Local contour based uncertainty visual perception augmentation method and system

Abstract

This invention discloses a method and system for enhancing visual perception based on local contour uncertainty, belonging to the field of machine vision automation technology. It includes real-time acquisition of local contour image sequences of a target workpiece for visual processing, feature extraction, and matching with pre-stored reference contour data; generating a state uncertainty vector based on the matching residual and feature stability calculation; online planning of the subsequent motion path of the acquisition device to generate an active perception path that optimizes the integrity of contour information; using the state uncertainty vector and the active perception path as joint adjustment parameters to drive a pre-trained visual target dynamic model, generating a predicted pose and predicting the visual evolution trend; and finally, integrating the predicted pose, evolution trend, state uncertainty, and active path to generate and output contour perception enhancement data. This invention achieves reliable estimation of target pose and proactive improvement of perception quality under local, dynamic observation conditions.

Description

Technical Field

[0001] This invention relates to the field of machine vision automation technology, specifically to a method and system for enhancing visual perception of uncertainty based on local contours. Background Technology

[0002] Visual perception and pose estimation techniques based on local contours are crucial for tasks such as workpiece recognition, grasping, and assembly in industrial automation. Their goal is to accurately determine the position and orientation of a target object in three-dimensional space by analyzing its contour information. In recent years, with the development of high-resolution cameras and efficient image processing algorithms, visual localization methods based on complete or mostly contoured contours have become relatively mature. Existing technologies typically rely on pre-stored standard workpiece models (reference contour data). Through feature extraction and template matching algorithms, real-time acquired image features are compared with the reference data to calculate the pose.

[0003] These methods perform well in scenarios where the target contour is clear and fully visible. However, in complex real-world industrial environments, such as grasping stacked workpieces within an assembly line frame or when the field of view of a robot end-effector (EOAT) is limited, imaging systems often only continuously acquire partial and incomplete contour image sequences of the target workpiece. Existing technologies face challenges in processing such local contour information, including limited local feature information, poor compatibility with traditional template matching methods based on complete contours, low reliability of direct matching, and high uncertainty in pose calculation results. Secondly, static, single-shot matching strategies cannot cope with the continuous evolution of contour information caused by target motion or changes in viewpoint, making it difficult to guarantee tracking robustness. More importantly, existing methods lack the ability to quantitatively assess the reliability of their own perception results; the system cannot know how confident it is in making pose judgments based on local contours. This results in subsequent robotic arm grasping or visual servo control decisions being based on a perception result with unknown reliability, which can easily lead to operational failures or safety accidents when uncertainty is high. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for enhancing visual perception of uncertainty based on local contours.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention discloses a method for enhancing visual perception of uncertainty based on local contours, comprising the following steps:

[0007] Real-time acquisition of local contour image sequences of the target workpiece;

[0008] The local contour image sequence is visually processed. Based on image feature extraction and visual registration algorithms, the target contour features are extracted and registered and matched with the pre-stored reference contour data in real time. Based on the matching residual and feature stability calculation, a state uncertainty vector for quantifying the reliability of perception is generated.

[0009] Based on the state uncertainty vector, the subsequent motion path of the acquisition device is planned online to generate an active sensing path aimed at optimizing the integrity of contour information;

[0010] Using the state uncertainty vector and the active perception path as joint adjustment parameters, the pre-trained visual target dynamic model is driven to generate a predicted pose and predict the visual evolution trend of the local contour image sequence.

[0011] Based on the predicted pose, the visual evolution trend, the state uncertainty vector, and the planning information of the active perception path, contour perception enhancement data is generated and output.

[0012] Secondly, the present invention discloses an uncertainty-based visual perception enhancement system using local contours, comprising:

[0013] The image acquisition module is used to acquire local contour image sequences of the target workpiece in real time.

[0014] The uncertainty generation module is used to perform visual processing on the local contour image sequence. Based on image feature extraction and visual registration algorithms, it extracts target contour features and performs real-time registration and matching with pre-stored reference contour data. Based on the matching residual and feature stability calculation, it generates a state uncertainty vector for quantifying perceived reliability.

[0015] The path perception module is used to plan the subsequent motion path of the acquisition device online based on the state uncertainty vector, and generate an active perception path aimed at optimizing the integrity of the contour information.

[0016] The modeling and prediction module is used to drive a pre-trained visual target dynamic model to generate a predicted pose and predict the visual evolution trend of the local contour image sequence by using the state uncertainty vector and the active perception path as joint adjustment parameters.

[0017] The output module is used to generate and output contour perception enhancement data based on the predicted pose, the visual evolution trend, the state uncertainty vector, and the planning information of the active perception path.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This invention provides a quantitative assessment of the reliability of the current local contour perception result by dynamically calculating and generating a state uncertainty vector. This enables the system to clearly know the confidence level of a single visual matching result in complex observation environments, thus providing a reliable basis for subsequent decision-making and avoiding the problem of high decision-making risk caused by neglecting uncertainty in traditional methods.

[0020] 2. This invention uses the state uncertainty vector to plan the active sensing path online, guiding the acquisition device towards the area with the least complete information. This transforms the vision system from passively receiving images to actively exploring targets, enabling it to acquire key contour information more efficiently within a limited observation time. This compensates for the inherent limitations of local field of view and improves the overall coverage and utilization efficiency of contour information.

[0021] 3. This invention combines the state uncertainty vector with the active perception path as adjustment parameters for the dynamic model of the visual target, enabling the model's prediction process to be closely integrated with the current perception quality and the system's exploration intent. This not only improves the accuracy of pose prediction but also enhances the adaptability to predicting future visual evolution trends. The final output contour perception enhancement data incorporates more comprehensive state information, providing more robust and accurate guidance for subsequent tasks such as robot grasping, assembly, or visual servoing. Attached Figure Description

[0022] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0023] Figure 1 This is a flowchart of the steps of the present invention;

[0024] Figure 2 This is a flowchart illustrating the working principle of the present invention;

[0025] Figure 3 This is a system module connection diagram of the present invention. Detailed Implementation

[0026] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0027] In existing technologies, visual perception and pose estimation of workpieces often rely on complete contour information or static template matching, making it difficult to balance real-time performance and robustness in complex environments. Traditional methods, when only able to acquire local, dynamic contour sequences, struggle to assess the reliability of visual registration, leading to significant uncertainties in pose calculation results. Current processes typically separate perception from decision-making, failing to proactively adjust observation strategies based on the reliability of current perception. Especially when information is severely insufficient or the target moves rapidly, static, single-shot matching models may experience performance degradation or even failure, making it difficult to meet the control requirements of high-precision automated operations.

[0028] To address the aforementioned issues, the study found that the matching residual distribution of visual registration and the stability of feature point tracking can effectively quantify the uncertainty of the current local perception results. Furthermore, it was discovered that using the quantified uncertainty to guide the active movement of the acquisition device can systematically fill information gaps and improve the completeness of perception. Further simulation and experimental verification demonstrated that using the uncertainty vector and the active perception path as joint inputs to a dynamic model enables pose and trend prediction to adaptively adapt to perception quality and system intent. Finally, an enhanced perception system with uncertainty as its feedback core, incorporating real-time re-evaluation and self-verification mechanisms, was constructed, achieving a shift from passive reception to active enhancement.

[0029] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Example 1:

[0031] like Figure 1 As shown, the uncertainty-based visual perception enhancement method based on local contours includes the following steps:

[0032] Real-time acquisition of local contour image sequences of the target workpiece;

[0033] Visual processing is performed on local contour image sequences. Based on image feature extraction and visual registration algorithms, target contour features are extracted and registered and matched with pre-stored reference contour data in real time. Based on matching residuals and feature stability calculations, a state uncertainty vector for quantifying perception reliability is generated.

[0034] Based on the state uncertainty vector, the subsequent motion path of the acquisition device is planned online to generate an active sensing path aimed at optimizing the integrity of contour information;

[0035] Using the state uncertainty vector and the active perception path as joint adjustment parameters, a pre-trained visual target dynamic model is driven to generate predicted poses and predict the visual evolution trend of local contour image sequences.

[0036] Based on the predicted pose, visual evolution trend, state uncertainty vector, and planning information of the active perception path, contour perception enhancement data is generated and output.

[0037] like Figure 2 The diagram shown illustrates the working principle of this invention. The working principle of this application is as follows: the acquisition device continuously captures images of the target workpiece at a preset frequency (e.g., 30 frames per second). Since the workpiece may be partially obscured, stacked, or the camera's field of view may be limited, the acquired image sequence typically only contains a portion of the overall outline of the workpiece. This partial outline image sequence is then transmitted in real-time to the data processing device.

[0038] After receiving the data, the data processing device invokes its built-in image feature extraction algorithms (such as SIFT, SURF, ORB feature point descriptors, or deep learning-based feature extraction networks) to extract key points or edge feature sets representing the workpiece contour from the current frame image, serving as the target contour features. Subsequently, a visual registration algorithm (such as the RANSAC-based PnP algorithm, ICP algorithm, or its variants) performs real-time matching and alignment between the extracted target contour features and the reference contour data pre-stored in the data processing device. The reference contour data is a feature set extracted from a complete and accurate 3D CAD model of the target workpiece or from standard image templates acquired from multiple perspectives. The matching process generates a set of matching residuals (i.e., the error in the spatial position of the corresponding feature points), while the program analyzes the trajectory stability of the same feature point in consecutive frames (e.g., consistency through optical flow tracking or feature point re-detection). Based on the statistical distribution (e.g., covariance) and the measure of feature stability (e.g., tracking success rate) of these matching residuals, a state uncertainty vector is generated through preset solution rules (e.g., normalizing the eigenvalues ​​of the residual covariance matrix and fusing them with the stability coefficient). The state uncertainty vector is a multi-dimensional data point used to quantify the reliability or confidence level of the current single-frame or short-sequence perception results across various dimensions such as position and attitude. It is defined as follows:

[0039] in, This represents the pose uncertainty component, a non-negative scalar. The larger the value, the lower the reliability of the current estimated workpiece pose (position and orientation).

[0040] This is represented as the contour completeness confidence component, a scalar with a value ranging from [0,1]. A larger value indicates more complete and reliable contour information. In practical systems, It can be represented as an array or structure containing the two components mentioned above.

[0041] Image feature extraction and visual registration algorithms are a general overview of techniques used in this field to identify objects from images and calculate their poses. Feature extraction can employ feature descriptors including, but not limited to: handcrafted features such as SIFT (Scale Invariant Feature Transform), SURF (Speed-Up Robust Features), ORB (Oriented Fast and Rotated BRIEF), and BRISK (Binary Robust Invariant Scalable Keypoints), or deep features learned from convolutional neural networks (CNNs).

[0042] Visual registration methods include, but are not limited to: feature matching-based PnP (perspective N-point) solving algorithms (such as those combined with RANSAC), iterative nearest point (ICP) algorithms and their variants, template matching-based methods, or image alignment algorithms directly based on pixel intensity.

[0043] By analyzing the values ​​of each component in the state uncertainty vector, the direction with the least reliable or scarcest information in the current perception is identified (e.g., abnormally high uncertainty in a certain degree of freedom, or extremely low contour feature matching in a certain local area). Combining the kinematic model and workspace constraints of the motion control device, an optimized subsequent motion trajectory is calculated online as the active sensing path. This path aims to guide the acquisition device to move to one or more new observation poses to complete or verify the currently missing or unreliable contour information from different angles, thereby optimizing the completeness of the overall contour information. The planned path is converted into control commands and sent to the motion control device.

[0044] Simultaneously, the state uncertainty vector and the actively perceived path generated by the planning are used as joint adjustment parameters and input into a pre-trained visual target dynamic model. This visual target dynamic model can be a physics-based kinematic / dynamic model, a statistical time series model (such as a family of Kalman filters), or a data-driven machine learning model (such as a recurrent neural network). The joint adjustment parameters refer to the two sources of information input into the pre-trained visual target dynamic model used to adjust its predictive behavior:

[0045] Perceived quality regulation: by The scaling factor for the model process noise covariance matrix derived from component transformation. Specifically, if the model's default process noise covariance is... The adjusted noise covariance is ,in The preset sensitivity coefficient, .

[0046] Observation intention adjustment: The expected future observation perspective, resolved from the active sensing path, is used to refine the model's observation model. For example, in a filter-based model, the predicted observation Jacobian matrix can be calculated in advance based on the future camera pose.

[0047] After considering these moderating factors, the visual target dynamic model infers the future state of the target object and outputs two core predictions: first, the predicted pose (position and orientation) of the target at future time; and second, the visual evolution trend. The prediction result of the local contour of the target workpiece changing on the image plane over a period of time, that is, predicting how the appearance of the local contour on the image plane will change over time (e.g., how the edges move, how the shape deforms). Let's assume we select... A stable key point indicates the trend of visual evolution. It can be represented as:

[0048]

[0049] in, It is the predicted position of the i-th key point in the image pixel coordinate system at a future time t;

[0050] The current moment;

[0051] To predict the length of the time window.

[0052] The training process for the pre-trained visual target dynamic model is completed offline. The training set comes from time-series data collected in historical tasks. Specifically, it includes: image sequences containing local contours, the corresponding workpiece's true pose trajectory synchronously acquired through high-precision external measurement equipment (such as a laser tracker), and intermediate state data calculated from the image sequences (such as feature matching residual sequences and optical flow vector fields). The raw data undergoes time alignment, denoising, and normalization.

[0053] A Long Short-Term Memory (LSTM) network is used as the basic architecture, comprising an input layer, a hidden layer, and an output layer. The dimension of the input layer corresponds to the dimension of the hybrid visual feature set, preferably 128 dimensions; the number of neurons in the hidden layer is preferably 64; the output layer corresponds to the pose changes (6 dimensions, including 3D translation and 3D rotation) over the next 5 time steps. The input features are the local image feature encoding at the current time, the historical pose sequence, and the historical residual statistics; the output labels are the pose changes or keypoint position sequences over the next few steps. Mean squared error is used as the loss function. The optimizer is Adam, with an initial learning rate of 0.001 and an exponential decay strategy. The training cycle is set to 100 rounds, with each round traversing the entire training set, and a batch size of 32. During training, approximately 20% of the data is used as a validation set to monitor overfitting and implement an early stopping strategy.

[0054] The system fuses and encapsulates the four types of information generated in the above steps—predicted pose, visual evolution trend, state uncertainty vector, and active perception path planning information—into contour-aware augmented data. Fusion can include data alignment, weighted synthesis, or rule-based logical integration. Contour-aware augmented data transcends a single pose value; it is a comprehensive information package containing optimal estimation, uncertainty range, future change predictions, and recommended observation actions.

[0055] Through the above methods, the system achieves a transformation from passively receiving local images to actively guiding the perception process. The introduction of a state uncertainty vector enables the system to self-assess the perception quality, while active path planning based on this vector allows the system to autonomously fill information gaps. The visual target dynamic model, after integrating uncertainty assessment and active intent, makes predictions more reliable and forward-looking. Therefore, this invention can provide more accurate, robust, and comprehensive guidance data for robot grasping, assembly, or visual servo control in challenging environments where only local, dynamic contour information is available, thereby improving the overall performance and adaptability of automated systems.

[0056] This application further proposes that, before acquiring the local contour image sequence of the target workpiece in real time, initialization and preparation work is also included. First, the system receives a task instruction containing the target workpiece model identifier. This is usually issued by the upstream production management system or operator interface. After receiving the instruction, the data processing equipment will initiate the initialization process.

[0057] Based on the received model identifier, the system accesses a pre-built visual feature database, which can be stored in local storage or a network server. The database stores corresponding reference contour data packets for each model of workpiece. The system retrieves and accurately loads the corresponding reference contour data from the visual feature database.

[0058] The reference contour data includes a standard local contour visual feature set and its corresponding ideal geometric model. The standard local contour visual feature set is not composed of features from a single complete view, but rather a multi-view feature library compiled by offline acquisition of high-resolution images of the workpiece from multiple typical perspectives (such as front view, side view, top view, and multiple tilt angles), processed using the same or compatible feature extraction algorithms (such as SIFT descriptors) as the online stage. This ensures that during online matching, regardless of which part of the workpiece is initially observed, the corresponding benchmark features can be found in the feature library for comparison. The ideal geometric model is typically a precise 3D CAD model (such as STEP or IGES format) or a high-precision mesh model reconstructed from 3D scanning. This ideal geometric model is not only used for visualization, but more importantly, it provides accurate geometric priors and physical constraints for subsequent pose calculations (such as the PnP algorithm), visibility analysis in active path planning, and dynamic models of visual targets.

[0059] Through the above initialization and preparation work, this application ensures the accuracy of the sensing benchmark by loading precise data based on the model identifier, avoids mismatch between different workpieces, and forms a complete and engineerable technical solution by closely integrating with the online real-time sensing enhancement process. It can be stably and reliably applied to flexible automated production lines with multiple varieties and small batches.

[0060] This application further proposes a step of extracting target contour features and performing real-time registration and comparison with pre-stored reference contour data. This overcomes the limitations of a single feature descriptor when facing different imaging conditions (such as illumination changes, local occlusion, surface reflection, or motion blur) by adopting a dynamic strategy that adapts to environmental changes.

[0061] During system initialization, various contour feature descriptors with different characteristics are pre-set in the data processing equipment, such as the SIFT descriptor which is robust to scale changes, the ORB descriptor which is robust to rotation changes, and the FAST keypoint combined with BRIEF descriptor which has higher computational efficiency.

[0062] When a new local contour image frame is input, the system dynamically selects a primary and secondary descriptor from multiple preset contour feature descriptors based on the imaging quality of the local contour image sequence, using a pre-defined decision logic. Imaging quality is based on computable image metrics, such as overall image contrast, average gradient magnitude (used to characterize edge sharpness), and noise level estimates calculated through local variance. The pre-defined decision logic can be a rule-based lookup table or a lightweight classifier. For example, if a low average gradient magnitude is detected (potentially indicating blur), SIFT, which is insensitive to brightness changes, is preferentially selected as the primary descriptor; if computational resources are limited or a high frame rate is required, ORB might be chosen as the primary descriptor.

[0063] The system extracts contour features from the current frame image using the main descriptor, and then analyzes the confidence distribution of the extraction results. Confidence can be estimated by the response intensity of feature points (e.g., Harris corner response values) or the uniqueness of feature descriptors within a local region (preliminarily determined by the ratio of nearest to second nearest neighbor distances). For local image regions with confidence levels below a preset threshold (e.g., response intensity below 60% of the median overall image response intensity) (which can be determined by partitioning the image into grids), the system calls a pre-selected auxiliary descriptor to supplement feature extraction for these local regions with insufficient confidence in the main descriptor. The auxiliary descriptor is typically chosen to complement the characteristics of the main descriptor; for example, when the main descriptor is SIFT, the auxiliary descriptor can be an edge fragment descriptor that is more effective for regions with less texture.

[0064] To ensure consistency in the comparison, the same dynamic selection logic and descriptor set were used when constructing the standard local contour visual feature set of the reference contour data offline, generating multiple sets of corresponding reference contour data and storing them in the database. During online matching, the system retrieves the reference contour data generated under the corresponding combination from the database based on the currently selected primary and secondary descriptor combination. Subsequently, a robust registration algorithm (e.g., using a PnP solver based on the RANSAC framework) is used to register the extracted mixed visual feature set with the reference contour data obtained using the same descriptor strategy, solving for the preliminary pose transformation relationship.

[0065] During the registration process, the system performs a more granular reliability analysis. It divides the image (or the model portion projected onto the image plane) into several sub-regions (e.g., a 10x10 grid). For each sub-region, the system calculates two key metrics: the proportion of successfully matched feature point pairs within that region relative to the total number of attempted matches (inlier ratio); and the statistical measure of the reprojection error of the inliers of these successfully matched pairs (e.g., root mean square error), i.e., the residual. Based on these two metrics, the system constructs a local visual registration reliability map corresponding to the image region. In this local visual registration reliability map, each sub-region is assigned a reliability score, which can be a normalized weighted combination of the inlier ratio and the residual. When generating the macroscopic state uncertainty vector, the sub-region statistics provided by this local visual registration reliability map (e.g., the variance of the reliability score, the proportion of areas below the reliability threshold of 0.5) are introduced as important additional weighting factors in the calculation of the generated state uncertainty vector. For example, the calculation of the pose uncertainty component not only considers the covariance of the overall matching residual but also amplifies it based on the variance of the reliability map.

[0066] This application can adaptively cope with complex imaging environments, improve the success rate of initial feature extraction and matching under adverse conditions such as uneven illumination and missing local features, and enable the generated state uncertainty vector to more sensitively reflect the reliability differences in different spatial directions in pose estimation. Overall, it enhances the adaptability of the sensing system under non-ideal observation conditions and the robustness of the final output.

[0067] This application further proposes that the state uncertainty vector includes at least a pose uncertainty component and a contour integrity confidence component; wherein, the generation of the pose uncertainty component includes:

[0068] After the hybrid visual feature set and the reference contour data are registered, the algorithm outputs an optimal pose transformation matrix (including rotation and translation). Based on the visual registration result, the visual residual of the target workpiece relative to the reference pose is calculated. The visual residual is represented as the reprojection error of feature points in three-dimensional space, which can be decomposed into translation residuals along the X, Y, and Z axes and rotation residuals about these three axes (e.g., using axis angle representation).

[0069] Covariance analysis is performed on the visual residuals to estimate their error distribution across the three-dimensional degrees of freedom. Specifically, the system collects the residual vectors of all inlier matching pairs in the current frame and calculates their sample covariance matrix. This covariance matrix effectively characterizes the distribution characteristics of the residuals across the six degrees of freedom and the correlation between errors in each dimension.

[0070] Based on the error distribution, a statistic representing dispersion is calculated. A typical method is to calculate the eigenvalues ​​of the covariance matrix. These eigenvalues ​​reflect the variance of the error along each principal component direction. Pose uncertainty components By normalizing and aggregating these eigenvalues, quantization can be performed using the following formula:

[0071]

[0072] in, It is the i-th eigenvalue of the residual covariance matrix;

[0073] parameter and The preset weighting coefficients (e.g., calibrated using historical data to be 0.7 and 0.3 respectively) are used to balance the impact of overall average uncertainty and worst-case directional uncertainty. ∈[0.5,0.9], = .

[0074] This application provides a comprehensive and computable reliability characterization of current perception results based on local contours, enabling subsequent active perception path planning to make more refined decisions (e.g., whether to prioritize reducing pose uncertainty or prioritizing the completion of missing contours). It also provides key adjustment signals for the visual target dynamic model, enabling its predictions to be based on a deep understanding of the quality of the current perception state, thereby improving the overall credibility of the system output and the rationality of decision-making in information-constrained environments.

[0075] This application further proposes that the step of generating a state uncertainty vector for quantifying perceived reliability based on the matching residual and feature stability calculations also includes:

[0076] The system segments the acquired local contour image sequence into multiple regular local regions, such as dividing it into an M-row, N-column grid (M and N can be preset according to the image resolution, with typical values ​​being 8x8 or 16x16). The purpose of this segmentation is to spatially locate specific areas where information is missing or poorly matched.

[0077] For each local region, the system performs a critical matching evaluation. First, all contour features extracted within the region (keypoints, edge fragments, etc.) are collected. Then, the system compares this region with the theoretical features of its corresponding spatial location in the registered reference contour data. The matching score between the extracted contour features and the corresponding region features in the reference contour data is calculated. The corresponding region features are determined by applying the currently estimated pose to an ideal geometric model, projecting the model onto the image plane, and calculating the grid into which the projected model contour falls.

[0078] For each local region, the system calculates the matching score. For example, it can be defined as the ratio of the number of successfully matched feature point pairs within a region to the total number of feature points that should theoretically exist in that region, multiplied by the average feature descriptor similarity (such as cosine similarity) of the points within those matched pairs. Matching score The value of is in the range of [0,1].

[0079] The system compares the matching scores of all local regions with a preset matching threshold. Compare the results. Preset matching threshold. The set value considers the characteristics of feature descriptors and the level of environmental noise, ranging from 0.3 to 0.7. The 15th to 25th percentiles of historical good image matching distributions are used as the preset matching threshold. .

[0080] Calculate the matching score for all local regions. Below the preset matching threshold The number of regions, and calculate their proportion of the total number of regions assessed. A low matching score could stem from two reasons: first, the region might actually be occluded or outside the field of view (lack of true information); second, the image quality might be too poor, causing feature extraction to fail (lack of false information). To distinguish between these two situations, confidence calculation needs to incorporate image quality metrics.

[0081] The system calculates the image sharpness index of the current frame in parallel. (For example, using the variance of the Laplace operator) and noise level indicators. (For example, through wavelet transform or block effect estimation). Based on the region proportion and fused with the acquired image sharpness and noise level indices, the contour integrity confidence component is calculated through a preset confidence mapping function.

[0082] Contour completeness confidence components It is calculated using a pre-defined confidence mapping function F, as shown in the following formula:

[0083]

[0084] in, It is a normalization constant representing the theoretically maximum acceptable proportion of low-matching regions (ranging from 0.6 to 0.8, for example, it can be set to 0.7).

[0085] Q is the image quality factor, derived from the image sharpness index. and noise level indicators The fusion was obtained. , To balance the weights, they were fitted using a subjective quality assessment dataset, ranging from 0.5 to 2.0.

[0086] It is an exponential parameter greater than 0 (ranging from 1.5 to 3.0), and when Q is low (poor image quality), It will decrease sharply, significantly lowering the confidence component of contour integrity. .

[0087] Through the above-mentioned refined matching degree evaluation based on spatial grid, this application can effectively distinguish between the loss of true information caused by physical occlusion and the loss of false information caused by image blurring and noise. This avoids giving a high-confidence misjudgment when the image quality is low, making the behavior of the entire perception enhancement system more intelligent and reasonable, and improving the ability to work continuously and stably in complex and ever-changing industrial environments.

[0088] This application further proposes that the step of generating a state uncertainty vector for quantifying sensing reliability based on the matching residual and feature stability calculation also includes a decision on the active sensing strategy, specifically including:

[0089] The system sets a low confidence threshold. (Range from 0 to 1, e.g., 0.5), the setting is based on extensive experimental statistics. When the confidence component of contour integrity is consistently below the preset low confidence threshold, the pose estimation error based on the current single view usually exceeds the system's tolerance range (e.g., position error > 5mm, attitude error > 3°), significantly increasing the risk of subsequent operations. "Consistently below" means when... Meanwhile, the system continuously monitors the contour completeness confidence component. The mean of the most recent N frames (range: 3-10 frames), when this mean is consistently below the low confidence threshold. If the current view information is insufficient, active exploration needs to be initiated.

[0090] Once triggered, the active perception strategy first constructs a reference set for optimizing the acquisition pose. Specifically, based on the imaging features reflected by the currently acquired local contour image sequence (such as the observed contour edge direction and texture region) and reference contour data (CAD model), a series of candidate acquisition poses are sampled and generated within the workspace accessible to the motion control device, and an optimized reference set for the acquisition pose used to improve the coverage of local contour information is calculated.

[0091] Based on the acquired pose optimization reference set, for each candidate pose, the system places the ideal geometric model under that pose and simulates rendering in the virtual camera model using ray casting or Z-buffer algorithms in computer graphics, calculating the visible model surface area or contour length from that viewpoint.

[0092] Predict the local contour visibility under different candidate acquisition poses by defining the visibility gain for each candidate pose i. The formula is as follows:

[0093]

[0094] in: It is the total model surface area visible from pose i;

[0095] It is the total surface area of ​​the model;

[0096] It is the area of ​​the model region that is visible from pose i and marked as low matching degree in the current view;

[0097] It is the total area of ​​the model regions marked as low matching degree in the current view;

[0098] and These are weighting coefficients (e.g., determined through simulation experiments). ∈[0.2,0.5], = =0.3, =0.7), emphasizing the priority of supplementing areas with missing information.

[0099] The system selects to increase visibility gain. The candidate pose with the largest (highest visibility) is used as the target acquisition pose.

[0100] After the motion control device is driven to the target acquisition pose, the acquisition device acquires a new local contour image sequence. Based on the target acquisition pose, the new local contour image sequence is acquired and combined with the initial local contour image sequence based on the estimated pose information. Contour features (point clouds or edges) extracted from different viewpoints are then stitched or fused in three-dimensional space to generate a contour reconstruction view.

[0101] Based on the reconstructed contour view, the system re-executes the feature extraction, registration, and uncertainty calculation processes to update the state uncertainty vector. Due to the increased amount of observation information and improved spatial coverage, the pose uncertainty component in the recalculated state uncertainty vector usually decreases, while the contour integrity confidence component usually increases.

[0102] By introducing an active perception, decision-making, and execution mechanism based on quantitative confidence assessment, this application can autonomously plan and execute the optimal observation action to make up for the deficiency when it perceives that it has insufficient information acquisition, thereby improving the task success rate and system robustness in scenarios with poor initial observation conditions (such as severe occlusion and poor initial viewing angle).

[0103] This application further proposes that, while generating and outputting contour-aware enhanced data, it also includes real-time monitoring and data re-evaluation steps, forming an independent monitoring loop, specifically including:

[0104] The system continuously acquires and processes new local contour image sequences. A new local contour image sequence refers to a series of frames that are captured in real time by the acquisition device according to the current motion state (which may follow an active sensing path or other task trajectory) immediately after the generation of the previous round of contour perception enhancement data.

[0105] Based on the new local contour image sequence, it enters a simplified processing pipeline at a slightly lower frequency than the main pipeline (e.g., the main pipeline processes one keyframe, while the monitoring loop processes 2-3 fast frames). This pipeline performs fast feature extraction and registration, and calculates a new state uncertainty vector based on the simplified model (e.g., using fixed feature descriptors and a fast solver), denoted as . .

[0106] Calculate the profile completeness confidence component in the new state uncertainty vector. The decrease in confidence level D between the historical contour integrity confidence component and the previously stored historical contour integrity confidence component, where the previously stored historical contour integrity confidence component typically refers to the confidence value used when the most recent stable contour-aware augmentation data was output. Or, the moving average of that value over a short time window. The magnitude of the decrease, D, is calculated as the relative rate of change, using the following formula:

[0107]

[0108] The system will compare the descent rate D with the preset emergency descent threshold. Comparison, Set within the range of 0.3 to 0.5 (for example, 0.4). If the decrease exceeds the preset emergency decrease threshold... The monitoring loop immediately sends an interruption or priority processing signal to the main process. Based on the new state uncertainty vector, updated predicted pose estimates and trend predictions are re-output. This includes: using the latest observations and significantly increased uncertainty inputs, re-driving the visual target dynamic model to quickly generate updated predicted pose estimates; simultaneously, the model recalculates the visual evolution trend based on deteriorating perception conditions, potentially predicting a more pessimistic range of motion or higher uncertainty growth. These updated predictions are immediately encapsulated as new contour perception enhancement data output, thereby replacing or supplementing previously inaccurate data.

[0109] By implementing parallel real-time monitoring and an emergency threshold-based reassessment mechanism, when unexpected and drastic changes occur in the environment (such as a new obstruction suddenly entering the field of view or a target workpiece being accidentally displaced by a collision), the system can quickly detect anomalies through a sharp drop in confidence and immediately initiate a recalculation process. This provides enhanced perception data reflecting the latest and worst-case scenarios in the shortest possible time (typically within several processing cycles). This effectively prevents robots and other actuators from making dangerous actions based on outdated or erroneous perception information, improving the safety and operational reliability of the entire automation system in dynamic, unstructured environments.

[0110] This application further proposes that after driving the pre-trained visual target dynamic model to generate predicted poses and predict the visual evolution trend of local contour image sequences, a trend self-verification step is also included. During the movement of the acquisition device along the active sensing path, the system continuously acquires new local contour image sequences at a preset frequency and uses these acquired images to perform real-time self-verification of the visual evolution trend.

[0111] During the generation and output of contour-aware augmented data, new local contour image sequences are continuously acquired at a preset frequency (e.g., 10Hz). These new local contour image sequences are then rapidly processed to extract a stable and easily trackable set of key contour points. This is achieved by selecting points with high response values ​​from the contour feature points identified in previous frames, points that can be quickly and reliably relocated in the current frame using optical flow or template matching.

[0112] Obtain visual evolution trends predicted by the visual target dynamic model The real-time location and visual evolution trend of the key contour point set. The trajectory tracking consistency error is calculated by comparing the expected position at the corresponding time point. The formula is as follows:

[0113]

[0114] Where N is the number of key points that were successfully matched;

[0115] and These are the current times. The observed and predicted positions of the i-th key point.

[0116] If the trajectory tracking consistency error exceeds the dynamic verification threshold for M consecutive sampling periods (where M is an integer greater than 1, or a fixed value determined experimentally, ranging from 2 to 5, for example, M=3), If the prediction of the current visual evolution trend is inaccurate, it is determined that the prediction may be inaccurate, for example, due to the target suddenly accelerating, turning, or interacting with an unexpected object, and a lightweight local re-prediction process is triggered.

[0117] Among them, dynamic verification threshold It is not a fixed value, but is dynamically calculated from the state uncertainty vector reflecting the current perceived quality. The calculation formula is as follows:

[0118]

[0119] in, It is the pose uncertainty component (scalar form) in the current state uncertainty vector;

[0120] It is a scaling factor (ranging from 10 to 20, for example, by fitting 100 sets of historical data, it represents mapping the normalized uncertainty to the pixel scale).

[0121] It is a basic tolerance pixel value ( (e.g., set to 2 pixels) to deal with unavoidable slight jitter and noise.

[0122] The local re-prediction process uses only the most recent few frames of images (e.g., the last 5 frames) and the current state (including the latest pose and increased uncertainty) to invoke a simpler, faster prediction model (such as a linear dynamic model or a shallow neural network) that is more focused on short-term predictions. This faster prediction model uses the latest short-term observation data to re-predict the short-term evolution trend over a small period of time in the future (e.g., 0.5 seconds).

[0123] Based on the latest several frames of image sequence, this newly generated short-term trend will be immediately used as a temporary substitute or linear interpolation correction for the upcoming part of the long-term trend output by the main model, thereby quickly correcting the predicted trajectory and realigning it with the observations.

[0124] This application improves the robustness and reliability of the visual target dynamic model in response to sudden changes in target motion patterns or dynamic environmental disturbances, ensuring that the predictions in contour perception enhancement data always maintain high practical value, thereby supporting safer and more accurate predictive control.

[0125] This application further proposes that, before generating contour-aware enhanced data, an uncertainty adaptive update step is included, specifically including:

[0126] After the system generates the predicted pose based on the dynamic model of the visual target, it continues to acquire subsequent frames from the local contour image sequence. Based on the registration results between the predicted pose and the newly acquired frames in the local contour image sequence, an update residual reflecting the prediction stability is calculated. This reflects the latest matching error between the actual observation and the model under the guidance of the predicted pose.

[0127] Based on the statistical properties of the updated residuals (e.g., their root mean square value) With the pose uncertainty component in the state uncertainty vector degree of deviation between The calculation formula is:

[0128]

[0129] in, It is a scaling factor. Used to normalize uncertainty Mapping to Comparable to actual pixels or millimeters.

[0130] Based on the degree of deviation The system dynamically adjusts the confidence weights of the state uncertainty vector. Specifically, the system maintains an internal weight. This is used to control the influence of the state uncertainty vector on subsequent modules (such as path planning and model tuning). Weights according to Update as follows: ,in , where is the attenuation coefficient. Degree of deviation. The smaller the value (the better the prediction error matches the expectation), the better. The closer to 1 (complete trust); the degree of deviation The larger, the better The lower.

[0131] When the deviation exceeds the preset adaptive adjustment threshold When the value is set between 0.3 and 0.6, such as 0.4, it indicates a significant systematic deviation between the prediction and the observation, which may mean that the observation information under the current path is insufficient to support high-quality prediction, or that the dynamic characteristics of the target have changed.

[0132] The local acquisition pose of the active sensing path is re-optimized based on the adjusted state uncertainty vector, thereby updating the input parameters used to generate contour sensing augmentation data. For example, on an ongoing active sensing path, if the uncertainty assessment of the next planned observation point is significantly increased, it can be immediately recalculated online. Under the premise of satisfying kinematic constraints, the pose of that point can be fine-tuned to focus more on reducing the uncertainty in a direction currently considered the least reliable, or to more directly target the region with the largest updated residual for verification observation.

[0133] By introducing an adaptive uncertainty update and online path optimization mechanism, this application not only generates a more conservative and accurate corrected uncertainty estimate, but also adjusts the active sensing strategy being implemented in real time, guiding the sensors to prioritize verifying or filling information gaps that lead to inaccurate predictions. This makes the entire system more adaptive and intelligent, maintaining effective sensing enhancement capabilities even under dynamic target changes or intensified environmental interference, thereby continuously outputting high-value contour sensing enhancement data and providing stable and reliable environmental sensing assurance for downstream precision operations.

[0134] The following is a specific embodiment of a method for enhancing visual perception of uncertainty based on local contours:

[0135] In a flexible assembly line for automotive parts, a six-axis robotic arm equipped with a 5-megapixel industrial camera at its end effector performs the task of gripping stacked aluminum alloy shell workpieces within a material box. Initially, the observation view could only capture the top right corner and one-third of the side profile of the workpiece. Due to insufficient local feature information, the traditional static template matching system repeatedly failed to grip due to pose estimation errors.

[0136] After system startup, the camera continuously acquires local contour image sequences at a sampling rate of 30Hz, and preprocesses them for denoising and distortion correction. By analyzing the current imaging quality (low average gradient magnitude, slight motion blur), SIFT is dynamically selected as the primary descriptor and ORB as the secondary descriptor: the primary descriptor extracts 32 key points on the workpiece edge, of which 8 feature points located in the blurred area have confidence levels below the threshold. The secondary descriptor supplements this by extracting 12 edge segment features from this area. The hybrid feature set is matched with a reference contour database (pre-stored multi-view SIFT / ORB feature library for this type of workpiece), and the six eigenvalues ​​of the matching residual covariance matrix are calculated to be 0.12, 0.08, 0.05, 0.03, 0.02, and 0.01. The pose uncertainty components are calculated using the formula: =0.7×(0.31 / 6)+0.3×0.12≈0.036+0.036=0.072 (normalized value); Simultaneously, the image is divided into an 8×8 grid, and the proportion of low-matching regions is statistically analyzed. =0.6 (17 out of 28 grids have a matching degree below the threshold), combined with the image quality factor Q=0.4 (weighted fusion of sharpness 0.3 and noise 0.5), the contour integrity confidence component is obtained through the formula: .

[0137] Because the confidence level of contour integrity is lower than the preset threshold of 0.9, an active perception strategy is triggered: based on the current pose estimation and the workpiece CAD model, 12 candidate observation poses are generated within the reachable space of the robotic arm, and the visibility gain of each pose is calculated through virtual rendering. The pose with the highest gain (observed from a 45° angle on the left side of the workpiece) is selected as the target acquisition pose, and an active sensing path is generated to drive the robotic arm to move.

[0138] The pose uncertainty vector (0.072) and the active sensing path are used as joint adjustment parameters and input into the pre-trained visual target dynamic model. Based on the observation data of the most recent 5 frames and the motion intention, the model predicts the pose change of the workpiece in the next 0.5 seconds (X-axis translation +2mm, Y-axis rotation -3°) and the visual evolution trend (the left edge will gradually become clearer). By fusing the predicted pose, evolution trend, uncertainty vector and active path planning information, contour perception enhancement data is generated: including the optimal grasping pose suggestion for the workpiece, the grasping success probability assessment (based on contour integrity confidence) and the warning prompt if the workpiece slides.

[0139] As the robotic arm moves along the active sensing path, the system monitors the visual evolution trend at a frequency of 10Hz. Self-verification is performed. When the trajectory tracking error exceeds the dynamic threshold calculated based on the current uncertainty for three consecutive cycles, the system triggers the local re-prediction model to quickly correct the short-term trend. At the same time, the system dynamically adjusts the weight of the state uncertainty vector according to the deviation between the registration residual of the newly acquired frame and the current uncertainty, and fine-tunes the subsequent acquisition pose.

[0140] Based on the augmented data, the robotic arm adjusts its end effector posture, moves to the actively sensing pose, and completes the left contour information of the workpiece. The uncertainty generation module recalculates the pose uncertainty component, reducing it to 0.035, and the contour integrity confidence level increases to 0.92. Ultimately, the robotic arm successfully grasps the workpiece and completes the assembly. The scenario of grasping failure caused by local contours is significantly reduced, improving the automation rate and stability of the production line. This solves the reliability assessment problem of traditional vision systems under local dynamic observation and provides robust perception support for high-precision operation of flexible assembly lines.

[0141] Example 2:

[0142] like Figure 3As shown, the uncertainty-based visual perception enhancement system based on local contours, using the aforementioned uncertainty-based visual perception enhancement method based on local contours, includes at least: one or more acquisition devices (such as industrial cameras, structured light sensors, or laser profilometers), a motion control device (such as a multi-axis robotic arm, linear module, or rotary platform) for driving the acquisition devices or the target workpiece itself, and a data processing device (such as an industrial computer, embedded processor, or programmable logic controller). The devices communicate with each other via a data bus or industrial network. The system includes physical entity components and a data processing core. Among them, it includes:

[0143] The image acquisition module is used to drive the acquisition device through an image acquisition card or gigabit network port, control its acquisition parameters (such as exposure and gain), and receive, buffer, and preprocess (such as noise reduction and distortion correction) image streams to obtain local contour image sequences of the target workpiece in real time for subsequent processing.

[0144] The uncertainty generation module performs operations such as dynamic feature selection, hybrid feature extraction, real-time registration with reference contour data, and local reliability map construction. It is used for visual processing of local contour image sequences. Based on image feature extraction and visual registration algorithms, it extracts target contour features and performs real-time matching with pre-stored reference contour data. Based on the matching residuals and feature stability calculations, it generates a state uncertainty vector for quantifying perceived reliability.

[0145] The path awareness module is used to plan the subsequent motion path of the acquisition device online based on the state uncertainty vector and combined with the kinematic model and workspace constraints of the motion control device. It executes online path planning algorithms (such as optimization algorithms based on information gain or uncertainty gradient) to generate an active sensing path aimed at optimizing the integrity of contour information, and converts the path points into motion control commands and sends them to the motion control device.

[0146] The modeling and prediction module loads and runs a pre-trained visual target dynamic model. It uses the state uncertainty vector and the active perception path as joint adjustment parameters to regulate the model's internal state or noise parameters, driving the pre-trained visual target dynamic model to generate predicted poses and predict the visual evolution trend of local contour image sequences.

[0147] The output module receives all output information from the uncertainty generation module, path perception module, and modeling and prediction module. Based on the predicted pose, visual evolution trend, state uncertainty vector, and planning information of the actively perceived path, it processes the data according to preset fusion logic (such as weighted averaging, Bayesian update, and rule synthesis) to generate and output contour perception enhanced data. This data is then sent to downstream robot controllers or monitoring systems via industrial Ethernet or fieldbus for guiding grasping, assembly, and other operations, or for visual monitoring.

[0148] In an application scenario where a robot randomly grasps stacked workpieces in a bin, after system initialization, the robot's end effector, carrying a camera, moves above the bin. The image acquisition module controls the camera to capture a sequence of local contour images (possibly only showing a top corner and part of the side of the workpiece). The uncertainty generation module extracts features and matches them with a reference database for this type of workpiece, making a preliminary pose estimate and simultaneously calculating a high-value state uncertainty vector (due to insufficient information). The path perception module analyzes this vector and plans an active perception path: controlling the robot to move around the side of the bin with the camera to observe the workpiece from a new angle. During the movement, the modeling and prediction module predicts the possible positional change trend of the workpiece based on the initial observation and the motion intention. When the camera reaches the new pose and acquires a new image, the uncertainty generation module fuses the old and new view information, recalculates, and obtains a significantly reduced state uncertainty vector and a more accurate predicted pose. Finally, the output module fuses all the information into contour perception enhanced data and sends it to the robot grasping planner. This data includes not only the estimated position of the workpiece, but also suggestions for the grasping posture, an assessment of the probability of successful grasping (based on the confidence level of contour integrity), and a warning that the workpiece may slide before grasping (based on visual evolution trends), thereby guiding the robot to successfully perform a grasping operation with a high success rate.

[0149] This application not only enables the execution of a complete automated processing flow from local images to augmented data, but also outputs contour-aware augmented data that integrates optimal estimation, uncertainty range, future prediction, and action suggestions. This allows downstream actuators or decision-making systems to make safer, smarter, and more accurate operations, thereby improving the level of automation while effectively reducing the risk of operational failures or equipment damage caused by inaccurate perception.

[0150] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for enhancing visual perception of uncertainty based on local contours, characterized in that: Includes the following steps: Real-time acquisition of local contour image sequences of the target workpiece; The local contour image sequence is visually processed. Based on image feature extraction and visual registration algorithms, the target contour features are extracted and registered and matched with the pre-stored reference contour data in real time. Based on the matching residual and feature stability calculation, a state uncertainty vector for quantifying the reliability of perception is generated. Based on the state uncertainty vector, the subsequent motion path of the acquisition device is planned online to generate an active sensing path aimed at optimizing the integrity of contour information; Using the state uncertainty vector and the active perception path as joint adjustment parameters, the pre-trained visual target dynamic model is driven to generate a predicted pose and predict the visual evolution trend of the local contour image sequence. Based on the predicted pose, the visual evolution trend, the state uncertainty vector, and the planning information of the active perception path, contour perception enhancement data is generated and output. The state uncertainty vector is used to quantify perceived reliability and includes at least a pose uncertainty component and a contour integrity confidence component. The pose uncertainty component is used to characterize the reliability of the target workpiece pose estimation; The contour integrity confidence component is used to characterize the integrity and reliability of the currently observed contour information.

2. The method for enhancing visual perception of uncertainty based on local contours according to claim 1, characterized in that: Before acquiring the local contour image sequence of the target workpiece in real time, the following steps are also included: Receive a task instruction containing the target workpiece model identifier; Based on the model identifier, retrieve and load the corresponding reference contour data from the visual feature database; The reference contour data includes a standard local contour visual feature set and a corresponding ideal geometric model.

3. The method for enhancing visual perception of uncertainty based on local contours according to claim 1, characterized in that: The steps of extracting target contour features and performing real-time registration and comparison with pre-stored reference contour data specifically include: Based on the imaging quality of the local contour image sequence, a primary descriptor and an auxiliary descriptor are dynamically selected from multiple preset contour feature descriptors; The main descriptor is used to extract contour features from the current frame image, and the auxiliary descriptor is used to extract supplementary features from local areas with insufficient confidence in the main descriptor, forming a hybrid visual feature set. The hybrid visual feature set is registered with the reference contour data; A local visual registration reliability map is constructed based on the registration process, and the corresponding statistical information is used as an additional weighting factor to generate the state uncertainty vector.

4. The method for enhancing visual perception of uncertainty based on local contours according to claim 3, characterized in that: The steps for generating the pose uncertainty components include: The visual residual of the target workpiece relative to the reference pose is calculated based on the visual registration results; the visual residual includes translation residual and rotation residual. Covariance analysis was performed on the visual residuals to obtain the error distribution in each degree of freedom in three-dimensional space; Based on the error distribution, a statistic characterizing the dispersion is calculated and normalized to obtain the pose uncertainty component.

5. The method for enhancing visual perception of uncertainty based on local contours according to claim 4, characterized in that: The steps for generating the contour integrity confidence component include: The local contour image sequence is segmented into multiple local regions; For each local region, calculate the matching score between the extracted contour features and the corresponding region features in the reference contour data; Calculate the proportion of regions in all local regions whose matching score is lower than the preset matching threshold; Based on the region ratio, the contour integrity confidence component is calculated by fusing it with the acquired image sharpness index and noise level index and using a preset confidence mapping function.

6. The method for enhancing visual perception of uncertainty based on local contours according to claim 5, characterized in that: The step of generating a state uncertainty vector for quantifying sensing reliability based on matching residuals and feature stability calculations also includes decisions on the proactive sensing strategy, specifically including: When the confidence component of the contour integrity is continuously lower than the preset low confidence threshold, an optimized reference set for acquisition pose is calculated based on the imaging features of the local contour image sequence and the reference contour data to improve the coverage of local contour information. Based on the acquisition pose optimization reference set, predict the local contour visibility under different candidate acquisition poses, and select the candidate pose with the highest visibility as the target acquisition pose. A new local contour image sequence based on the target acquisition pose is obtained, and then stitched or fused with the initial local contour image sequence to generate a contour reconstruction view. The state uncertainty vector is updated based on the contour reconstruction view.

7. The method for enhancing visual perception of uncertainty based on local contours according to claim 1, characterized in that: The process of generating and outputting contour-aware enhanced data also includes real-time monitoring and data reassessment steps, specifically including: Acquire and process a new local contour image sequence, and calculate and generate a new state uncertainty vector; Calculate the decrease magnitude between the contour integrity confidence component in the new state uncertainty vector and the pre-stored historical contour integrity confidence component; If the decrease exceeds the preset emergency decrease threshold, then based on the new state uncertainty vector, the updated predicted pose estimate and trend prediction are re-output.

8. The method for enhancing visual perception of uncertainty based on local contours according to claim 1, characterized in that: After driving the pre-trained visual target dynamic model to generate predicted poses and predict the visual evolution trend of the local contour image sequence, a trend self-verification step is also included, specifically including: During the process of generating and outputting contour-aware enhanced data, new local contour image sequences are continuously acquired at a preset frequency. The new local contour image sequence is processed quickly to extract the key contour point set; The real-time position of the key contour point set is compared with the expected position of the same key point at the corresponding time point in the visual evolution trend, and the trajectory tracking consistency error is calculated. If the trajectory tracking consistency error exceeds the dynamic verification threshold within M consecutive sampling periods, a local re-prediction process is triggered. The dynamic verification threshold is calculated from the state uncertainty vector.

9. The method for enhancing visual perception of uncertainty based on local contours according to claim 1, characterized in that: Before generating the contour-aware enhanced data, an uncertainty adaptive update step is also included, specifically including: Based on the registration results between the predicted pose and the newly acquired frames in the local contour image sequence, the updated residual is calculated; The confidence weight of the state uncertainty vector is dynamically adjusted based on the degree of deviation between the updated residual and the pose uncertainty component in the state uncertainty vector. When the deviation exceeds a preset adaptive adjustment threshold, the local acquisition pose of the active sensing path is re-optimized based on the adjusted state uncertainty vector.

10. A visual perception enhancement system for uncertainty based on local contours, characterized in that: Using the uncertainty-based visual perception enhancement method based on any one of claims 1 to 9, comprising: The image acquisition module is used to acquire local contour image sequences of the target workpiece in real time. The uncertainty generation module is used to perform visual processing on the local contour image sequence. Based on image feature extraction and visual registration algorithms, it extracts target contour features and performs real-time registration and matching with pre-stored reference contour data. Based on the matching residual and feature stability calculation, it generates a state uncertainty vector for quantifying perceived reliability. The path perception module is used to plan the subsequent motion path of the acquisition device online based on the state uncertainty vector, and generate an active perception path aimed at optimizing the integrity of the contour information. The modeling and prediction module is used to drive a pre-trained visual target dynamic model to generate a predicted pose and predict the visual evolution trend of the local contour image sequence by using the state uncertainty vector and the active perception path as joint adjustment parameters. The output module is used to generate and output contour perception enhancement data based on the predicted pose, the visual evolution trend, the state uncertainty vector, and the planning information of the active perception path.

Images