Intelligent data acquisition method and system based on AI multi-modal fusion RFID
By using a multimodal fusion of RFID and machine vision, camera extrinsic drift is corrected in real time, solving the problem of misalignment between workpiece identification and process parameters in mixed-flow assembly lines, and achieving stable acquisition and correct distribution of workpiece information.
Patent Information
- Application Number
- CN202611120385.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-11-10
AI Technical Summary
In mixed-flow assembly lines, when RFID is combined with machine vision, camera extrinsic parameter drift caused by conveyor belt vibration, workpiece posture changes, and interference from the metal environment makes it impossible to stably align the 3D coordinates retrieved by RFID with the visual corner point heat map, resulting in misalignment or loss of workpiece identification and process parameter binding.
By synchronously acquiring complex domain echo signal matrices and image corner heatmaps, tag positions are inverted using dynamic phase gradient fields. Combined with unscented transformations and confidence propagation of non-Gaussian state space and manifold tangent space, camera extrinsic drift is corrected in real time, generating a process parameter overlay layer bound to RFID identification.
This achieves continuous and stable binding of RFID identification tags to visual inspection areas during workpiece movement, reducing assembly errors, lowering production line maintenance costs, and ensuring the correct issuance of differentiated process parameters.
Smart Images

Figure CN123070152A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial Internet of Things and multi-source information fusion technology, specifically to an intelligent data acquisition method and system based on AI multimodal fusion RFID. Background Technology
[0002] In mixed-flow assembly lines, smart warehousing, and industrial automation, it is often necessary to simultaneously acquire the workpiece's identification and appearance or location information to execute differentiated work instructions. Radio Frequency Identification (RFID) technology reads electronic tags attached to workpieces non-contactly, providing a unique identification code, but it struggles to obtain the workpiece's precise spatial location or contour features. Machine vision technology can extract geometric information such as the workpiece's shape and corners through image acquisition, but it cannot directly identify the workpiece's internal model or process parameters. Therefore, multimodal data acquisition methods combining RFID and vision are gaining increasing attention.
[0003] During mixed-flow assembly production, the continuous drift of camera extrinsic parameters caused by factors such as random vibration of the conveyor belt, changes in workpiece posture, and multipath interference in the metal environment causes the pre-calibrated homography projection relationship to fail. As a result, it is impossible to continuously and stably align the 3D coordinates retrieved by RFID with the visual corner point heat map at sub-pixel accuracy. This leads to the problem of misalignment or loss of the identification and process parameter binding of workpieces with the same appearance but different internal models. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent data acquisition method and system for AI multimodal RFID to solve the problems mentioned above.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] The AI-based multimodal RFID-integrated intelligent data acquisition method includes the following steps: S1: During the workpiece transfer process, a complex domain echo signal matrix and an image corner heat map are simultaneously acquired. The complex domain echo signal matrix is composed of the IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by a high-speed camera. S2: Solve the phase interference fringes of each tag based on the complex domain echo signal matrix, generate a dynamic phase gradient field through multidimensional phase unwrapping, and use the dynamic phase gradient field to invert the instantaneous position coordinates of each tag in three-dimensional space. S3: Input the instantaneous position coordinates and the image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic parameter drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. S4: Based on the predicted projected residual flow, confidence propagation with constraint enhancement is performed in the tangent space of the manifold. The update increment is regressed to the special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field. S5: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field to generate a process parameter overlay layer bound with RFID identification, and output as differentiated operation instructions for subsequent execution devices.
[0007] As a further aspect of the present invention: the step of using a dynamic phase gradient field to invert the instantaneous position coordinates of each label in three-dimensional space specifically includes: Instantaneous phase extraction is performed on the complex domain echo signal matrix along the antenna array direction, and the phase difference between adjacent antennas is calculated to generate an initial interference fringe pattern; Multidimensional phase unwinding constructs an optimal path graph based on phase gradient confidence and uses a minimum discontinuity pruning algorithm to eliminate phase jumps, outputting a continuously distributed dynamic phase gradient field. The inversion of instantaneous position coordinates further includes: performing line integration of the dynamic phase gradient field along the three spatial axes, and iteratively solving for the three-dimensional coordinates of each tag by combining it with the pre-calibrated antenna spatial coordinates.
[0008] As a further aspect of the present invention: the continuously distributed dynamic phase gradient field output specifically includes: Calculate the phase gradient magnitude and neighborhood orientation entropy of each pixel in the initial interference fringe pattern, and mark the pixels with small magnitude and low orientation entropy as high confidence seed points; Starting from all seed points, the connection path is expanded along the gradient descent direction. A closed path graph is formed when the confidence difference between the two ends of the path is lower than the threshold. The minimum discontinuity pruning algorithm is as follows: Identify phase transition edges along the optimal path graph, calculate the continuous phase difference on both sides of each transition edge, prune the transition edges that result in the largest cumulative discontinuity, and fill the pruned region with radial basis interpolation of adjacent continuous phases to output a dynamic phase gradient field.
[0009] As a further aspect of the present invention: obtaining the predicted projection residual flow specifically includes: After aligning the instantaneous position coordinates with the timestamps of the image corner heatmap, a state augmentation vector for the non-Gaussian state space is constructed, which includes position components and corner residual components. Apply asymmetric scaling to the state augmentation vector to generate an initial finite set of granular points, and calculate the prediction projection Jacobian matrix for each granular point; The time-varying noise covariance is estimated based on the instantaneous vibration acceleration of the conveyor belt. The scale parameters of the particle set are dynamically adjusted. The projected residuals of the adjusted particle set are weighted and statistically analyzed to output the predicted projected residual flow.
[0010] As a further aspect of the present invention: the process of obtaining the predicted projection Jacobian matrix is as follows: Calculate the local skewness coefficients of each dimension of the state augmentation vector, and generate positive and negative scaling factors according to the sign of the local skewness coefficients. Using the current estimate of the state augmentation vector as the center, Gaussian perturbations of unequal amplitudes are applied to the positive and negative deviation directions according to the positive and negative scaling factors, respectively, to obtain the initial finite particle set; For each particle in the initial finite particle set, the partial derivatives of the particle with respect to each dimension of the state are calculated using the numerical difference method, and then assembled into a prediction projection Jacobian matrix.
[0011] As a further aspect of the present invention: obtaining the continuously dynamically corrected homography mapping field specifically includes: After segmenting the predicted projection residual flow along the time axis, it is mapped to the tangent space of the special Euclidean group Lie algebra to obtain the residual vector field in the tangent space; In the tangent space, a manifold curvature regularization constraint is applied to the residual vector field, and the posterior probability of each residual vector is iteratively updated using the belief propagation algorithm. After each iteration, the update result is regressed to the tangent space manifold through the projection operator. An exponential mapping is performed on the converged tangent space increment vector to transform it into a rigid body transformation matrix on a special Euclidean group. The rigid body transformation matrix is accumulated frame by frame and then the continuously dynamically corrected homography field is output.
[0012] As a further aspect of the present invention: the process of composing the residual vector field is as follows: Temporal differencing is performed on the predicted projected residual flow within each time axis segment to obtain the instantaneous rate of change tensor at each time point; The rotational and translational components are decomposed from the instantaneous rate of change tensor. The rotational components are converted into a three-dimensional antisymmetric matrix form, and the three independent elements of the antisymmetric matrix are extracted as rotation vectors. The rotation vector and translation components are arranged sequentially and spliced into a six-dimensional Lie algebra vector, which forms a residual vector field in the tangent space in time order.
[0013] As a further aspect of the present invention: the differentiated operation instructions specifically include: Based on the mapping coefficients at the corresponding time in the homography mapping field, the instantaneous position coordinates of each tag are transformed into sub-pixel coordinates on the pixel plane; Centered on sub-pixel coordinates, extract the gray-level gradient weighted values in the neighborhood, and use the gray-level gradient weighted values to perform local calibration on the mapping position to eliminate discretization error; The calibrated pixel coordinates are bound to the RFID identification of the corresponding tag and the pre-stored process parameters. An overlay layer containing identity-parameter pairs is generated by superimposing them according to the time sequence and encoded as a differentiated operation instruction.
[0014] As a further aspect of the present invention: the local calibration of the mapping position using grayscale gradient weighting specifically includes: Using sub-pixel coordinates as the center, calculate the gray-level gradient magnitude and gradient direction angle of each pixel in a small neighborhood around it, and use the gradient magnitude as the weight to construct a weighted centroid offset vector. The weighted centroid offset vector is normalized according to the distribution density of gradient direction angles in the neighborhood, and decomposed along the row and column directions of the image to obtain the row calibration amount and the column calibration amount. The row and column calibration values are multiplied by a preset relaxation factor and then superimposed onto the original subpixel coordinates. This process is iterated until the calibration value is less than half a pixel spacing, and then the calibrated subpixel coordinates are output.
[0015] This invention also provides an AI multimodal RFID-integrated intelligent data acquisition system, comprising: Multi-source synchronous acquisition module: During the workpiece transfer process, it synchronously acquires complex domain echo signal matrix and image corner heat map. The complex domain echo signal matrix is composed of IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by high-speed camera. Phase unwinding and 3D positioning module: The phase interference fringes of each tag are calculated based on the complex domain echo signal matrix. A dynamic phase gradient field is generated through multi-dimensional phase unwinding, and the instantaneous position coordinates of each tag in 3D space are inverted using the dynamic phase gradient field. Non-Gaussian state space prediction module: Input the instantaneous position coordinates and image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. Manifold constraint belief propagation module: Based on the predicted projected residual flow, it performs constraint-enhanced belief propagation in the manifold tangent space, and regresses the update increment to a special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field; Projection mapping and instruction generation module: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field, generating a process parameter overlay layer bound with RFID identification, and outputting differentiated operation instructions for subsequent execution devices.
[0016] The beneficial effects of this invention are: (1) By fusing the three-dimensional coordinates of the RFID phase difference inversion with the image corner heatmap in the non-Gaussian state space, and using unscented transformation prediction and belief propagation iterative correction, the influence of dynamic factors such as conveyor belt vibration, metal reflection interference and visual partial occlusion on data association can be effectively overcome, so that the RFID identification mark and visual detection area of the same workpiece maintain a continuous and stable binding relationship during the movement process, reducing assembly errors caused by identity jump or loss.
[0017] (2) By utilizing the manifold curvature regularization constraint of the predicted projection residual flow in the special Euclidean group Lie algebra tangent space, combined with the time-varying noise covariance estimated in real time by the accelerometer, online adaptive correction of camera extrinsic drift is realized. There is no need for frequent machine stoppages for manual offline calibration, thereby reducing the maintenance cost of the production line. At the same time, it ensures that the different process parameters such as torque value and tightening sequence corresponding to different types of workpieces in mixed flow assembly can be correctly sent to the execution device. Attached Figure Description
[0018] The invention will now be further described with reference to the accompanying drawings.
[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the process of outputting a continuously distributed dynamic phase gradient field in this invention; Figure 3 This is a flowchart of the process for obtaining the predicted projection residual flow in this invention; Figure 4 This is a system block diagram of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Please see Figure 1 As shown, this invention is an intelligent data acquisition method for AI multimodal fusion RFID, comprising the following steps: S1: During the workpiece transfer process, a complex domain echo signal matrix and an image corner heat map are simultaneously acquired. The complex domain echo signal matrix is composed of the IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by a high-speed camera. S2: Solve the phase interference fringes of each tag based on the complex domain echo signal matrix, generate a dynamic phase gradient field through multidimensional phase unwrapping, and use the dynamic phase gradient field to invert the instantaneous position coordinates of each tag in three-dimensional space. S3: Input the instantaneous position coordinates and the image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic parameter drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. S4: Based on the predicted projected residual flow, confidence propagation with constraint enhancement is performed in the tangent space of the manifold. The update increment is regressed to the special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field. S5: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field to generate a process parameter overlay layer bound with RFID identification, and output as differentiated operation instructions for subsequent execution devices.
[0022] In S1, a complex domain echo signal matrix and an image corner heatmap are simultaneously acquired during the workpiece transfer process. The complex domain echo signal matrix consists of the IQ components reflected from RFID tags received by multiple antennas, and the image corner heatmap consists of a set of sub-pixel coordinates of the workpiece contour acquired by a high-speed camera, specifically including: At the start of the mixed-flow assembly conveyor or at the workstation entrance, a linear or planar array consisting of at least four UHF RFID antennas is deployed, each antenna connected to a multi-channel RFID reader. When a workpiece with an RFID tag enters the coverage area of the antenna array along the conveyor belt, the reader excites each antenna in a cyclical sequence. During excitation, each antenna receives the radio frequency signal reflected from the tag and demodulates the voltage amplitudes of the in-phase and quadrature components, denoted as I and Q values, respectively. For a single workpiece tag, each antenna acquires a pair of I and Q values simultaneously. The I and Q values of all antennas are arranged in order of antenna number, forming a primitive complex array of 2 rows multiplied by the number of antennas. This array is denoted as the complex domain echo signal matrix. When multiple workpieces are present simultaneously, the reader separates the responses of different tags using time-division multiple access or frequency hopping mechanisms, generating a corresponding complex domain echo signal matrix independently for each tag.
[0023] Simultaneously, a high-speed industrial camera is fixed on the same physical mounting bracket as the RFID antenna array. The camera's optical axis forms an angle of 15 to 30 degrees with the normal direction of the conveyor belt plane to ensure clear acquisition of the contour features of the workpiece's upper surface. The camera continuously captures images at a constant frame rate, triggered by a hardware synchronization pulse signal shared with the RFID reader, ensuring that the deviation between the image acquisition time and the RFID data sampling time for each frame does not exceed 1 millisecond. After acquiring the raw image, it is first converted to grayscale. Then, a corner detection operator based on sub-pixel precision, such as the improved Forstner operator or the Harris-Plessey operator, is used to extract the geometric corner coordinates of the workpiece from the image. The coordinate value of each corner point is recorded as a floating-point number at the pixel row and column position. The coordinates of all corner points detected at the same time and their corresponding grayscale gradient amplitudes are normalized to form a two-dimensional heatmap structure. Each pixel position in the heatmap stores the corner response intensity at that location, while the response intensity at non-corner positions is set to zero. This heatmap is the image corner heatmap. The acquired complex domain echo signal matrix and the image corner heatmap are associated with a shared timestamp tag and used as input data pairs for subsequent processing.
[0024] Please see Figure 2 As shown, in S2, the phase interference fringes of each tag are calculated based on the complex domain echo signal matrix. A dynamic phase gradient field is generated through multidimensional phase unwrapping, and the instantaneous position coordinates of each tag in three-dimensional space are inverted using the dynamic phase gradient field. Specifically, this includes: For the obtained complex domain echo signal matrix, instantaneous phase extraction is performed on the complex data received by each antenna along the antenna array arrangement direction. The extraction method is as follows: calculate the ratio of the in-phase component to the quadrature component of each antenna, and perform an arctangent operation on this ratio to obtain the principal phase value, which is limited to the range between negative and positive pi. The difference between the principal phase values of adjacent antennas is calculated sequentially. If the absolute value of the difference is greater than pi, it is adjusted to the range between negative and positive pi by adding or subtracting twice the value of pi, thus forming the phase difference between adjacent antennas. The calculated phase differences of all adjacent antenna pairs are arranged in order of antenna position, and each difference is assigned a corresponding pixel coordinate, where the horizontal coordinate of the pixel corresponds to the position number of the antenna pair, and the vertical coordinate corresponds to the sampling time number, thereby constructing an initial interference fringe pattern.
[0025] Multidimensional phase unwrapping is performed on the initial interference fringe pattern. An optimal path map based on phase gradient confidence is constructed: For each pixel in the initial interference fringe pattern, the phase difference between that pixel and its adjacent pixels in the horizontal and vertical directions is calculated, and the root mean square of the two directional differences is taken as the phase gradient magnitude. A 3x3 pixel neighborhood window is taken centered on that pixel, and a histogram of the phase value distribution of all pixels within the window is plotted. The histogram is divided into 16 equally spaced intervals, and the proportion of pixels in each interval to the total number of pixels is calculated. Each proportion is multiplied by its base-2 logarithm, summed, and the negative value is taken to obtain the neighborhood direction entropy. Pixels with a phase gradient magnitude less than 30% of the average magnitude of the entire image and a neighborhood direction entropy less than 1.2 bits are marked as high-confidence seed points. Then, starting from all seed points, the connection path is expanded pixel by pixel along the direction of the fastest phase gradient descent (i.e., the direction of the smallest phase gradient magnitude among adjacent pixels). When the absolute value of the confidence difference between the starting and ending seed points of a path is less than 0.1, the path is closed to form the optimal path graph. Next, the minimum discontinuity pruning algorithm is executed: traversing all adjacent pixel pairs along the optimal path graph, identifying edges where the phase value jump exceeds the magnitude of pi, and recording each jump edge; for each jump edge, calculating the difference in average phase values in the continuous regions on both sides, and using this difference as the jump magnitude of the edge; accumulating the sum of the jump magnitudes of all jump edges as the current discontinuity total value; pruning the edge with the largest jump magnitude, and recalculating the discontinuity total value; repeating the above pruning operation until the discontinuity total value no longer decreases. For the region enclosed by the clipped edges, radial basis interpolation is used for filling. Specifically, all un-clipped continuous phase values on the boundary of the region are used as known points. The continuous phase value of each pixel in the region is calculated by interpolation using the radial basis function (whose shape parameter is 0.5 times the average distance between boundary points). Finally, a dynamic phase gradient field in which the phase values of all pixels are continuously distributed is output.
[0026] The instantaneous 3D spatial coordinates of each tag are retrieved using a dynamic phase gradient field. Specifically, the 3D spatial coordinates of each antenna are pre-calibrated using a laser tracker. For each row in the dynamic phase gradient field (corresponding to a sampling moment), the phase gradient value is linearly integrated along the horizontal axis (corresponding to the antenna array's arrangement direction), with the integration starting at the position of the first antenna. The integration result gives the cumulative amount of phase change with antenna position. This cumulative amount curve is projected along three perpendicular directions in space (conveyor belt travel direction, conveyor belt lateral direction, and vertical height direction), with the projection relationship in each direction determined by the antenna's spatial coordinates. A system of equations is established: the cumulative phase received by each antenna is equal to the distance from the tag to that antenna multiplied by the reciprocal of twice the wavelength (after modulo). This system of equations is solved using an iterative least squares method, with the initial values set to the 3D coordinates of the previous sampling moment. The number of iterations is set to 5, and the distance calculation is updated after each iteration. Finally, the system converges to obtain the instantaneous 3D spatial coordinates of each tag at the current sampling moment.
[0027] Please see Figure 3 As shown, in S3, the instantaneous position coordinates and the image corner heatmap are input into the non-Gaussian state space. In the non-Gaussian state space, an unscented transformation is performed to predict the camera extrinsic parameter drift caused by the conveyor belt motion, resulting in the predicted projection residual flow. Specifically, this includes: After obtaining the instantaneous three-dimensional spatial coordinates of each label and the image corner heatmap at the same moment, these two types of data need to be fused in a non-Gaussian state space to cope with the continuous drift of camera extrinsic parameters caused by conveyor belt vibration.
[0028] Align the timestamp of the instantaneous location coordinates with the timestamp of the image corner heatmap. The alignment accuracy should be less than 0.5 milliseconds. If the time difference exceeds this value, linear interpolation should be used to fill in the gap.
[0029] After alignment, construct a state augmentation vector, which is composed of two concatenated parts: The first part consists of the three-dimensional instantaneous position coordinates of all tags at the current moment, arranged in order of tag number, forming a position component of 3 times the total number of tags; The second part consists of the residual coordinates of all detected corner points in the image corner heatmap, which is the difference between the actual image coordinates of each corner point and the projected coordinates predicted based on the extrinsic parameters of the previous frame, arranged in order of corner point number. The two parts are concatenated to form a one-dimensional column vector, denoted as the state augmentation vector.
[0030] Asymmetric scaling is applied to the aforementioned state augmentation vector to generate an initial finite granular point set. Specifically: The local skewness coefficient is calculated for each dimension of the augmented state vector. The skewness coefficient is calculated as follows: A sample sequence is formed by taking the values of the 10 most recent sampling times for that dimension. The third central moment of the sample sequence is divided by the cube of the sample standard deviation; the resulting ratio is the local skewness coefficient. Positive and negative scaling factors are determined based on the sign of the local skewness coefficient: when the skewness coefficient is positive, the positive scaling factor is 0.8 and the negative scaling factor is 1.2; when the skewness coefficient is negative, the positive scaling factor is 1.2 and the negative scaling factor is 0.8; when the skewness coefficient is zero, both are 1.0. Then, using the current estimated value of the augmented state vector as the center point, a Gaussian perturbation is applied to each dimension: for the positive deviation direction, the standard deviation of the perturbation is equal to the square root of the historical variance of that dimension multiplied by the positive scaling factor; for the negative deviation direction, the standard deviation of the perturbation is equal to the square root of the historical variance multiplied by the negative scaling factor. A perturbation is generated in each dimension, and the perturbations in all dimensions are combined to form an offset vector of a particle relative to the center point. The above process is repeated to generate a total of 2 times the number of dimensions of the state augmentation vector, with each particle corresponding to an offset vector. The center point itself is also considered a particle, thus forming an initial finite set of particles.
[0031] For each point in the initial finite set of points, calculate the predicted projection Jacobian matrix. The calculation method uses the numerical difference method: for the state value represented by the current point, its i-th... Each dimension's value is incremented by a tiny increment, which is one-thousandth of the historical value range for that dimension. Then, the residual vector between the predicted projected coordinates of all labels at that point and the actual corner coordinates in the image corner heatmap is recalculated. The change in the residual vector is divided by the tiny increment to obtain an approximate value of the partial derivative of the residual vector with respect to the i-th dimension. After traversing all dimensions, the partial derivatives of each dimension are arranged column-wise to form a matrix with (2 times the number of corners) rows and the dimension of the state augmentation vector. This matrix is the prediction-projection Jacobian matrix. This matrix describes the linear mapping relationship between tiny changes in state and the projection residuals.
[0032] The time-varying noise covariance is estimated based on the instantaneous vibration acceleration of the conveyor belt. A triaxial accelerometer is fixed near the camera mounting bracket on the conveyor belt, and vibration acceleration values in three directions are collected in real time at a sampling rate of 1000 times per second. The root mean square (RMS) values of the acceleration values at the most recent 10 sampling moments are calculated for each of the three directions. The RMS value of each direction is multiplied by 0.05 to obtain the standard deviation of the process noise in that direction. The squares of the standard deviations of the process noise in the three directions are used as the diagonal elements to construct a 3x3 diagonal matrix. This matrix is then transformed to fit the dimensions of the state augmented vector through coordinate transformation to obtain the time-varying noise covariance matrix.
[0033] The scale parameter of the particle set is dynamically adjusted: the scale parameter of each particle is initially set to 1.0, and then updated according to the relationship between the predicted projection Jacobian matrix and the time-varying noise covariance matrix corresponding to that particle. The update rule is as follows: calculate the product of the predicted projection Jacobian matrix and its transpose, and then multiply it by the current scale parameter of the particle to obtain a temporary matrix; calculate the trace of this temporary matrix and compare it with the trace of the time-varying noise covariance matrix; if the former is greater than the latter, the scale parameter is multiplied by 0.9, otherwise it is multiplied by 1.1, and the scale parameter is restricted to between 0.2 and 5.0.
[0034] The projected residuals of the adjusted particle set are weighted and statistically analyzed. The projected residual of each particle is calculated using the following formula: ;in, Indicates the first Each particle at time The projection residual vector, For a moment The observation vector is constructed from the actual corner coordinates extracted from the image corner heatmap. For the first The state augmentation vector of each particle. The observation function maps the state augmentation vector to the predicted projected coordinates. This function consists of the camera pinhole model and the current extrinsic parameter estimates. The weight of each particle is determined based on its projected residual and time-varying noise covariance matrix: First, the squared Mahalanobis distance of the projected residual vector is calculated, which is the transpose of the projected residual vector multiplied by the inverse of the time-varying noise covariance matrix, then multiplied by the projected residual vector. The negative half of this squared value is then taken as the exponent, multiplied by a normalization constant, to obtain the unnormalized weight of that particle. The unnormalized weights of all particles are summed, and the weight of each particle is divided by this sum to obtain the normalized weight. Finally, a weighted statistical analysis is performed on the normalized particle set: the weighted average of the projection residuals of all particles is calculated, i.e., the projection residual vector of each particle is multiplied by its normalization weight and then summed to obtain the weighted average residual vector; simultaneously, the weighted covariance matrix is calculated, i.e., the projection residual vector of each particle is subtracted from the weighted average residual vector, multiplied by its transpose, and then multiplied by its normalization weight and summed. These weighted average residual vectors are arranged in chronological order, with each time point corresponding to one residual vector. The sequence of residual vectors from all time points constitutes the predicted projection residual flow. This residual flow reflects the evolution of the systematic deviation between the predicted projection and the actual corner point caused by camera extrinsic parameter drift over time in the current non-Gaussian state space.
[0035] In S4, based on the predicted projected residual flow, constraint-enhanced confidence propagation is performed within the manifold tangent space. The update increment is regressed to a special Euclidean group manifold via exponential mapping to obtain a continuously dynamically corrected homography mapping field, specifically including: The predicted projection residual flow is segmented along the time axis. The segmentation is based on the dominant frequency period of the conveyor belt vibration: the vibration frequency is measured in real time using an accelerometer mounted on the conveyor belt support, and the reciprocal of the vibration frequency in the most recent second is taken as the segment length. Each segment contains residual vectors from at least three sampling times. For each time period, the predicted projection residual vectors from all sampling times within that time period are arranged in chronological order to form a residual sequence.
[0036] Map this residual sequence to the tangent space of a special Euclidean group Lie algebra: perform temporal differencing on the residual vectors of two adjacent time points, i.e., subtract the residual vector of the previous time point from the residual vector of the later time point, to obtain the instantaneous rate of change tensor at each time point. This tensor is essentially a six-dimensional vector, where the first three dimensions represent the rotational change component and the last three dimensions represent the translational change component. Separate the rotational change component from the instantaneous rate of change tensor and organize it into a three-dimensional column vector. Then, convert this column vector into a three-dimensional antisymmetric matrix form, i.e., construct a 3x3 matrix with diagonal elements of zero. The elements in the first row and second column are equal to the negative third value of the rotational change component, the elements in the first row and third column are equal to the second value of the rotational change component, the elements in the second row and first column are equal to the third value of the rotational change component, the elements in the second row and third column are equal to the negative first value of the rotational change component, the elements in the third row and first column are equal to the negative second value of the rotational change component, and the elements in the third row and second column are equal to the first value of the rotational change component. Then, three independent elements are extracted from this antisymmetric matrix (i.e., the elements in the first row, second column, first row, third column, and second row, third column), and arranged sequentially to form a new three-dimensional rotation vector. This rotation vector is then concatenated end-to-end with the translational components (i.e., the latter three dimensions of the instantaneous rate of change tensor) to form a six-dimensional vector, called the six-dimensional Lie algebra vector. The six-dimensional Lie algebra vectors from all sampling moments within this time period are arranged in chronological order to form a vector field, which is the residual vector field in the tangent space.
[0037] Manifold curvature regularization constraints are applied to the residual vector field within the tangent space. Specifically, for each six-dimensional Lie algebra vector in the residual vector field, the difference between it and its two adjacent vectors is calculated to obtain the discrete second derivative at that vector. The sum of the squares of the discrete second derivatives of all vectors is defined as the manifold curvature cost.
[0038] A smoothing regularization factor is defined, with a value of 0.3 multiplied by the average magnitude of the vectors in the residual vector field. The product of the manifold curvature cost and the smoothing regularization factor is used as a regularization constraint term and added to the objective function of subsequent confidence propagation. Next, the posterior probability of each residual vector is iteratively updated using the confidence propagation algorithm: the confidence score of each residual vector is initialized with its normalized weight, which is obtained by weighted statistics of the granular point set. In each iteration, for each residual vector, the current confidence scores of its preceding and following neighboring vectors are collected, and the influence weight of the neighboring vectors on the current vector is calculated using a Gaussian kernel function. The width parameter of the kernel function is taken as 0.8 times the standard deviation of the distance between all vectors in the residual vector field. The confidence scores of the neighboring vectors are multiplied by their influence weights and accumulated, then multiplied by the observation likelihood of the current vector itself (determined by the Mahalanobis distance of the corresponding residual vector in the predicted projected residual flow) to obtain a new non-normalized confidence score. Repeat the above calculation for all vectors, then normalize them so that the sum of the confidence scores of all vectors equals 1. After each iteration, the residual vectors corresponding to the updated confidence scores are regressed to the tangent space manifold using a projection operator: the projection operator first converts the rotated part of the six-dimensional vector after the confidence weighted average into a three-dimensional antisymmetric matrix, and then uses an exponential mapping to keep it within the Lie algebraic tangent space of the special Euclidean group, ensuring that the vectors always lie on the tangent space manifold. The iterative process continues until the absolute value of the maximum change in the confidence scores of all vectors between two adjacent iterations is less than 0.001, at which point convergence is considered achieved.
[0039] An exponential mapping is performed on the converged tangent space increment vector. The specific implementation of the exponential mapping is as follows: For each six-dimensional Lie algebra increment vector, the magnitude of its rotation part (first three dimensions) is calculated. If the magnitude is less than 0.001 radians, the rotation matrix is approximately an identity matrix; otherwise, the rotation vector is converted into a three-dimensional rotation matrix using the Rodriguez formula, i.e., an antisymmetric matrix is constructed using the rotation vector, and then the matrix exponent is calculated through series expansion. The translation part (last three dimensions) remains unchanged. The resulting rotation matrix and translation vectors are combined into a 4x4 rigid body transformation matrix, with the top-left 3x3 column representing the rotation matrix, the top-right 3x1 column representing the translation vector, and the last row containing values of 0, 0, 0, and 1. The above exponential mapping is performed independently for each sampling moment within each time period to obtain the rigid body transformation matrix at each moment. Starting from the initial moment, the rigid body transformation matrix at the current moment is multiplied by the cumulative transformation matrix of the previous moment (using matrix multiplication) to obtain the cumulative rigid body transformation matrix at that moment. The cumulative rigid body transformation matrices at each sampling time are arranged in chronological order to form a sequence, which is the homography mapping field for continuous dynamic correction. Each matrix in this mapping field describes the projection relationship from the label's 3D coordinates to the image pixel coordinates, and the extrinsic parameter drift caused by conveyor belt vibration has been eliminated through manifold constraints and confidence propagation.
[0040] In S5, the instantaneous position coordinates of each tag are projected onto the image pixel domain in real time using a homography mapping field, generating a process parameter overlay layer bound to RFID identification. This overlay is then output as differentiated operation instructions for subsequent execution devices, specifically including: Using a continuously dynamically corrected homography field, the instantaneous position coordinates of each tag in 3D space are projected onto the image pixel plane in real time. For the current sampling moment, the rigid transformation matrix corresponding to that moment is extracted from the homography field. This matrix is a 4x4 numerical square matrix, where the first three rows and first three columns constitute the rotation part, and the first three rows and fourth column constitute the translation part. The instantaneous position coordinates of each tag in 3D space (represented as a 3D column vector) are multiplied by the rotation part of this rigid transformation matrix, and the translation part is added to the product to obtain a 3D intermediate vector. The first dimension of this 3D intermediate vector is divided by the third dimension to obtain the x-coordinate on the pixel plane, and the second dimension is divided by the third dimension to obtain the y-coordinate on the pixel plane. Since the results of division and multiplication operations are usually floating-point numbers, the obtained x-coordinate and y-coordinate are sub-pixel precision values, denoted as the original sub-pixel x-coordinate and the original sub-pixel y-coordinate, respectively.
[0041] Local calibration based on grayscale gradient weighting is performed on the aforementioned sub-pixel coordinates to eliminate minor offsets caused by discretization sampling and mapping approximation. The calibration process is as follows: A square neighborhood with sides of 5 pixels is taken as the center of the original sub-pixel coordinates (i.e., a rectangular area with 2 pixels to the left and right of the horizontal coordinate and 2 pixels up and down of the vertical coordinate). For each integer pixel position within this neighborhood, the grayscale value of that pixel is read from the original grayscale image captured by the high-speed camera. The grayscale gradient of that pixel in the horizontal direction is calculated as the difference between the grayscale value of the pixel to the right and the grayscale value of the pixel to the left, divided by 2. If a pixel on one side is missing, it is replaced by the pixel itself. Similarly, the grayscale gradient in the vertical direction is calculated as the difference between the grayscale value of the pixel below and the grayscale value of the pixel above, divided by 2. The square root of the sum of the squares of the horizontal and vertical gradients gives the grayscale gradient magnitude of that pixel. The arctangent of the ratio of the horizontal to the vertical gradient (taken between negative and positive pi) gives the gradient direction angle of that pixel. The grayscale gradient magnitude of each pixel in the neighborhood is used as the weight of that pixel. The integer coordinates of the pixel are multiplied by the weight, summed, and then divided by the sum of all weights to obtain the weighted centroid x-coordinate and weighted centroid y-coordinate. The original sub-pixel x-coordinate is subtracted from the weighted centroid x-coordinate to obtain the horizontal component of the weighted centroid offset vector; the original sub-pixel y-coordinate is subtracted from the weighted centroid y-coordinate to obtain the vertical component of the weighted centroid offset vector. Then, the gradient orientation angle distribution of all pixels in the neighborhood is statistically analyzed: the range from negative to positive pi is divided into 36 equal intervals, each interval being 1 / 18th the width of pi. The number of pixels contained in each interval is counted to form an orientation angle histogram. The interval with the most pixels in the histogram is identified, and the center orientation angle of that interval is taken as the principal orientation angle. The absolute cosine of the difference between the gradient orientation angle of each pixel and the principal orientation angle is calculated, and this absolute value is used as the density normalization coefficient. The horizontal and vertical components of the weighted centroid offset vector are multiplied by the density normalization coefficient corresponding to the pixel, and then averaged over all pixels to obtain the row and column calibration values. A relaxation factor of 0.35 is preset. The row calibration value is multiplied by the relaxation factor and added to the original sub-pixel horizontal coordinate; the column calibration value is multiplied by the relaxation factor and added to the original sub-pixel vertical coordinate to obtain the updated sub-pixel horizontal and vertical coordinates. This entire calibration process is repeated (starting from neighborhood selection). After each iteration, the absolute values of the row and column calibration values are calculated. Iteration stops when both are less than 0.49 pixel spacing (i.e., less than half a pixel). Generally, the number of iterations does not exceed 3. The sub-pixel horizontal and vertical coordinates at the end of the final iteration are output as the calibrated pixel coordinates.
[0042] The calibrated pixel coordinates are bound to the corresponding RFID identification and pre-stored process parameters. Each tag's RFID identification is a unique serial number written at the factory. The pre-stored process parameters include the torque value (in Newton-meters, floating-point) and tightening sequence number (integer sequence) for the workpiece model corresponding to that tag. An overlay data structure is constructed, containing a list where each entry corresponds to a tag. The entry content is as follows: timestamp of the sampling time, calibrated sub-pixel horizontal coordinate, calibrated sub-pixel vertical coordinate, RFID identification, torque value, and tightening sequence number. The overlay entries for each sampling time are arranged sequentially to form a time-series overlay. This overlay is converted into differentiated operation instructions that can be recognized by subsequent execution devices: for the first step to be executed in the tightening sequence number of each entry, an instruction string is generated in the format "Locate to horizontal coordinate value, vertical coordinate value, apply torque value to the workpiece corresponding to the RFID identification," where the horizontal and vertical coordinate values are rounded to the nearest integer pixel coordinates, and the torque value is retained to one decimal place. The instruction strings of all entries are concatenated into an instruction stream according to the timestamp order and sent to the robotic arm controller via industrial Ethernet to complete the output of differentiated operation instructions.
[0043] The working principle of this invention is as follows: During workpiece conveying, a complex domain echo signal matrix and an image corner heatmap are simultaneously acquired; phase interference fringes are calculated based on the complex domain echo signal matrix; a dynamic phase gradient field is generated through multi-dimensional phase unwrapping, and the three-dimensional instantaneous position coordinates of each label are retrieved; the instantaneous position coordinates and the image corner heatmap are input into a non-Gaussian state space, constructing a state augmented vector containing position components and corner residual components; asymmetric scaling is applied to generate an initial finite particle set; the predicted projection Jacobian matrix is calculated; the time-varying noise covariance is estimated based on the conveyor belt vibration acceleration, and the particle set scale parameters are dynamically adjusted; weighted statistics are then used to obtain the predicted projection. The projection residual flow is segmented along the time axis and mapped to the tangent space of a special Euclidean group Lie algebra to obtain a residual vector field. Manifold curvature regularization constraints are applied and the posterior probability is updated iteratively using belief propagation. The converged tangent space increment vector is transformed into a rigid body transformation matrix through exponential mapping, and a continuously dynamically corrected homography mapping field is obtained by accumulating frame by frame. The instantaneous position coordinates of each tag are projected onto the image pixel domain using the homography mapping field to obtain sub-pixel coordinates. The mapping position is locally calibrated using gray-level gradient weighting. The calibrated pixel coordinates are bound to RFID identification and pre-stored process parameters to generate an overlay layer. After encoding, differentiated operation instructions are output.
[0044] Please see Figure 4 As shown, the AI multimodal RFID-integrated intelligent data acquisition system includes: Multi-source synchronous acquisition module: During the workpiece transfer process, it synchronously acquires complex domain echo signal matrix and image corner heat map. The complex domain echo signal matrix is composed of IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by high-speed camera. Phase unwinding and 3D positioning module: The phase interference fringes of each tag are calculated based on the complex domain echo signal matrix. A dynamic phase gradient field is generated through multi-dimensional phase unwinding, and the instantaneous position coordinates of each tag in 3D space are inverted using the dynamic phase gradient field. Non-Gaussian state space prediction module: Input the instantaneous position coordinates and image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. Manifold constraint belief propagation module: Based on the predicted projected residual flow, it performs constraint-enhanced belief propagation in the manifold tangent space, and regresses the update increment to a special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field; Projection mapping and instruction generation module: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field, generating a process parameter overlay layer bound with RFID identification, and outputting differentiated operation instructions for subsequent execution devices.
[0045] The working principle of this invention is as follows: During workpiece conveying, a complex domain echo signal matrix and an image corner heatmap are simultaneously acquired; phase interference fringes are calculated based on the complex domain echo signal matrix; a dynamic phase gradient field is generated through multi-dimensional phase unwrapping, and the three-dimensional instantaneous position coordinates of each label are retrieved; the instantaneous position coordinates and the image corner heatmap are input into a non-Gaussian state space, constructing a state augmented vector containing position components and corner residual components; asymmetric scaling is applied to generate an initial finite particle set; the predicted projection Jacobian matrix is calculated; the time-varying noise covariance is estimated based on the conveyor belt vibration acceleration, and the particle set scale parameters are dynamically adjusted; weighted statistics are then used to obtain the predicted projection. The projection residual flow is segmented along the time axis and mapped to the tangent space of a special Euclidean group Lie algebra to obtain a residual vector field. Manifold curvature regularization constraints are applied and the posterior probability is updated iteratively using belief propagation. The converged tangent space increment vector is transformed into a rigid body transformation matrix through exponential mapping, and a continuously dynamically corrected homography mapping field is obtained by accumulating frame by frame. The instantaneous position coordinates of each tag are projected onto the image pixel domain using the homography mapping field to obtain sub-pixel coordinates. The mapping position is locally calibrated using gray-level gradient weighting. The calibrated pixel coordinates are bound to RFID identification and pre-stored process parameters to generate an overlay layer. After encoding, differentiated operation instructions are output.
[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.
Claims
1. An intelligent data acquisition method based on AI multimodal RFID, characterized in that, Includes the following steps: S1: During the workpiece transfer process, a complex domain echo signal matrix and an image corner heat map are simultaneously acquired. The complex domain echo signal matrix is composed of the IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by a high-speed camera. S2: Solve the phase interference fringes of each tag based on the complex domain echo signal matrix, generate a dynamic phase gradient field through multidimensional phase unwrapping, and use the dynamic phase gradient field to invert the instantaneous position coordinates of each tag in three-dimensional space. S3: Input the instantaneous position coordinates and the image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic parameter drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. S4: Based on the predicted projected residual flow, confidence propagation with constraint enhancement is performed in the tangent space of the manifold. The update increment is regressed to the special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field. S5: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field to generate a process parameter overlay layer bound with RFID identification, and output as differentiated operation instructions for subsequent execution devices.
2. The intelligent data acquisition method for AI multimodal fusion RFID according to claim 1, characterized in that, The process of retrieving the instantaneous position coordinates of each label in three-dimensional space using a dynamic phase gradient field specifically includes: Instantaneous phase extraction is performed on the complex domain echo signal matrix along the antenna array direction, and the phase difference between adjacent antennas is calculated to generate an initial interference fringe pattern; Multidimensional phase unwinding constructs an optimal path graph based on phase gradient confidence and uses a minimum discontinuity pruning algorithm to eliminate phase jumps, outputting a continuously distributed dynamic phase gradient field. The inversion of instantaneous position coordinates further includes: performing line integration of the dynamic phase gradient field along the three spatial axes, and iteratively solving for the three-dimensional coordinates of each tag by combining it with the pre-calibrated antenna spatial coordinates.
3. The intelligent data acquisition method for AI multimodal fusion RFID according to claim 2, characterized in that, The continuously distributed dynamic phase gradient field output specifically includes: Calculate the phase gradient magnitude and neighborhood orientation entropy of each pixel in the initial interference fringe pattern, and mark the pixels with small magnitude and low orientation entropy as high confidence seed points; Starting from all seed points, the connection path is expanded along the gradient descent direction. A closed path graph is formed when the confidence difference between the two ends of the path is lower than the threshold. The minimum discontinuity pruning algorithm is as follows: Identify phase transition edges along the optimal path graph, calculate the continuous phase difference on both sides of each transition edge, prune the transition edges that result in the largest cumulative discontinuity, and fill the pruned region with radial basis interpolation of adjacent continuous phases to output a dynamic phase gradient field.
4. The intelligent data acquisition method of AI multimodal fusion RFID according to claim 1, characterized in that, The process of obtaining the predicted projection residual flow specifically includes: After aligning the instantaneous position coordinates with the timestamps of the image corner heatmap, a state augmentation vector for the non-Gaussian state space is constructed, which includes position components and corner residual components. Apply asymmetric scaling to the state augmentation vector to generate an initial finite set of granular points, and calculate the prediction projection Jacobian matrix for each granular point; The time-varying noise covariance is estimated based on the instantaneous vibration acceleration of the conveyor belt. The scale parameters of the particle set are dynamically adjusted. The projected residuals of the adjusted particle set are weighted and statistically analyzed to output the predicted projected residual flow.
5. The intelligent data acquisition method of AI multimodal fusion RFID according to claim 4, characterized in that, The process of obtaining the predicted projection Jacobian matrix is as follows: Calculate the local skewness coefficients of each dimension of the state augmentation vector, and generate positive and negative scaling factors according to the sign of the local skewness coefficients. Using the current estimate of the state augmentation vector as the center, Gaussian perturbations of unequal amplitudes are applied to the positive and negative deviation directions according to the positive and negative scaling factors, respectively, to obtain the initial finite particle set; For each particle in the initial finite particle set, the partial derivatives of the particle with respect to each dimension of the state are calculated using the numerical difference method, and then assembled into a prediction projection Jacobian matrix.
6. The intelligent data acquisition method of AI multimodal fusion RFID according to claim 1, characterized in that, The process of obtaining the continuously dynamically corrected homography mapping field specifically includes: After segmenting the predicted projection residual flow along the time axis, it is mapped to the tangent space of the special Euclidean group Lie algebra to obtain the residual vector field in the tangent space; In the tangent space, a manifold curvature regularization constraint is applied to the residual vector field, and the posterior probability of each residual vector is iteratively updated using the belief propagation algorithm. After each iteration, the update result is regressed to the tangent space manifold through the projection operator. An exponential mapping is performed on the converged tangent space increment vector to transform it into a rigid body transformation matrix on a special Euclidean group. The rigid body transformation matrix is accumulated frame by frame and then the continuously dynamically corrected homography field is output.
7. The intelligent data acquisition method for AI multimodal fusion RFID according to claim 6, characterized in that, The process of composing the residual vector field is as follows: Temporal differencing is performed on the predicted projected residual flow within each time axis segment to obtain the instantaneous rate of change tensor at each time point; The rotational and translational components are decomposed from the instantaneous rate of change tensor. The rotational components are converted into a three-dimensional antisymmetric matrix form, and the three independent elements of the antisymmetric matrix are extracted as rotation vectors. The rotation vector and translation components are arranged sequentially and spliced into a six-dimensional Lie algebra vector, which forms a residual vector field in the tangent space in time order.
8. The intelligent data acquisition method of AI multimodal fusion RFID according to claim 1, characterized in that, The differentiated operation instructions specifically include: Based on the mapping coefficients at the corresponding time in the homography mapping field, the instantaneous position coordinates of each tag are transformed into sub-pixel coordinates on the pixel plane; Centered on sub-pixel coordinates, extract the gray-level gradient weighted values in the neighborhood, and use the gray-level gradient weighted values to perform local calibration of the mapping position to eliminate discretization error; The calibrated pixel coordinates are bound to the RFID identification of the corresponding tag and the pre-stored process parameters. An overlay layer containing identity-parameter pairs is generated by superimposing them according to the time sequence and encoded as a differentiated operation instruction.
9. The intelligent data acquisition method of AI multimodal fusion RFID according to claim 8, characterized in that, The local calibration of the mapped position using gray-level gradient weighting specifically includes: Using sub-pixel coordinates as the center, calculate the gray-level gradient magnitude and gradient direction angle of each pixel in a small neighborhood around it, and use the gradient magnitude as the weight to construct a weighted centroid offset vector. The weighted centroid offset vector is normalized according to the distribution density of gradient direction angles in the neighborhood, and decomposed along the row and column directions of the image to obtain the row calibration amount and the column calibration amount. The row and column calibration values are multiplied by a preset relaxation factor and then superimposed onto the original subpixel coordinates. This process is iterated until the calibration value is less than half a pixel spacing, and then the calibrated subpixel coordinates are output.
10. An AI-based multimodal RFID-integrated intelligent data acquisition system, characterized in that, The intelligent data acquisition method for performing the AI multimodal fusion RFID according to any one of claims 1-9 includes: Multi-source synchronous acquisition module: During the workpiece transfer process, it synchronously acquires complex domain echo signal matrix and image corner heat map. The complex domain echo signal matrix is composed of IQ components reflected by RFID tags received by multiple antennas, and the image corner heat map is composed of the set of sub-pixel coordinates of the workpiece contour acquired by high-speed camera. Phase unwinding and 3D positioning module: The phase interference fringes of each tag are calculated based on the complex domain echo signal matrix. A dynamic phase gradient field is generated through multi-dimensional phase unwinding, and the instantaneous position coordinates of each tag in 3D space are inverted using the dynamic phase gradient field. Non-Gaussian state space prediction module: Input the instantaneous position coordinates and image corner heatmap into the non-Gaussian state space, and perform unscented transformation prediction on the camera extrinsic drift caused by the conveyor belt motion in the non-Gaussian state space to obtain the predicted projection residual flow. Manifold constraint belief propagation module: Based on the predicted projected residual flow, it performs constraint-enhanced belief propagation in the manifold tangent space, and regresses the update increment to a special Euclidean group manifold through exponential mapping to obtain a continuously dynamically corrected homography mapping field; Projection mapping and instruction generation module: The instantaneous position coordinates of each tag are projected onto the image pixel field in real time using a homography mapping field, generating a process parameter overlay layer bound with RFID identification, and outputting differentiated operation instructions for subsequent execution devices.
