. During the training process, the parameters of the driving trajectory planning knowledge donor model are frozen, and only the relevant parameters of the driving trajectory planning knowledge recipient model are updated. Then, through multiple rounds of iterative training until the model loss converges, a driving trajectory planning model adapted to the vehicle edge computing platform is obtained.

[0048] The driving trajectory planning method provided in this application can effectively avoid problems such as feature extraction failure caused by noise interference and trajectory drift caused by blindly inheriting incorrect knowledge in traditional models. At the same time, through a lightweight model architecture, it adapts to the computing power, power consumption and real-time requirements of the vehicle edge computing platform, ensuring that the generation of current driving trajectory planning information is both accurate and efficient, and greatly improving the trajectory planning capability of the end-to-end autonomous driving system in complex unstructured scenarios.

[0049] Figure 3 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment 2 of this application is shown. Its difference from Embodiment 1 described above lies in:

[0050] The driving trajectory planning donor-recipient alignment model includes a driving scene image feature alignment sub-model and a driving trajectory planning semantic generation sub-model.

[0051] The driving trajectory planning donor-recipient alignment information includes driving trajectory planning donor-recipient image feature alignment loss information and driving trajectory planning donor-recipient semantic feature alignment loss information.

[0052] Step S1103 specifically includes:

[0053] Step S201: Generate driving trajectory planning donor and recipient image alignment feature information based on the driving trajectory planning donor knowledge information, driving trajectory planning recipient knowledge information, and driving scene image feature alignment sub-model.

[0054] In this embodiment, the visual feature information contained in the knowledge information of the driving trajectory planning donor and the visual feature information contained in the knowledge information of the driving trajectory planning recipient can be input into the driving scene image feature alignment sub-model. Then, the sub-model performs visual feature-level matching and alignment processing on the two types of visual feature information, adaptively filters out environmental noise features in the knowledge information of the driving trajectory planning recipient, and then fuses and reconstructs the feature information that has completed noise filtering and visual feature matching to generate driving trajectory planning donor and recipient image alignment feature information.

[0055] Step S202: Based on the alignment feature information of the driving trajectory planning donor and recipient images and the knowledge information of the driving trajectory planning donor, calculate the alignment loss information of the driving trajectory planning donor and recipient images.

[0056] In this embodiment, the original visual feature information in the knowledge information of the driving trajectory planning donor can be extracted, and the feature similarity and deviation of the donor and recipient images of the driving trajectory planning can be calculated. Then, by quantifying the consistency offset of the two types of feature information in the semantic topological space, the deviation quantization result at the feature level is obtained. Then, the deviation quantization result is normalized to calculate the feature alignment loss information of the driving trajectory planning donor and recipient images.

[0057] Step S203: Calculate the driving trajectory planning donor knowledge weight information based on the driving trajectory planning donor knowledge information and the preset driving trajectory planning donor knowledge filtering rules.

[0058] In this embodiment, the preset driving trajectory planning donor knowledge screening rule can be preset by humans. This rule takes the prediction uncertainty of driving trajectory planning donor knowledge information as the core screening basis. First, the confidence level of each planning logic in the driving trajectory planning donor knowledge information can be quantified. Then, according to the preset driving trajectory planning donor knowledge screening rule, the planning logics with different confidence levels are weighted. High confidence planning logics are given high weights, and low confidence planning logics are given low weights or even zero weights. Then, the weight information after allocation is integrated and normalized to calculate the driving trajectory planning donor knowledge weight information.

[0059] Step S204: Generate driving trajectory planning semantic information based on the image alignment feature information of the driving trajectory planning donor and recipient, the knowledge information of the driving trajectory planning donor, the knowledge weight information of the driving trajectory planning donor, the knowledge information of the driving trajectory planning recipient, and the semantic generation sub-model of driving trajectory planning.

[0060] In this embodiment, the planning logic information contained in the donor and recipient image alignment feature information of driving trajectory planning, the planning logic information contained in the donor knowledge information of driving trajectory planning, the weight information of the donor knowledge information of driving trajectory planning, and the planning logic information contained in the recipient knowledge information of driving trajectory planning can be jointly input into the semantic generation sub-model of driving trajectory planning. Then, the sub-model filters the planning logic of the donor knowledge information of driving trajectory planning based on the weight information of the donor knowledge information of driving trajectory planning, retaining only the planning logic information with high confidence. The filtered planning logic information is then semantically fused and reasoned with the planning logic information of the donor and recipient image alignment feature information of driving trajectory planning and the recipient knowledge information of driving trajectory planning. Finally, the fused semantic reasoning result is optimized and generated to generate the semantic information of driving trajectory planning.

[0061] Step S205: Based on the driving trajectory planning semantic information and the driving trajectory planning donor knowledge information, calculate the driving trajectory planning donor-recipient semantic feature alignment loss information.

[0062] In this embodiment, core planning semantic feature information can be extracted from the knowledge information of the driving trajectory planning donor, and the deviation quantification calculation can be performed between the core feature information and the driving trajectory planning semantic information. Then, the consistency and deviation degree of the two types of semantic information in terms of planning logic and decision tendency are analyzed to obtain the deviation quantification result at the semantic level. Then, the deviation quantification result is standardized to calculate the driving trajectory planning donor and recipient semantic feature alignment loss information.

[0063] The driving trajectory planning method provided in this application realizes hierarchical alignment and loss quantification of image features and semantic features, improves the accuracy of visual feature alignment and the safety of dual-model planning logic transfer, and at the same time, the refined weight allocation of donor knowledge effectively avoids the transmission of low-confidence knowledge, making the feature extraction and logical reasoning capabilities of the driving trajectory planning model better, thus making it more suitable for complex and unstructured autonomous driving scenarios.

[0064] Figure 4 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment 3 of this application is shown. Its difference from Embodiment 2 described above lies in:

[0065] The driving trajectory planning receptor knowledge information includes driving trajectory planning receptor knowledge content information and driving trajectory planning receptor noise information;

[0066] Step S201 specifically includes:

[0067] Step S301: Based on the knowledge information of the driving trajectory planning donor, the knowledge content information of the driving trajectory planning recipient, the noise information of the driving trajectory planning recipient, and the alignment model of the driving scene image features, generate the variance distribution information and the mean distribution information of the driving scene image features.

[0068] In this embodiment, visual feature information from the knowledge information of the driving trajectory planning donor, effective visual feature information from the knowledge information of the driving trajectory planning recipient, and noise feature information from the noise information of the driving trajectory planning recipient are jointly input into the driving scene image feature alignment sub-model. Then, the sub-model performs probabilistic modeling and distribution analysis on the visual feature information from the knowledge information of the driving trajectory planning donor, the effective visual feature information from the knowledge information of the driving trajectory planning recipient, and the noise feature information from the noise information of the driving trajectory planning recipient. It quantifies the degree of interference of the noise information of the driving trajectory planning recipient on the effective visual features, and then calculates the mean and variance of the feature distribution that matches the visual features of the driving trajectory planning donor. The calculated mean and variance information are then accurately quantified and recorded to generate the variance distribution information and mean distribution information of the driving scene image features.

[0069] In this embodiment, the knowledge donor model for driving trajectory planning can be a teacher model, and the knowledge recipient model for driving trajectory planning can be a student model. The knowledge information from the knowledge donor model can be teacher features output by the teacher model, and the knowledge information from the knowledge recipient model can be student features output by the student model.

[0070] The AdaVIB module can be deployed after the feature projection layer of the student model. Its primary task is to transform the deterministic feature vectors into probability distributions, thereby introducing a measure of uncertainty. Specifically, let the student features after dimension alignment be... ,in For batch size, For sequence length, It is the feature dimension. Unlike traditional methods that directly use... Subsequent calculations can be performed to construct a probabilistic encoder to predict the latent distribution of features. .

[0071] Hypothetical latent variables It follows a multidimensional Gaussian distribution. To capture the nonlinear relationships of the features, the encoder employs a multilayer perceptron (MLP) structure:

[0072]

[0073] in:

[0074] The mean of the potential distribution carries the main semantic information of the features (such as lane line location and obstacle category).

[0075] This represents the log-variance of the latent distribution, which quantifies the uncertainty or noise level of the feature. (Predictions are made here.) Instead of direct prediction or This is to ensure numerical stability. The variance must be positive, while the range of values ​​for the output of a neural network is typically... Through exponential transformation This naturally guarantees the non-negativity of variance.

[0076] Step S302: Based on the variance distribution information of the driving scene image features and the mean distribution information of the driving scene image features, reconstruction processing is performed to generate driving trajectory planning donor-recipient image alignment feature information.

[0077] In this embodiment, the core semantic benchmark for feature alignment can be determined based on the mean distribution information of driving scene image features. Noise features that exceed the reasonable distribution range are filtered out based on the variance distribution information of driving scene image features. Then, the feature information is reconstructed and optimized according to the semantic benchmark, retaining effective features that are semantically consistent with the knowledge information of the driving trajectory planning donor, and eliminating redundant features caused by noise information of the driving trajectory planning recipient. Finally, the reconstructed feature information is integrated and normalized to generate driving trajectory planning donor and recipient image alignment feature information.

[0078] In this embodiment, in the variational autoencoder (VAE) class structure, the data is directly derived from the distributed... Mid-sampling is a stochastic process, which is non-differentiable and can prevent gradient backpropagation. To address this issue, this invention introduces a stochastic manifold sampling technique. An auxiliary noise variable, independent of the model parameters, is introduced. It follows a standard normal distribution: Then latent features It can be obtained through deterministic transformation:

[0079]

[0080] in This represents the Hadamard product (element-wise multiplication). This formula removes randomness from the sampling operation and places it in an external variable. Up. During backpropagation, the gradient can flow smoothly through the deterministic path (i.e., and This enables end-to-end training. The generated... It contains the core semantics And then superimposed by The controlled random perturbation forces subsequent networks to learn feature representations that are robust to noise.

[0081] The driving trajectory planning method provided in this application separates effective features from noisy features in the recipient knowledge, and combines driving scene image features with an alignment sub-model to complete the accurate modeling and reconstruction of feature distribution. This efficiently filters out feature interference caused by environmental noise, improves the robustness and accuracy of feature alignment, and makes the trained driving trajectory planning model more capable of feature extraction in harsh environments and complex road conditions, thereby improving the accuracy and safety of driving trajectory planning.

[0082] Figure 5 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment 4 of this application is shown. The difference between this method and Embodiment 2 is that step S203 specifically includes:

[0083] Step S401: Perform probability distribution transformation based on the driving trajectory planning donor knowledge information to obtain driving trajectory planning donor knowledge probability distribution information.

[0084] In this embodiment, the planning logic output information in the knowledge information of the driving trajectory planning donor can be extracted and converted into an unnormalized logarithmic probability form. Then, the logarithmic probability form is normalized by a preset probability transformation rule and mapped into numerical information that conforms to the probability distribution characteristics. Finally, the distribution characteristics of the normalized numerical information are verified to generate the probability distribution information of the driving trajectory planning donor.

[0085] Step S402: Calculate the driving trajectory planning donor knowledge entropy value based on the probability distribution information of the driving trajectory planning donor knowledge and the preset driving trajectory planning donor knowledge entropy calculation function.

[0086] In this embodiment, the preset function for calculating the entropy of knowledge of the driving trajectory planning donor can be preset by the user. This function is based on the Shannon entropy calculation logic in information theory. The probability distribution information of the knowledge of the driving trajectory planning donor can be input into the function. Then, the function calculates the self-information and expectation of each probability value in the probability distribution information, quantifies the overall uncertainty of the knowledge of the driving trajectory planning donor, and then accurately quantifies and records the expected result obtained from the calculation, thereby calculating the entropy value of the knowledge of the driving trajectory planning donor.

[0087] Step S403: Calculate the driving trajectory planning donor knowledge weight information based on the entropy value of the driving trajectory planning donor knowledge and the preset driving trajectory planning donor knowledge filtering rules.

[0088] In this embodiment, the preset rule for filtering knowledge of driving trajectory planning donors can be preset by humans. This rule uses the information entropy value of driving trajectory planning donors as the core uncertainty measurement basis. First, the information entropy value of driving trajectory planning donors can be mapped to a fixed relative value range. Then, the uncertainty index is transformed into the corresponding confidence index through a reversal operation. Next, a sensitivity coefficient is introduced to perform a nonlinear power transformation on the confidence index to enhance the filtering effect on high uncertainty planning knowledge. Then, the transformed confidence index is normalized to transform it into a weight value that can be directly used for distillation training, thereby calculating the weight information of driving trajectory planning donors.

[0089] In this embodiment, the preset knowledge screening rules for driving trajectory planning can be based on a Bayesian uncertainty analysis framework, introducing Shannon Entropy as the core physical indicator for measuring the confidence of the teacher model's predictions. This allows the system to not only focus on the single-point prediction results generated by the teacher model during knowledge transfer, but also to dynamically quantify the epistemological uncertainty of the teacher model under specific driving conditions by analyzing the geometric characteristics of the probability distribution, thereby achieving a refined evaluation of the reliability of the teaching signal. Specifically, a dynamic gating mechanism located at the output logic layer (LogitsLevel) can be constructed to quantify the teacher's uncertainty in real time and automatically adjust the weight of distillation loss accordingly, thereby blocking the transfer of low-quality knowledge.

[0090] Set at time step The unnormalized log probabilities (Logits) output by the teacher model are: ,in This refers to the vocabulary size. First, it is determined by adding a temperature coefficient. The Softmax function transforms it into a probability distribution. :

[0091]

[0092] in:

[0093] Indicates the first word in the vocabulary. Index of each token.

[0094] This is the temperature coefficient, typically set to 1. When... When the probability distribution tends to smooth out, a phenomenon known as "distribution softening," it increases the probability weight of non-dominant categories; when At this time, the distribution exhibits a clear "sharpening" characteristic, which further amplifies the confidence response intensity of the dominant category. In this embodiment, to ensure high-fidelity reproduction of the original confidence distribution characteristics of the teacher model and to avoid interference with the accurate quantification of epistemological uncertainty due to artificially introduced distribution bias, we take... .

[0095] According to information theory principles, entropy is a core indicator for measuring the uncertainty of random variables. The entropy calculation function for the pre-defined knowledge information of the driving trajectory planning donor can be based on a probability distribution. Shannon entropy Defined as the expectation of the self-information of all possible events:

[0096]

[0097] Mathematical property analysis:

[0098] Cognitive deterministic maximum (zero entropy state): when the prediction distribution of the teacher model exhibits unit impulse characteristics (i.e., the probability of a certain token is 1, and the probability of the rest is 0). ,at this time This means that teachers have absolute certainty about the decision.

[0099] Cognitive confusion maximum (maximum entropy state): when the prediction distribution exhibits a discrete uniform distribution (i.e., the probability of all tokens is equal). ),at this time This indicates that the system is currently in a state of lack of discriminative power, that is, it has generated serious epistemological uncertainty and cannot provide effective planning guidance signals.

[0100] Quantitative Measurement Definition: Therefore, information entropy constitutes a continuous and monotonic quantitative index of epistemological uncertainty, whose physical value space is confined to... .

[0101] Traditional screening methods typically employ "hard thresholding," such as discarding samples when the entropy exceeds 1.5. This approach leads to instability during the training process. This invention proposes a smooth, non-linear softweighting mechanism.

[0102] Because the numerical range of entropy is related to the size of the vocabulary. To make the algorithm more universal, the entropy value is first mapped to... The relative interval. Defining relative uncertainty. :

[0103]

[0104] at this time, . Indicates low uncertainty. This indicates high uncertainty.

[0105] In distillation tasks, we want weights Entropy is positively correlated with "confidence level," while it measures "uncertainty." Therefore, a reversal operation is needed to define a base confidence level. :

[0106]

[0107] Enhanced numerical stability: introduced here This is to handle the tiny floating-point errors that may occur in numerical calculations and ensure that the confidence level is non-negative.

[0108] Nonlinear confidence shaping: based on the base confidence level It is linear, and its discrimination accuracy for "high cognitive confusion" samples is limited when representing decision confidence, making it difficult to effectively suppress low-fidelity teaching signals in complex long-tail scenarios. To construct a highly discriminative adaptive screening mechanism, this invention introduces a sensitivity coefficient. (SensitivityFactor, usually taken as...) The final dynamic weights are obtained by performing a nonlinear power transform on the original signal. :

[0109]

[0110] The geometric meaning of this formula:

[0111] This is a convex function curve.

[0112] High confidence response zone: When the teacher model exhibits high cognitive certainty (e.g. )and At that time, the calculated dynamic weights Within this range, the weight decay coefficient is extremely small, and the system aims to retain the vast majority of the learning signal with high fidelity, ensuring that the student model can be completely decoupled from the deterministic prior of the teacher model.

[0113] Uncertainty transition zone: When the teacher model exhibits moderate cognitive confusion (e.g.) )and At that time, dynamic weight The weights rapidly drop to 0.25. Within this range, the weights exhibit a significant non-linear drop, indicating that the system begins to actively suppress potential low-fidelity signals, significantly weakening the gradient contribution of the fuzzy decision logic to the student model.

[0114] High-entropy noise suppression region: when the teacher model is in a state of high cognitive confusion (e.g.) When γ = 2, the dynamic weights Approaching zero (approximately 0.01). At this point, the system achieves a gradual "gradient masking" of anomalous samples at the loss function level through an extremely small gain coefficient, effectively blocking the risk of negative transfer caused by the "illusion" of the teacher model.

[0115] By adjusting We can control the strictness of the screening: The larger the number of hallucinations, the lower the system's tolerance for teacher hallucinations.

[0116] The driving trajectory planning method provided in this application embodiment achieves accurate quantification of the uncertainty of the donor knowledge, and then completes nonlinear weight allocation based on the preset driving trajectory planning donor knowledge screening rules, making the donor knowledge screening process smoother and the training more stable. At the same time, it can accurately identify and suppress low-confidence planning knowledge, significantly reducing the probability of erroneous knowledge transmission, thereby improving the logical reasoning ability and decision safety of the driving trajectory planning model in long-tail driving scenarios, and better adapting to the complex and ever-changing actual autonomous driving road conditions.

[0117] Figure 6 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment 5 of this application is shown. The difference between this method and Embodiment 2 described above is that step S204 specifically includes:

[0118] Step S501: Calculate the weighted information of the driving trajectory planning donor knowledge based on the driving trajectory planning donor knowledge information and the driving trajectory planning donor knowledge weight information.

[0119] In this embodiment, the planning logic features and visual semantic features in the knowledge information of the driving trajectory planning donor can be extracted and weighted and fused with the weight information of the driving trajectory planning donor in a dimension-by-dimensional manner. Then, high-confidence planning knowledge is given a high-weight enhanced feature expression, and low-confidence planning knowledge is given a low-weight weakened feature expression. Finally, the weighted and fused feature information is integrated and normalized to generate the weighted information of the driving trajectory planning donor.

[0120] Step S502: Generate driving trajectory planning semantic information based on the image alignment feature information of the driving trajectory planning donor and recipient, the weighted information of driving trajectory planning donor knowledge, the knowledge information of driving trajectory planning recipient, and the semantic generation sub-model of driving trajectory planning.

[0121] In this embodiment, the image alignment feature information of the driving trajectory planning donor and recipient, the weighted information of the driving trajectory planning donor knowledge, and the knowledge information of the driving trajectory planning recipient can be jointly input into the semantic generation sub-model of driving trajectory planning. Then, the sub-model performs semantic fusion and reasoning on the three types of information. It prioritizes learning the high-confidence planning logic in the weighted information of the driving trajectory planning donor knowledge, and then combines it with the visual semantics of the image alignment feature information of the driving trajectory planning donor and recipient. It adapts the lightweight feature expression of the driving trajectory planning recipient knowledge information, and then optimizes and generates the semantic information after fusion and reasoning, thereby generating the semantic information of driving trajectory planning.

[0122] The driving trajectory planning method provided in this application first performs weighted processing on the driving trajectory planning donor knowledge information before participating in semantic generation. This makes high-confidence planning knowledge dominate the semantic fusion reasoning, weakens the influence of low-confidence knowledge, improves the generation quality of driving trajectory planning semantic information, and makes subsequent semantic feature alignment more accurate. The trained driving trajectory planning model can inherit high-quality planning knowledge more efficiently, thereby enhancing the reliability of the model's decision-making in complex scenarios.

[0123] Figure 7 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment Six of this application is shown. The difference between this method and Embodiment Two described above is that step S1104 specifically includes:

[0124] Step S601: Based on the driving trajectory planning donor and recipient image feature alignment loss information, train the driving scene image feature alignment sub-model to generate the trained driving scene image feature alignment sub-model.

[0125] In this embodiment, the loss information of the driver trajectory planning donor and recipient image feature alignment can be used as the training supervision basis to backpropagate the driver scene image feature alignment sub-model. Then, the internal parameters of the sub-model are continuously adjusted to reduce the deviation value in the feature alignment process and improve the feature matching ability of the driver trajectory planning donor knowledge information and the driver trajectory planning recipient knowledge information. Then, through multiple rounds of iterative training until the loss of the sub-model converges, the trained driver scene image feature alignment sub-model is generated.

[0126] In this embodiment, the teacher model features of the frozen parameters can be utilized. As a "semantic anchor," the teacher model, having been pre-trained on massive amounts of data, possesses features with extremely strong anti-interference capabilities and semantic representation abilities.

[0127] First, calculate the potential distribution center of students. With teacher characteristics Orientation consistency in high-dimensional feature space. To eliminate the influence of feature magnitude on similarity calculation, we employ vector-level cosine similarity:

[0128]

[0129] in For a minimal constant (e.g.) ( ), to prevent the denominator from being zero.

[0130] Step S602: Generate a driving trajectory planning model to be trained based on the alignment sub-model and the driving trajectory planning semantic generation sub-model of the trained driving scene image features.

[0131] In this embodiment, the training-based driving scene image feature alignment sub-model and the driving trajectory planning semantic generation sub-model can be hierarchically fused. The output of the training-based driving scene image feature alignment sub-model is connected to the input of the driving trajectory planning semantic generation sub-model, thereby constructing a complete inference link from image feature alignment to semantic logic generation. Then, the fused model architecture is initialized with parameters and the link is debugged to ensure smooth information transmission between modules, thereby generating the driving trajectory planning model to be trained.

[0132] Step S603: Calculate the comprehensive loss of driver trajectory planning donor and recipient features based on the driver trajectory planning donor and recipient image feature alignment loss information, the driver trajectory planning donor and recipient semantic feature alignment loss information, and the preset driver trajectory planning feature alignment loss calculation weights.

[0133] In this embodiment, the preset weights for calculating the feature alignment loss of driving trajectory planning can be preset manually. These weights include image feature alignment loss weights and semantic feature alignment loss weights. The image feature alignment loss information of the driving trajectory planning donor and recipient can be multiplied with the corresponding image feature alignment loss weights, and the semantic feature alignment loss information of the driving trajectory planning donor and recipient can be multiplied with the corresponding semantic feature alignment loss weights. Then, the two weighted loss results are summed, and the summed result is quantified as a whole to calculate the comprehensive feature loss of the driving trajectory planning donor and recipient.

[0134] In this embodiment, a weighted Kullback-Leibler (KL) divergence can be constructed as the final logistic distillation loss function. For a length of... The driving planning sequence, total distillation loss Defined as:

[0135]

[0136] in This is the predicted distribution for the student model.

[0137] During backpropagation, we need to update the parameters of the student model. Find the relationship between the above loss function and the loss function. gradient:

[0138]

[0139] Physical explanation: This acts as an "Adaptive Gradient Gain Modulator". Because... It is calculated solely from the frozen teacher model, and it applies to the student model parameters. In this context, it is a constant and does not participate in the calculation of the second derivative during backpropagation. It is only used as a scaling factor for the gradient vector to achieve fine-tuning of the intensity of knowledge transfer at different time steps. Blocking effect: When the teacher experiences hallucinations... (High) ,lead to This means that, for this specific time step The fact that the parameters of the student model are hardly updated ensures that the parameter evolution of the student model is in an instantaneous stagnation state at specific outlier time steps or under high confusion conditions, thereby avoiding contamination by erroneous knowledge.

[0140] The final training objective is unified through a global gradient aggregator. Total loss function. It is composed of a weighted average of task supervision loss, manifold divergence loss, and logical transition loss:

[0141]

[0142] in A standard cross-entropy loss based on ground truth trajectory labels is used to ensure the acquisition of basic driving skills. During training, the teacher model parameters are frozen throughout, and the gradient flow only updates the lightweight marginal neural planner and its interface parameters. Through the synergy of these multiple mechanisms, the system achieves adaptive noise filtering at the feature layer and adaptive blocking of illusions at the logic layer, achieving highly robust end-to-end capability resonance.

[0143] Step S604: Train the driving trajectory planning model to be trained according to the comprehensive loss of the driving trajectory planning donor and recipient features to obtain the driving trajectory planning model.

[0144] In this embodiment, the comprehensive loss of the driving trajectory planning features can be used as the overall training supervision basis. The driving trajectory planning model to be trained after the architecture fusion has been completed can be jointly optimized and trained end-to-end. Then, during the training process, the relevant parameters of each module of the model are updated synchronously to continuously reduce the comprehensive alignment loss of image features and semantic features. Then, through multiple rounds of iterative training, the comprehensive loss of the model converges and reaches the preset accuracy requirements, thereby obtaining a driving trajectory planning model adapted to the vehicle edge computing platform.

[0145] The driving trajectory planning method provided in this application improves the independent learning ability and accuracy of the feature alignment part, and then trains the overall model by fusing the comprehensive loss of image and semantics, so that the model can achieve collaborative optimization in image feature extraction and semantic logical reasoning, avoid the model capability imbalance caused by single-level loss training, and greatly improve the overall performance of the driving trajectory planning model, so as to enhance the accuracy and robustness in feature alignment and planning decision-making at the same time.

[0146] Figure 8 The flowchart illustrating the implementation of the driving trajectory planning method provided in Embodiment Seven of this application is shown. The difference between this method and Embodiment Six is ​​that step S601 specifically includes:

[0147] Step S701: Calculate the driving trajectory planning mapping feature coefficients based on the alignment loss information of the driving trajectory planning donor and recipient image features and the preset driving trajectory planning mapping feature coefficient calculation function.

[0148] In this embodiment, the preset driving trajectory planning mapping feature coefficient calculation function can be preset by humans. This function takes feature similarity as the core calculation basis. The feature alignment loss information of the driving trajectory planning donor and recipient images can be input into the function. Then, the function quantifies the consistency offset between the knowledge information of the driving trajectory planning donor and the knowledge information of the driving trajectory planning recipient in the semantic topological space. Then, a negatively correlated nonlinear mapping relationship is constructed based on the offset. The coefficient value that adapts to the current feature alignment state is dynamically generated. Then, the generated coefficient value is accurately quantified and verified to calculate the driving trajectory planning mapping feature coefficient.

[0149] In this embodiment, a negatively correlated nonlinear mapping function can be constructed based on vector-level cosine similarity to generate the adaptive information bottleneck throttling coefficient, i.e., the driving trajectory planning mapping feature coefficient. :

[0150]

[0151] This formula incorporates the following key design considerations:

[0152] Function: The hyperbolic tangent function maps the input to... The interval provides a smooth nonlinear transition.

[0153] Negative correlation: in the formula The structure establishes an adaptive compression coefficient. With manifold consistency measure The negative correlation mapping relationship between them was used to construct an adaptive suppression mechanism based on feature signal-to-noise ratio.

[0154] when Low (indicating severe environmental noise interference) Output negative adjustment signal The responsiveness is enhanced. When the system detects a significant deviation between student and teacher characteristics in the manifold space, it automatically strengthens the compression threshold of the information bottleneck and forcibly removes high-variance noise components to ensure the robustness of core driving semantics in complex perception environments.

[0155] when High time (representation semantic space height alignment) Output positive adjustment signal Consequently, the constraint strength of the information bottleneck decreases. When the feature manifold exhibits high-fidelity aligned features, the system adaptively reduces the constraint strength of the information bottleneck, allowing richer detailed features to pass through the encoding path, thus supporting the perception requirements of higher-order planning tasks for subtle changes in the environment and improving prediction accuracy.

[0156] Adjust parameters:

[0157] (BaseScale): Control The overall amplitude benchmark.

[0158] (CenterThreshold): The center of the similarity threshold determines the inflection point of the adjustment curve.

[0159] (SmoothingFactor): The steepness of the control curve determines the system's sensitivity to changes in similarity.

[0160] In this embodiment, the optimization objective of AdaVIB is to minimize the variational upper bound. In the distillation scenario, the teacher feature distribution can be considered as the "target distribution" or "prior distribution." Let the target distribution be... (in It is usually set to a fixed, small value, such as 1).

[0161] Total loss function It consists of two parts: a weighted sum of Kullback-Leibler (KL) divergences. Since both distributions are Gaussian, the KL divergence has an analytical solution (Closed-form Solution).

[0162]

[0163] Step S702: Based on the driving trajectory planning mapping feature coefficients, train the driving scene image feature alignment sub-model to generate the trained driving scene image feature alignment sub-model.

[0164] In this embodiment, the driving trajectory planning mapping feature coefficients can be used as the basis for dynamic adjustment and integrated into the training process of the driving scene image feature alignment sub-model. Then, the feature compression intensity of the sub-model is dynamically adjusted according to the coefficients. When the feature alignment deviation is large, the compression is enhanced to filter out noise, and when the feature alignment deviation is small, the compression is reduced to retain details. Then, the backpropagation parameters are updated by combining the driving trajectory planning feeder image feature alignment loss information. Then, through multiple rounds of dynamic adjustment iterative training until the sub-model loss converges, a trained driving scene image feature alignment sub-model is generated.

[0165] The driving trajectory planning method provided in this application enables the feature alignment process to adaptively adjust the compression strategy according to the actual feature deviation and noise conditions, breaking the limitations of fixed parameter adjustment in the prior art. This significantly improves the adaptability of the sub-model to driving scenarios of different complexity, achieving a better balance between filtering environmental noise and retaining key semantics. The trained model has significantly improved feature extraction and alignment accuracy in unstructured scenarios such as severe weather and dynamic lighting, thereby enhancing the environmental robustness of the driving trajectory planning model.

[0166] Corresponding to the method in the above embodiments, Figure 9 A structural block diagram of the driving trajectory planning device provided in the embodiments of this application is shown. For ease of explanation, only the parts related to the embodiments of this application are shown. Figure 9 The driving trajectory planning device in the example can be the execution subject of the driving trajectory planning method provided in the aforementioned embodiment 1.

[0167] Reference Figure 9 The driving trajectory planning device includes:

[0168] The current driving scene image information acquisition module 810 is used to acquire the current driving scene image information;

[0169] The current driving trajectory planning information generation module 820 is used to generate current driving trajectory planning information based on the current driving scene image information and the driving trajectory planning model.

[0170] For details on how each module in the driving trajectory planning device provided in this application implements its respective function, please refer to the foregoing. Figure 1 The description of Embodiment 1 shown will not be repeated here.

[0171] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0172] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0173] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0174] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0175] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. It should also be understood that although the terms "first," "second," etc., are used in the text to describe various elements in some embodiments of this application, these elements should not be limited by these terms. These terms are merely used to distinguish one element from another. For example, a first table may be named a second table, and similarly, a second table may be named a first table, without departing from the scope of the various described embodiments. Both the first table and the second table are tables, but they are not the same table.

[0176] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0177] The driving trajectory planning method provided in this application can be applied to terminal devices such as mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality / virtual reality devices, laptops, super mobile personal computers, netbooks, and personal digital assistants. This application does not impose any restrictions on the specific type of terminal device.

[0178] Figure 10 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. For example... Figure 10 As shown, the terminal device 9 of this embodiment includes: at least one processor 90 ( Figure 10 (Only one is shown in the image) a memory 91, which stores a computer program 92 that can run on the processor 90. When the processor 90 executes the computer program 92, it implements the steps in the various driving trajectory planning method embodiments described above, for example... Figure 1 Steps S101 to S102 are shown. Alternatively, when the processor 90 executes the computer program 92, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 9 The functions of modules 810 to 820 are shown.

[0179] The terminal device 9 can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art will understand that... Figure 10 This is merely an example of terminal device 9 and does not constitute a limitation on terminal device 9. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal device may also include input transmission devices, network access devices, buses, etc.

[0180] The processor 90 may be a central processing unit, or it may be other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0181] In some embodiments, the memory 91 may be an internal storage unit of the terminal device 9, such as a hard disk or memory of the terminal device 9. The memory 91 may also be an external storage device of the terminal device 9, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the terminal device 9. Furthermore, the memory 91 may include both internal and external storage units of the terminal device 9. The memory 91 is used to store operating systems, applications, bootloaders, data, and other programs, such as the program code of computer programs. The memory 91 can also be used to temporarily store data that has been sent or will be sent.

[0182] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0183] This application also provides a terminal device, which includes at least one memory, at least one processor, and a computer program stored in the at least one memory and executable on the at least one processor. When the processor executes the computer program, it causes the terminal device to implement the steps in any of the above method embodiments.

[0184] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0185] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0186] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0187] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0188] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0189] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0190] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A driving trajectory planning method, characterized in that, include: Obtain image information of the current driving scene; Based on the current driving scene image information and the driving trajectory planning model, generate current driving trajectory planning information; The driving trajectory planning model is obtained through the following steps: Acquire historical driving scene image information, driving trajectory planning knowledge donor model, and driving trajectory planning knowledge recipient model; Based on the historical driving scene image information, the driving trajectory planning knowledge donor model, and the driving trajectory planning knowledge recipient model, driving trajectory planning donor knowledge information and driving trajectory planning recipient knowledge information are generated. Based on the driving trajectory planning donor knowledge information, driving trajectory planning recipient knowledge information, and driving trajectory planning donor-recipient alignment model, driving trajectory planning donor-recipient alignment information is obtained. Based on the alignment information of the driving trajectory planning donor and recipient, the driving trajectory planning knowledge recipient model is trained to obtain the driving trajectory planning model; The driving trajectory planning donor-recipient alignment model includes a driving scene image feature alignment sub-model and a driving trajectory planning semantic generation sub-model. The driving trajectory planning donor-recipient alignment information includes driving trajectory planning donor-recipient image feature alignment loss information and driving trajectory planning donor-recipient semantic feature alignment loss information. The driving trajectory planning receptor knowledge information includes driving trajectory planning receptor knowledge content information and driving trajectory planning receptor noise information; The step of obtaining the driving trajectory planning donor-recipient alignment information based on the driving trajectory planning donor knowledge information, the driving trajectory planning recipient knowledge information, and the driving trajectory planning donor-recipient alignment model specifically includes: Based on the knowledge information of the driving trajectory planning donor, the knowledge content information of the driving trajectory planning recipient, the noise information of the driving trajectory planning recipient, and the driving scene image feature alignment sub-model, the variance distribution information and the mean distribution information of the driving scene image features are generated. Based on the variance distribution information and mean distribution information of the driving scene image features, reconstruction processing is performed to generate driving trajectory planning source image alignment feature information. Based on the alignment feature information of the donor and recipient images in the driving trajectory planning and the knowledge information of the donor in the driving trajectory planning, the feature alignment loss information of the donor and recipient images in the driving trajectory planning is calculated. Based on the driving trajectory planning donor knowledge information, a probability distribution transformation is performed to obtain the driving trajectory planning donor knowledge probability distribution information; Based on the probability distribution information of the knowledge donor for driving trajectory planning and the preset information entropy calculation function for the knowledge donor for driving trajectory planning, the information entropy value of the knowledge donor for driving trajectory planning is calculated. Based on the entropy value of the knowledge information of the driving trajectory planning donor and the preset knowledge filtering rules for driving trajectory planning, the weight information of the knowledge of the driving trajectory planning donor is calculated. Based on the image alignment feature information of the driving trajectory planning donor and recipient, the knowledge information of the driving trajectory planning donor, the knowledge weight information of the driving trajectory planning donor, the knowledge information of the driving trajectory planning recipient, and the semantic generation sub-model of driving trajectory planning, semantic information of driving trajectory planning is generated. Based on the semantic information of driving trajectory planning and the knowledge information of driving trajectory planning donor, the semantic feature alignment loss information of driving trajectory planning donor and recipient is calculated.

2. The driving trajectory planning method as described in claim 1, characterized in that, The step of generating driving trajectory planning semantic information based on the driving trajectory planning donor and recipient image alignment feature information, driving trajectory planning donor knowledge information, driving trajectory planning donor knowledge weight information, driving trajectory planning recipient knowledge information, and driving trajectory planning semantic generation sub-model specifically includes: Based on the driving trajectory planning donor knowledge information and the driving trajectory planning donor knowledge weight information, the driving trajectory planning donor knowledge weighted information is calculated; Based on the image alignment feature information of the driving trajectory planning donor and recipient, the weighted information of driving trajectory planning donor knowledge, the knowledge information of driving trajectory planning recipient, and the semantic generation sub-model of driving trajectory planning, semantic information of driving trajectory planning is generated.

3. The driving trajectory planning method as described in claim 1, characterized in that, The step of training the driving trajectory planning knowledge receptor model based on the driving trajectory planning donor-recipient alignment information to obtain the driving trajectory planning model specifically includes: Based on the driving trajectory planning donor and recipient image feature alignment loss information, the driving scene image feature alignment sub-model is trained to generate the trained driving scene image feature alignment sub-model. Based on the training-post driving scene image feature alignment sub-model and the driving trajectory planning semantic generation sub-model, a driving trajectory planning model to be trained is generated. Based on the image feature alignment loss information of the driving trajectory planning donor and recipient, the semantic feature alignment loss information of the driving trajectory planning donor and recipient, and the preset driving trajectory planning feature alignment loss calculation weights, the comprehensive loss of driving trajectory planning donor and recipient features is calculated. The driving trajectory planning model is trained based on the comprehensive loss of the donor and recipient features of the driving trajectory planning model to obtain the driving trajectory planning model.

4. The driving trajectory planning method as described in claim 3, characterized in that, The step of training the driving scene image feature alignment sub-model based on the driving trajectory planning donor and recipient image feature alignment loss information to generate the trained driving scene image feature alignment sub-model specifically includes: Based on the alignment loss information of the driving trajectory planning donor and recipient image features and the preset driving trajectory planning mapping feature coefficient calculation function, the driving trajectory planning mapping feature coefficients are calculated. Based on the driving trajectory planning mapping feature coefficients, the driving scene image feature alignment sub-model is trained to generate the trained driving scene image feature alignment sub-model.

5. A driving trajectory planning device, characterized in that, For performing the driving trajectory planning method as described in claim 1, including: The current driving scene image information acquisition module is used to acquire current driving scene image information; The current driving trajectory planning information generation module is used to generate current driving trajectory planning information based on the current driving scene image information and the driving trajectory planning model.

6. A terminal device, characterized in that, The terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 4.

Images