Intelligent identification type terminal sorting system, device and method

Abstract

The invention provides an intelligent identification type terminal sorting system, device and method, relates to the field of logistics automation, and solves the technical problems of sorting process interruption, efficiency reduction and excessive dependence on manual intervention caused by wear, fouling or unreadable article barcodes in the prior art. The method comprises the steps that multi-dimensional feature information of articles to be sorted is collected, and a multi-dimensional feature vector is generated; storing the multi-dimensional feature vector and the identification information of the to-be-sorted article in a feature database; when it is detected that the bar code of the article is unreadable, sequentially obtaining the physical dimension feature, the path topology feature and the visual feature of the current article, and sequentially carrying out feature matching on the multi-dimensional features of the current article by using a feature database; and identity information of the current article is determined according to a matching result, and sorting operation is completed.

Description

Technical Field

[0001] This invention belongs to the field of logistics automation, specifically an intelligent identification-based terminal sorting system, device, and method. Background Technology

[0002] In modern logistics sorting systems, barcode-based automatic identification technology is the core support for achieving efficient and accurate sorting. However, existing sorting methods heavily rely on the clarity and readability of barcodes. In actual operation, package barcodes are easily blurred due to wear, stains, tears, or printing quality issues during transportation, causing scanning equipment to fail to read them properly. When this happens, traditional sorting systems often come to a standstill: either problematic packages need to be removed from the automatic sorting line and transferred to manual processing stations for visual identification and manual input; or package information needs to be determined directly based on user feedback. Both significantly increase labor costs, drastically reduce overall sorting efficiency, and become bottlenecks for the throughput of logistics hubs. Essentially, existing technologies have failed to effectively solve the problem of continuous identification of items after physical tags fail. They either revert to inefficient manual modes or attempt to optimize on a single visual feature dimension, lacking a deep utilization of the inherent laws of logistics operations themselves. Summary of the Invention

[0003] This application provides an intelligent identification-based terminal sorting system, apparatus, and method, which solves the technical problem of sorting process interruption and efficiency reduction caused by wear, dirt, or unreadable item barcodes in the prior art.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, an intelligent identification-based terminal sorting method is provided, including: Collect multidimensional feature information of items to be sorted and generate multidimensional feature vectors; the multidimensional features include physical dimension features, path topology features and visual features. The multidimensional feature vector and the identification information of the items to be sorted are stored in the feature database; When an item barcode is detected to be unreadable, the physical dimension features, path topology features, and visual features of the current item are obtained in sequence, and the feature database is used to perform feature matching on the multidimensional features of the current item in sequence. The identity information of the current item is determined based on the matching results, and the sorting operation is completed.

[0005] Based on the above technical solutions, the intelligent identification-based terminal sorting method provided in this application effectively solves the problem of recognition failure caused by barcode wear in traditional sorting systems by constructing a multi-dimensional feature vector that includes physical dimensions, path topology, and visual features, and adopting a hierarchical and progressive matching strategy. This method incorporates logistics path topology into the feature system, using the movement patterns of items in the logistics network as the key identification basis, significantly reducing the limitations of relying solely on visual features. By performing feature matching sequentially from low to high computational complexity, both recognition efficiency and accuracy are ensured, achieving a balance between computational resource optimization and sorting efficiency. Furthermore, this method possesses strong fault tolerance and self-recovery capabilities. Even when the barcode is completely damaged, it can still accurately restore the item's identity through multi-dimensional feature cross-validation, significantly reducing the frequency of manual intervention and improving the automation level and robustness of the sorting system.

[0006] Furthermore, the collection of multidimensional feature information of the items to be sorted includes: The physical dimension features of the items to be sorted are collected by measuring equipment. The physical dimension features include at least length, width, height and weight information. The original measurement values ​​are dynamically calibrated. The dynamic calibration process includes temperature compensation, equipment error correction and transportation deformation prediction compensation. Based on the sequence of logistics path nodes, path topology features are collected, the starting node, the sequence of nodes passed through and the timestamp of the item are recorded, and the transition probability between nodes is calculated to generate a path topology feature vector. Visual features of objects are acquired through image acquisition devices, and a hierarchical feature extraction strategy is adopted. Lightweight visual features are extracted first, and depth visual features are extracted when the lightweight visual features fail to match. The lightweight visual features include color histograms, texture features, and shape features, and the depth visual features are high-dimensional feature vectors extracted based on convolutional neural networks. Physical dimension features, path topology features, and visual features are combined into a multi-dimensional feature vector, which is then associated with and stored with item identification information.

[0007] Furthermore, the step of sequentially performing feature matching on the multidimensional features of the current item using the feature database includes: The first set is obtained by performing a first-level match between the physical dimension features of the current item and the physical dimension features in the feature database; When there are multiple identifiers in the first set, a second-level matching is performed based on the path topology features of the current item and the path topology features in the first set to obtain the second set; When there are multiple identification information in the second set, a third-level matching is performed based on the visual features of the current item and the visual features in the second set to obtain the third set; If multiple identifiers exist in the third set, a manual intervention process is initiated; otherwise, the barcode of the current item is determined based on the unique identifier in the first, second, or third set.

[0008] Furthermore, the first-level matching employs a fast-computation screening algorithm, including: Calculate the physical attribute similarity between the current item and candidate items in the feature database. The physical attribute similarity includes volume similarity, density similarity, and dimensional ratio similarity obtained based on the physical dimension features. The dimensional ratio similarity refers to the degree of similarity between the current item and the candidate item in terms of three-dimensional size ratio. The physical similarity scores are obtained by weighted summation of the values ​​of each physical attribute similarity; where volume similarity has a weight of 0.4, density similarity has a weight of 0.3, and dimensionality similarity has a weight of 0.3. Set a similarity threshold. When the physical similarity score is higher than the preset similarity threshold, the corresponding candidate item will be included in the first set.

[0009] Furthermore, the specific process of the second-level matching includes: Extract the historical path topology feature sequence of each first candidate item in the first set; based on the directed graph of the logistics network, verify whether there is a directed edge from the latest record node of each first candidate item to the node where the current item is located; if yes, then obtain the node transition probability from the previous node of the first candidate item to the current node and the travel time weight of the directed edge based on historical logistics data; if no, determine that the matching has failed. The path consistency score PC of the first candidate item is calculated based on the historical path topological feature sequence; the path consistency score is used to characterize the path continuity and rationality of the candidate item from the historical path node to the current node, and reflects the spatiotemporal consistency of the path sequence. Based on the real-time traffic congestion index, traffic impact coefficient, and travel time weight, the theoretical transmission time of the current item from the previous node to the current node is calculated. Calculate the deviation between the current node's time and the theoretical transmission time to obtain the time deviation. Calculate the similarity between the topological path feature sequence of the current item and the historical path topological feature sequence of each first candidate item to obtain the path similarity. The path matching degree is calculated based on the node transition probability, the path consistency, the theoretical transmission time, and the path similarity. Set a path matching threshold. When the path matching degree exceeds the path matching threshold, the corresponding first candidate item will be included in the second set. The candidate items in the second set are sorted in descending order of path matching degree.

[0010] Furthermore, the formula for calculating the path consistency score PC is: Where k represents the node index in the historical path topology feature sequence, and n represents the number of nodes. Indicates the time of the current node. This represents the time of the k-th node, where τ is the decay time constant. This represents the k-th node of the first candidate item. Let P represent the current node, and let P represent the node transition probability.

[0011] Furthermore, the theoretical transmission time The formula for calculation is: ;in, The travel time weight of the directed edge. Traffic impact coefficient, Real-time traffic congestion index; The time deviation The formula for calculation is: ; This is the timestamp of the most recently recorded node in the historical path topological feature sequence. This is the timestamp of the current node.

[0012] Furthermore, the path matching degree The formula for calculation is: ;in, Let represent the node transition probability, PS represent the path similarity, α represent the time deviation penalty coefficient, and β represent the path consistency weight coefficient.

[0013] Furthermore, the inherent matching logic of the second-level matching is as follows: Check the connectivity between the current node and its predecessor node in the logistics network topology, and use a graph traversal algorithm to verify whether there is a reasonable connection in the path; The feasibility of the time window is calculated based on the distance between nodes, the type of transportation, and the real-time traffic conditions to determine whether the time it takes for an item to travel from the previous node to the current node is within a reasonable range. To assess path consistency, a weighted sum of transition probabilities from each historical node to the current node is calculated based on the historical path sequence, and a time decay factor is introduced to emphasize the importance of recent path nodes. To calculate path pattern similarity, a dynamic time warping algorithm is used to compare the morphological consistency between the current path feature sequence and the historical path sequence, thereby capturing the dynamic patterns of path changes. By considering path connectivity, time window feasibility, path consistency, and path pattern similarity, a path matching degree is generated through a nonlinear fusion formula. The first candidate items with matching degrees exceeding the path matching threshold are selected to form a second set.

[0014] Furthermore, the third-level matching employs a multimodal feature fusion mechanism, including: Based on the lightweight visual features of each second candidate item in the second set and the current item, calculate the color histogram similarity, texture feature similarity, and shape feature similarity to obtain a lightweight visual feature similarity vector. When the similarity value in the lightweight visual feature similarity vector is lower than a preset threshold, the high-dimensional feature vector extracted by the convolutional neural network is used to calculate the similarity and obtain the deep visual feature similarity; the convolutional neural network is embedded with an attention module to improve the comparison of key features of the items. The lightweight visual feature similarity vector or the deep visual feature similarity of each second candidate item is weighted and fused with the physical similarity score and path matching degree to obtain the final matching confidence. A third set is generated based on the matching confidence scores.

[0015] Furthermore, the manual intervention process includes: The current item is transferred to the manual processing area for temporary storage, and the operator is shown detailed information of all candidate items in the third set. The detailed information includes at least the identification information, matching score and key feature comparison obtained by deep visual feature matching of each candidate item. The update status of the path topology feature sequence of each candidate item in the third set is monitored in real time. When a record of any candidate item being successfully scanned in a subsequent node of the logistics network is received, the non-association of the candidate item with the current item is determined and it is removed from the third set. Candidate items are dynamically filtered based on the update status of the path topology feature sequence. When only one candidate item remains in the third set, the system automatically determines the identification information of that candidate item as the final barcode of the current item. After the final barcode of the item is confirmed through a manual intervention process, the corresponding barcode label is regenerated and affixed using a barcode printing device. Record the entire process data and confirmation results of this manual intervention, which will be used to update the feature database and optimize the parameters of the matching model.

[0016] Secondly, an intelligent identification terminal sorting device is provided, comprising: a communication unit and a processing unit; The communication unit is used to interact with the feature database, image acquisition equipment, measurement sensors and logistics path information system, receive multi-dimensional feature information of items to be sorted, and send sorting control instructions. The processing unit is connected to the communication unit and is used to control the communication unit to collect multi-dimensional feature information of the items to be sorted and generate multi-dimensional feature vectors. The multi-dimensional features include physical dimension features, path topology features and visual features. The multidimensional feature vector is associated with the identification information of the items to be sorted and stored in the feature database through the communication unit; When the communication unit detects that the item barcode is unreadable, the physical dimension features, path topology features, and visual features of the current item are obtained in sequence, and the multidimensional features of the current item are matched sequentially using the feature database; the identity information of the current item is determined according to the matching result, and an instruction is sent through the communication unit to complete the sorting operation.

[0017] Thirdly, this application provides an intelligent identification-based terminal sorting device, comprising: a processor and a storage medium; the storage medium includes instructions, and the processor is used to execute the instructions to implement the method described in the first aspect and any possible implementation thereof. This terminal sorting and scanning device can be an electronic device or a chip within an electronic device.

[0018] Fourthly, this application provides an intelligent identification-based terminal sorting system, comprising: a feature acquisition module, a storage management module, a hierarchical matching module, and a sorting execution module; wherein, The feature acquisition module is used to acquire multi-dimensional feature information of the items to be sorted and generate multi-dimensional feature vectors. The multi-dimensional features include physical dimension features, path topology features and visual features. The storage management module is connected to the feature acquisition module and is used to associate the multidimensional feature vector with the identification information of the items to be sorted and store it in the feature database. The hierarchical matching module is connected to the storage management module. When the barcode of an item is detected to be unreadable, it sequentially acquires the physical dimension features, path topology features, and visual features of the current item, and uses the feature database to sequentially perform feature matching on the multidimensional features of the current item, and determines the identity information of the current item based on the matching results. The sorting execution module is connected to the hierarchical matching module and is used to complete the sorting operation of the current item based on the determined identity information.

[0019] Fifthly, this application provides a computer-readable storage medium storing instructions that, when executed on an intelligent identification terminal sorting device, cause the intelligent identification terminal sorting device to perform the method described in the first aspect and any possible implementation thereof.

[0020] In a sixth aspect, this application provides a computer program product containing instructions that, when the computer program product is run on an intelligent identification terminal sorting device, cause the intelligent identification terminal sorting device to perform the methods described in the first aspect and any possible implementation thereof.

[0021] Compared with the prior art, the beneficial effects of this application are: (1) By using a multi-dimensional feature fusion and hierarchical matching mechanism, the success rate and efficiency of item identification in the case of barcode damage are improved. Specifically, physical dimension features are used for fast and low-computational-cost first-level matching, which can quickly narrow down the candidate range; then, path topology features are used for efficient second-level matching, and identity inference is performed by verifying path connectivity, time window feasibility and path pattern similarity, and by using the constraints of the logistics network itself; finally, computationally intensive visual feature matching is only initiated when necessary. This strategy of progressing from simple to complex avoids the waste of resources in traditional schemes that directly perform high-cost image matching, and achieves the best balance between recognition accuracy and processing efficiency.

[0022] (2) This application introduces and deeply utilizes path topology features to construct a spatiotemporal context-aware sorting model, enhancing the robustness and intelligence of the system. This application not only uses path information as a feature but also quantifies the rationality of the item's movement trajectory through a mathematical model. This method does not rely on the integrity of the item's appearance; even if the barcode and surface information are completely worn, reliable reasoning can be performed based on the logistics rules of where the item comes from and where it should go, solving the failure problem of traditional vision solutions when severely soiled.

[0023] (3) Through dynamic adaptive strategies and algorithm design, the system resource allocation and performance were optimized. For example, the subsequent matching strategy was dynamically adjusted according to the number of candidates in the first set, avoiding unnecessary computational overhead; dynamic calibration of physical measurement values ​​for temperature, deformation, etc., improved the reliability of basic data; real-time traffic data was introduced into path matching, making the calculation of theoretical transmission time more consistent with the actual situation; and a hierarchical strategy was adopted for visual feature matching, prioritizing the use of lightweight features. These methods enable the system to intelligently adapt to different sorting scenarios, package types, and network conditions, achieving adaptive optimization between resource consumption and performance requirements. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A system architecture diagram of an intelligent identification terminal sorting system provided in this application embodiment; Figure 2 A flowchart illustrating an intelligent identification-based terminal sorting method provided in an embodiment of this application; Figure 3 A flowchart illustrating another intelligent identification terminal sorting method provided in an embodiment of this application; Figure 4 A flowchart illustrating another intelligent identification terminal sorting method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an intelligent identification terminal sorting device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the hardware structure of an intelligent identification terminal sorting device provided in an embodiment of this application. Detailed Implementation

[0026] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0027] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0028] The intelligent identification terminal sorting method provided in this application embodiment can be applied to, for example... Figure 1 In an intelligent identification-based terminal sorting system, such as... Figure 1 As shown, the system includes: a feature acquisition module, a storage management module, a hierarchical matching module, and a sorting execution module, all connected in communication; wherein, The feature acquisition module is used to collect multi-dimensional feature information of the items to be sorted and generate multi-dimensional feature vectors, including physical dimension features, path topology features and visual features. The storage management module is used to associate multidimensional feature vectors with the identification information of the items to be sorted and store them in the feature database; The hierarchical matching module is used to sequentially obtain the physical dimension features, path topology features, and visual features of the current item when the item barcode is detected to be unreadable. It then uses the feature database to perform feature matching on the multidimensional features of the current item in sequence, and determines the identity information of the current item based on the matching results. The sorting execution module is used to complete the sorting operation of the current item based on the determined identity information.

[0029] To address the technical problem in existing technologies where item identification fails when barcodes are worn, soiled, or obscured, leading to sorting process stagnation and low efficiency, this application provides an intelligent identification terminal sorting system, apparatus, and method. The method includes: Collect multidimensional feature information of items to be sorted and generate multidimensional feature vectors, including physical dimension features, path topology features and visual features; The multidimensional feature vectors are associated with the identification information of the items to be sorted and stored in the feature database; During the sorting process, when an item's barcode is detected to be unreadable, a hierarchical matching process is initiated. This process sequentially calls features from different levels of the feature database, in ascending order of computational complexity, to match the multidimensional feature vector of the current item. The hierarchical matching process includes: The first-level matching based on physical dimension features uses the physical attributes of items for rapid screening. The second-level matching based on path topology features is validated using logistics path network constraints. The third-level matching based on visual features uses visual feature information for accurate identification. The identity information of the current item is determined based on the matching results, and the sorting operation is completed.

[0030] Based on this, the method breaks through the single dependence of traditional barcode recognition by using multi-dimensional feature complementarity and hierarchical matching strategies. It can still efficiently and accurately identify the identity of items in barcode failure scenarios, which can reduce the cost of manual intervention and improve the automation level and robustness of sorting systems.

[0031] like Figure 2 As shown in the embodiment of this application, an intelligent identification terminal sorting method includes: S1. Collect multi-dimensional feature information of items to be sorted and generate multi-dimensional feature vectors.

[0032] Among them, multidimensional feature vectors include, but are not limited to, physical dimension features, path topology features, and visual features. Physical dimension features are the inherent quantifiable attributes of an item, such as length, width, height, and weight. These can be quickly collected using sensor technology, and the differences in these attributes can be used for preliminary and rapid screening of items. Path topology features are the trajectory information formed by the flow of items in the logistics network, including the starting node, the sequence of nodes passed through, and the timestamps of each node. Based on this information, the identity of the item can be initially verified by leveraging the flow patterns of the logistics network. Visual features are the appearance representation information of the item, including color, shape, and texture. They provide identification basis from the visual presentation dimension of the item, supplementing the identification limitations of physical and path features.

[0033] In some implementations, existing conventional measurement equipment such as weighing sensors and laser rangefinders can be used to collect physical dimensional features, the path topology features can be recorded in real time through the logistics management system, and visual features can be collected using ordinary cameras or industrial cameras.

[0034] S2. Associate the multidimensional feature vector with the identification information of the items to be sorted and store it in the feature database.

[0035] The feature database stores the correspondence between multi-dimensional feature vectors and unique identification information for all items to be sorted. This identification information includes, but is not limited to, barcodes, order numbers, and recipient information. During logistics and transportation, barcodes can easily become unreadable due to various common factors such as transportation friction, humid environments, and human-caused damage. In such cases, the associated data stored in the feature database comes into play, allowing the identification of the item corresponding to the invalid barcode to be confirmed by matching the feature information of the current item.

[0036] For example, the length, width, height, weight, path node sequence, and visual feature data of the package are bound to the package's barcode and stored in a MySQL database, with each barcode corresponding to a unique set of multidimensional feature vectors.

[0037] S3. When the item barcode is detected to be unreadable, the physical dimension features, path topology features and visual features of the current item are obtained in sequence, and the feature database is used to perform feature matching on the multi-dimensional features of the current item in sequence.

[0038] Before feature matching, it's crucial to constantly monitor the scanned information at the current node to systematically eliminate items whose identities have already been verified, preventing duplicate matching. Furthermore, performing feature matching sequentially is for computational optimization. Matching physical dimension features has the lowest computational complexity and is the fastest. In most cases, if physical dimension feature matching can directly identify a unique item, subsequent path and visual feature matching processes are unnecessary, reducing computational resource consumption and sorting time.

[0039] In some implementations, the matching process can be triggered by the signal feedback from the scanning device. According to the preset feature calling order, the physical features of the current item are first extracted and compared with the data in the feature database. Then, based on the matching result, it is decided whether to start the matching of subsequent levels.

[0040] For example, when the barcode of a package is blurry and cannot be read, its physical characteristics such as weight and size are first obtained and compared with the physical characteristics stored in the feature database to filter out a set of candidate packages with the same physical attributes. If the candidate set contains only one package, its identity is directly confirmed without further matching. If there are still multiple candidates, the path topology features are called for secondary filtering.

[0041] S4. Determine the identity information of the current item based on the matching results and complete the sorting operation.

[0042] Once the identity information is confirmed, the item status data in the feature database needs to be updated in a timely manner, including the current sorting node and matching results. At the same time, the corresponding valid barcode can be regenerated and pasted through existing barcode printing equipment to facilitate the subsequent logistics of the items.

[0043] Based on the above technical solutions, the intelligent identification terminal sorting method provided in this application constructs a multi-dimensional identity recognition system by collecting multi-dimensional and complementary item features, which can improve the recognition success rate in barcode failure scenarios; by using a low-complexity feature screening and refinement method, the computational overhead is reduced while ensuring recognition accuracy; the whole method can solve the technical problem of reduced sorting efficiency caused by missing barcode information in traditional sorting systems, and improve the automation level and overall efficiency of logistics sorting.

[0044] In one possible implementation of this application embodiment, the above-mentioned S1 can be specifically implemented by the following S101, S102, S103 and S104, which are described in detail below: S101. Collect the physical dimension characteristics of the items to be sorted.

[0045] Among them, physical dimension features are the inherent quantifiable attributes of the items to be sorted. They are the basic features for subsequent rapid item screening. By collecting the physical dimension features of the items, a large number of unsuitable candidate items can be eliminated based on simple attribute differences, reducing the computational pressure of subsequent matching.

[0046] In some implementations, the physical dimensional characteristics of the items to be sorted can be collected using measuring devices such as weighing sensors and laser rangefinders, including at least length, width, height, and weight information. At the same time, the collected raw measurement values ​​need to be dynamically calibrated, specifically including temperature compensation, equipment error correction, and transportation deformation prediction compensation. Among them, transportation deformation prediction compensation refers to predicting and correcting possible dimensional deviations based on the material characteristics of the items and common deformation patterns in the transportation process. Deformation prediction methods based on material classification and pressure models can usually be used for compensation.

[0047] In some implementations, deformation prediction compensation may include the following steps: 1. Material type identification.

[0048] Using a hyperspectral camera or an infrared sensor of a specific wavelength installed at the measurement station, the outer packaging material of the item is quickly scanned and analyzed.

[0049] Using pre-trained material classification models, such as image classification models based on convolutional neural networks (CNNs) or classifiers based on spectral features, packaging materials are identified into predefined categories, such as rigid cardboard boxes, corrugated cardboard, foam boxes, woven bags, plastic bags, and metal cans. Each material category corresponds to different deformation characteristic parameters.

[0050] 2. Obtaining deformation characteristic parameters.

[0051] The system has a pre-defined "Material-Deformation Parameter" lookup table. This table defines a set of deformation parameters for each material category, for example: Vertical compression deformation coefficient (Kv): represents the expected deformation ratio of an item in the height direction under stacking pressure.

[0052] Lateral deformation coefficient (Kh): represents the deformation potential of an item in the length and width directions when subjected to lateral compression or when the internal filling is not full.

[0053] Deformation recovery time constant (T): represents the estimated time required for an item to return to its original size after the pressure is released.

[0054] 3. Deformation calculation and compensation.

[0055] Estimate the static pressure F that the item is currently bearing based on its condition on the sorting line; The deformation ΔD in the main deformation direction is predicted using a formula. For example, a simplified linear prediction model is: ΔD = Kv F; where Kv is the vertical compression deformation coefficient corresponding to the current material of the item, and F is the estimated static pressure; The original measurements are compensated to obtain dimensions that more closely approximate the item's dimensions under standard, stress-free conditions. D_compensated = D_measured + ΔD; where D_compensated represents the corrected physical dimension feature value, and D_measured represents the original physical dimension value obtained from the acquisition.

[0056] S102. Collect the path topology features of the items to be sorted.

[0057] Among them, path topology features are the spatiotemporal trajectory information formed by the flow of goods in the logistics network. They can reflect the source and flow patterns of goods and are the core features that break through the dependence on the appearance of goods and realize scene-based recognition. Even if the barcode or appearance of the goods is damaged, its movement path in the logistics network still has uniqueness and regularity.

[0058] In some implementations, collecting path topology features based on the sequence of logistics path nodes may include: The logistics management system records the origin, sequence of nodes visited, and timestamp of each node for each item. Then, based on historical logistics data, the transfer probability between each node is calculated, and finally, the starting node, the sequence of passing nodes, the timestamp, and the node transfer probability are integrated to generate a path topology feature vector.

[0059] In some implementations, the calculation process for node transition probabilities is as follows: First, extract all item circulation data for the specified statistical period from the historical database of the logistics management system, and exclude abnormal data; Then, it is clarified that the calculation object of the node transfer probability is the transfer pair of source node-target node, denoted as i->j, where i represents the starting node of the item and j represents the next node that the item arrives at after starting from i; for example, "Guangzhou warehouse (G01) → Shenzhen transit center (S02)" and "Shenzhen transit center (S02) → Hangzhou sorting center (H03)" are independent node transfer pairs; Next, for each node transition pair i->j, two core data points are calculated: first, the total number of all items originating from the source node i within the statistical period, denoted as N. i Second, the number of items that start from source node i and eventually reach target node j within the statistical period, denoted as N. ij ; According to statistics N i With N ij The node transition probability P(i->j) is calculated by estimating the probability through frequency, and the formula is: P(i->j) = N ij / N i Where, if N i=0, indicating that no items originated from source node i during the statistical period. This is usually a newly enabled node, so P(i->j) = 0.01 is set to a minimum value to avoid calculation errors with a denominator of 0, and also indicates that the transfer path has not yet been verified; if N ij If P(i->j) = 0, then P(i->j) = 0, indicating that there is no actual flow record for this transfer path within the statistical period.

[0060] For example, a clothing package departs from the origin node "Guangzhou Warehouse (G01)" and passes through "Shenzhen Transit Center (S02)" and "Hangzhou Sorting Center (H03)" in sequence. The timestamps of each node are "T1=2024-05-01 08:30" (departing from G01), "T2=2024-05-01 12:15" (arriving at S02), "T3=2024-05-01 18:40" (departing from S02), and "T4=2024-05-02 09:20" (arriving at H03). Based on historical data statistics, the transfer probability of G01→S02 is 0.92, and the transfer probability of S02→H03 is 0.85. The final generated path topology feature vector contains core information such as "G01→S02→H03", "T1-T4", and "0.92, 0.85".

[0061] S103. Collect visual characteristics of items to be sorted.

[0062] In some implementations, visual features of objects can be acquired using image acquisition devices such as industrial cameras and high-definition cameras, and a hierarchical feature extraction strategy can be adopted: First, lightweight visual features are extracted, including color histograms, texture features, and shape features. When lightweight visual features are insufficient for matching, a high-dimensional feature vector is extracted using a Convolutional Neural Network (CNN) to obtain deep visual features for subsequent feature matching. The CNN needs to embed an attention module to enhance the extraction and comparison of features in key areas of the item (such as packaging logos and label positions), thereby improving feature discriminative power.

[0063] S104. Combine multidimensional features and associate them with identification information.

[0064] In some implementations, the physical dimension features obtained in S101, the path topology features obtained in S102, and the visual features obtained in S103 are combined into a multi-dimensional feature vector according to a preset format; then, the multi-dimensional feature vector is associated with the identification information of the items to be sorted and stored in the feature database through a data interface; wherein, the identification information includes, but is not limited to, barcodes, order numbers, unique item codes, etc.

[0065] Based on the above technical solution, step S1 constructs a comprehensive and reliable item feature system. Among them, physical dimension features provide a foundation for rapid screening, path topology features break through appearance dependence, and visual features supplement detailed differences. The three work together to cover the feature requirements under different sorting scenarios. At the same time, dynamic calibration and hierarchical extraction strategies respectively ensure feature accuracy and computational efficiency, laying a data foundation for the efficient and accurate execution of the subsequent hierarchical matching process.

[0066] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Standardize and normalize the collected multidimensional feature vectors to form standard feature vectors.

[0067] Standardization aims to eliminate differences in the dimensions and orders of magnitude of different features, while normalization aims to map feature values ​​to a unified numerical range, laying the foundation for subsequent feature matching calculations.

[0068] In some implementations, Z-Score normalization is used for physical dimension features. For path topology features, such as node transition probabilities, which are already probability values ​​within the range [0,1], no additional processing is needed. For visual feature vectors, L2 normalization is used. That is, for a vector V=[v1,v2,...,vn], the normalization formula is: V_norm=V / ||V||2, where ||V||2 is the L2 norm of the vector. This transforms the visual feature vector into a unit vector, making subsequent similarity calculations more accurate and efficient.

[0069] S202. Associate the standard feature vector with the item identification information and construct an efficient index structure.

[0070] In this context, association refers to binding and storing the unique identifier of an item with processed standard physical dimension feature vectors, path topology feature vectors, lightweight visual feature vectors, etc., in a single record in the database.

[0071] In some implementations, to support fast retrieval in subsequent hierarchical matching, specialized indexes need to be built for different types of features. For example, a composite B+ tree index is built for physical dimension features to quickly filter candidate items whose volume and weight are within tolerance ranges during the first level of S3 matching. A hash index is built for the latest record node field in path topology features to quickly locate all candidate items whose latest position is the same as the current node during the second level of matching. Locality-sensitive hash indexes or quantization indexes are built for lightweight visual feature vectors to achieve fast near-neighbor search for high-dimensional vectors.

[0072] It should be noted that for deep visual feature vectors, which have high computational and storage costs, a hierarchical storage strategy can be adopted. That is, lightweight visual feature vectors are directly associated with the identification information and reside in memory or a high-speed database, while deep visual feature vectors are stored in object storage or a distributed file system, and are only loaded based on the identification information when needed for third-level matching. This approach can significantly reduce the storage pressure and access latency of the main feature database.

[0073] For example, in a relational database, a table named parcel_features can be designed, containing fields such as waybill_id (primary key), phys_dim_vector (physical feature vector), route_topo_vector (path topology vector), light_visual_vector (lightweight visual vector), and deep_visual_vector_blob_id (reference to the storage location of the deep feature vector), and the aforementioned indexes can be created for the key fields in phys_dim_vector and route_topo_vector, as well as light_visual_vector.

[0074] S203. Persistently store the associated and indexed feature data to the feature database, and record the data version and metadata.

[0075] Persistent storage ensures that data is not lost after a system restart. Recording metadata ensures data traceability and manageability, and provides a basis for lifecycle management of feature data.

[0076] In some implementations, metadata should at least include: data entry timestamp, data version number, and validity period of feature data. For example, route topology features can be set to be retained for 30 days after successful item delivery, after which they are automatically archived or deleted.

[0077] It should be noted that the selection of feature databases can be flexibly chosen according to the data scale and performance requirements. For example, a hybrid architecture combining relational databases and vector databases can be adopted, using relational databases to manage structured metadata and identification information, and using vector databases to efficiently process the similarity search of high-dimensional feature vectors.

[0078] Based on the above technical solution, through standardization processing, associated index construction, and persistent storage in steps S201 to S203, this application establishes a well-structured and highly efficient feature database. This database not only comprehensively records the multidimensional identity information of items, but more importantly, its index structure provides core support for the rapid execution of the hierarchical matching process in S3. This enables the rapid location of candidate items even with massive amounts of data through a coarse-to-fine retrieval strategy, thereby significantly improving sorting efficiency.

[0079] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302, and S303, which are explained in detail below: S301. Use physical dimension features to perform first-level matching and quickly screen the candidate item set.

[0080] In some implementations, matching logic based on a fast calculation screening algorithm is used, specifically including the following steps: The first step is to calculate the physical attribute similarity between the current item and the candidate items in the feature database, including volume similarity, density similarity, and dimensional ratio similarity. Volume similarity is calculated by the ratio of the volume difference between the current item and the candidate item to the volume of the candidate item. Density similarity is calculated by the ratio of the density difference between the two items to the density of the candidate item (density = weight / volume). Dimensional ratio similarity refers to the similarity between the current item and the candidate item in the three-dimensional dimension ratio of length:width:height, and is calculated by the mean of the ratio differences. The second step is to perform a weighted summation of the various values ​​of the physical attribute similarity to obtain the physical similarity score. The calculation formula is as follows: Physical similarity score = volume similarity × 0.4 + density similarity × 0.3 + dimensionality similarity × 0.3; among which, the volume similarity weight is set to 0.4 because volume is less affected by the environment and has high stability, and contributes the most to the distinction of items. The density similarity and dimensionality similarity weights are each set to 0.3. The third step is to set a preset similarity threshold, typically between 0.8 and 0.9. When the physical similarity score is higher than this threshold, the corresponding candidate item is included in the first set. Simultaneously, the subsequent matching strategy can be adjusted based on the number of candidate items in the first set: if the number of candidates is less than the preset value, path topology feature verification can be directly initiated; if the number of candidates is greater than the preset value, lightweight visual feature matching is initiated first to further narrow down the scope.

[0081] It should be noted that the core advantage of the fast screening algorithm lies in its low cost and high efficiency. Compared with path topology or visual features, the calculation of physical attributes only requires basic numerical operations and does not require complex sequence analysis or image processing. It can complete the screening of a large number of candidate items in milliseconds. The weight allocation design is based on the stability difference of physical features, ensuring that features with higher distinguishability of items play a greater role and avoiding invalid matching.

[0082] S302. Perform a second-level matching based on the path topology features and narrow down the scope using logistics network constraints.

[0083] Among them, path topology features reflect the flow trajectory and patterns of goods in the logistics network and have spatiotemporal uniqueness—even if the physical attributes of goods are similar, their path nodes and timestamps will be different. Therefore, as the second-level matching link, its core role is to use the constraints of the logistics network to further filter candidate items in the first set, improve matching accuracy, and avoid the limitations of relying solely on physical features.

[0084] In some implementations, second-level matching includes the following steps: The first step is to extract the historical path topological feature sequence for each candidate item in the first set. ,in For each candidate item in the first set, The timestamp of reaching the latest record node; The second step involves verifying the latest record node of candidate items based on a directed graph of the logistics network (nodes being sorting centers / transfer stations, and directed edges being logistics channels between nodes). Go to the node where the current item is located Check if there is a directed edge; if not, the match fails and the element is removed from the first set. The third step is to calculate the path consistency score (PC), using the following formula: Where k is the node index in the historical path sequence; The timestamp of the current node. τ is the timestamp of the kth historical node, and τ is the decay time constant, usually set to 24-48 hours, to reduce the impact of early nodes on the score and emphasize the importance of recent paths; The probability of transitioning from the k-th historical node to the current node is given. It should be noted that the path consistency score is used to quantify the continuity and rationality of candidate items from historical paths to the current node. The higher the PC value, the more coherent the path. The fourth step is to calculate the theoretical transmission time. Time deviation from the current node time : The formula is Where W(e) is the directed edge travel time weight from the latest node of the candidate item to the current node, which is the statistically derived transmission time under normal road conditions. The traffic impact coefficient, typically set between 0.1 and 0.3, reflects the degree to which traffic congestion affects travel time. This is a real-time traffic congestion index, with values ​​ranging from 0 to 1, where 0 represents no congestion and 1 represents extreme congestion. The formula is Its function is to measure the deviation between the current arrival time of an item and the theoretical arrival time. The smaller the value, the more reasonable the timing. Step 5: Calculate the path matching score PScore, using the following formula: ;in, This represents the transition probability from the latest node to the current node for candidate items. This is the time deviation penalty coefficient, usually set to 1~2. The larger the deviation, the heavier the penalty. is the path consistency weight coefficient, usually set to 0.5~1, used to improve the contribution of PC to the matching degree, and PS is the path similarity between the current item and the candidate item; The sixth step is to set a path matching threshold, usually 0.7 to 0.8, and include candidate items with PS scores higher than the threshold into the second set, and sort them in descending order of PS score.

[0085] In some implementations, path similarity can be calculated in the following ways: Let the path topological feature sequence of the current item be denoted as The historical path topological feature sequence of candidate items is... ; The two sequences were then preprocessed to remove invalid node records and standardize the node encoding format. For preprocessed and Construct an m×n distance matrix D, with matrix elements d ij express The i-th "node-timestamp" pair and The j-th node-timestamp pair The overall distance between them is calculated as follows: ;in: Let be the node distance, if = That is, if the two nodes are the same logistics node, then =0; if the two nodes are different, then =1, that is, quantifying whether nodes match by encoding with 0-1, highlighting the impact of node consistency on path similarity; For time distance, time difference normalization is used for calculation: ,in The maximum value of all timestamps in the two sequences. It is the minimum value of all timestamps in the two sequences; Node Distance Weight Weighting based on time distance The preset coefficients, and + =1, usually set =0.6、 =0.4.

[0086] The Dynamic Time Warping (DTW) algorithm is used to find the distance matrix D from the top left corner d. 11 Go to the bottom right corner d mn The optimal matching path must satisfy monotonically increasing conditions; let the elements on the optimal path be... Let k be the number of elements contained in the path, then the total path distance is: ; Total DTW distance Normalized to the 0-1 interval, it is converted into path similarity PS, and the calculation formula is as follows: ;in, The maximum possible length of the optimal path is m+n-1; This represents the maximum possible value of the distance matrix element; through this formula, The smaller the value, the smaller the difference between the two sequences, and the closer the PS is to 1; conversely, the larger the value, the closer it is to 0.

[0087] S303. Third-level matching based on visual features to accurately match the identity of items.

[0088] Among them, visual features are detailed representations of the appearance of an item, which can capture differences that physical and path features cannot cover. Therefore, as the third-level matching link, accurate identification is achieved through multimodal feature fusion. At the same time, when visual feature matching still cannot lock in a unique identity, a manual intervention process is triggered to ensure that no one is missed in the sorting and to ensure the accuracy of the matching.

[0089] In some implementations, third-level matching may include the following steps: The first step is lightweight visual feature priority matching: extract lightweight visual features between the current item and candidate items in the second set, and calculate the similarity of each feature, including color histogram similarity calculated by Bach distance, texture feature similarity calculated by gray-level co-occurrence matrix, and shape feature similarity calculated by Hu moment, to form a lightweight visual feature similarity vector. The second step is to supplement and match deep visual features: If any value in the lightweight visual feature similarity vector is lower than a preset threshold (such as 0.75), a convolutional neural network is used to extract the deep visual features of the current item and the candidate items. Among them, an attention module can be embedded in the CNN to focus on key areas of the item, such as packaging logo, label, color distribution, object shape, etc., to improve feature discrimination. Then, the deep visual feature similarity is calculated through cosine similarity. The third step is to calculate the matching confidence score through multimodal feature fusion: The lightweight / deep visual feature similarity score is weighted and fused with the physical similarity score in S301 and the path matching score PScore in S302 to obtain the matching confidence score. The calculation formula is as follows: Match confidence score = physical similarity score × 0.2 + path matching score × 0.3 + visual feature similarity score × 0.5; where, since visual features have the highest distinguishability of item details, the visual feature similarity score is set to 0.5; candidate items are sorted in descending order according to match confidence score to generate a third set; The fourth step is manual intervention to trigger judgment: If there is only one candidate item in the third set, its identification information is directly identified as the current item's identity; if there are multiple candidate items, the manual intervention process is initiated—the current item is transferred to the manual processing area for temporary storage, and the identification information, matching score, and key comparison of deep visual features of the candidate items are displayed to the operator; at the same time, the path topology feature update status of the candidate items is monitored in real time. If a candidate item is successfully scanned in a subsequent node, it is determined that it is not related to the current item and is removed from the third set; when there is only one candidate item left in the third set, the system automatically confirms its identity, or the operator confirms it manually, and then the barcode is regenerated and pasted through the barcode printing device.

[0090] Based on the above technical solution, S301 to S303 constitute a progressive hierarchical matching process with increasing computational cost and coarse-to-fine discrimination granularity. This process optimizes system resource allocation to the maximum extent, and while ensuring high recognition accuracy, it can improve the processing efficiency of item sorting in the case of damaged barcodes, effectively solving the technical problem mentioned in the background art of the difficulty in balancing recognition rate and efficiency.

[0091] In one possible implementation of this application embodiment, the above-mentioned S4 can be specifically implemented by the following S401, S402, and S403, which are described in detail below: S401. Determine the unique identity information of the current item based on the matching results.

[0092] In some implementations, identity verification needs to be completed according to different scenarios: In the first scenario, when the third set generated by S3 contains only one candidate item, the system automatically extracts the identification information of the candidate item as the unique identity information of the current item, and can start the feature cross-validation process: compare the physical dimension features, path topology features, and visual features of the current item with the features of the candidate item stored in the feature database. If the similarity of each feature is higher than the preset verification threshold, the identity verification is valid. In the second scenario, when S3 triggers the manual intervention process, the system first monitors the path topology feature sequence update status of all candidate items in the third set in real time. If it detects that a candidate item has been scanned and sorted at other nodes, it is determined that the candidate item is not related to the current item and is automatically removed from the third set. When the third set has only one candidate item left due to the removal of irrelevant candidate items, the system automatically confirms its identification information. If the operator still needs to make a judgment after manual intervention, the operator confirms the unique identity by combining the detailed information of the candidate item displayed by the system and then enters the identity information into the system through the operation terminal.

[0093] S402. Update the feature database and complete the item barcode information.

[0094] After identity verification, the matching results and current status of the item need to be added to the database. This provides the latest data support for the subsequent circulation of items and accumulates samples for the parameter optimization of the matching model. Barcode completion is the core action to solve the problem of subsequent node identification of the current item, avoiding secondary sorting obstacles caused by missing barcodes.

[0095] In some implementations, step S402 may include the following operations: Barcode reprinting and affixing: Control the barcode printer to immediately print a new barcode label containing the correct waybill number. After printing, the new label can be securely affixed to a prominent position on the item via a robotic arm or manual prompting. Path topology feature update: Append the current node and its timestamp to the end of the path topology feature sequence of the item in the feature database; Event Log Recording: The system log records detailed information such as the barcode damage event, matching process, final judgment result, operator ID, and timestamp.

[0096] S403. Perform sorting operations based on identity information and complete the process loop.

[0097] In some implementations, step S403 may include the following operations: The first step is the generation and execution of sorting instructions. After receiving the unique identification information of the current item, the system's sorting execution module retrieves the target sorting path corresponding to that identification information from the logistics management system, generates sorting control instructions, and sends the instructions to the sorting equipment through the communication unit. The sorting equipment executes the actions according to the instructions, transferring the current item to the corresponding target sorting channel to complete the physical sorting.

[0098] The second step is recording and feedback of sorting results. After the sorting equipment completes its work, it sends a sorting success signal to the system. The system then associates the sorting results with the current item's identification information, forming a closed-loop sorting data record.

[0099] The third step is to use closed-loop data for system optimization. Data such as matching time, frequency of manual intervention, and equipment execution efficiency of this sorting process are summarized and analyzed regularly to identify patterns: if a certain type of item frequently triggers manual intervention, the weight of its physical dimension features or the matching threshold is optimized; if a sorting device repeatedly reports execution delays, maintenance personnel are prompted to check the equipment status.

[0100] Based on the above technical solutions, S401 ensures the uniqueness and accuracy of identity information through multi-mechanism verification, S402 ensures smooth subsequent flow through database updates and barcode completion, and S403 improves the long-term performance of the system through sorting execution and closed-loop optimization. Step S4 not only solves the core problem of how to determine identity, but also takes into account the long-term needs of subsequent reuse and system iteration. It works in synergy with the layered matching strategy of S1-S3 to overcome the pain point of barcode dependence in traditional sorting systems, greatly improve the automation level, robustness and optimizability of sorting systems, reduce manual intervention costs, and meet the high efficiency and stability requirements of logistics sorting scenarios.

[0101] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as an intelligent identification terminal sorting device, includes at least one of the hardware structures and software modules corresponding to each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in a hardware or computer software-driven hardware manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0102] This application embodiment can divide a smart identification terminal sorting device into functional units based on the above method example. For example, each function can be divided into separate functional units, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0103] When using integrated units, Figure 5 The above embodiment shows a possible structural schematic diagram of an intelligent identification terminal sorting device (referred to as an intelligent identification terminal sorting device 50). The intelligent identification terminal sorting device 50 includes a processing unit 501 and a communication unit 502, and may also include a storage unit 503. Figure 5 The structural diagram shown can be used to illustrate the structure of an intelligent identification terminal sorting device involved in the above embodiments.

[0104] when Figure 5 The schematic diagram shown illustrates the structure of an intelligent identification terminal sorting device involved in the above embodiments. The processing unit 501 is used to control and manage the operation of the intelligent identification terminal sorting device, the communication unit 502 is used for the intelligent identification terminal sorting device to communicate with other devices, and the storage unit 503 is used to store the program code and data of the intelligent identification terminal sorting device.

[0105] For example, the communication unit 502 is used to interact with the feature database, image acquisition equipment, measurement sensors and logistics path information system, receive physical dimension features, path topology features and visual feature data of the items to be sorted, and send sorting control instructions to the sorting execution equipment and transmit item identification information to the barcode printing equipment to trigger barcode generation. The processing unit 501 is used to control the communication unit to complete the collection and reception of various feature information, integrate physical dimension features, path topology features and visual features to generate multi-dimensional feature vectors, establish a correlation between the multi-dimensional feature vectors and the identification information of the items to be sorted, and store the associated data in the feature database through the communication unit. When the item barcode is detected to be unreadable, the hierarchical matching process is started and the communication unit is controlled to retrieve the corresponding data in the feature database. After completing the feature matching calculation, the item identity information is determined.

[0106] In one possible implementation, the processing unit 501 is also used to calibrate the collected physical dimension raw data, dynamically adjust the subsequent matching strategy according to the number of candidate items in the feature matching process, trigger the manual intervention process when multiple candidate identities still exist after multiple rounds of matching, and control the communication unit to update the item status data in the feature database after the identity information is confirmed.

[0107] In one possible implementation, the communication unit 502 is further configured to receive real-time traffic status information and node flow data of the logistics network, send detailed information of candidate items in the third set to the manual processing terminal, and receive the unique identity information of the items after manual confirmation; the processing unit 501 is further configured to optimize path matching related judgments based on the real-time traffic data received by the communication unit, dynamically screen candidate items in the manual intervention process, and after receiving the identity information confirmed by the manual unit, control the communication unit to send a reprint instruction to the barcode printing device, and at the same time generate the control parameters required for the sorting operation and send them to the sorting execution device through the communication unit.

[0108] The processing unit 501 can be a processor or a controller, and the communication unit 502 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 503 can be a memory. When an intelligent identification terminal sorting device 50 is a chip, the processing unit 501 can be a processor or a controller, and the communication unit 502 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 503 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.).

[0109] The communication unit can also be called a transceiver unit. An antenna with transceiver functions and a control circuit in an intelligent identification terminal sorting device 50 can be considered as a communication unit 502 of the intelligent identification terminal sorting device 50, and a processor with processing functions can be considered as a processing unit 501 of the intelligent identification terminal sorting device 50. Optionally, the device in the communication unit 502 that implements the receiving function can be considered as a communication unit, which is used to execute the receiving steps in the embodiments of this application. The communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 502 that implements the transmitting function can be considered as a transmitting unit, which is used to execute the transmitting steps in the embodiments of this application. The transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0110] Figure 5If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0111] Figure 5 The units in the process can also be called modules; for example, a processing unit can be called a processing module.

[0112] This application also provides a hardware structure diagram of an intelligent identification terminal sorting device (referred to as an intelligent identification terminal sorting device 60), see [link to diagram]. Figure 6 The intelligent identification terminal sorting device 60 includes a processor 601, and optionally, a memory 602 connected to the processor 601.

[0113] In the first possible implementation, see Figure 6 A smart identification terminal sorting device 60 further includes a transceiver 603. The processor 601, memory 602, and transceiver 603 are connected via a bus. The transceiver 603 is used to communicate with other devices or communication networks. Optionally, the transceiver 603 may include a transmitter and a receiver. The device in the transceiver 603 that implements the receiving function can be considered as a receiver, which is used to perform the receiving steps in the embodiments of this application. The device in the transceiver 603 that implements the transmitting function can be considered as a transmitter, which is used to perform the transmitting steps in the embodiments of this application.

[0114] Based on the first possible implementation method Figure 6 The structural diagram shown can be used to illustrate the structure of an intelligent identification terminal sorting device involved in the above embodiments.

[0115] in, Figure 6 This can also be illustrated as a system chip in an intelligent identification terminal sorting device. In this case, the actions performed by the aforementioned intelligent identification terminal sorting device can be implemented by this system chip. The specific actions performed can be found above and will not be repeated here.

[0116] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0117] This application also provides a computer-readable storage medium including instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0118] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0119] This application also provides a chip including a processor and an interface circuit. The interface circuit is coupled to the processor. The processor is used to run computer programs or instructions to implement the above-described method. The interface circuit is used to communicate with other modules outside the chip.

[0120] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0121] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. An intelligent identification terminal sorting method, characterized by, include: Collect multidimensional feature information of items to be sorted and generate multidimensional feature vectors; The multidimensional features include physical dimension features, path topology features, and visual features; The multidimensional feature vector and the identification information of the items to be sorted are stored in the feature database; When an item barcode is detected to be unreadable, the physical dimension features, path topology features, and visual features of the current item are obtained in sequence, and the feature database is used to perform feature matching on the multidimensional features of the current item in sequence. The identity information of the current item is determined based on the matching results, and the sorting operation is completed.

2. The intelligent identification type terminal sorting method according to claim 1, characterized in that, The step of sequentially matching the multidimensional features of the current item using a feature database includes: The first set is obtained by performing a first-level match between the physical dimension features of the current item and the physical dimension features in the feature database; When there are multiple identifiers in the first set, a second-level matching is performed based on the path topology features of the current item and the path topology features in the first set to obtain the second set; When there are multiple identification information in the second set, a third-level matching is performed based on the visual features of the current item and the visual features in the second set to obtain the third set; If multiple identifiers exist in the third set, a manual intervention process is initiated; otherwise, the barcode of the current item is determined based on the unique identifier in the first, second, or third set.

3. The intelligent identification type terminal sorting method according to claim 2, characterized in that, The first-level matching uses a fast-computation screening algorithm, including: Calculate the physical attribute similarity between the current item and candidate items in the feature database. The physical attribute similarity includes volume similarity, density similarity, and dimensional ratio similarity obtained based on the physical dimension features. The dimensional ratio similarity refers to the degree of similarity between the current item and the candidate item in terms of three-dimensional size ratio. The physical similarity scores are obtained by weighted summation of the values ​​of each physical attribute similarity; where volume similarity has a weight of 0.4, density similarity has a weight of 0.3, and dimensionality similarity has a weight of 0.

3. Set a similarity threshold. When the physical similarity score is higher than the preset similarity threshold, the corresponding candidate item will be included in the first set.

4. The intelligent identification terminal sorting method according to claim 2, characterized in that, The specific process of the second-level matching includes: Extract the historical path topology feature sequence of each first candidate item in the first set; based on the directed graph of the logistics network, verify whether there is a directed edge from the latest record node of each first candidate item to the node where the current item is located; if yes, then obtain the node transition probability from the previous node of the first candidate item to the current node and the travel time weight of the directed edge based on historical logistics data; if no, determine that the matching has failed. The path consistency score PC of the first candidate item is calculated based on the historical path topological feature sequence; the path consistency score is used to characterize the path continuity and rationality of the candidate item from the historical path node to the current node, and reflects the spatiotemporal consistency of the path sequence. Based on the real-time traffic congestion index, traffic impact coefficient, and travel time weight, the theoretical transmission time of the current item from the previous node to the current node is calculated. Calculate the deviation between the current node's time and the theoretical transmission time to obtain the time deviation. Calculate the similarity between the topological path feature sequence of the current item and the historical path topological feature sequence of each first candidate item to obtain the path similarity. The path matching degree is calculated based on the node transition probability, the path consistency, the theoretical transmission time, and the path similarity. Set a path matching threshold. When the path matching degree exceeds the path matching threshold, the corresponding first candidate item will be included in the second set. The candidate items in the second set are sorted in descending order of path matching degree.

5. The intelligent identification terminal sorting method according to claim 4, characterized in that, The formula for calculating the path consistency score (PC) is: Where k represents the node index in the historical path topology feature sequence, and n represents the number of nodes. Indicates the time of the current node. This represents the time of the k-th node, where τ is the decay time constant. This represents the k-th node of the first candidate item. Let P represent the current node, and let P represent the node transition probability.

6. The intelligent identification terminal sorting method according to claim 5, characterized in that, The theoretical transmission time The formula for calculation is: ;in, The travel time weight of the directed edge. Traffic impact coefficient, Real-time traffic congestion index; The time deviation The formula for calculation is: ; This is the timestamp of the most recently recorded node in the historical path topological feature sequence. This is the timestamp of the current node.

7. The intelligent identification terminal sorting method according to claim 4, characterized in that, The path matching degree The formula for calculation is: ;in, Let P be the node transition probability, PS be the path similarity, PC be the path consistency score, α be the time deviation penalty coefficient, and β be the path consistency weight coefficient.

8. The intelligent identification terminal sorting method according to claim 2, characterized in that, The third-level matching employs a multimodal feature fusion mechanism, including: Based on the lightweight visual features of each second candidate item in the second set and the current item, calculate the color histogram similarity, texture feature similarity, and shape feature similarity to obtain a lightweight visual feature similarity vector. When the similarity value in the lightweight visual feature similarity vector is lower than a preset threshold, the high-dimensional feature vector extracted by the convolutional neural network is used to calculate the similarity and obtain the deep visual feature similarity; the convolutional neural network is embedded with an attention module to improve the comparison of key features of the items. The lightweight visual feature similarity vector or the deep visual feature similarity of each second candidate item is weighted and fused with the physical similarity score and path matching degree to obtain the final matching confidence. A third set is generated based on the matching confidence scores.

9. An intelligent identification-based terminal sorting device, characterized in that, include: Communication unit and processing unit; The communication unit is used to interact with the feature database, image acquisition equipment, measurement sensors and logistics path information system, receive multi-dimensional feature information of items to be sorted, and send sorting control instructions. The processing unit is connected to the communication unit and is used to control the communication unit to collect multi-dimensional feature information of the items to be sorted and generate multi-dimensional feature vectors. The multi-dimensional features include physical dimension features, path topology features and visual features. The multidimensional feature vector is associated with the identification information of the items to be sorted and stored in the feature database through the communication unit; When the communication unit detects that the item barcode is unreadable, the physical dimension features, path topology features, and visual features of the current item are obtained in sequence, and the multidimensional features of the current item are matched sequentially using the feature database; the identity information of the current item is determined according to the matching result, and an instruction is sent through the communication unit to complete the sorting operation.

10. An intelligent identification-based terminal sorting system and device, characterized in that, include: The system comprises a feature acquisition module, a storage management module, a hierarchical matching module, and a sorting execution module. in, The feature acquisition module is used to acquire multi-dimensional feature information of the items to be sorted and generate multi-dimensional feature vectors. The multi-dimensional features include physical dimension features, path topology features and visual features. The storage management module is connected to the feature acquisition module and is used to associate the multidimensional feature vector with the identification information of the items to be sorted and store it in the feature database. The hierarchical matching module is connected to the storage management module. When the barcode of an item is detected to be unreadable, it sequentially acquires the physical dimension features, path topology features, and visual features of the current item, and uses the feature database to sequentially perform feature matching on the multidimensional features of the current item, and determines the identity information of the current item based on the matching results. The sorting execution module is connected to the hierarchical matching module and is used to complete the sorting operation of the current item based on the determined identity information.

Images