Cross-weather unmanned target robust detection training method and system, and storage medium

Abstract

This invention discloses a robust target detection training method, system, and storage medium for unmanned vehicles operating across different weather conditions. The method includes: S1, acquiring multiple weather image sample groups of the same scene, inputting them into the encoding network model and mapping them as approximate posterior distributions of latent variables, and obtaining a detection loss based on the difference between the detection results and the ground truth labels; generating a conditional information bottleneck loss based on the difference between the approximate posterior distribution and the joint distribution of the priors of each anchor point; S2, applying independence constraints to weather variables and latent variables through HSIC to generate an independence decoupling loss; S3, applying alignment constraints to the approximate posterior distributions of different image samples through symmetric KL divergence to generate a representation-level invariance loss; S4, using the four losses as a joint optimization objective, and collaboratively updating the parameters of the encoding network model and the detection model through backpropagation. This invention adopts a progressive approach to construct a robust cross-weather detection framework around the goal of fully representing causality.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent transportation technology, and in particular to a robust target detection training method, system and storage medium for unmanned vehicles operating across weather conditions. Background Technology

[0002] The safety of autonomous driving relies heavily on the accurate detection and recognition of road targets by onboard vision systems. Deep learning-based visual target detection methods have achieved good performance under clear weather and sufficient lighting conditions. However, under conditions such as rain, fog, snow, and nighttime, image degradation occurs, including decreased contrast, blurred textures, atmospheric particle occlusion, dynamic rain streaks, and increased sensor noise, resulting in a significant decrease in detection performance.

[0003] like Figure 1 As shown, from the perspective of causal structure, the scene variable S of the real road and the weather variable W jointly determine the observed image X, while the target label Y is determined only by the scene variable S and has no direct causal dependence on the weather condition W. It is evident that the weather variable W is a confounding variable; any feature sensitive to weather variables will directly lead to performance degradation when the weather distribution shifts during the testing phase. Therefore, feature representations for cross-weather robustness should satisfy the following requirements: retaining all scene structure information needed to predict the target label, and leaving no residual information that can be used to distinguish weather types.

[0004] To address the above issues, existing methods can be mainly divided into three categories: 1. Image Domain Restoration Method: Image restoration networks such as deraining, dehazing, and low-light enhancement are cascaded before the detection model. Typically, image reconstruction quality is used as the optimization metric. Degraded images are restored pixel-by-pixel before being fed into the downstream detection model to reduce the impact of input degradation on detection performance.

[0005] 2. Data augmentation and adversarial perturbation methods: Based on atmospheric scattering models or generative adversarial networks, samples such as rain, fog, snow, and nighttime are synthesized to expand the coverage of training data, so that the detection model can be exposed to image distributions under various weather conditions at the input level; some also further introduce consistency regularization, requiring similar predictions for the original samples and augmented samples.

[0006] 3. Feature Layer Information Bottleneck Method: This method introduces an information bottleneck mechanism at the feature encoding stage, using the Kullback-Leibler divergence (relative entropy) between latent variables and a preset prior (usually a standard normal distribution) to constrain information capacity, thus implicitly suppressing degraded information at the feature layer. Some methods further use clear-sky prediction as a fixed reference to impose consistency constraints on degraded samples. This improves cross-weather prediction consistency at the output layer.

[0007] The above method still has shortcomings in the following three aspects: 1. Task discrimination information and degradation information are compressed equally: Image domain restoration methods tend to suppress high-frequency components, while the discrimination of distant small targets and low-contrast targets relies precisely on subtle high-frequency structures and edge textures. The unconditional KL constraint used by the feature layer information bottleneck method compresses both "target label-related" and "target label-independent" information indiscriminately. Because it compresses the entire dependence of latent variables on the input image, rather than the "residual redundancy under known target label conditions," it may inadvertently damage key structures such as the edge morphology of distant small targets and the internal texture of low-contrast targets, resulting in a decrease in detection recall.

[0008] 2. Lack of active independence constraints on weather variables at the feature layer: Latent variables still implicitly encode weather information. Data augmentation and adversarial perturbation methods only operate on the input layer, without imposing any constraints on the information representation in the feature space. While feature layer information bottleneck methods passively suppress some redundant information through information capacity limitations, none of them impose explicit constraints on the statistical dependence between latent variables and weather variables at the feature layer. This is because compressing the information between latent variables and the input image merely passively expects weather information to be discarded, rather than actively ensuring the independence of latent variables from weather variables. This results in latent variables potentially still encoding weather distribution-related information in multiple dimensions. When the weather distribution changes during the testing phase, the feature distributions in these dimensions drift, leading to unstable detection results.

[0009] 3. Gaps exist in the modeling of representation-level intervention invariance: Existing feature-layer information bottleneck methods impose cross-weather consistency constraints on the output layer of the detection head, but whether intermediate feature representations are truly invariant to weather changes is not explicitly guaranteed. Furthermore, this constraint typically uses a one-way gradient truncation method, taking clear weather predictions as a fixed reference, lacking causal symmetry. This is because output layer consistency approximates the result of representation invariance, rather than directly guaranteeing "latent variables remain unchanged under different weather conditions in the same scenario" at the representation level. This leads to potential drift in latent variable representations when different weather conditions are applied to the same road segment, resulting in systematic shifts in detection outputs in scenarios such as cross-vehicle equipment deployment and cross-regional migration. Summary of the Invention

[0010] To address the limitations of existing methods in three aspects—information compression discrepancies, explicit decoupling of weather interference variables, and representation-level intervention invariance—this invention proposes a robust target detection training method, system, and storage medium for unmanned vehicles operating across weather conditions, aiming to solve the following core technical problems: 1. Task-related information preservation problem: How to compress redundant information between latent variables and input samples during the feature encoding stage while avoiding indiscriminate compression of discriminative information related to the target detection task, so that the encoding network can prioritize the compression of the remaining redundant information (mainly composed of weather-related degradation patterns and non-discriminative background textures) under the "given target label" condition, thereby preserving key discriminative structures such as the edge morphology of distant small targets, the internal texture of low-contrast targets at night, and the subtle spatial relationship between the target and the surrounding road environment while suppressing degradation.

[0011] 2. Explicit Decoupling of Weather Disturbance Variables in Feature Layer: How to impose explicit constraints on the statistical dependence between latent variables and weather variables during the model feature encoding process, so that the encoded latent variables are statistically independent of weather conditions, thereby avoiding the appearance of weather-sensitive feature dimensions in the latent variable space. This addresses the problem that existing methods cannot eliminate weather information at the feature layer by simply perturbing the input space or constraining the output layer consistency, and achieves decoupling of the feature layer from weather disturbance variables.

[0012] 3. The problem of cross-weather representation invariance under intervention conditions: How to establish a constraint mechanism based on the intervention perspective in causal inference to ensure that the latent variable representation remains unchanged under operational intervention conditions (i.e., when different weather conditions are applied to the same road scene), so that the model can guarantee a consistent response to the same road scene under different weather conditions from the representation level rather than just the output level. This solves the problem that existing methods limit consistency constraints to the output layer and use unidirectional gradient truncation references without causal backing, thereby achieving a systematic improvement from detection accuracy under a single weather condition to detection consistency and repeatability across weather, day and night, regions, and sensor scenarios.

[0013] Firstly, the robust target detection training method for unmanned vehicles across weather conditions proposed in this invention includes the following steps: S1. Obtain a group of multiple weather image samples of the same scene, input them into the encoding network model and map them into an approximate posterior distribution of latent variables, output the latent variables to the detection model to obtain the detection results, and obtain the detection loss based on the difference between the detection results and the true labels; establish the prior distribution of the corresponding local latent variables according to the category labels of each anchor point in the image samples; generate the conditional information bottleneck loss based on the difference between the approximate posterior distribution and the joint distribution of the prior of each anchor point. S2. Using the Hilbert-Schmidt independence criterion, independence constraints are applied to the weather variables and latent variables corresponding to the multiple weather image sample groups of the same scene, generating an independence decoupling loss. S3. By applying alignment constraints to the approximate posterior distributions corresponding to different image samples through symmetric KL divergence, a representation-level invariant loss is generated. S4. Using detection loss, conditional information bottleneck loss, independence decoupling loss, and representation level invariance loss as joint optimization objectives, the parameters of the encoding network model and the detection model are updated collaboratively through backpropagation.

[0014] Furthermore, the same scene multiple weather image sample group includes original road scene image samples, and multiple same scene weather image samples generated by performing operational intervention on weather variables while keeping the road scene structure corresponding to the original road scene image samples unchanged; the sample distribution in the same scene multiple weather image sample group remains consistent under the operational intervention.

[0015] Furthermore, the difference between the approximate posterior distribution and the joint distribution of the priors at each anchor point is measured by KL divergence.

[0016] Furthermore, in S1, the joint distribution of the priors of each anchor point is the product of the prior distributions of the local latent variables of each anchor point; wherein, for anchor points whose category label belongs to the foreground, the prior distribution is the class-conditional prior, and for anchor points whose category label belongs to the background, the prior distribution is the fixed-edge prior.

[0017] Furthermore, in S1, the latent variable is obtained by reparameterizing the approximate posterior distribution, which is modeled as a diagonal covariance Gaussian distribution.

[0018] Furthermore, before applying independence constraints, the latent variables in S2 are projected into low-dimensional latent variables through a projection model, and S4 also includes a collaborative update of the parameters of the projection model.

[0019] Furthermore, the independence constraint is expressed as calculating the empirical estimates of the low-dimensional latent variables and the weather variables, and using the empirical estimates as the independence decoupling loss.

[0020] Furthermore, the weather variable is a hybrid vector containing a type component and an intensity component; the type component is a uniquely thermally encoded vector indicating the weather category, covering discrete weather categories; the intensity component is a continuous real-valued variable used to characterize weather physical parameters; the weather physical parameters include atmospheric scattering coefficient, rain line density, rain line length, rain line transparency, brightness scaling factor, gamma correction exponent, and rain line tilt angle.

[0021] In a second aspect, the present invention also provides a robust target detection training system for unmanned vehicles operating across weather conditions, for performing the method described in the first aspect, comprising: The conditional information bottleneck encoding module is used to acquire multiple weather image sample groups of the same scene, input them into the encoding network model to map them into an approximate posterior distribution of latent variables, and output the latent variables to the detection model to obtain the detection results; establish the prior distribution of the corresponding local latent variables according to the category label of each anchor point in the image sample; and generate the conditional information bottleneck loss based on the difference between the joint distribution of the approximate posterior distribution and the prior distribution of each anchor point. The weather disturbance variable feature layer decoupling module is used to apply independence constraints to the weather variables and latent variables corresponding to the multiple weather image sample groups of the same scene through the Hilbert-Schmidt independence criterion, and generate independence decoupling loss. The representation-level invariance constraint module is used to apply alignment constraints to the approximate posterior distributions corresponding to different image samples through symmetric KL divergence, thereby generating representation-level invariant loss. The multi-objective joint optimization module is used to jointly update the parameters of the encoding network model and the detection model through backpropagation, using detection loss, conditional information bottleneck loss, independence decoupling loss, and representation-level invariance loss as joint optimization objectives; wherein, the detection loss is calculated based on the difference between the detection result and the true annotation.

[0022] Thirdly, the present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described weather-resistant unmanned target robust detection training method.

[0023] The present invention achieves the following beneficial effects compared to the prior art: 1. Differentiated Preservation of Task-Specific Information. This invention uses category labels as conditions to model priors down to the anchor point granularity. Foreground anchors employ class-conditional priors, while background anchors employ fixed-edge priors. This ensures that redundancy compression only applies to task-independent dimensions. While suppressing weather degradation, it preserves key discriminative structures such as small target edges and low-contrast textures, achieving sufficient information sufficiency of latent variables for the detection task and avoiding the misjudgment of discriminative information by existing methods.

[0024] 2. Mechanism-level decoupling of weather disturbance variables. Based on information sufficiency, this invention further imposes explicit independence constraints on latent variables and weather variables at the feature layer, eliminating weather-sensitive dimensions in latent variables, eliminating arbitrary-order nonlinear dependencies, and achieving statistical independence between latent variables and weather variables. This is equally effective for complex operating conditions and does not introduce inference overhead.

[0025] 3. Causal mechanism endorsement for cross-weather consistency. Based on information sufficiency and statistical independence, this invention synthesizes paired samples through operational intervention and applies symmetric KL divergence alignment to the representation layer, thereby elevating cross-weather consistency from output equivalence to representation equivalence. This achieves the representation invariance of latent variables under intervention conditions, overcomes the problem of causal asymmetry in existing methods, and does not require reliance on real cross-weather paired data. Attached Figure Description

[0026] Figure 1 This is a causal diagram showing the effects of weather and scene variables on observed images and target labels. Figure 2 This is a block diagram of the robust target detection and training system for unmanned vehicles operating across weather conditions proposed in this invention. Figure 3 This is a flowchart of S1 in the training method of the present invention; Figure 4 This is a flowchart of S2 in the training method of the present invention; Figure 5 This is a flowchart of step S3 in the training method of the present invention; Figure 6 This is a flowchart of S3 in the training method of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] See Figure 2 The present invention proposes a robust target detection training method for unmanned vehicles operating across weather conditions, comprising the following steps: S1, such as Figure 3 As shown, multiple weather image samples of the same scene are obtained as input samples and input into the encoding network model to be mapped to the approximate posterior distribution of latent variables. The latent variables are then output to the detection model to obtain the detection results. The detection loss is obtained based on the difference between the detection results and the true labels. The prior distribution of the corresponding local latent variables is established according to the category labels of each anchor point in the image samples. The conditional information bottleneck loss is generated based on the difference between the approximate posterior distribution and the joint distribution of the prior of each anchor point.

[0029] 1.1 Acquisition of multiple weather image sample groups for the same scene Multiple weather image sample groups for the same scene are represented as follows: ,in, These are original road scene image samples. To generate K weather image samples with the same scene by performing an operational intervention (do(W) intervention) on weather variables while keeping the road scene structure corresponding to the original road scene image samples unchanged.

[0030] Specifically, operational intervention refers to maintaining scenario variables. Under the premise of remaining unchanged, the values ​​of weather variables are manually specified. This allows for the pairing of original road scene image samples to generate corresponding observation images.

[0031] In the formula, This represents a scene variable, ensuring consistency across a set of multiple weather image samples for the same scene. Indicates the first An operational intervention specifies a particular weather value instance, the form of which is the same as that of the weather variable. Consistent; Represents the image generation function of the observed image as characterized by a physical imaging model or other generative model; This represents the corresponding observed image; This represents paired weather image samples for each scene, excluding the original road scene image samples. Specifically, Corresponding to the original road scene image sample, i.e. ,in Take values ​​from sunny days or other baseline weather conditions to make the original road scene image samples and A sample of weather images of the same scene Unified expression within the framework; ,correspond A sample of weather images of the same scene obtained from the intervention.

[0032] In specific embodiments, operational interventions include, but are not limited to, the following implementation methods: fog synthesis based on an atmospheric scattering model, generating fog images of different concentrations by adjusting the atmospheric scattering coefficient and scene depth parameters; rain synthesis based on a rain pattern rendering model, generating rain images of different intensities by adjusting parameters such as rain line tilt angle, density, length, and transparency; night / low-light synthesis based on illumination adjustment, generating night or low-light images by brightness scaling, gamma correction, and superposition of sensor noise terms; cross-domain weather migration based on generative adversarial networks; conditional weather image generation based on a diffusion model; or directly using real-world cross-weather paired datasets. The above implementation methods can be arbitrarily superimposed (e.g., concurrent rain and fog, nighttime rain) to cover complex conditions. The specific implementation method of operational intervention does not constitute a limitation of the present invention; any method that can maintain scene variables... Unchanged, only weather variables are intervened All data acquisition methods can be used as inputs to the encoding network model.

[0033] Compared to existing technologies that simply sample the observed distribution, the method used in this invention to generate multiple weather map samples of the same scene through operational intervention cuts off scene variables. Weather variables The natural generation mechanism can better represent causal relationships.

[0034] 1.2 Approximate posterior distribution of latent variables Multiple weather image samples from the same scene are feature-mapped using the same encoding network model with identical parameters. After linear mapping, the encoding network model outputs at each anchor point of the image sample. A dimensional vector, where the first dimensional as the mean vector of the latent variable distribution ,back After activation by the softplus function, the vector is used as the standard deviation vector of the local latent variable distribution at the anchor point. This ensures that the value is positive. From this, the distribution parameters of the latent variables are obtained:

[0035] In the formula, Indicates the first image in a group of multiple weather image samples from the same scene. One sample; This indicates the number of weather image samples in each scene, excluding the original road scene image samples, in the current training round; Represents the encoding network model; Represents the trainable parameters of the encoding network model; These represent multiple weather image samples from the same scene. The mean vector and standard deviation vector corresponding to the latent variable distribution; This represents the dimension of the hidden variable space.

[0036] Among them, the coding network model Convolutional or self-attention structures based on backbone networks such as ResNet (residual network) and Swin Transformer (a hierarchical visual transformation network based on shift windows) can be used. The purpose of using the same encoding network model with the same parameters is to ensure that input samples under different weather conditions are encoded under the same representation rules, so that subsequent conditional information bottleneck constraints, independence decoupling constraints, and representation-level invariance constraints are comparable in the same representation space.

[0037] After obtaining the distribution parameters of the latent variables, the encoding network model maps the input samples to an approximate posterior distribution of the latent variables, which is expressed as a conditional probability distribution in the form of a diagonal covariance Gaussian:

[0038] In the formula, Indicates sample Corresponding latent variables after encoding The approximate posterior distribution; Indicates a multivariate Gaussian distribution; This represents a diagonal matrix formed by using vector elements as diagonal elements; This represents the element-wise square of the corresponding standard deviation vector, which is the variance vector of the latent variable distribution.

[0039] The above probability distribution modeling replaces traditional deterministic features with random latent variable representations, so that the coding network model not only outputs the feature center location, but also the uncertainty range of the feature, providing a mathematical basis for subsequent processing.

[0040] It should be noted that in the coding network model Will Sample The mapping yields the corresponding approximate posterior distribution. During the process, The distribution of each sample remains consistent under operational intervention, that is, for any have:

[0041] This consistency constraint limits the equivalence at the level of approximate posterior distributions, rather than just the detection consistency at the output level.

[0042] 1.3 Reparameterized Sampling To enable the random sampling process to participate in the backpropagation training from input to output, the approximate posterior distribution is reparameterized to obtain latent variables, which are then input into the detection model. The reparameterization formula is as follows:

[0043] In the formula, Indicates sample Latent variables; This represents a random noise vector sampled from an approximate posterior distribution, along with the parameters of the encoding network model. Irrelevant; This indicates that the mean is zero and the covariance matrix is... 3D unit array The standard multivariate Gaussian distribution; This indicates element-wise multiplication.

[0044] After reparameterization sampling, randomness is transferred to the parameters of the encoding network model. irrelevant auxiliary variables In this way, the mean of the latent variable distribution is thus reduced. with standard deviation The gradient of the total loss function can be backpropagated normally, enabling training from the input to the output. In practical applications, to avoid the influence of randomness on the detection results, the mean value can be directly taken. As a deterministic latent variable, it is input into the subsequent detection model to complete the detection.

[0045] 1.4 Prior distribution of each anchor point Object detection in autonomous driving scenarios is a multi-object dense prediction task, where a single image contains multiple foreground objects and background regions simultaneously. Specifically, a single input image contains a total of [number missing] anchor points. The anchor points are categorized into foreground anchor points and background anchor points based on their location. The target label (i.e., the ground truth label for the detection task) Y of the image is composed of the category labels of all anchor points. For each anchor point in the image Let its category label be ,in Indicates anchor point Located in the background area, it does not correspond to any specific target. Indicates anchor point Located in the foreground area, belonging to the first Class of targets (such as vehicles, pedestrians, etc.) This represents the total number of category labels for the foreground. Correspondingly, this is a latent variable of the image. It consists of the local latent variables of all anchor points, i.e. ,in Indicates anchor point Local latent variables. Based on whether the anchor point is foreground, their corresponding local latent variables are... Define a prior distribution.

[0046] (1) For the foreground anchor point, establish class conditional priors for learnable local latent variables.

[0047] Specifically, for anchors with category labels as the foreground, in their local latent variables... The a priori conditions of the defined class are:

[0048] In the formula, Indicates category label The corresponding learnable mean vector; Indicates category label The corresponding learnable variance vector; This represents a diagonal matrix with its input vector as its diagonal elements; the class-conditional prior model is a Gaussian distribution. Parameter set. During training, the parameters of the encoded network model are compared. Joint optimization, class-conditional priors are used for each class label in the latent variable space. Define independent Gaussian distributions, where the mean vector is... The variance vector determines the central location of the local latent variables in the feature space that determine the foreground anchor point of this category. This controls the allowed distribution range of the category label. Since the class conditional distribution is established at the anchor point granularity, different category labels in the feature space have their own independent distribution range, which is conducive to forming a representation structure of similar clusters and dissimilar separation at the feature layer.

[0049] (2) For background anchor points, establish fixed edge priors that are independent of category labels.

[0050] For anchor points with category labels as background, establish a fixed marginal prior that is category-independent and does not participate in training. :

[0051] In the formula, Represents a fixed marginal prior that is independent of category label; It is a vector with zero mean; is the unit covariance matrix of the latent variable space dimension.

[0052] Under the above correspondence, the joint distribution of the priors of each anchor point is the product of the prior distributions of the local latent variables of each anchor point:

[0053] That is, the foreground anchor point is based on its class condition prior, and the background anchor point is based on the fixed edge prior. .

[0054] 1.5. Conditional Information Bottleneck Loss Based on the difference between the joint distribution of the approximate posterior distribution and the prior distribution of each anchor point, a conditional information bottleneck loss is generated. This loss is a category-conditional feature compression regularization term used to limit the redundant dependence of latent variables on the input image at the feature encoding layer.

[0055] In the formula, This indicates a bottleneck loss in conditional information; Indicates KL divergence; This represents the approximate posterior distribution of the mapping of the encoding network model; This represents the joint distribution of the priors of the above anchor points (the foreground anchor points are conditional Gaussian priors, and the background anchor points are fixed edge priors). This represents the expectation operator.

[0056] Since both the approximate posterior distribution and the prior distribution are independently decomposed at anchor points, the above KL divergence term is strictly decomposed according to anchor point type (foreground / background) as follows:

[0057] In the formula, the first summation term is the foreground condition compression term, which differentially compresses redundant information under the condition that the foreground anchor point category label is known. That is, it only compresses redundant components such as weather degradation that are not related to the category, while retaining useful discriminative structures such as the boundary and texture of the target in that category. The second summation term is the background edge compression term. Since the background anchor point does not correspond to any specific target, it directly uses the fixed edge prior. As a reference, we actively suppress weather degradation information that may be carried by the background area to prevent it from remaining in the latent variables.

[0058] Since the number of background anchor points is usually much larger than that of foreground anchor points, if the two are summed with equal weights, the background term will numerically overwhelm the foreground term, weakening the preservation of foreground discrimination information. Therefore, the background edge compression term is multiplied by a weighting factor. To balance the imbalance in the number of foreground and background anchor points, the actual optimization bottleneck loss is:

[0059] when When the background edge compression term weight is exactly the same as the foreground conditional compression term, when... At that time, the background edge compression item is reduced in weight accordingly. This means that the average of the KL divergence over all image samples is used as the final conditional information bottleneck loss.

[0060] It should be noted that this conditional information bottleneck loss only compresses redundant information and does not include task supervision. Specifically, the conditional information bottleneck loss constrains conditional mutual information. That is, given the category label of each anchor point, compress redundant information unrelated to the category in the latent variables; while ensuring detection accuracy is the task of detection loss. bear.

[0061] From an information theory perspective, this module, given the anchor point category labels, compresses redundant information in the latent variables that is irrelevant to detection. This prompts the model to prioritize the retention of discriminative information that is relevant to the detection task (such as the boundary shape of the target, internal texture, and spatial relationship between the target and the road environment) within a limited information capacity. Information lacking detection capability for the category labels (mainly composed of weather degradation patterns such as rain streaks, haze scattering, nighttime noise, and dynamic rain streak bands, as well as non-discriminative background textures) is compressed first. Specifically, the foreground conditional prior modeling down to the anchor point granularity avoids the indiscriminate loss of key discriminative structures such as distant small targets and low-contrast targets due to the unconditional information bottleneck at the whole-image semantic granularity. The fixed background edge prior ensures that the weather degradation redundancy parasitized in the background region is explicitly compressed, rather than remaining in the background latent variables, thereby suppressing background false positive drift when weather distribution shifts.

[0062] S2, such as Figure 4As shown, the Hilbert-Schmidt independence criterion (HSIC) is used to impose independence constraints on the statistical dependencies between weather variables and latent variables corresponding to multiple weather image sample groups in the same scene, so as to generate independence decoupling loss.

[0063] 2.1 Dimensionality Reduction of Latent Variables Due to the latent variables output by the encoding network model Typically has high dimensionality (such as) Direct computation of independence measures using kernel methods can lead to a decrease in sensitivity due to the curse of dimensionality. Therefore, before imposing independence constraints, latent variables are evaluated through a projection model. The projection is a low-dimensional latent variable.

[0064] Specifically, the projection model It consists of two multilayer perceptrons connected after the encoding network model. This projection model The dimensionality reduction representation of the latent variables is as follows:

[0065] In the formula, This represents the low-dimensional latent variables after projection; Represents the trainable parameters of the projection model; Represents the projection dimension, satisfying Its value can be adjusted according to training needs; a typical value is... .

[0066] 2.2 Coding of Weather Variables Weather variables are mixed vectors containing type and intensity components, represented as:

[0067] In the formula, type classification A uniquely thermally encoded vector indicating weather categories, covering discrete weather categories such as fog / rain / night / combined weather; intensity component These are continuous real-valued variables used to characterize the values ​​of weather physical parameters. Weather physical parameters include, but are not limited to, atmospheric scattering coefficients. (Characterizing the intensity of light scattering by atmospheric particles), rain line density (Characterizing the coverage of rain streaks per unit area), rain line length (Characterizing the length of a single rain streak in pixel space), rain streak transparency (Characterizing the visual penetration of rain streaks relative to the background image), brightness scaling factor (Characterizing the ratio of total light intensity to standard daytime light intensity), gamma correction index (Characterizing the intensity of nonlinear compression of pixel dynamic range) and rain line tilt angle (Characterizing the angle of deviation of the rain line from the vertical direction under the action of wind) etc.

[0068] The hybrid vectorization representation of weather variables can support the kernel function in processing discrete and continuous components, providing a unified vector input for the construction of the subsequent kernel matrix. This ensures that the weather coding of different weather variants of the same original road scene image sample can accurately reflect the differences in weather type and intensity, providing a reliable weather variable input for HSIC measurement.

[0069] 2.3 Independence Decoupling Loss The Hilbert-Schmidt Independence Criterion (HSIC) used in this step is a nonparametric independence measure based on the regenerative kernel Hilbert space. The independence constraint is expressed as calculating the empirical estimates of low-dimensional latent variables and weather variables through HSIC, and using the empirical estimates as the independence decoupling loss.

[0070] Specifically, for a single scene's original road scene image sample and its corresponding K weather image samples of the same scene, the set of projected latent variables for all samples is taken. Corresponding weather variables The collection of value instances The empirical estimate is calculated as follows:

[0071] In the formula, This indicates the number of weather image samples in a given scene, excluding the original road scene image samples. This indicates the total number of samples included in a group of multiple weather image samples for the same scene; This represents the kernel matrix corresponding to the low-dimensional latent variables, and its elements... ; This represents the kernel matrix corresponding to the weather variables, and its elements... A sample index for multiple weather image sample groups in the same scene; , These represent kernel functions on the low-dimensional latent variable space and the weather variable space, respectively; Denotes a centered matrix, where for 3D identity matrix It is a column vector consisting entirely of 1s; Represents the trace of a matrix.

[0072] In one embodiment, the kernel function may use an RBF kernel:

[0073] In the formula, This represents the sample vector to be compared; Denotes the Euclidean norm; The kernel bandwidth controls the rate at which sample vector similarity decays. It can be solved adaptively using the median heuristic, which involves taking the distance between all sample pairs within a group of multiple weather image samples from the same scene in each calculation. The median as .

[0074] When the kernel function uses a characteristic kernel such as the RBF kernel, an empirical estimate of zero implies a low-dimensional latent variable. Weather variables They are statistically completely independent. Based on this property, the empirical estimate can be directly used as the independence decoupling loss:

[0075] HSIC measures the statistical dependence between latent variables and weather variables. When using a feature kernel, HSIC has the following properties: the empirical estimate is zero only when the two variables are completely independent; if any dependence exists (whether linear or nonlinear), the empirical estimate is greater than zero. Therefore, minimizing the independence decoupling loss is equivalent to actively eliminating any nonlinear dependence of latent variables on weather variables.

[0076] In S1, weather information is passively expected to be discarded by limiting the total amount of information. However, this compression applies to the total amount of information rather than specifically targeting weather information, so weather information may still remain in several dimensions. This step further constrains the dependency relationship between latent variables and weather variables, actively suppressing residual information. This ensures that the latent variable space no longer contains feature dimensions sensitive to weather distribution, thus preventing the phenomenon of latent variable distribution drifting with weather distribution shifts when the detection model is deployed across weather, day / night, and regions.

[0077] S3, such as Figure 5 As shown, by using symmetric KL divergence, alignment constraints are applied to the approximate posterior distributions corresponding to different image samples in a multi-weather image sample group of the same scene, generating a representation-level invariant loss.

[0078] Specifically, the representation-level invariance loss is calculated as follows:

[0079] In the formula, This represents the loss of characterization level invariance; This indicates the number of samples contained in a group of multiple weather image samples of the same scene; Indicates from The number of combinations of any two samples selected from the sample is used to calculate the arithmetic mean of the summation term; It means that for all satisfying Summing unordered index pairs; Representation of encoding network model Acting on the sample The approximate posterior distribution obtained.

[0080] Since the approximate posterior distribution is configured in this invention as a diagonal covariance Gaussian form, the aforementioned symmetric KL divergence has a direct calculation formula and can be directly obtained from the mean output by the encoding network model. With variance Calculations show that Monte Carlo sampling is not required.

[0081] In real-world autonomous driving scenarios across different weather conditions, even though S1 and S2 achieve differentiated compression of redundant information and decoupling of the independence of the marginal distribution of weather variables, these two constraints only apply to the marginal distribution level. This indicates that even if the latent variable Z as a whole is independent of the weather variable W, the specific value of Z may still drift when different weather conditions are applied to the same scenario.

[0082] This step establishes intervention invariance by imposing representation-level alignment constraints on latent variables, ensuring that the latent variable representations remain consistent across different weather conditions in the same scenario. This reduces the interference of weather distribution changes on the detection results. Since latent variables are the carriers of sufficient causal representation, this mechanism can suppress representation drift caused by differences in weather data collection equipment across the same scenario, improve detection stability when equipment is deployed across fleets or migrates across regions, and enhance the detection model's adaptability to weather distribution shifts.

[0083] Unlike existing methods that impose consistency constraints on the detection model output, this step uses symmetric KL divergence instead of unidirectional KL divergence. It characterizes the bidirectional equivalence between any two sample distributions, rather than the unidirectional proximity, thus explaining the rationality of "the same scene should produce the same detection output under different weather conditions" at the causal level. Symmetric KL divergence penalizes the distribution differences in both the "information gain" and "information loss" directions simultaneously, rather than using the unidirectional detection (such as using sunny weather prediction as a fixed reference in existing unidirectional gradient truncation methods) as a fixed reference, thereby avoiding the loss of causal symmetry.

[0084] S4, such as Figure 6 As shown, detection loss, conditional information bottleneck loss, independence decoupling loss, and representation level invariance loss are used as joint optimization objectives, and the parameters of the encoding network model, detection model, and projection model are updated collaboratively through backpropagation.

[0085] Based on the preceding steps, the total loss function is defined as follows:

[0086] Among them, detection loss It consists of classification loss and bounding box regression loss:

[0087] In the formula, This represents the overall optimization objective of the model; This represents the detection loss, including the classification loss. (e.g., Focal Loss or Standard Cross Entropy) and Bounding Box Regression Loss (e.g., Smooth L1 or GIoU); This indicates a bottleneck loss in conditional information; This represents the loss due to independence decoupling. This represents the loss of characterization level invariance; These represent the weight coefficients of the conditional information bottleneck loss, independence decoupling loss, and representation-level invariance loss relative to the detection loss, respectively, used to balance the contributions of different optimization objectives during the training process.

[0088] The four types of loss terms mentioned above apply to different levels of the model and form a complementary relationship during the optimization process. Among them, the detection loss... This is used to provide supervisory signals from manually labeled data, enabling the entire detection system to have accurate target detection capabilities. This corresponds to maximizing the mutual information between latent variables and class labels within the information bottleneck; conditional information bottleneck loss. It only undertakes feature compression based on category, suppressing the overexpression of non-discriminatory redundant information in the encoding network model, corresponding to the goal of limiting mutual information in the information bottleneck; independence decoupling loss Used to explicitly eliminate the arbitrary nonlinear dependence of latent variables on weather variables from the characteristic marginal distribution layer; representational level invariance loss. This is used to constrain the coding results of the same scene under different weather conditions from the point-to-point layer of the representation to maintain consistency.

[0089] These four constraints correspond to "detection accuracy constraint", "differential compression constraint", "edge independence constraint" and "point-to-point invariance constraint", respectively. They form a collaborative optimization during the training process, enabling the entire training system to simultaneously meet multiple performance requirements.

[0090] During backpropagation, the gradients of the above loss terms are propagated back through different paths and converge at the encoding network model. The encoding network model simultaneously receives gradient signals from the four types of losses, forming a dynamic balance between information compression, edge independence, point-to-point invariance, and task discrimination, thereby achieving unified parameter updates.

[0091] Specifically, The gradient is first detected by the model. The data is then fed back and further passed to the coding network model. ; The gradient directly affects the parameters of the encoding network model. Used to constrain the distribution parameters of latent variables and the class condition prior parameters of the anchor point ; The gradient is obtained through the projection model. Feedback to the encoding network model parameters ; The gradient is encoded by the network model. It applies to the approximate posterior distribution parameters corresponding to all paired samples.

[0092] The model parameter update process can be represented as:

[0093] In the formula, Indicates the parameters of the encoding network model; Indicates the detection model parameters; Indicates the parameters of the projection model; The learnable parameters representing the prior conditions of the anchor point class; Indicates the learning rate; This represents the gradient of the total loss with respect to each parameter.

[0094] Through the above joint optimization process, the detection loss ensures that the training system has accurate target detection capabilities; the conditional information bottleneck constraint enables differentiated preservation of task discrimination information; the independence decoupling loss can eliminate the marginal dependence of latent variables on weather variables; the representation level invariance loss can ensure the consistency of point-to-point representation across weather conditions in the same scene; through the synergistic effect of the four, this step enables the training model to gradually converge to the goal of "causally sufficient representation of task labels and statistically independent of weather variables".

[0095] Compared with traditional single-loss optimization methods, this step improves the overall robustness by using multi-objective joint optimization, which enables the training system to meet constraints at multiple levels simultaneously.

[0096] This invention also provides a robust target detection training system for unmanned vehicles operating across weather conditions, used to implement the above-mentioned training method. The training system includes a conditional information bottleneck encoding module, a weather interference variable feature layer decoupling module, a representation-level invariance constraint module, and a multi-objective joint optimization module.

[0097] The conditional information bottleneck encoding module maps the input image into a latent variable representation with limited information capacity and conditional on the class label. During the encoding process, a conditional information bottleneck constraint based on the class label is introduced, allowing the model to prioritize the retention of discriminative information related to target detection within its limited information representation capacity, while suppressing redundant weather-related information in a differentiated manner, thus generating the conditional information bottleneck loss. This module, through differential compression of information, suppresses weather degradation information while preserving key discriminative structures such as the edge morphology of distant small targets and the internal texture of low-contrast targets.

[0098] The weather disturbance variable feature layer decoupling module imposes independence constraints on the statistical dependence between latent variable Z and weather variable W at the feature layer, eliminating weather-sensitive feature dimensions in the latent variable space. Specifically, a projection model is cascaded after the output of the encoding network model, and a nonparametric independence measure between the latent variable and the weather variable is calculated based on the Hilbert-Schmidt independence criterion to generate an independence decoupling loss. This module upgrades passive compression to active decoupling, eradicating weather-sensitive dimensions in the latent variable space from the feature representation mechanism level.

[0099] The representation-level invariance constraint module applies KL divergence alignment constraints to the encoding results of the same road scene under different weather conditions at the latent variable level, generating a representation-level invariant loss. This ensures that the model maintains a consistent response to the same scene under different weather conditions from the representation level, rather than just the output level. Specifically, unlike existing methods that apply alignment constraints to the detection model output and use unidirectional gradient truncation references, this module moves the alignment constraint forward to the approximate posterior distribution level and uses symmetric KL divergence as the distribution alignment metric, simultaneously penalizing both information gain and information loss between the two distributions in an information-theoretic sense.

[0100] The multi-objective joint optimization module is used to uniformly model the detection loss, conditional information bottleneck loss, independence decoupling loss, and representation-level invariance loss based on the previously defined loss terms, constructing a total loss function. Through the backpropagation mechanism, gradient signals from different constraints are merged in the encoding network model to achieve collaborative updates of the parameters of each model, enabling the training system to achieve a balance between detection accuracy, information compression capability, feature layer independence, and cross-weather consistency.

[0101] The following uses the BDD100K dataset as an example to illustrate the execution process of the method of this invention.

[0102] The BDD100K dataset is a large-scale, diverse autonomous driving dataset released in 2018 by the AI ​​Lab at UC Berkeley (BAIR). It contains approximately 100,000 real-world driving video frames and corresponding 2D bounding box annotations for various targets such as vehicles, pedestrians, and traffic signs. The scenarios cover a variety of weather and lighting conditions, including sunny, rainy, foggy, and nighttime, providing a realistic and multi-weather-annotated data foundation for the training and validation of this invention. The following description focuses on the changes in data format and the interrelationships between various processing stages; specific parameter values ​​can be flexibly adjusted according to the training set distribution and target operating conditions.

[0103] I. Input Data Video frames depicting sunny days were selected from the BDD100K dataset as training input. First, all images were resampled at a uniform spatial resolution to eliminate resolution differences caused by different acquisition devices and shooting protocols. Then, the images were cropped and normalized according to the requirements of vehicle-mounted visual object detection; for example, this resulted in two-dimensional image samples with a fixed size of 720×1280×3. Each sample was accompanied by corresponding two-dimensional bounding box annotations for multiple object classes as supervision signals.

[0104] II. Acquisition of Multiple Weather Image Samples from the Same Scene For each sunny day sample, while preserving the scene variables Under the premise of remaining unchanged, synthesize a paired sample set according to the operational intervention. .in This is the original sunny day sample, followed by... Each sample corresponds to different weather and working conditions. This embodiment takes... These correspond to four representative operating conditions: light fog, dense fog, moderate rain, and nighttime.

[0105] The light fog and dense fog levels are generated by an atmospheric scattering model, and the scene depth is given by the depth channel of the vehicle dataset or by inference from a monocular depth estimation network. The atmospheric scattering coefficient is... Set according to two typical concentration levels, for example, set to The images correspond to light fog and dense fog, respectively. The transmittance and global atmospheric light intensity are superimposed in the image domain to obtain fog images of the corresponding concentrations.

[0106] Moderate rain is generated by a rain pattern rendering model—within the tilt angle range corresponding to light to moderate winds (e.g., The sampling direction of the rain line is set according to the typical rainfall intensity. Parametric rain texture layers are overlaid on a clear sky image; Nighttime brightness is adjusted based on ambient light levels – overall brightness is based on typical nighttime illuminance ratios. Scaling (e.g.) Gamma-corrected index (like The pixel dynamic range is nonlinearly compressed, and sensor noise that conforms to the true imaging characteristics is superimposed. The above four operating conditions can also be superimposed in pairs (such as nighttime rainy days, rain and fog concurrently) to expand the coverage of composite operating conditions.

[0107] After the above synthesis, Its corresponding Each paired weather sample is completely identical in spatial size and target annotation, differing only in appearance, thus forming a group of multiple weather image samples of the same scene.

[0108] Specifically, each sample At the same time, obtain the corresponding weather variable value examples. ,in The unique heat vector indicating the weather category to which the sample belongs. The actual physical parameters used during sample synthesis ( A vector formed by arranging (etc.) in a predetermined order. As a function for generating observed images The input is used for weather image sample synthesis and also serves as the weather variable input for subsequent HSIC independence measurement.

[0109] III. Constraints in Conditional Information Group multiple weather image samples from the same scene Encoding network model with shared input parameters Feature extraction is performed. Through layer-by-layer convolution and spatial downsampling, the spatial resolution of the feature map decreases with each layer while the channel dimension increases with each layer, thus compressing the image space into high-level semantic features. The backbone output passes through a linear mapping layer to obtain the distribution parameters of the corresponding local latent variables at each anchor point. Then, their respective latent variables are sampled using a reparameterization method. This completes the information compression process from images to latent variables.

[0110] During the compression process, the joint distribution of priors at each anchor point is used. For reference, based on the approximate posterior distribution The KL divergence between the joint distribution and the conditional information bottleneck loss constitutes the conditional information bottleneck loss. .

[0111] IV. Independence Decoupling Loss Latent variables for each sample Projected model Compressed to a low-dimensional space (such as the projected dimension) ), to obtain the projected latent variables Take the previously constructed weather variable value examples. Weather variables are input as HSIC measurements.

[0112] Within the same scene sample group, projective latent variable kernel matrices are constructed based on the RBF kernel. With weather variable kernel matrix After centralization matrix After the action, the empirical estimate of HSIC is obtained, which yields the independence decoupling loss. .

[0113] V. Characterization-level invariance loss Multiple weather image sample groups for the same scene Extract their approximate posterior distributions respectively Since the approximate posterior distribution is modeled as a diagonal covariance Gaussian form, the pairwise symmetric KL divergence has a direct calculation formula and can be directly derived from the latent variable distribution parameters. The calculation yields the representation-level invariant loss by pairwise combining the approximate posterior distributions of all paired samples in the same scene, calculating the symmetric KL curve, and normalizing it. .

[0114] VI. Joint Optimization Detect loss Conditional information bottleneck loss Independence decoupling loss With representational level invariance loss By weighting factor Weighted summation constitutes the total loss function for multi-objective joint optimization. During backpropagation, the gradients of the four types of losses are propagated back to the encoding network model via their respective paths. Projection model With detection model At the encoding network model, all trainable parameters are merged and collaboratively updated, so that the entire training system achieves a balance between detection accuracy, retention of task-related information, decoupling of weather variables, and consistency of cross-weather representation.

[0115] This invention adopts a progressive approach, with causal sufficiency as the theoretical goal, and constructs a robust cross-weather detection framework layer by layer from three dimensions: information sufficiency, statistical independence, and representation invariance. 1. Differentiated preservation of task-specific discriminative information ensures sufficient information sufficiency of latent variables for the detection task. This invention introduces an information bottleneck constraint based on category labels and models class-conditional priors to foreground anchors. Background anchors are unconditionally compressed using fixed edge priors, actively suppressing weather degradation redundancy in the background area. This ensures that the compression of redundant information only applies to dimensions irrelevant to the task, effectively preserving key discriminative structures such as the edge morphology of distant small targets and the internal texture of low-contrast targets at night while suppressing weather degradation information. Compared to image domain restoration methods that sacrifice high-frequency suppression for fidelity reconstruction and the one-size-fits-all compression of task-related and irrelevant dimensions by unconditional information bottlenecks, this mechanism moves the protection of discriminative dimensions from the later stages of the detection model to the earlier stages of the encoding network model, avoiding the inherent tension between "noise compression" and "discriminative information loss." Simultaneously, the class-conditional priors at the anchor granularity allow latent variables of the same type of anchor to naturally cluster and dissimilar types to separate, resulting in more stable training convergence relative to the baseline.

[0116] 2. Mechanism-level decoupling of weather disturbance variables achieves statistical independence between latent variables and weather variables. While preserving the key discriminant structure, this invention eliminates the sensitivity of latent variables to weather distribution at the feature expression mechanism level by imposing explicit independence constraints on latent variables and weather variables at the feature layer. Since the independence metric used is identifiable for nonlinear dependencies of any order, it eliminates not only linear correlations but is also effective for complex conditions such as concurrent rain and fog, and rainy nights, overcoming the shortcomings of existing methods that rely solely on input perturbations or edge compression for explicit decoupling. Furthermore, the independence constraint only takes effect during the training phase, requiring no additional computation in practical applications. It can be seamlessly integrated with mainstream detectors without introducing any additional inference overhead.

[0117] 3. A causal mechanism for cross-scenario consistency ensures the invariance of latent variables under intervention conditions. While preserving key discriminative structures and eliminating the influence of weather variables, this invention elevates cross-weather consistency assurance from output equivalence to representational equivalence by applying symmetric KL alignment at the approximate posterior distribution level. This provides explicit endorsement at the causal level that the same scenario should produce the same detection output under different weather conditions, overcoming the causal asymmetry problem caused by existing methods using sunny weather prediction as a fixed reference. This mechanism significantly reduces the output drift of the perception system in three scenarios: repeated detection on the same road segment, equipment migration across vehicle fleets, and cross-regional deployment. Furthermore, since paired data is synthesized through operational intervention, it does not rely on real cross-weather paired data collection, thus eliminating the bottleneck of cross-domain training data collection and annotation.

[0118] This invention also proposes a storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described cross-weather unmanned target robust detection training method.

[0119] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A robust target detection training method for unmanned vehicles operating across different weather conditions, characterized in that, Includes the following steps: S1. Obtain a group of multiple weather image samples of the same scene, input them into the encoding network model and map them into an approximate posterior distribution of latent variables, output the latent variables to the detection model to obtain the detection results, and obtain the detection loss based on the difference between the detection results and the real labels; establish the prior distribution of the corresponding local latent variables according to the category labels of each anchor point in the image samples, and generate the conditional information bottleneck loss based on the difference between the approximate posterior distribution and the joint distribution of the prior of each anchor point. S2. Using the Hilbert-Schmidt independence criterion, apply independence constraints to the weather variables and latent variables corresponding to the multiple weather image sample groups of the same scene, and generate independence decoupling loss. S3. By applying alignment constraints to the approximate posterior distributions corresponding to different image samples through symmetric KL divergence, a representation-level invariant loss is generated. S4. Using detection loss, conditional information bottleneck loss, independence decoupling loss, and representation level invariance loss as joint optimization objectives, the parameters of the encoding network model and the detection model are updated collaboratively through backpropagation.

2. The cross-weather unmanned target robust detection training method according to claim 1, characterized in that, The same scene multiple weather image sample group includes original road scene image samples, and multiple same scene weather image samples generated by performing operational intervention on weather variables while keeping the road scene structure corresponding to the original road scene image samples unchanged; the sample distribution in the same scene multiple weather image sample group remains consistent under the operational intervention.

3. The cross-weather unmanned target robust detection training method according to claim 1, characterized in that, In S1, the difference between the approximate posterior distribution and the joint distribution of the priors of each anchor point is measured by KL divergence.

4. The cross-weather unmanned target robust detection training method according to claim 1, characterized in that, In S1, the joint distribution of the priors of each anchor point is the product of the prior distributions of the local latent variables of each anchor point; wherein, for anchor points whose category label belongs to the foreground, the prior distribution is the class conditional prior, and for anchor points whose category label belongs to the background, the prior distribution is the fixed edge prior.

5. The cross-weather unmanned target robust detection training method according to claim 1, characterized in that, In S1, the latent variable is obtained by reparameterizing the approximate posterior distribution, which is modeled as a diagonal covariance Gaussian distribution.

6. The cross-weather unmanned target robust detection training method according to claim 1, characterized in that, Before applying independence constraints, the latent variables in S2 are projected into low-dimensional latent variables through a projection model, and S4 also includes a collaborative update of the parameters of the projection model.

7. The cross-weather unmanned target robust detection training method according to claim 6, characterized in that, The independence constraint is expressed as calculating the empirical estimates of the low-dimensional latent variables and the weather variables, and using the empirical estimates as the independence decoupling loss.

8. The cross-weather unmanned target robust detection training method according to claim 7, characterized in that, The weather variable is a hybrid vector containing a type component and an intensity component; the type component is a uniquely thermally encoded vector indicating the weather category, covering discrete weather categories; the intensity component is a continuous real-valued variable used to characterize weather physical parameters; the weather physical parameters include atmospheric scattering coefficient, rain line density, rain line length, rain line transparency, brightness scaling factor, gamma correction exponent, and rain line tilt angle.

9. A robust target detection training system for unmanned vehicles operating across weather conditions, characterized in that, include: The conditional information bottleneck encoding module is used to acquire multiple weather image sample groups of the same scene, input them into the encoding network model to map them into an approximate posterior distribution of latent variables, and output the latent variables to the detection model to obtain the detection results; establish the prior distribution of the corresponding local latent variables according to the category label of each anchor point in the image sample; and generate the conditional information bottleneck loss based on the difference between the joint distribution of the approximate posterior distribution and the prior distribution of each anchor point. The weather disturbance variable feature layer decoupling module is used to apply independence constraints to the weather variables and latent variables corresponding to the multiple weather image sample groups of the same scene through the Hilbert-Schmidt independence criterion, and generate independence decoupling loss. The representation-level invariance constraint module is used to apply alignment constraints to the approximate posterior distributions corresponding to different image samples through symmetric KL divergence, thereby generating representation-level invariant loss. The multi-objective joint optimization module is used to jointly update the parameters of the encoding network model and the detection model through backpropagation, using detection loss, conditional information bottleneck loss, independence decoupling loss, and representation-level invariance loss as joint optimization objectives; wherein, the detection loss is calculated based on the difference between the detection result and the true annotation.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the cross-weather unmanned target robust detection training method according to any one of claims 1-8.

Images