Automatic driving model updating method and device based on CTTA framework

Abstract

The invention provides an automatic driving model updating method and device based on a CTTA framework. The method comprises the following steps: constructing an automatic driving model based on a Bayesian neural network structure and automatic driving source data; constructing a variational average teacher model based on the automatic driving model, wherein the variational average teacher model comprises a student model and a teacher model; the student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method. According to the CTTA framework-based automatic driving model updating method and device provided by the invention, a rapid Bayesian model construction method is constructed, and a pre-trained deterministic CNN model is converted into a BNN. And meanwhile, a variational average teacher structure is constructed, so that the prediction accuracy can be improved under distribution change, and reliable uncertainty estimation can also be provided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an automatic driving model updating method and device based on a CTTA framework. Background Art

[0002] In autonomous driving tasks, intelligent vehicles need to adapt to changing scenarios in an open test environment in a timely manner to ensure that the model remains advanced. Continuous Test-Time Adaptation (CTTA) aims to enable the model to adapt to a series of different distribution changes during testing, making it suitable for various risk-sensitive applications in open environments, such as autonomous driving and medical imaging. However, non-stationary test data in the real world exhibit high uncertainty in its temporal dynamics, bringing challenges related to error accumulation. Previous CTTA studies have relied on methods that enforce prediction confidence, such as entropy minimization. However, these methods often lead to overconfident and undercalibrated predictions, thereby limiting the model's ability to quantify risk in predictions.

[0003] Variational Bayesian methods are widely used in machine learning and statistics. They are a statistical method for approximate inference, often applied to probabilistic graphical models, especially when dealing with large-scale data and complex models, because they can effectively approximate complex posterior distributions. In Bayesian inference, researchers want to calculate the posterior probability distribution, that is, the probability distribution of parameters given the observed data. However, for many models, the posterior distribution is difficult to calculate analytically, so approximate inference methods are needed. Variational Bayesian methods solve this problem by approximating the posterior distribution to an easy-to-handle distribution, such as a Gaussian distribution or an exponential family distribution. It obtains an optimal approximation by minimizing the difference between the approximate distribution and the true posterior distribution. This method uses variational inference techniques, in which the variational parameters are adjusted to minimize the difference between the two distributions, usually measured by the KL divergence (Kullback-Leibler divergence). Summary of the invention

[0004] The present invention provides an automatic driving model updating method and device based on the CTTA framework, so as to solve the technical problem of low prediction accuracy caused by excessive prior drift in the CTTA framework in the prior art in automatic driving tasks.

[0005] The present invention provides an automatic driving model updating method based on a continuous testing time domain adaptation (CTTA) framework, comprising the following steps.

[0006] Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data; Constructing a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method.

[0007] In some embodiments, building an autonomous driving model based on the Bayesian model structure and autonomous driving source data includes: Determine prior knowledge based on the Bayesian neural network structure; Determining a source prior distribution of the autonomous driving model parameters based on the prior knowledge and the autonomous driving source data; The autonomous driving model is determined based on the source prior distribution.

[0008] In some embodiments, updating the student model based on variational inference includes: Determine a teacher model prior and an autonomous driving model prior based on the teacher model and the autonomous driving model, respectively; Determining a hybrid prior based on the teacher model prior and the autonomous driving model prior; The student model is updated based on the mixed prior.

[0009] In some embodiments, updating the student model based on the mixed prior comprises: A KL divergence based on a symmetric cross entropy loss and a hybrid prior between the student model and the teacher model; The student model is updated based on the symmetric cross entropy loss and the KL divergence.

[0010] In some embodiments, the method further comprises: An autonomous driving prediction result is determined based on the hybrid prior.

[0011] In some embodiments, updating the teacher model based on the exponential moving average method includes: The teacher model is updated based on the mean and standard deviation of the student model and the mean and standard deviation of the teacher model.

[0012] The present invention also provides an automatic driving model updating device based on a continuous testing time domain adaptation CTTA framework, comprising the following modules: A first building module is used to build an autonomous driving model based on a Bayesian neural network structure and autonomous driving source data; A second building module is used to build a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The first updating module is used to update the student model based on variational inference and to update the teacher model based on an exponential moving average method.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, an autonomous driving model updating method based on the continuous testing time domain adaptive CTTA framework as described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for updating an autonomous driving model based on the continuous testing time domain adaptive CTTA framework as described in any one of the above is implemented.

[0015] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the automatic driving model updating method based on the continuous testing time domain adaptive CTTA framework as described in any one of the above.

[0016] The automatic driving model updating method and device based on the CTTA framework provided by the present invention constructs a fast Bayesian model construction method to convert the pre-trained deterministic CNN model into a BNN. At the same time, a variational average teacher structure is constructed, which can not only improve the prediction accuracy under distribution changes, but also provide reliable uncertainty estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a flowchart of the automatic driving model updating method based on the CTTA framework provided by the present invention.

[0019] Figure 2 is the classification error rate of the standard cifar10 to cifar10c continuous test time adaptation task provided by the present invention.

[0020] Figure 3 is the classification error rate of the standard ImageNet-to-ImageNetC continuous test time adaptation task provided by the present invention.

[0021] Figure 4The present invention provides a fast Bayesian model construction method and symmetric cross entropy when the severity is 5.

[0022] Figure 5 It is a mixed prior with different weights provided by the present invention when the severity is 5.

[0023] Figure 6 It is a structural schematic diagram of an automatic driving model updating device based on a continuous testing time domain adaptive CTTA framework provided by the present invention.

[0024] Figure 7 It is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Figure 1 is a flow chart of the automatic driving model updating method based on the CTTA framework provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 101: Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data.

[0027] Specifically, a fast Bayesian model construction method is a method for converting an existing pre-trained convolutional neural network (CNN) into a pre-trained Bayesian neural network (BNN).

[0028] At the source stage, a pre-trained deterministic vehicle model is first converted into a Bayesian neural network (BNN) through the variational fast Bayesian model building method, and the uncertainty is injected into the autonomous driving model using the local reparameterization technique.

[0029] Step 102: construct a variational average teacher model based on the autonomous driving model, where the variational average teacher model includes a student model and a teacher model.

[0030] Specifically, based on the BNN model, a variational average teacher model is constructed for intelligent vehicles, and the mixing coefficients of teacher priors and autonomous driving priors are calculated in combination with data augmentation.

[0031] Step 103: Update the student model based on variational inference, and update the teacher model based on the exponential moving average method.

[0032] Specifically, a variational inference method based on the variational average teacher model in an open test environment is constructed for the on-board average teacher architecture. According to the test environment data collected by the vehicle in real time, the student model is updated using variational inference (VI), in which symmetric cross entropy is used to calculate the entropy term, and prior mixing is used to calculate the KL term.

[0033] The teacher model is updated by the exponential moving average (EMA) method.

[0034] The automatic driving model updating method based on the CTTA framework provided by the present invention constructs a fast Bayesian model construction method to convert the pre-trained deterministic CNN model into a BNN. At the same time, a variational average teacher structure is constructed, which can not only improve the prediction accuracy under distribution changes, but also provide reliable uncertainty estimation.

[0035] In some embodiments, building an autonomous driving model based on the Bayesian model structure and autonomous driving source data includes: Determine prior knowledge based on the Bayesian neural network structure; Determining a source prior distribution of the autonomous driving model parameters based on the prior knowledge and the autonomous driving source data; The autonomous driving model is determined based on the source prior distribution.

[0036] Specifically, we use the local parameter reparameterization technique to add random parameters and perform a fast Bayesian model building method on the pre-trained model. and pre-trained models , determine the pre-trained deterministic model. Convert an existing pre-trained CNN model to a pre-trained BNN model.

[0037] The prior distribution Initialize to the distribution of the pre-trained model . Thus, the parameters of the pre-trained model are used as prior knowledge.

[0038] Using the variational inference formula:

[0039] in, is the variational distribution, For the distribution space, are model parameters, The zoom factor is set.

[0040] By source data and the prior distribution , approximate update of the prior distribution for . will output Get the source prior distribution Get source data The updated prior distribution is outputted and used in subsequent tasks or domain adaptation.

[0041] The autonomous driving model updating method based on the CTTA framework provided by the present invention adopts real-time VI in the Bayesian inference framework to construct a variational model adaptation method for continuous domain changes in an open test environment, which can reduce the prior drift in continuous model adaptation in an open test environment and obtain significant performance improvements.

[0042] In some embodiments, the method further comprises: An autonomous driving prediction result is determined based on the hybrid prior.

[0043] Specifically, we use the average teacher model architecture and a hybrid prior as an inference reference for each domain variation.

[0044] Calculate the mixing coefficient of the teacher model prior and the autonomous driving model prior to obtain the mixed prior. The formula is as follows:

[0045] Among them, H represents the calculation of entropy, U represents unlabeled data, and I represents the enhancement type. and are the parameters of the source model and the teacher model. τ represents the temperature factor.

[0046] Use the mixed prior to predict the class label by mixing the source model and the teacher model. For a test data point , the inference of the model is expressed as:

[0047] The autonomous driving model updating method based on the CTTA framework provided by the present invention adopts real-time VI in the Bayesian inference framework to construct a variational model adaptation method for continuous domain changes in an open test environment, which can reduce the prior drift in continuous model adaptation in an open test environment and obtain significant performance improvements.

[0048] In some embodiments, updating the student model based on variational inference includes: Determine a teacher model prior and an autonomous driving model prior based on the teacher model and the autonomous driving model, respectively; Determining a hybrid prior based on the teacher model prior and the autonomous driving model prior; The student model is updated based on the mixed prior.

[0049] Specifically, the student model is updated using the VI method , performing adaptation and approximation simultaneously. VI defines a variational distribution To approximate the posterior distribution The approximate process is as follows. According to the formula: ]

[0050] Among them, U represents unlabeled data, Q represents the distribution search space, represents a normalization hyperparameter that is difficult to compute.

[0051] The whole process can be minimized Optimize. According to the formula: L( )+(1- )KL( )

[0052] in is the current enhanced teacher model, and , .

[0053] The autonomous driving model updating method based on the CTTA framework provided by the present invention adopts real-time VI in the Bayesian inference framework to construct a variational model adaptation method for continuous domain changes in an open test environment, which can reduce the prior drift in continuous model adaptation in an open test environment and obtain significant performance improvements.

[0054] In some embodiments, updating the student model based on the mixed prior comprises: A KL divergence based on a symmetric cross entropy loss and a hybrid prior between the student model and the teacher model; The student model is updated based on the symmetric cross entropy loss and the KL divergence.

[0055] Specifically, the calculation of symmetric cross entropy (SCE) loss is:

[0056] Among them This is equivalent to the final minimization formula .

[0057] A variational inference method in an open test environment based on a variational average teacher model is used to update the student model, where the entropy term is calculated using symmetric cross entropy and the KL term is calculated using prior mixing.

[0058] The value of this parameter comes from:

[0059] in = , = ,in, (·) is the confidence function, represents the confidence bound, and 1(·) is the indicator function. It can be seen that a reliable teacher model is represented by the average of its augmented data with higher confidence than the original data.

[0060] To determine the mixing ratio, the average entropy obtained given strict data enhancement is used to represent the confidence of the two prior models, and the two priors are mixed by a modulation factor. The calculation method is as follows: α

[0061] In VI, the present invention uses the upper bound to update the KL divergence term:

[0062] The autonomous driving model updating method based on the CTTA framework provided by the present invention adopts real-time VI in the Bayesian inference framework to construct a variational model adaptation method for continuous domain changes in an open test environment, which can reduce the prior drift in continuous model adaptation in an open test environment and obtain significant performance improvements.

[0063] In some embodiments, updating the teacher model based on the exponential moving average method includes: The teacher model is updated based on the mean and standard deviation of the student model and the mean and standard deviation of the teacher model.

[0064] Specifically, the teacher model is updated through EMA. The specific method is: and are the mean and standard deviation of the student model and the teacher model respectively. Real-time updates: ,

[0065] in, Represents the preset EMA coefficient.

[0066] The autonomous driving model updating method based on the CTTA framework provided by the present invention adopts real-time VI in the Bayesian inference framework to construct a variational model adaptation method for continuous domain changes in an open test environment, which can reduce the prior drift in continuous model adaptation in an open test environment and obtain significant performance improvements.

[0067] The above method is described below with reference to a specific example.

[0068] The validation experiments of this paper use CIFAR10C, CIFAR100C and ImageNetC datasets as benchmarks to evaluate the robustness of the classification model. Each dataset contains 15 different types of damage, each of which is applied at five different severity levels (from 1 to 5). These damages are systematically applied to the test images of the original CIFAR10 and CIFAR100 datasets, as well as the validation images of the original ImageNet dataset.

[0069] For the sake of simplicity of the table, the present invention uses C1 to C15 to represent 15 types of damage, namely C1: Gaussian, C2: Shot, C3: Impulse, C4: Defocus, C5: Glass, C6: Motion, C7: Zoom, C8: Snow, C9: Frost, C10: Fog, C11: Brightness, C12: Contrast, C13: Elastic, C14: Pixelate, C15: Jpeg. For the pre-trained model, following previous studies, the present invention uses a pre-trained WideResNet28 model for CIFAR10 to CIFAR10C, a pre-trained ResNeXt-29 for CIFAR100 to CIFAR100C, and a standard pre-trained ResNet-50 for ImageNet to ImageNetC. The present invention further utilizes a fast Bayesian model construction method to obtain the random dynamics of each data set model. Similar to CoTTA, the present invention updates all trainable parameters in all experiments. The enhancement number of all compared methods was set to 32 and then the enhancement strategy was used.

[0070] For the validation methods, the present invention is compared with a variety of state-of-the-art methods. The SOURCE method represents a baseline pre-trained model without any adaptation. The BN method represents keeping the network parameters frozen but only updating the batch normalization. The TENT method updates the unlabeled test data by Shannon entropy. The CoTTA method establishes the MT structure and uses random recovery parameters to the source model. The SATA method uses source-anchored self-distillation to modify the batch normalization affine parameters to ensure that the model avoids catastrophic forgetting while integrating newly encountered domain knowledge. The SWA method improves the pseudo-label learning process from the perspective of the instantaneous and long-term effects of noisy pseudo-labels. The PETAL method attempts to estimate the uncertainty in CTTA, similar to BNN, but it ignores the prior offset problem. All compared methods use the same backbone, pre-trained model and hyperparameters.

[0071] exist Figure 2 and Figure 3A comparison with other state-of-the-art methods is mainly conducted in this paper. Figure 2 is the classification error rate of the standard cifar10 to cifar10c continuous test time adaptation task provided by the present invention, Figure 3 It is the classification error rate of the standard ImageNet-to-ImageNetC continuous test time adaptation task provided by the present invention. At the same time, it is compared with other advanced methods. All results are evaluated in real time with the maximum damage severity level 5.

[0072] The following results can be observed based on this experiment. First, no adaptation at test time (SOURCE) suffers from severe domain shift, which shows the necessity of CTTA. Second, traditional TTA methods that ignore continuous transfer at test time perform poorly, such as TENT and BN. We find that simple Shannon entropy is effective in the first few domain transfers, especially in the complex 1000-class ImageNetC, but shows a significant performance drop in subsequent transfers. Third, the mean teacher structure is very useful in CTTA, such as COTTA and PETAL, which means that pseudo-labels are useful in domain transfer. In previous methods, error accumulation leads to unreliable pseudo-labels, and then the model may get more negative transfers in CTTA on the timeline. We outperform on all three datasets, such as 13.1% vs. 15.3% (SWA) on CIFAR10C, 28.4% vs. 30.0% (SWA) on CIFAR100C, and 64.2% vs. 66.7% (COTTA) on ImageNetC. The present invention mitigates the process of prior drift through prior mixing, and this advantage is more obvious in the early adaptation stage.

[0073] This example evaluates two components in an ablation experiment. The example results can be seen in Figure 4 , Figure 4 The present invention provides a fast Bayesian model building method (VWU) and symmetric cross entropy under the condition of severity 5, namely, a fast Bayesian model building method (VWU) and symmetric cross entropy (SCE). The ablation experiment results show that both components are important to the present invention. First, VWU is used to introduce random dynamics into the ready-made pre-trained model. Without VWU, the performance of the present invention drops from 13.9% to 18.4% of CIFAR10C, from 28.8% to 31.5% of CIFAR100C, and from 64.2% to 68.1% of ImageNetC. In addition, SCE can further improve the performance of CIFAR10C and CIFAR100C because SCE balances the gradients of high-confidence and low-confidence predictions. The present invention also found that SCE is not effective for complex ImageNetC.

[0074] Prior Mixture introduces a Gaussian mixture strategy where the current prior is approximated as a weighted sum of the source prior and the teacher prior. These weights are determined by computing entropy over multiple augmentations of the two models. To evaluate the effectiveness of these weights, this approach is compared to three simple weighting configurations: using only the source model, using only the teacher model, and using a simple average with equal weights for both models. Figure 5 is a mixed prior with different weights provided by the present invention when the severity is 5, such as Figure 5 The results shown reveal that relying solely on the source model or the teacher model (i.e., weighted as (1, 0) and (0, 1)) leads to suboptimal performance. In addition, a simple weighting of both models with equal contribution (i.e., (0.5, 0.5)) performs poorly in the model adaptation task facing continuous domain changes in an open test setting due to the inherent uncertainty of the two models. From the above three comparisons, it can be seen that the adaptive weights of the Gaussian mixture in the open test setting prove their effectiveness. This highlights the importance of striking a balance between the two prior models in an unsupervised setting.

[0075] The comparison of the above experimental results shows that: the present invention proposes a variational model adaptation method for continuous domain changes in an open test environment, and the proposed method can alleviate the problem of prior drift.

[0076] Figure 6 60 is a schematic diagram of the structure of an automatic driving model updating device based on a continuous test time domain adaptation CTTA framework provided by the present invention. The automatic driving model updating device based on a continuous test time domain adaptation CTTA framework provided by the present invention includes a first building module 601, a second building module 602 and a first updating module 603: A first building module 601 is used to build an autonomous driving model based on a Bayesian neural network structure and autonomous driving source data; A second construction module 602 is used to construct a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The first updating module 603 is used to update the student model based on variational inference and to update the teacher model based on an exponential moving average method.

[0077] Specifically, the above-mentioned autonomous driving model updating device based on the continuous testing time domain adaptive CTTA framework provided by the present invention can implement all the method steps implemented by the above-mentioned autonomous driving model updating method embodiment based on the continuous testing time domain adaptive CTTA framework, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as the method embodiment will not be described in detail here.

[0078] Figure 7is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communications interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the automatic driving model update method based on the continuous test time domain adaptation CTTA framework, and the method includes: Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data; Constructing a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method.

[0079] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0080] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can execute the automatic driving model update method based on the continuous test time domain adaptive CTTA framework provided by the above methods, the method comprising: Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data; Constructing a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method.

[0081] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the automatic driving model update method based on the continuous test time domain adaptive CTTA framework provided by the above methods, the method comprising: Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data; Constructing a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method.

[0082] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0083] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for updating an autonomous driving model based on a continuous testing temporal adaptation (CTTA) framework, characterized in that: include: Build an autonomous driving model based on the Bayesian neural network structure and autonomous driving source data; Constructing a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The student model is updated based on variational inference, and the teacher model is updated based on an exponential moving average method.

2. The method for updating the autonomous driving model based on the CTTA framework according to claim 1, characterized in that: The method of building an autonomous driving model based on the Bayesian model structure and autonomous driving source data includes: Determine prior knowledge based on the Bayesian neural network structure; Determining a source prior distribution of the autonomous driving model parameters based on the prior knowledge and the autonomous driving source data; The autonomous driving model is determined based on the source prior distribution.

3. The automatic driving model updating method based on the CTTA framework according to claim 1 is characterized in that: The updating of the student model based on variational inference comprises: Determine a teacher model prior and an autonomous driving model prior based on the teacher model and the autonomous driving model, respectively; Determining a hybrid prior based on the teacher model prior and the autonomous driving model prior; The student model is updated based on the mixed prior.

4. The automatic driving model updating method based on the CTTA framework according to claim 3 is characterized in that: The updating of the student model based on the mixed prior comprises: A KL divergence based on a symmetric cross entropy loss and a hybrid prior between the student model and the teacher model; The student model is updated based on the symmetric cross entropy loss and the KL divergence.

5. The method for updating the autonomous driving model based on the CTTA framework according to claim 3, characterized in that: The method further comprises: An autonomous driving prediction result is determined based on the hybrid prior.

6. The method for updating the automatic driving model based on the CTTA framework according to claim 1, characterized in that: The updating of the teacher model based on the exponential moving average method comprises: The teacher model is updated based on the mean and standard deviation of the student model and the mean and standard deviation of the teacher model.

7. An automatic driving model updating device based on a continuous testing time domain adaptation CTTA framework, characterized in that: include: A first building module is used to build an autonomous driving model based on a Bayesian neural network structure and autonomous driving source data; A second building module is used to build a variational average teacher model based on the autonomous driving model, wherein the variational average teacher model includes a student model and a teacher model; The first updating module is used to update the student model based on variational inference and to update the teacher model based on an exponential moving average method.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the automatic driving model updating method based on the CTTA framework as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the automatic driving model updating method based on the CTTA framework as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the automatic driving model updating method based on the CTTA framework as described in any one of claims 1 to 6 is implemented.

Images