System for automatic recognition of packaging assets based on feature point cloud

Abstract

The present application relates to computer vision, digital twin and intelligent logistics automation field, specifically to a packaging asset automatic identification system based on feature point cloud, comprising: a data acquisition module for acquiring scene depth point cloud of target scene; an ideal reference reconstruction module for constructing an ideal geometric reference as noiseless reference data and generating ideal reference point cloud; a parameter injection simulation module for constructing an active simulation engine and generating simulation state point cloud with specific defect labels; a double-track difference extraction module for calculating the normal vector difference distribution between point clouds; a topological coupling decision module for calculating the numerical matching degree between the two and outputting the identification result and specific state according to the matching degree; the present application effectively shields the interference caused by the change of viewing angle, solves the problem of invalidation of traditional rigid matching algorithm, and realizes high-precision matching of damage fingerprints.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, digital twins and intelligent logistics automation, specifically to an automatic identification system for packaging assets based on feature point clouds. Background Technology

[0002] Automatic identification of packaging assets is a crucial link in the material flow and warehousing management of smart logistics systems. Its identification accuracy and robustness directly determine the level of automation and overall efficiency of logistics operations. Automatic identification systems primarily utilize 3D sensors to collect surface depth information of targets such as pallets, turnover boxes, and cartons, achieving asset location and identification through geometric comparison with a preset model. However, in actual logistics conditions, packaging assets often experience non-rigid deformations such as bulging, collapse, and corner shrinkage due to heavy stacking or violent handling, and their surfaces are often covered with transparent film wrapping or strapping. Interference causes the collected point cloud data to contain a large amount of highly reflective noise or spatially discontinuous features. Existing technologies mainly rely on template matching or fixed geometric feature extraction methods based on rigid assumptions, which cannot effectively decouple physical deformation from sensor noise. It is difficult to construct a dynamic reference system containing theoretical deformation laws using multi-source data, and lacks the ability to actively simulate and adapt to complex deformation patterns. As a result, when facing assets with severe non-rigid deformation or strong environmental interference, the rejection rate of the recognition algorithm is high and the anti-interference ability is insufficient. Therefore, there is an urgent need for a solution to address the problems existing in the current technology. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an automatic identification system for packaged assets based on feature point clouds. Specifically, the technical solution of this invention includes:

[0004] The data acquisition module is used to acquire the scene depth point cloud of the target scene and retrieve the preset standard parametric model. The scene depth point cloud includes three-dimensional coordinate information and reflection intensity information. The standard parametric model is the pre-constructed three-dimensional geometric data of the target object.

[0005] The ideal reference reconstruction module is used to construct an ideal geometric reference as noise-free reference data based on the scene depth point cloud and the standard parameterized model, estimate the six-dimensional pose of the standard parameterized model in the target scene through coarse registration, and perform virtual ray projection in this pose to generate the ideal reference point cloud.

[0006] The parameter injection simulation module is used to build an active simulation engine. Based on the standard parameterized model, it uses preset deformation factors and environmental interference factors to adjust or superimpose noise on the geometric surface of the model to generate a simulated point cloud with specific defect labels. The simulated point cloud is used to simulate the non-rigid deformation and surface occlusion features of the target object in the physical environment.

[0007] The dual-track difference extraction module is used to calculate the normal vector difference distribution between point clouds, and to calculate the real residual feature set of the scene depth point cloud relative to the ideal reference point cloud, which is defined as the real residual space, and the theoretical residual feature set of the simulated point cloud relative to the ideal reference point cloud, which is defined as the theoretical residual space.

[0008] The topology coupling decision module is used to identify target objects based on topological similarity, extract high-dimensional features of the real residual space and the theoretical residual space, calculate the numerical matching degree between the two, and output the recognition result and specific status according to the matching degree.

[0009] Optionally, the data acquisition module includes:

[0010] The scene perception unit is used to collect the scene depth point cloud in real time through an industrial 3D sensor. The scene depth point cloud is a discrete point set containing geometric information of the object surface and reflection intensity information.

[0011] The knowledge base calling unit is used to store and provide the standard parametric model, which is the ideal geometric data of the target object and is associated with a common deformation topology library, which contains deformation pattern data of the object under pressure and impact physical conditions.

[0012] Optional, the ideal baseline reconstruction module includes:

[0013] The pose estimation unit is used to calculate the spatial transformation matrix between the scene depth point cloud and the standard parametric model, and to determine the position and rotation angle of the standard parametric model in the current coordinate system.

[0014] A clean projection unit is used to eliminate sensor measurement noise. Under the determined six-dimensional pose, the standard parameterized model is resampled using a virtual ray projection algorithm to generate the ideal reference point cloud, which has a theoretically smooth surface and complete geometric structure.

[0015] Optionally, the parameter injection simulation module includes:

[0016] The deformation simulation unit is used to simulate changes in the physical structure of an object. Free deformation mesh control points are introduced on the mesh of the standard parametric model. According to the preset collapse model or bulge model, the spatial position of the control points is adjusted to generate an intermediate model that undergoes non-rigid deformation.

[0017] The noise superposition unit is used to simulate surface covering interference. It superimposes a layer of random noise or fluctuation at a preset frequency on the surface of the intermediate state model to simulate the discontinuity of point cloud caused by film reflection or strapping. The unit then combines the intermediate state model to generate the simulated point cloud.

[0018] Optionally, the dual-track differential extraction module includes:

[0019] The real-world computation unit is used to extract the mixed features of real damage and environmental noise, align the scene depth point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the real-world residual space.

[0020] The theoretical field calculation unit is used to extract pure physical damage features, align the simulated point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the theoretical residual space, which does not contain sensor random noise.

[0021] Optionally, the topology coupling decision module includes:

[0022] The feature mapping unit is used to transform the residual space into feature vectors, and to encode the real residual space and the theoretical residual space respectively using a manifold harmonic analysis algorithm or a three-dimensional convolutional neural network to generate real feature vectors and theoretical feature vectors.

[0023] The similarity calculation unit is used to quantify the matching degree between two residual spaces, calculate the cosine similarity or Euclidean distance between the real feature vector and the theoretical feature vector, and obtain the coupling similarity value.

[0024] Optionally, the topology coupling decision module further includes: a logic decision unit, configured to preset a first decision threshold and a second decision threshold, wherein the first decision threshold is greater than the second decision threshold; the logic decision unit executes the following decision logic:

[0025] If the coupling similarity value is greater than or equal to the first determination threshold, then the object corresponding to the scene depth point cloud is determined to be the target object, and the object has undergone a specific deformation corresponding to the simulated point cloud, and a recognition success signal and deformation type are output.

[0026] If the coupling similarity value is less than or equal to the second determination threshold, it is determined that the real residual space is mainly composed of environmental noise or non-target objects, and noise interference or unknown object signals are output and a filtering operation is performed.

[0027] If the coupling similarity value is between the second determination threshold and the first determination threshold, it is marked as a suspected target, and the system is triggered to request multi-angle re-acquisition or manual review.

[0028] Optional, also includes:

[0029] An adaptive feedback module is used to optimize simulation parameters. When the topology coupling decision module outputs a successful recognition signal, it feeds back the feature parameters of the real residual space to the parameter injection simulation module.

[0030] The parameter injection simulation module responds to the feedback feature parameters and corrects the weight of the preset deformation factor so that the subsequently generated simulated point cloud more closely approximates the deformation law of objects in real logistics scenarios.

[0031] Optionally, the target object is a logistics packaging asset, including pallets, crates, or cartons;

[0032] The non-rigid deformations include bulges, collapses, or corner shrinkage caused by stacking and compression.

[0033] The environmental interference factors include surface reflections and point cloud defects caused by transparent film wrapping or strapping.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention constructs an active simulation engine that includes preset deformation factors and environmental interference factors, which can simulate the non-rigid deformation and surface occlusion features of objects in the physical environment. The system uses free deformation technology and noise superposition model to transform uncontrollable physical deformation into a controllable mathematical intermediate state. This method enables the recognition algorithm to have the ability to actively adapt to non-rigid deformations such as bulging and collapse, and significantly reduces the rejection rate caused by packaging box deformation in logistics scenarios.

[0036] 2. This invention introduces an ideal benchmark reconstruction module, utilizing virtual ray projection technology to generate a noise-free ideal benchmark point cloud. By constructing both a real residual space and a theoretical residual space, the system successfully decouples real physical deformation features from sensor random noise. Even in complex environments with reflective transparent films or interference from strapping, it can effectively filter out mismatches caused by viewpoint occlusion, ensuring that feature extraction purely reflects the physical changes on the object's surface, rather than sampling errors.

[0037] 3. This invention employs a dual-track differential extraction and topological coupling decision mechanism, transforming the traditional absolute coordinate comparison into a comparison of difference morphologies. By calculating the difference in normal vectors and Euclidean distance between point clouds, it extracts and compares the high-dimensional topological features of the real and theoretical residual spaces, achieving high-precision matching of damaged fingerprints. This method focuses on the essential similarity of topological structures, effectively shielding the interference caused by changes in perspective, and solving the problem of failure of traditional rigid matching algorithms.

[0038] 4. This invention sets up an adaptive feedback module and establishes a closed-loop digital twin verification mechanism. When the recognition is successful, the system uses the statistical characteristics of the real residual space to dynamically correct the deformation factor weight of the simulation model. This online learning mechanism enables the simulation parameters to converge to the real working conditions over time, allowing the system to continuously approximate the deformation law of objects in the real logistics scenario, thereby continuously improving the system's recognition accuracy and adaptability to specific stacking methods. Attached Figure Description

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 An automatic identification system for packaged assets based on feature point clouds includes:

[0044] The data acquisition module is used to acquire the scene depth point cloud of the target scene and retrieve the preset standard parametric model. The scene depth point cloud includes three-dimensional coordinate information and reflection intensity information. The standard parametric model is the pre-constructed three-dimensional geometric data of the target object.

[0045] The ideal reference reconstruction module is used to construct an ideal geometric reference as noise-free reference data based on the scene depth point cloud and the standard parameterized model, estimate the six-dimensional pose of the standard parameterized model in the target scene through coarse registration, and perform virtual ray projection in this pose to generate the ideal reference point cloud.

[0046] The parameter injection simulation module is used to build an active simulation engine. Based on the standard parameterized model, it uses preset deformation factors and environmental interference factors to adjust or superimpose noise on the geometric surface of the model to generate a simulated point cloud with specific defect labels. The simulated point cloud is used to simulate the non-rigid deformation and surface occlusion features of the target object in the physical environment.

[0047] The dual-track difference extraction module is used to calculate the normal vector difference distribution between point clouds, and to calculate the real residual feature set of the scene depth point cloud relative to the ideal reference point cloud, which is defined as the real residual space, and the theoretical residual feature set of the simulated point cloud relative to the ideal reference point cloud, which is defined as the theoretical residual space.

[0048] The topology coupling decision module is used to identify target objects based on topological similarity, extract high-dimensional features of the real residual space and the theoretical residual space, calculate the numerical matching degree between the two, and output the recognition result and specific status according to the matching degree.

[0049] This embodiment details the overall architecture and workflow of an automatic packaging asset identification system based on feature point clouds. The system employs the core idea of ​​synthetic analysis to construct a closed-loop digital twin verification mechanism; its execution flow is as follows:

[0050] The data acquisition module performs the task of perceiving the physical world, acquiring the scene depth point cloud of the target scene through a depth camera, denoted as... And create an index in the database;

[0051] The ideal reference reconstruction module is introduced to establish a noise-free mathematical reference system. This module estimates the six-dimensional pose of the standard parametric model in the scene using a coarse registration algorithm. Then, under this determined pose, virtual scanning technology is used to resample the model to generate the ideal reference point cloud. ;

[0052] The parameter injection simulation module constructs an active simulation engine, transforming the chaos of the physical world into a computable model. This module uses preset deformation factors and environmental interference factors to adjust the model, generating a series of simulated point clouds with specific defect labels. Specifically, the parameter injection simulation module does not generate a single deformation model, but performs Monte Carlo sampling based on the parameter space of the deformation factor to generate simulated point cloud clusters covering different deformation locations and degrees in parallel, thereby ensuring that the topology coupling decision module can match the topological feature that is closest to the current real deformation within the preset search space.

[0053] Based on this, the dual-track difference extraction module performs parallel computation on two paths, respectively constructing the real residual space. With theoretical residual space This transforms the comparison of absolute coordinates into a comparison of differences in morphology.

[0054] The topological coupling decision module identifies the target object based on topological similarity, extracts high-dimensional features and calculates the numerical matching degree, and outputs the final recognition result.

[0055] This embodiment successfully decouples physical deformation from sensor noise by introducing a dual reference of ideal benchmark and simulated state. Even if the packaging box undergoes severe non-rigid deformation, as long as the deformation conforms to physical laws, the system can identify it as the target object, thereby reducing the rejection rate in logistics scenarios and effectively filtering out irregular noise interference caused by reflective strips or films.

[0056] Example 2:

[0057] The data acquisition module includes:

[0058] The scene perception unit is used to collect the scene depth point cloud in real time through an industrial 3D sensor. The scene depth point cloud is a discrete point set containing geometric information of the object surface and reflection intensity information.

[0059] The knowledge base calling unit is used to store and provide the standard parametric model, which is the ideal geometric data of the target object and is associated with a common deformation topology library, which contains deformation pattern data of the object under pressure and impact physical conditions.

[0060] This embodiment details the hardware and data structure of the data acquisition module;

[0061] The scene perception unit activates an industrial-grade structured light camera or The sensor acts as a data acquisition terminal, capturing depth information of the physical space in real time.

[0062] This unit generates scene depth point clouds. Its mathematical expression is a set. ;

[0063] in, The source is discrete sampling from sensors, and its physical meaning is single-point data in space;

[0064] : The source is the echo detection of laser or infrared light, and its physical meaning is the material reflection intensity, with the unit being the normalized intensity value [0,1].

[0065] The knowledge base calling unit not only retrieves static CAD models, but also synchronously loads the associated common deformation topology library; this library is a dataset based on finite element analysis pre-calculation, containing deformation mode data of objects under physical conditions such as compression and impact;

[0066] This embodiment introduces reflectivity information and pre-set deformation topology knowledge, enabling the system to have multi-dimensional perception capabilities at the acquisition end. Information from high reflectivity areas assists in the subsequent removal of high-light noise, providing physical-level prior data support for subsequent simulation and decision-making.

[0067] Example 3:

[0068] The ideal benchmark reconstruction module includes:

[0069] The pose estimation unit is used to calculate the spatial transformation matrix between the scene depth point cloud and the standard parametric model, and to determine the position and rotation angle of the standard parametric model in the current coordinate system.

[0070] A clean projection unit is used to eliminate sensor measurement noise. Under the determined six-dimensional pose, the standard parameterized model is resampled using a virtual ray projection algorithm to generate the ideal reference point cloud, which has a theoretically smooth surface and complete geometric structure.

[0071] This embodiment describes in detail the algorithm implementation of the ideal benchmark reconstruction module;

[0072] The pose estimation unit employs a two-stage registration strategy: coarse registration using FPFH features, followed by fine adjustment using a point-to-surface ICP algorithm to calculate the spatial transformation matrix. ;

[0073] The pure projection unit executes a virtual ray casting algorithm to generate an ideal reference point cloud that perfectly matches the sensor's viewpoint. Construct the following projection model:

[0074]

[0075] in, The source is a knowledge base calling unit, and the physical meaning is a standard parameterized model;

[0076] The source is the calculation result of the pose estimation unit, and the physical meaning is the pose of the model in the current sensor coordinate system;

[0077] The source is the physical sensor calibration parameters, and the physical meaning is the intrinsic parameter matrix of the virtual camera;

[0078] The system uses this projection model to resample the standard parametric model, generating an ideal reference point cloud with a theoretically smooth surface and complete geometry;

[0079] This embodiment eliminates mismatches caused by viewpoint occlusion and resolution differences by using virtual resampling from the same viewpoint, thus constructing a perfect differential benchmark. This allows the subsequently extracted residuals to purely reflect the physical changes on the object's surface, rather than sampling errors.

[0080] Example 4:

[0081] The parameter injection simulation module includes:

[0082] The deformation simulation unit is used to simulate changes in the physical structure of an object. Free deformation mesh control points are introduced on the mesh of the standard parametric model. According to the preset collapse model or bulge model, the spatial position of the control points is adjusted to generate an intermediate model that undergoes non-rigid deformation.

[0083] The noise superposition unit is used to simulate surface covering interference. It superimposes a layer of random noise or fluctuation at a preset frequency on the surface of the intermediate state model to simulate the discontinuity of point cloud caused by film reflection or strapping. The unit then combines the intermediate state model to generate the simulated point cloud.

[0084] This embodiment details how the parameter injection simulation module generates realistic simulated point clouds;

[0085] The deformation simulation unit uses free deformation (FFD) technology to construct a control lattice around the standard parameterized model;

[0086] The system uses a deformation mapping function to calculate the displacement of mesh vertices:

[0087]

[0088] in, This represents a weighted summation along the dimensions of the three-dimensional mesh control points;

[0089] The source is the calculation output, and the physical meaning is the coordinates of the mesh vertices after deformation;

[0090] The source is a pre-set collapse model or bulge model, and the physical meaning is the coordinates of the control points and their displacement.

[0091] The source is a mathematical definition; its physical meaning is the Bernstein basis function, and its calculation formula is... , respectively corresponding , , Direction, among which, Represent , , , They represent the order respectively , , , Representing indexes respectively , , , used to determine the geometric influence weight of the control points on the target vertex;

[0092] The source is the definition of the local coordinate system, and its physical meaning is the normalized parameter coordinates of the grid vertices within the control lattice, with values ​​ranging from [0,1].

[0093] The source is preset parameters, and its physical meaning is to control the order of the crystal lattice in the X, Y, and Z directions;

[0094] The noise superposition unit does not directly perform scalar superposition, but instead performs vector superposition based on normal displacement; for each vertex of the intermediate model and its corresponding unit normal vector Calculate the vertex of the simulation state :

[0095]

[0096] in, The source is a random number generator, and its physical meaning is to simulate the Gaussian noise component of tiny wrinkles in a thin film, which satisfies a probability distribution. , These are preset thin film roughness parameters;

[0097] The source is a periodic function calculation, and its physical meaning is the oscillating layer component simulating the indentation or regular texture of a strapping tape. To avoid confusion with the symbol f representing a scalar field in subsequent embodiments, the symbol f is used here. Represents spatial frequency, with the unit being the reciprocal of the millimeter. Used to offset projection distance The length dimension is used to ensure that the input parameter of the sine function is a dimensionless phase angle, and the calculation formula is as follows: ,in, For preset amplitude, For spatial frequency, As vertices Unit vector in the preset texture extension direction The projected distance on, i.e. , This is the initial phase;

[0098] The final simulated point cloud is generated by combining deformation and noise characteristics;

[0099] This embodiment transforms uncontrollable physical deformation into controllable mathematical parameters. Through FFD technology and a normal noise superposition model, the system can generate various intermediate states such as slight bulging, severe compression, and surface texture interference, making the recognition algorithm adaptable to non-rigid deformation and high-frequency noise.

[0100] Example 5:

[0101] The dual-track differential extraction module includes:

[0102] The real-world computation unit is used to extract the mixed features of real damage and environmental noise, align the scene depth point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the real-world residual space.

[0103] The theoretical field calculation unit is used to extract pure physical damage features, align the simulated point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the theoretical residual space, which does not contain sensor random noise.

[0104] This embodiment details the calculation logic of the dual-track differential extraction module;

[0105] Both the real-field calculation unit and the theoretical-field calculation unit use the difference in normal vectors as the core metric.

[0106] For any point in the reference point cloud Find its nearest neighbor in the target point cloud. Calculate the difference in normal vectors between the two. The calculation formula is as follows:

[0107]

[0108] in, The source is CAD model calculation, and the physical meaning is the ideal reference point cloud midpoint. The normal vector;

[0109] The source is local plane fitting, and its physical meaning is the corresponding point in the target point cloud. The normal vector;

[0110] The source is calculated using the above formula, and its physical meaning is the surface normal difference value, ranging from [0,1].

[0111] At the same time, calculate the Euclidean distance between corresponding point pairs. And perform normalization:

[0112]

[0113] The system constructs both the real residual space and the theoretical residual space, and merges the two scalar features mentioned above into a residual feature vector. The calculation formula is as follows:

[0114]

[0115] Definition here The calibration method is as follows: Based on a pre-constructed sample dataset containing typical defects such as wrinkles and damage, the system uses logistic regression analysis or principal component analysis to calculate the difference in normal vectors. Euclidean distance For the correlation coefficients of manually labeled results, the normalized correlation coefficients are assigned to... and To characterize the contribution of different dimensional features to the final residual; among them, The preset weighting coefficients, is the hyperbolic tangent function, used to normalize the Euclidean distance to the interval [-1, 1]; The distance scale factor is set according to the sensor accuracy and object size, typically taking 5% to 10% of the average side length of the object; this residual space set The former includes both actual damage and environmental noise, while the latter only includes purely physical damage characteristics.

[0116] This embodiment utilizes the characteristic that the difference in normal vectors is more sensitive to the minute bumps and depressions on the surface, and combines it with the Euclidean distance to measure macroscopic deformation. Compared with the simple distance difference, it can more effectively capture details such as cardboard creases and film textures. In fact, it is extracting damaged fingerprints, laying the foundation for subsequent high-precision matching.

[0117] Example 6:

[0118] The topology coupling decision module includes:

[0119] The feature mapping unit is used to transform the residual space into feature vectors, and to encode the real residual space and the theoretical residual space respectively using a manifold harmonic analysis algorithm or a three-dimensional convolutional neural network to generate real feature vectors and theoretical feature vectors.

[0120] The similarity calculation unit is used to quantify the matching degree between two residual spaces, calculate the cosine similarity or Euclidean distance between the real feature vector and the theoretical feature vector, and obtain the coupling similarity value.

[0121] This embodiment describes the feature extraction process of the topological coupling decision module;

[0122] The feature mapping unit uses the manifold harmonic analysis algorithm (MHA) to process the input residual space;

[0123] The residual values ​​are treated as a scalar field defined on the surface of the manifold, and decomposed using the Laplace-Beltramian operator to obtain the spectral coefficient vector. The specific calculation steps are as follows:

[0124] Constructing the Discrete Laplace Operator ,in, Vertex blending Area matrix; Let be the cotangent weight matrix, for connecting vertices and The edge, its weight ,in, and Let be the angles of the two angles opposite the side; D is a degree matrix whose diagonal elements satisfy . That is, for vertices The weights of all adjacent edges are summed.

[0125] Solve the characteristic equation Before extraction The eigenvector basis corresponding to the smallest non-zero eigenvalues ;in, For the corresponding eigenvalues, The preset feature dimension is usually set to 64 or 128 to ensure a balance between feature expressiveness and computational efficiency.

[0126] The characteristic scalar field in the residual space Projected onto the substrate, The residual eigenvectors corresponding to each point in the aforementioned real or theoretical residual space. The L2 norm, i.e., the modulus Generate feature vectors : ,in, That is, when considering area weight Calculate scalar field under the condition With basis functions Discrete inner product;

[0127] As an alternative to parallel processing, when the system configuration uses a three-dimensional convolutional neural network, the feature mapping unit first voxels the discrete point set in the residual space to construct a three-dimensional tensor mesh, and the value of each voxel is filled with the residual feature in that spatial location. The statistical mean; for empty voxels that do not contain any point cloud data, their values ​​are filled with zero vectors. To maintain the integrity of the tensor data, the three-dimensional tensor is input into a pre-trained deep network. This three-dimensional convolutional neural network is pre-trained using a publicly available point cloud classification dataset. The network structure contains four three-dimensional convolutional layers, each followed by a batch normalization layer. The activation function includes alternating 3D max-pooling layers to extract spatial hierarchical features, and is connected to fully connected layers at the ends to compress the high-dimensional tensor into a one-dimensional feature vector. ;

[0128] The system generates real-world feature vectors. With theoretical eigenvectors ;

[0129] The similarity calculation unit calculates the cosine similarity between the two and outputs a coupled similarity value. :

[0130]

[0131] in, The source is the real-world residual space coding, and its physical meaning is the frequency domain characteristics of the real scene;

[0132] The source is theoretical residual space coding, and the physical meaning is the frequency domain characteristics of the simulation scene;

[0133] This embodiment extracts features through frequency domain analysis, making the feature vector rotation invariant and robust to local noise. It can effectively shield interference caused by changes in viewing angle and focus on the essential similarity of topological structures.

[0134] Example 7:

[0135] The topology coupling decision module also includes:

[0136] The logic decision unit is configured to preset a first decision threshold and a second decision threshold, wherein the first decision threshold is greater than the second decision threshold;

[0137] The logic adjudication unit executes the following judgment logic: if the coupling similarity value is greater than or equal to the first judgment threshold, then the object corresponding to the scene depth point cloud is determined to be the target object, and the object has undergone a specific deformation corresponding to the simulated point cloud, and the recognition success signal and deformation type are output.

[0138] If the coupling similarity value is less than or equal to the second determination threshold, it is determined that the real residual space is mainly composed of environmental noise or non-target objects, and noise interference or unknown object signals are output and a filtering operation is performed.

[0139] If the coupling similarity value is between the second determination threshold and the first determination threshold, it is marked as a suspected target, and the system is triggered to request multi-angle re-acquisition or manual review.

[0140] This embodiment is a further specification of the logical decision-making process after feature comparison in Embodiment 6;

[0141] The logic decision unit is configured with a preset first decision threshold. With the second judgment threshold The threshold was selected based on the receiver operating characteristic (TPR) curve: the similarity value corresponding to a true positive rate (TPR) of 98% or higher was chosen as the first judgment threshold. The similarity value corresponding to a false positive rate (FPR) of less than 1% was selected as the second judgment threshold. ;

[0142] Response to coupling similarity value Greater than or equal to The system determines that the object corresponding to the scene depth point cloud is the target object, and confirms that the object has undergone specific deformation consistent with the simulated point cloud, and outputs a recognition success signal.

[0143] Response to coupling similarity value Less than or equal to The system determines that the real residual space is mainly composed of environmental noise or non-target objects, outputs noise interference signals and performs filtering;

[0144] Response to coupling similarity value lie in and In between, the system marks it as a suspected target and triggers a system request for multi-angle re-collection or manual review;

[0145] The dual-threshold logic used in this embodiment effectively balances the false alarm rate and the false detection rate. It can identify real objects that have been deformed, and resolutely eliminate false objects caused by reflection, thus achieving dual protection of identification accuracy and safety in industrial settings.

[0146] Example 8:

[0147] Also includes:

[0148] An adaptive feedback module is used to optimize simulation parameters. When the topology coupling decision module outputs a successful recognition signal, it feeds back the feature parameters of the real residual space to the parameter injection simulation module.

[0149] The parameter injection simulation module responds to the feedback feature parameters and corrects the weight of the preset deformation factor so that the subsequently generated simulated point cloud more closely approximates the deformation law of objects in real logistics scenarios.

[0150] This embodiment introduces an adaptive feedback module to enable the system's self-evolution;

[0151] When the topology coupling decision module outputs a successful recognition signal, the adaptive feedback module is activated; during system initialization, the initial value of the deformation factor weight is set. Simultaneously set a safety threshold. For example, 10% of the object's characteristic dimensions, if the calculated If the threshold is exceeded, it will be forcibly truncated. To prevent the simulation model from becoming distorted due to excessive deformation;

[0152] This module extracts the statistical characteristics of the real residual space. Specifically, this includes the peak positions of the residual histogram. With half-width ;

[0153] The aforementioned feature parameters are fed back to the parameter injection simulation module to correct the weights of the preset deformation factors; to ensure the uniqueness of the symbols throughout the text and to avoid confusion with the geometric weight matrix W in Example 6, symbolic representation is used here. This represents scalar weights, and simultaneously extracts statistical features of the current simulation state point cloud residual space, namely the peak positions of the residual histogram. With half-width System construction scalar error index Defined as ,in, These are the preset weighted normalization coefficients; used to balance the influence weights of mean deviation and variance deviation, and satisfying the following conditions: + =1; The correction process uses the following incremental proportional feedback update formula:

[0154]

[0155] in, The source is calculated using the scalar error formula mentioned above, and its physical meaning is the scalar of the comprehensive deviation of the degree of deformation between reality and simulation;

[0156] The source is the deformation factor weight at the current time step, such as the maximum displacement parameter of the bulge model;

[0157] The source is a preset parameter, and its physical meaning is the learning rate parameter. The initial value is set to 0.05, and it can be dynamically adjusted according to the system's operational stability. To ensure iterative stability, The value range is set between [0.01, 0.1], and its dimension is defined as millimeters (mm), used to represent dimensionless error indicators. The displacement correction is converted into a geometric space; at the same time, the system sets the upper limit of the single correction step size to no more than 1% of the object feature size to prevent feedback oscillation.

[0158] The source is the residual statistical characteristics of the simulated point cloud generated under the current parameters;

[0159] The source is the difference calculation, and its physical meaning is the deviation in the degree of deformation between reality and simulation;

[0160] The parameter injection simulation module responds to the calculated The deformation amplitude will be adjusted in the next simulation cycle.

[0161] This embodiment realizes online learning of the algorithm. Through the above feedback loop, as the running time increases, the simulation model parameters of the system will converge to the actual working condition parameters of the logistics site, thereby continuously improving the recognition accuracy and speed and adapting to the specific deformation law caused by the specific stacking method.

[0162] Example 9:

[0163] The target object is a logistics packaging asset, including: pallets, turnover boxes or cartons; the non-rigid deformation includes bulging, collapse or corner crushing caused by stacking and compression;

[0164] The environmental interference factors include surface reflections and point cloud defects caused by transparent film wrapping or strapping.

[0165] This embodiment clarifies the specific application objects and scenario boundaries of the system;

[0166] The system defines the target objects as logistics packaging assets, specifically including pallets, turnover boxes, and cartons;

[0167] For the aforementioned assets, the non-rigid deformation identified by the system specifically refers to changes in the physical structure caused by stacking and compression, such as outward bulging of sidewalls or collapse of corners, rather than manufacturing tolerances.

[0168] The environmental interference factors processed by the system specifically refer to the laser penetration or total reflection caused by the wrapping of transparent films, which is common in logistics scenarios, as well as the point cloud segmentation caused by strapping.

[0169] This embodiment clarifies the applicable boundaries of the technical solution and makes targeted optimizations for specific pain points in the logistics industry, such as flexible packaging and highly reflective materials, ensuring the actual usability of the system in complex industrial environments and solving the failure problem of traditional rigid matching algorithms in such scenarios.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automatic identification system for packaged assets based on feature point clouds, characterized in that, include: The data acquisition module is used to acquire the scene depth point cloud of the target scene and retrieve the preset standard parametric model. The scene depth point cloud includes three-dimensional coordinate information and reflection intensity information. The standard parametric model is the pre-constructed three-dimensional geometric data of the target object. The ideal reference reconstruction module is used to construct an ideal geometric reference as noise-free reference data based on the scene depth point cloud and the standard parameterized model, estimate the six-dimensional pose of the standard parameterized model in the target scene through coarse registration, and perform virtual ray projection in this pose to generate the ideal reference point cloud. The parameter injection simulation module is used to build an active simulation engine. Based on the standard parameterized model, it uses preset deformation factors and environmental interference factors to adjust or superimpose noise on the geometric surface of the model to generate a simulated point cloud with specific defect labels. The simulated point cloud is used to simulate the non-rigid deformation and surface occlusion features of the target object in the physical environment. The dual-track difference extraction module is used to calculate the normal vector difference distribution between point clouds, and to calculate the real residual feature set of the scene depth point cloud relative to the ideal reference point cloud, which is defined as the real residual space, and the theoretical residual feature set of the simulated point cloud relative to the ideal reference point cloud, which is defined as the theoretical residual space. The topology coupling decision module is used to identify target objects based on topological similarity, extract high-dimensional features of the real residual space and the theoretical residual space, calculate the numerical matching degree between the two, and output the recognition result and specific status according to the matching degree.

2. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The data acquisition module includes: The scene perception unit is used to collect the scene depth point cloud in real time through an industrial 3D sensor. The scene depth point cloud is a discrete point set containing geometric information of the object surface and reflection intensity information. The knowledge base calling unit is used to store and provide the standard parametric model, which is the ideal geometric data of the target object and is associated with a common deformation topology library, which contains deformation pattern data of the object under pressure and impact physical conditions.

3. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The ideal benchmark reconstruction module includes: The pose estimation unit is used to calculate the spatial transformation matrix between the scene depth point cloud and the standard parametric model, and to determine the position and rotation angle of the standard parametric model in the current coordinate system. A clean projection unit is used to eliminate sensor measurement noise. Under the determined six-dimensional pose, the standard parameterized model is resampled using a virtual ray projection algorithm to generate the ideal reference point cloud, which has a theoretically smooth surface and complete geometric structure.

4. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The parameter injection simulation module includes: The deformation simulation unit is used to simulate changes in the physical structure of an object. Free deformation mesh control points are introduced on the mesh of the standard parametric model. According to the preset collapse model or bulge model, the spatial position of the control points is adjusted to generate an intermediate model that undergoes non-rigid deformation. The noise superposition unit is used to simulate surface covering interference. It superimposes a layer of random noise or fluctuation at a preset frequency on the surface of the intermediate state model to simulate the discontinuity of point cloud caused by film reflection or strapping. The unit then combines the intermediate state model to generate the simulated point cloud.

5. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The dual-track differential extraction module includes: The real-world computation unit is used to extract the mixed features of real damage and environmental noise, align the scene depth point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the real-world residual space. The theoretical field calculation unit is used to extract pure physical damage features, align the simulated point cloud with the ideal reference point cloud, calculate the difference in normal vectors and Euclidean distance between corresponding point pairs, and construct the theoretical residual space, which does not contain sensor random noise.

6. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The topology coupling decision module includes: The feature mapping unit is used to transform the residual space into feature vectors, and to encode the real residual space and the theoretical residual space respectively using a manifold harmonic analysis algorithm or a three-dimensional convolutional neural network to generate real feature vectors and theoretical feature vectors. The similarity calculation unit is used to quantify the matching degree between two residual spaces, calculate the cosine similarity or Euclidean distance between the real feature vector and the theoretical feature vector, and obtain the coupling similarity value.

7. The automatic packaging asset identification system based on feature point cloud according to claim 6, characterized in that, The topology coupling decision module further includes: a logic decision unit, configured to preset a first decision threshold and a second decision threshold, wherein the first decision threshold is greater than the second decision threshold; the logic decision unit executes the following decision logic: If the coupling similarity value is greater than or equal to the first determination threshold, then the object corresponding to the scene depth point cloud is determined to be the target object, and the object has undergone a specific deformation corresponding to the simulated point cloud, and a recognition success signal and deformation type are output. If the coupling similarity value is less than or equal to the second determination threshold, it is determined that the real residual space is mainly composed of environmental noise or non-target objects, and noise interference or unknown object signals are output and a filtering operation is performed. If the coupling similarity value is between the second determination threshold and the first determination threshold, it is marked as a suspected target, and the system is triggered to request multi-angle re-acquisition or manual review.

8. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, Also includes: An adaptive feedback module is used to optimize simulation parameters. When the topology coupling decision module outputs a successful recognition signal, it feeds back the feature parameters of the real residual space to the parameter injection simulation module. The parameter injection simulation module responds to the feedback feature parameters and corrects the weight of the preset deformation factor so that the subsequently generated simulated point cloud more closely approximates the deformation law of objects in real logistics scenarios.

9. The automatic packaging asset identification system based on feature point cloud according to claim 1, characterized in that, The target object is a logistics packaging asset, including pallets, turnover boxes, or cartons; The non-rigid deformations include bulges, collapses, or corner shrinkage caused by stacking and compression; The environmental interference factors include surface reflections and point cloud defects caused by transparent film wrapping or strapping.

Images