A pose recognition method based on adaptive uncertainty-aware meta-learning

Abstract

The application relates to the technical field of intelligent identification, and discloses a gesture recognition method based on adaptive uncertainty-aware meta learning, which formalizes a gesture small sample regression framework, extracts features by using a linearized neural network and a neural tangent kernel, establishes a Gaussian process regression model by combining Bayesian inference, projects a covariance matrix in a reduced dimension by using a Fisher information matrix in view of high-dimensional characteristics, simultaneously constructs an adaptive weight generator, generates a dynamic threshold value by means of historical loss moving average, training progress and an uncertainty correction term, and assigns specific weights to tasks; a posterior mean value is output as a predicted angle in a meta test stage, and the posterior covariance is used to quantize uncertainty. The application can effectively cope with object symmetry ambiguity and feature loss, focuses the model on difficult tasks through an adaptive mechanism, significantly improves prediction accuracy and reliability in a small sample scene, and reduces computational complexity.

Description

Technical Field

[0001] This invention relates to the field of intelligent recognition technology, and in particular to a posture recognition method based on adaptive uncertainty-aware meta-learning. Background Technology

[0002] With the rapid development of embodied intelligence and computer vision technologies, 3D object pose estimation has become a core task in robot environmental perception. Its prediction accuracy and reliability directly affect downstream applications such as robot grasping, autonomous driving obstacle avoidance, and augmented reality interaction. Monocular vision-based pose regression methods can directly infer 3D geometric attributes from 2D images and have become one of the mainstream technical approaches. By extracting contours, textures, and semantic cues from images, this method can effectively regress object rotation angles, providing reliable geometric priors for path planning and action execution.

[0003] In practical applications and baseline scenarios for few-shot regression, object pose estimation faces challenges of data scarcity and geometric heterogeneity. On the one hand, the real world contains a wide variety of objects, and collecting and labeling full-view pose data for each new object is costly, typically providing only a very small number of reference viewpoint images. On the other hand, different categories of objects have significant differences in geometric topology. For example, "airplanes" typically have streamlined features, while categories such as "cars" and "sofas" often contain complex structures and symmetries, resulting in a highly nonlinear mapping from image features to pose angles. Therefore, 3D object pose regression is essentially a typical few-shot regression problem, urgently requiring a technical framework that can quickly adapt to new objects based on a small number of viewpoint images and effectively resolve geometric complexity.

[0004] 3D target vehicle attitude estimation has become a core task in autonomous driving environmental perception. Its prediction accuracy and reliability directly affect downstream applications of autonomous driving systems, such as trajectory prediction, automatic emergency braking (AEB), and obstacle avoidance decision-making. Visual attitude regression methods based on onboard monocular, binocular, or surround-view cameras can directly infer the 3D geometric attributes (such as yaw angle, pitch angle, and roll angle) of target vehicles from 2D traffic scene images, providing reliable geometric prior data for path planning and action execution.

[0005] In real-world, complex autonomous driving scenarios on open roads, on the one hand, there is a wide variety of vehicles on real roads, including not only common passenger cars but also frequently irregularly shaped engineering vehicles, various modified vehicles, and vehicles that are severely obscured or damaged. Collecting and labeling full-view pose data for each of these rare vehicle types is extremely costly, typically providing only a very small number of reference view images, thus constituting a typical small-sample regression problem for vehicle pose. On the other hand, different vehicle categories exhibit significant differences in geometric topology. Some specialized vehicles often contain complex structures and possess high front-to-back or left-to-right symmetry, which can easily lead to ambiguities in visual features under adverse lighting conditions such as low light, backlight, or halo effects. This results in a highly nonlinear mapping from image features to the vehicle's three-dimensional pose angles.

[0006] Most existing autonomous driving visual perception models are based on static deep learning frameworks, which have poor generalization ability when faced with rare and unseen vehicle types. While introducing Model-Independent Meta-Learning (MAML) allows for rapid fine-tuning of the model on new tasks by learning shared parameters, it still faces significant limitations when applied to complex visual regression tasks: First, it lacks an explicit uncertainty quantification mechanism, and traditional methods cannot effectively address cognitive uncertainties caused by sudden changes in lighting or missing features in vehicle images. Second, the task weight allocation strategy is rigid. In real traffic flow, the difficulty of identifying regular vehicles differs greatly from that of rare and dangerous vehicles. Indiscriminate weight allocation leads to insufficient learning resources for difficult tasks (such as special vehicles with complex geometries), ultimately limiting the accuracy of pose estimation in challenging scenarios. This inaccuracy in pose recognition is directly passed to the downstream behavior prediction module, causing the system to fail to accurately determine the target vehicle's lane-changing or cutting intentions, leading to premature or delayed intervention of the Automatic Emergency Braking (AEB) system, seriously threatening driving safety.

[0007] Meta-learning, as a core learning paradigm in few-shot scenarios, optimizes the learning algorithm itself by extracting prior knowledge of the task distribution, significantly reducing the need for retraining for new object categories. Modern meta-learning methods are mainly divided into three categories: optimization-based, model-based, and metric-based. Among them, MAML, a classic optimization-based method, learns shared parameter initialization points, allowing the model to be quickly fine-tuned with only a small amount of gradient descent on new tasks, demonstrating some effectiveness in pose regression prediction of similar objects. Furthermore, meta-learning can be combined with various methods, such as probabilistic meta-learning frameworks and dynamic weight generation methods.

[0008] However, applying these traditional meta-learning methods to pose prediction tasks in visual regression still faces significant limitations, including a lack of explicit uncertainty quantification and rigid task weight allocation strategies. Existing meta-learning methods assume that all object categories are equally important during the meta-training phase, employing a static equal-weight allocation strategy. However, in real life, the pose regression difficulty of different objects varies significantly. This indiscriminate allocation leads to insufficient learning resources for difficult tasks, failing to fully learn their complex geometric manifolds, ultimately limiting the pose estimation accuracy for highly difficult objects and making it difficult to meet the requirements of high-reliability visual perception. Summary of the Invention

[0009] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a pose recognition method based on adaptive uncertainty-aware meta-learning.

[0010] A pose recognition method based on adaptive uncertainty-aware meta-learning. S0: Acquire images, identify target objects in the images and acquire target images, classify target objects, and establish the pose regression task for target objects. S1, obtain the pose regression task of the target object and its associated support set and query set, and construct a Gaussian process regression model based on a linearized neural network to map the target image features to the function space, and obtain the predicted probability distribution of the query set based on the support set; S2, based on the predicted probability distribution obtained in S1 and the actual output of the query set, calculate the negative log-likelihood loss function value of the task, wherein the divergence measure is used to measure the difference between the latent distribution and the predicted distribution followed by the actual output of the query set; S3, based on the negative log-likelihood loss function value calculated by S2, is used to calculate task weights for the task through an adaptive weight generator; S4. Based on the negative log-likelihood loss function values ​​calculated for multiple tasks and the corresponding task weights, the meta-parameters of the Gaussian process regression model are optimized and updated. S5 identifies the pose of the target object based on the optimized and updated meta-parameters.

[0011] Furthermore, the Gaussian process regression model uses a neural tangent kernel to measure the correlation between different inputs and outputs.

[0012] Furthermore, the meta-parameters of the Gaussian process regression model include the covariance matrix; The method further includes: performing dimensionality reduction processing on the high-dimensional characteristics of the target image, the steps of which include: constructing a projection matrix using the principal eigenvectors of the Fisher information matrix, and projecting the covariance matrix onto a low-dimensional subspace to reduce computational complexity.

[0013] Furthermore, the optimized and updated meta-parameters include the learnable variance over the low-dimensional subspace.

[0014] Furthermore, the adaptive weight generator calculates the task weights by constructing an adaptive threshold mechanism.

[0015] Furthermore, the generation of the adaptive threshold includes: establishing a baseline threshold using the exponential moving average of historical task losses over a sliding window, calculating the standard deviation of the loss set in the current training phase to quantify loss fluctuations, and introducing a linear decay factor based on the training progress.

[0016] Furthermore, the generation of the adaptive threshold also includes setting an uncertainty correction term based on the difference between the validation loss and the training loss of the task query set.

[0017] Furthermore, when the negative log-likelihood loss function value of the support set of a task is higher than the adaptive threshold, the task is assigned a higher task weight.

[0018] Furthermore, the steps of identifying target objects in an image and acquiring the target image include image preprocessing steps: Dynamically adjust image resolution, focal length, ISO, or exposure parameters to ensure image clarity; The target image is formed by extracting the region containing the feature points and bounding boxes of the target object from the image.

[0019] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: Compared with the prior art, the present invention has the following significant advantages: This invention solves the problem of accuracy and reliability of object pose regression in small sample scenarios by introducing an adaptive weighting mechanism and probabilistic modeling.

[0020] The main manifestations include: The core adaptive weight generator can automatically identify and assign higher weights to "hard tasks" that are undertrained or have complex geometries by evaluating the difference in loss between the query set and the support set, ensuring that the model can effectively capture features in long-tail scenarios.

[0021] By combining Bayesian inference with linearized neural networks, a Gaussian process is constructed at the function level, which can effectively capture cognitive uncertainty caused by symmetry ambiguity or texture loss, thereby significantly improving the reliability of posture recognition and providing a high-confidence reference for safety-sensitive application scenarios such as obstacle avoidance in autonomous driving.

[0022] By using the Fisher information matrix to perform dimensionality reduction projection on the covariance matrix, the uncertainty of task-sensitive parameters is preserved, and the system resource consumption is significantly reduced, making it suitable for scenarios with limited computing power, such as automotive embedded devices.

[0023] Experimental data show that, compared with existing meta-learning methods (such as MAML), the present invention has smaller prediction angle errors with limited data points and can adapt to new categories of 3D objects more quickly.

[0024] This method can be used not only for predicting the behavior of target vehicles and making AEB decisions in autonomous driving, but also for industrial robots to accurately pick up and recognize the posture of fragile or high-precision workpieces.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort.

[0027] Figure 1 This is a flowchart of the pose recognition method based on adaptive uncertainty-knowable meta-learning of the present invention.

[0028] Figure 2 This invention provides angle prediction results for pose recognition tasks of different types of objects.

[0029] Figure 3 This invention provides the angle prediction results for the attitude recognition task of a specific vehicle.

[0030] Figure 4 This is a comparison chart of the angle error index results for predicting the pose of a 3D object using the present invention and classical algorithms. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0032] The core challenge in the field of 3D object pose regression lies in constructing a unified framework that can efficiently adapt to new object categories, accurately quantify pose recognition uncertainty, and provide reliable angle prediction. Against this backdrop, this invention proposes an uncertainty-aware meta-learning framework based on an adaptive weight generator. This framework combines Bayesian inference with a linearized neural network to construct a Gaussian process function, effectively capturing cognitive uncertainty caused by object symmetry or feature sparsity, as well as task-specific geometric variations. Simultaneously, it adaptively adjusts task weights based on historical task losses and training progress, ensuring that difficult or ambiguous tasks with complex geometric structures receive more attention, thereby achieving robust pose estimation with limited viewpoint samples.

[0033] The purpose of this invention is to provide a meta-learning method for modeling cross-task cognitive uncertainty in the meta-learning process in a small-sample regression scenario of 3D object pose, and adaptively adjusting task weights to achieve accurate prediction of the pose of unseen category objects.

[0034] This method acquires image information for specific usage scenarios, identifies target objects in the images and obtains target images, classifies the target objects, and establishes a pose regression task for the target objects. Since target objects in the images may be unclear or affected by the angle of light, image preprocessing is required when acquiring target images. This can be achieved by dynamically adjusting image resolution, focal length, ISO, or exposure parameters to ensure image clarity. The method also extracts the regions in the image containing feature points and bounding boxes of the target objects to form the target images.

[0035] like Figure 1 The diagram shows a flowchart of the pose recognition method based on adaptive uncertainty-knowable meta-learning of the present invention.

[0036] The specific steps of this method are as follows: Step 1: Formalize the meta-learning framework for small-sample regression of 3D object pose, construct a linear neural network and neural tangent kernel to extract high-dimensional image features, and establish a Gaussian process regression model using a Bayesian linear regression model. Step 1 includes: a formal meta-learning framework, and sampling a single task. ,in It is a task-specific index. Each task Corresponding to a specific type of object, and a dataset Relatedly, this dataset contains two-dimensional images of such objects in different poses, which are then divided into support sets. and query set ,have (1) Each set consists of a series of input-output pairs, where the input is a feature vector extracted from a 2D image, and the output is its corresponding angle label. and .

[0037] Construct a linearized neural network and a neural tangent kernel. Consider a neural network. ,in Represents the learnable parameters of a neural network. It is input. It corresponds to the output, that is At a fixed point Nearby, neural network Linearization can be achieved using the first-order Taylor approximation: (2) in, Is The Jacobian matrix at that location. For a set of Input The corresponding stacked Jacobian matrix is: (3) Linearization models allow the variation in network output to be represented as: (4) in, To output the increment, It is a correction of the parameters.

[0038] When parameter When the distribution follows a Gaussian pattern, construct the tangential nucleus of the nerve. It measures the correlation between pose features of images of different object categories: (5) in, Representation based on linearization points The tangential nucleus of the nerve, and Representing different inputs, and This represents the stacked Jacobian matrix corresponding to these two sets of inputs.

[0039] A Gaussian process regression model is established using a Bayesian linear regression model. For a set of... Input and stacked output vectors According to the Bayesian linear regression model, we have (6) in, , It is a feature mapping. It is a parameter vector. Let represent the noise matrix, where , It is the variance of Gaussian noise. It is an identity matrix.

[0040] Assuming neural network weights ,in, The mean, Represents covariance. Due to the output... The distribution is based on Gaussian priors. Decision-made, and each Corresponding to a function Therefore, the focus is on the modeling parameters. The distribution shifts to the characterization function The distribution of this Gaussian process is expressed as... (7) (8) in, Let the prior mean of the function be . Let be the prior covariance between functions with different inputs.

[0041] For new input The posterior predicted distribution still conforms to the Gaussian process distribution. (9) (10) in, The posterior mean is... For posterior covariance, The kernel function of a Gaussian process describes the input vector. and Similarity between functions , , Calculation and The calculation method is the same, the only difference is the input vector, which will not be elaborated here.

[0042] Step 2: Define the objective function to be optimized and quantify the difference between the model's predicted distribution and the theoretical distribution. Step 2 includes: due to theoretical distribution It cannot be obtained directly; the meta-training dataset is sampled from the theoretical distribution. As the foundational data pool for extracting contextual data, among which... This represents the total number of meta-training tasks, and the dataset for each specific task is... , and Indicates task The Each input and output.

[0043] The KL (Kullback Leibler) divergence is used as the optimization objective to minimize the theoretical distribution. With model prediction of agent distribution The differences between them (11) in, This represents the learnable parameters of the surrogate model, ensuring the interpretability of the model. The KL divergence is expressed in the following mathematical form: (12) in and Indicates from theoretical distribution The query input and output are sampled in the middle. Indicates the input in the query. Heyuan training dataset Query output under conditions Theoretical distribution, Indicates the input in the query. Heyuan training dataset Query output under conditions The proxy distribution.

[0044] because With optimization parameters Unrelated, minimizing the KL divergence is equivalent to maximizing the log-likelihood of the surrogate distribution on the true data. This is equivalent to minimizing the negative logarithmic likelihood (NLL), i.e. (13) Step 3: To address the high-dimensionality of the image input, construct a dimension-reduced subspace within the linearized probability model. Step 3 includes: calculating the principal eigenvectors of the Fisher Information Matrix (FIM) using the Jacobian matrix representing the pose features of the 3D object, and then applying the parameters... covariance matrix Projected to This is achieved by using a subspace of dimensionality, thereby reducing computational complexity while preserving task-sensitive parameter uncertainties. The specific mathematical expression is as follows: (14) in, Represents the main feature vector (previous) containing FIM. The projection matrix (in each direction), If the learnable variance is along these directions, then during meta-training, the effective meta-parameters become... Therefore, it can be deduced that for a pose regression task of a certain type of object... The specific form of the Gaussian prior prediction distribution follows the model established in step 1: (15) (16) in, The prior covariance of the input data for this task. This represents the prior mean of the input data for this task. , During the meta-testing process, the posterior prediction distribution can be expressed as: (17) (18) Among them, the posterior mean Used to generate prediction results The covariance matrix This quantifies the uncertainty introduced by limited contextual data. Based on the above derivation, the specific form of the loss function in step 2 can be written as: (19) in, Indicates task The negative log-likelihood function, The dimension of the output vector. This represents the number of output vectors.

[0045] Step 4: Construct an adaptive weight generator to generate task-specific weights through a dynamic threshold mechanism. Step 4 includes: During meta-training, the instantaneous loss value of a single pose regression task is easily affected by factors such as image noise. To reduce the instability caused by noise-sensitive instantaneous loss values, an adaptive threshold mechanism for generating weights for pose regression tasks of different object categories is designed. This method first uses a sliding window. The exponential moving average of the historical task loss is used to establish a baseline threshold, and then the calculation is performed during the current training phase. loss set Standard deviation This method quantifies the loss fluctuation of pose regression tasks for different object categories. Furthermore, it introduces a linear decay factor based on the training progress. The specific expression is: (20) in, This represents the total number of training steps. Furthermore, to address the distributional discrepancy between the training and validation sets in a specific task, an uncertainty correction term is added. : (twenty one) in, Indicating during the training phase Previous mission The set of all validation losses for the query set. This represents the set of all training losses for the same query set. Combining the above factors, a loss structure is constructed for the task... Adaptive threshold as follows: (twenty two) in, Indicates a sliding window Historical loss values ​​within, It is an empirical scaling factor, which can usually be taken as... .

[0046] Given a support set training loss and the query set after model update Corresponding verification loss Then the task weight Defined as: (twenty three) Normalization process yields This yields the normalized weights for the pose regression task of each specific object category. Based on the loss function calculation in step 3, the weights can then be used... To optimize parameters .

[0047] Step 5: Run the algorithm to identify the pose of the target object and analyze the results.

[0048] Step 5 includes: This invention uses the ShapeNet-1D benchmark dataset for simulation experiments. This dataset contains renderings of 3D objects of various categories, primarily focusing on one-dimensional orientation regression. Figure 2 and Figure 3 The prediction results of this invention are shown. Figure 4 The comparison of prediction angle errors between traditional algorithms and the present invention under different contextual data quantities is presented.

[0049] Through the above process, the model can quickly adapt to pose regression tasks for new categories of 3D objects and achieve accurate predictions even with limited data points. The advantages of this invention are mainly twofold. First, the adaptive weight generator allows it to focus more on tasks with poor training performance, enhancing the model's generalization ability. Second, by integrating Bayesian inference methods, it effectively captures the common structures and uncertainties among pose regression tasks for different categories of objects, thereby achieving accurate pose prediction in small-sample scenarios.

[0050] The method of this invention can be applied to various industrial scenarios, such as the identification of workpieces by robots in industrial production, which requires precise angle recognition to select a suitable gripping interface. This is important for preventing damage to high-precision machined surfaces or preventing the application of mechanical force to fragile areas.

[0051] It also has great application value in intelligent driving. By acquiring images of a moving car and accurately identifying its posture, driving behavior can be predicted by continuously identifying the direction and rate of posture change.

[0052] The following section selects the intelligent driving scenario from the above application scenarios to introduce an embodiment of the method. The basic process of other application scenarios will not be listed in detail. The difference is that the acquisition and preprocessing of the target object image need to be adjusted according to the working conditions. After identifying the angle of the target object, these angle data are used to perform subsequent processing for specific working conditions in order to make control operations that adapt to the working conditions.

[0053] The specific steps of this method are as follows: Traffic environment image acquisition and target extraction preprocessing based on vehicle-mounted cameras.

[0054] The vehicle uses its onboard intelligent forward-facing camera or surround-view wide-angle camera to capture raw video streams of the road ahead in real time. Due to the complex driving environment, in-depth preprocessing is required by the onboard image processing unit. First, the resolution, focal length, color depth, ISO, and exposure value of the original image are extracted. The system evaluates the current lighting conditions. If poor ambient light or strong backlighting is detected, the above parameters are dynamically adjusted to preset resolution, focal length, color depth, ISO, and exposure value to enhance image features and generate a new image. Subsequently, image distortion correction, color adjustment, white balance, and contrast equalization are performed to eliminate feature extraction interference caused by the physical environment.

[0055] Based on this, dynamic object detection is performed, target recognition and classification are performed on the preprocessed new image, feature points and bounding boxes containing specific target vehicles are extracted, and these are cropped and normalized into region of interest (ROI) feature images, which are used as input images for target pose recognition.

[0056] Input the pose meta-learning framework model to perform pose recognition tasks for specific vehicle models based on vehicle images.

[0057] A linear neural network is used to extract high-dimensional features of the vehicle, and a first-order Taylor approximation is performed near a fixed point.

[0058] Calculate the correlation between images of different vehicle models; establish a Gaussian process using Bayesian linear regression, and output vehicle angle labels through posterior prediction distribution.

[0059] By minimizing the difference between the theoretical distribution and the surrogate distribution using the KL divergence function, this function can simultaneously measure the accuracy of angle prediction and the confidence (uncertainty) of the model's attitude judgment for the vehicle model.

[0060] To address the high-dimensionality of in-vehicle image inputs and avoid consuming limited in-vehicle computing resources with a large number of redundant parameters, a dynamic thresholding mechanism is used after dimensionality reduction processing of vehicle images to allocate more resources to challenging vehicle pose recognition tasks that are geometrically complex and susceptible to light and shadow interference.

[0061] When autonomous driving models encounter uncommon or geometrically complex vehicles (i.e., difficult tasks), their conventional validation loss often exceeds the baseline threshold. In this case, the adaptive weight generator prioritizes the task and assigns extremely high normalized weights to its negative log-likelihood loss function. This forces the onboard model to focus its computational power on extracting the pose geometry of the rare vehicle within a limited recognition period, significantly enhancing the model's feature capture capability in long-tail traffic scenarios.

[0062] In other words, task weights are calculated based on the loss value; for special vehicles with larger recognition errors, the model will automatically assign higher weights for focused reinforcement learning. For such vehicles, the model can quickly regress the accurate angle through meta-learning parameters.

[0063] After completing the model update and obtaining high-precision real-time three-dimensional attitude angles of the target vehicle (especially the yaw angle that provides driving direction indication) and the corresponding uncertainty confidence, the system executes physical obstacle avoidance control operations.

[0064] To better illustrate the application scenario, an example of obstacle avoidance based on changes in vehicle posture is given below.

[0065] First, the system acquires signals sent by the positioning sensors, calculates the position coordinates of the sensors to establish a preset relative coordinate system, and calculates the relative attitude vector between the vehicle and the target vehicle.

[0066] Secondly, the system inputs the continuously output target vehicle attitude angle sequence into the vehicle behavior strategy prediction model. By analyzing the abrupt change trend of the yaw angle, it predicts the target vehicle's behavioral intentions, such as lane changing, cutting in, or sudden braking. Based on this predicted behavior, lane constraints provided by the high-precision map, and the target vehicle's current speed information, the system calculates the target vehicle's future trajectory coordinates.

[0067] Next, based on the current hydraulic and mechanical state of the Automatic Emergency Braking (AEB) system, and combined with the vehicle's trajectory and the future coordinates of the target vehicle, the minimum collision distance between the vehicle and the obstacle is calculated. If the cognitive uncertainty of the posterior covariance output is high (such as encountering severe halo), the system will automatically compensate and extend the safety judgment threshold.

[0068] Finally, decision information for automatic emergency braking is generated based on the calculated minimum collision distance. If a collision risk exists, an obstacle avoidance algorithm is used to calculate the vehicle's next movement vector (including braking deceleration or steering offset), and the movement vector is corrected according to lane constraints. Then, the command is sent to the drive-by-wire chassis and steering actuator to control the vehicle to achieve safe obstacle avoidance.

[0069] The computational or control model used in obstacle avoidance can adopt existing technical solutions, which will not be elaborated here. The main purpose of this invention is to improve the object posture recognition method under the condition of scarce samples.

[0070] Compared with existing technologies, the specific beneficial technical effects of the invention are as follows: In a 3D object pose regression scenario based on the ShapeNet-1D dataset, an adaptive weight generator prioritizes different object regression tasks. This generator calculates task-specific weights by evaluating the difference in query and support set losses for different tasks, and assigns higher weights to under-trained tasks using a dynamic thresholding mechanism. This design ensures that difficult or feature-fuzzy tasks receive more attention during meta-optimization.

[0071] To address the issue that visual features of 3D objects are susceptible to texture loss and geometric symmetry, leading to unstable feature extraction, a method combining Bayesian inference with a Gaussian process defined on a linearized neural network function is used to effectively capture cognitive uncertainty and task-specific geometric variations in pose regression.

[0072] By constructing a Gaussian process from the perspective of a function, the probabilistic behavior of the function is directly simulated. The covariance matrix is ​​reduced in dimensionality using the Fisher information matrix, which avoids the influence of high-dimensional parameters, significantly improves computational efficiency, and reduces the consumption of system resources.

[0073] Experiments based on the ShapeNet-1D dataset show that, compared with existing meta-learning methods, the present invention significantly improves prediction accuracy and reduces prediction angle error.

[0074] These results highlight the ability of this invention to model the uncertainty of 3D object pose regression tasks and adaptively adjust task learning priorities, demonstrating that the framework can still provide accurate predictions and reliable uncertainty assessments under complex task distribution conditions.

[0075] This approach focuses on maintaining the learner's high efficiency and accuracy in prediction even when considering uncertainty and complex task distribution. It introduces an adaptive weighting mechanism in the meta-learning process and uses Bayesian inference for probabilistic modeling to achieve task-specific adaptation. On the one hand, the adaptive weight generator enables it to pay more attention to tasks with poor training performance, enhancing the model's generalization ability. On the other hand, by integrating Bayesian inference methods, it effectively captures the common structure and uncertainty features among pose regression tasks of different object categories, thereby achieving effective regression prediction of 3D object poses in small sample scenarios.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pose recognition method based on adaptive uncertainty-aware meta-learning, characterized in that, S0: Acquire images, identify target objects in the images and acquire target images, classify target objects, and establish the pose regression task for target objects. S1, obtain the pose regression task of the target object and its associated support set and query set, and construct a Gaussian process regression model based on a linearized neural network to map the target image features to the function space, and obtain the predicted probability distribution of the query set based on the support set; S2, based on the predicted probability distribution obtained in S1 and the actual output of the query set, calculate the negative log-likelihood loss function value of the task, wherein the divergence measure is used to measure the difference between the latent distribution and the predicted distribution followed by the actual output of the query set; S3, based on the negative log-likelihood loss function value calculated by S2, is used to calculate task weights for the task through an adaptive weight generator; S4. Based on the negative log-likelihood loss function values ​​calculated for multiple tasks and the corresponding task weights, the meta-parameters of the Gaussian process regression model are optimized and updated. S5, based on the optimized and updated meta-parameters, identifies the pose of the target object; The adaptive weight generator calculates the task weights by constructing an adaptive threshold; The adaptive threshold is constructed by combining the baseline threshold, the standard deviation of the loss set in the current training phase, the linear decay factor, and the uncertainty correction term. When the negative log-likelihood loss function value of the support set of a task is higher than the adaptive threshold, the task is assigned a higher task weight.

2. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 1, characterized in that, The Gaussian process regression model measures the correlation between different inputs and outputs using a neural tangent kernel.

3. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 1, characterized in that, The meta-parameters of the Gaussian process regression model include the covariance matrix; The method further includes: performing dimensionality reduction processing on the high-dimensional characteristics of the target image, the steps of which include: constructing a projection matrix using the principal eigenvectors of the Fisher information matrix, and projecting the covariance matrix onto a low-dimensional subspace to reduce computational complexity.

4. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 3, characterized in that, The optimized and updated meta-parameters include the learnable variance over the low-dimensional subspace.

5. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 1, characterized in that, The generation of the adaptive threshold includes: establishing a baseline threshold using the exponential moving average of historical task losses over a sliding window, calculating the standard deviation of the loss set in the current training phase to quantify loss fluctuations, and introducing a linear decay factor based on the training progress.

6. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 5, characterized in that, The generation of the adaptive threshold also includes setting an uncertainty correction term based on the difference between the validation loss and the training loss of the task query set.

7. The pose recognition method based on adaptive uncertainty-based meta-learning according to claim 1, characterized in that, The steps for identifying target objects in an image and acquiring the target image include image preprocessing steps: Dynamically adjust image resolution, focal length, ISO, or exposure parameters to ensure image clarity; The target image is formed by extracting the region containing the feature points and bounding boxes of the target object from the image.

Images