Express item comprehensive sorting identification method and system based on multi-feature fusion
By integrating multi-dimensional features of goods with user identity information into the logistics system, the problem of low efficiency and high identification error rate in traditional logistics has been solved, realizing efficient and automated goods sorting and user experience optimization in unmanned stations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MODULUS ZHIXING (BEIJING) TECHNOLOGY CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114790A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of logistics technology, specifically to a comprehensive sorting and identification method and system for express delivery items based on multi-feature fusion. Background Technology
[0002] In the field of logistics sorting and last-mile delivery, traditional centralized distribution warehouses generally use barcodes or QR codes to identify items. This technology requires manual or machine scanning at each node of the item flow, such as outbound, sorting, inbound, and pickup. However, this method, which relies on physical identifiers, has many drawbacks: multiple scans lead to low efficiency, barcodes are easily damaged, soiled, or lost, leading to recognition failures, printing and affixing barcodes increases operating costs, and users still need to scan the code or enter a pickup code to pick up their items, resulting in a poor user experience. In non-standard last-mile scenarios such as rural stations, the high mobility of personnel and the complex environment make traditional secondary scanning and recognition error rates high, and it is difficult to effectively bind specific pickup personnel, making it difficult to meet the needs of automated and highly robust operations.
[0003] Existing technologies rely on externally damaged physical identifiers, which cannot achieve contactless, automated, and highly reliable item identification. They are particularly unsuitable for scenarios such as unmanned rest stations that require low-cost and high-efficiency operation. Therefore, there is an urgent need for an intelligent identification solution that does not require physical identifiers and can integrate item characteristics with user identity information. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a comprehensive sorting and identification method and system for express delivery items that can quickly and accurately identify items by combining multi-dimensional features of the items with user identity information without the need for physical identifiers, so as to improve logistics sorting efficiency, reduce operating costs, and optimize the user pick-up experience in scenarios such as regional unmanned stations.
[0005] To solve the above-mentioned technical problems, the technical solution provided by the present invention is: a comprehensive sorting and identification method for express items based on multi-feature fusion, including initial feature input, end-point identification, return and exchange processing flow, and unmanned collection process; The initial feature entry is completed in the centralized distribution warehouse. The visual features of the items are collected from multiple angles through the feature acquisition terminal, and after being bound to the user ID, a unique feature fingerprint is generated and stored in the database. The terminal identification is completed at the regional unmanned station, including three steps: warehouse identification and shelving, user notification, and user pickup. The return and exchange process revolves around fingerprint verification to verify the identity of returned goods. The unmanned package collection process creates a unique fingerprint for new packages and binds the sender's information, thereby automating the collection process.
[0006] Furthermore, in the initial feature input, the visual features of the item are extracted using a ResNet-50 or EfficientNet-B4 model trained on the package image. After removing the last classification layer, the output of the penultimate layer is taken as a single-view feature vector. The top view and side view features are then fused by concatenating the vectors to obtain a 4096-dimensional visual feature vector.
[0007] Furthermore, the process of generating the characteristic fingerprint is as follows: The visual feature vector, the size and volume features extracted from the 3D point cloud, the point cloud distribution histogram, and the weight features are Z-score standardized. After being stitched together, the dimensions are reduced to a 512-dimensional dense vector by principal component analysis or autoencoder, which is the feature fingerprint. The feature fingerprint is then bound to the user ID corresponding to the logistics order. In the unmanned package collection process, the multi-dimensional physical characteristics of the items to be sent are collected and bound to the sender's user ID to generate a new feature fingerprint representing that the items to be sent belong to the sender. In the return and exchange process, the return order is uniquely linked to the original order and the corresponding feature fingerprint record.
[0008] Furthermore, in the final identification process of warehouse entry and shelving: All the feature fingerprint vectors that have been entered in the cloud feature fingerprint database are used to construct an IndexIVFFlat index type vector index using the Faiss library, and the vector space is divided into 100 units using IndexFlatL2 as the quantizer. After collecting the features of the package to be identified, a 512-dimensional query vector is generated; By using vector indexing to perform an approximate nearest neighbor search on the query vector generated from the package to be identified, the 5 most similar result vectors are returned. Calculate the cosine similarity between the query vector and these 5 result vectors; Different matching results are given based on the cosine similarity value; In the outbound verification stage of the unmanned package collection process, the characteristics of the package with the label attached are collected to generate a query vector, which is then matched with the feature fingerprint vector of the pending shipment status in the database. In the final verification stage of the return and exchange process, real-time features of returned goods are collected to generate query vectors, which are then compared one-to-one with the original feature fingerprints associated with the return order.
[0009] Furthermore, the matching determination result is: The recognition is considered successful when the cosine similarity is ≥0.95; When 0.85 ≤ cosine similarity < 0.95, a second weight comparison is triggered. When the cosine similarity is less than 0.85, the exception handling process is initiated. In the verification logic of the return and exchange process, if the cosine similarity is ≥0.95, the verification is considered successful, confirming that this item is the same as that item; if the cosine similarity is <0.85, the matching is considered to have failed, indicating that the goods are inconsistent, generating an abnormal record and issuing an alarm. In the outbound verification of the unmanned package collection process, a successful match is determined when the cosine similarity is ≥0.95, and the package status is updated to "in transit".
[0010] Furthermore, during the user pickup process, the user completes identity verification by facial recognition or QR code scanning. The system automatically queries all packages under that user's name that are pending pickup and their storage locations, retrieves the package, and updates the package status to "picked up" after the user confirms, and empties the storage space. In the scenario of manual intervention in the return and exchange process, if the matching fails due to severe wear and tear of the product or replacement of the packaging, but the manual inspection confirms that it is genuine, the administrator can force the verification to pass, the system records the reason for the manual intervention, and can choose to update the feature fingerprint of the product.
[0011] Furthermore, in the last-mile identification process, only after the package is successfully stored and assigned a storage location will a text message or app push notification containing the station name, storage location number, and pickup method be sent to the corresponding user. In the unmanned package collection process, users fill in the recipient information in advance through the touch screen in the station or through a mobile app / app, select the express delivery service and pay the fee, and the system generates a shipping order and a corresponding QR code. In the return and exchange process, users submit a return and exchange application on the e-commerce platform. After the system approves the application, a return order is generated. After verification, the product information and status are updated to "pending return for quality inspection," and the subsequent refund or exchange process is triggered. If the verification fails, the refund / exchange is rejected and the user is notified.
[0012] The present invention also provides a comprehensive sorting and identification system for express items based on multi-feature fusion, including a feature acquisition terminal, a user management module, a feature extraction and fusion module, a comprehensive feature database, and an identification and matching engine; The feature acquisition terminal is deployed in centralized distribution warehouses and regional unmanned stations. It includes a high-definition camera, a 3D structured light sensor and a weight sensor, and is used to collect information on the color, texture, shape, size and weight of items. The feature extraction and fusion module associates and fuses the visual features of the item with the user identity information in the user management module to generate a feature fingerprint of the item and user combination and stores it in the comprehensive feature database. The identification and matching engine collects the features of the package to be identified in real time and performs similarity calculations with a comprehensive feature database to complete the identification of the item and its owner.
[0013] Furthermore, the feature acquisition terminal of the centralized distribution warehouse includes at least two global shutter industrial cameras, a 3D camera, a dynamic weighing module, and an industrial control computer. The industrial cameras are equipped with polarizing filters to eliminate reflections.
[0014] Furthermore, the comprehensive feature database adopts a cloud-native distributed database, which is capable of reading, writing, and storing and retrieving vector data.
[0015] The advantages of this invention compared to the prior art are: This invention eliminates the cumbersome operation of multiple barcode scanning in traditional methods, achieving "one-time collection and full-process identification". No manual intervention is required for scanning during the warehousing and retrieval of goods, greatly improving the efficiency of logistics sorting and last-mile delivery. This invention integrates multi-dimensional features such as object vision, 3D geometry, and weight with user identity information. It has rich feature dimensions and strong anti-interference ability. Even if some features of the object are occluded, it can still achieve accurate matching through other features and user-related information, effectively reducing recognition errors. This invention eliminates the need for manual contact with item labels or manual input of information throughout the entire process, achieving a high degree of automation from warehousing to retrieval. It also supports contactless retrieval methods such as facial recognition, enhancing user experience and epidemic prevention safety. This invention eliminates the material costs of barcode printing and affixing, as well as the purchase and wear and tear costs of barcode scanning equipment, and is especially suitable for large-scale, low-cost operation scenarios such as unmanned rest stations in the region. The invention features refined configuration on the centralized distribution warehouse side to ensure accurate feature collection, and simplified configuration on the delivery station side to control costs. It can be flexibly adapted to various logistics scenarios such as centralized distribution warehouses, regional unmanned delivery stations, and smart express cabinets, making it highly practical. Attached Figure Description
[0016] Figure 1 This is a system block diagram of the centralized distribution warehouse side.
[0017] Figure 2 This is a system block diagram of the unmanned rest station in the area.
[0018] Figure 3 This is a flowchart of the initial feature entry stage.
[0019] Figure 4 This is a flowchart of the end-point identification stage.
[0020] Figure 5 This is a schematic diagram of a comprehensive feature data structure.
[0021] Figure 6 This is a flowchart of the return and exchange process.
[0022] Figure 7 This is a flowchart of the unmanned package collection process. Detailed Implementation
[0023] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0024] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.
[0025] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.
[0026] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.
[0027] The following detailed description, in conjunction with the accompanying drawings, provides a more comprehensive description of the multi-feature fusion-based comprehensive sorting and identification method and system for express delivery items according to the present invention.
[0028] Combined with appendix Figure 1-7 This invention will be described in detail below.
[0029] This invention relates to a comprehensive sorting and identification method for express delivery items based on multi-feature fusion, comprising two core stages: initial feature input and end-point identification. The process for each stage is as follows: Initial Feature Entry Phase: This phase is completed in the centralized distribution warehouse, establishing a unique feature fingerprint for each item. The specific steps are as follows: Step 1: After a user places an order, the items enter the distribution warehouse. The system obtains the order information by scanning the shipping label or using OCR recognition, and retrieves the corresponding user information (such as user ID) from the user management module based on the order number, thus completing the binding between the order and the user.
[0030] Step 2: The feature acquisition terminal starts working, and takes images of the object from the top and side views using an industrial camera. The 3D sensor collects point cloud data of the object, and the dynamic weighing module obtains stable weight data of the object after the conveyor belt pauses for 1-2 seconds, thus comprehensively capturing the visual features, 3D geometric features and weight features of the object.
[0031] Step 3: Feature extraction and fusion processing, specifically including: Image feature extraction: A ResNet-50 or EfficientNet-B4 pre-trained model, finely tuned from tens of thousands of wrapped images, is used as the backbone network. The last classification layer of the model is removed, and the output of the second to last layer (2048-dimensional vector for ResNet-50) is taken as the single-view image feature vector. The top view and side view feature vectors are fused by vector concatenation to obtain a 4096-dimensional visual feature vector. 3D point cloud feature extraction: Voxelization downsampling and statistical filtering are performed on the original point cloud data to remove noise. The minimum bounding cube of the object is calculated, and the dimensional volume features such as length, width, height, and volume are extracted. A point cloud distribution histogram is also extracted (the space is divided into multiple grids, and an n-dimensional vector is formed by the point density in each grid). Feature normalization and fusion: Z-score normalization is performed on all dimensions of features, such as visual feature vectors, 3D point cloud features, and weight features, to eliminate the influence of different units; all normalized feature vectors are concatenated to form an ultra-long original vector; the original vector is dimensionality reduced by principal component analysis or autoencoder to obtain a 512-dimensional dense vector, which is the unique "feature fingerprint" of the item. Associate user information: Associate the generated 512-dimensional feature fingerprint with the bound user ID, order number and other information to form a complete data record.
[0032] Step 4: Upload the data record containing information such as feature fingerprint, user ID, order number, and collection timestamp to the comprehensive feature database to complete the initial feature entry. At this time, the item status is marked as "in transit".
[0033] The final identification stage: This stage is completed at the regional unmanned station and includes three core steps: inventory identification and shelving, user notification, and user pickup. The specific process is as follows: Step 1: Warehouse Identification and Shelving Step 1: After the unmanned delivery vehicle arrives at the unmanned station in the area, the automated equipment or robots in the station will send the packages into the identification area one by one; Step 2: The identification and matching engine is started. Multi-dimensional features of the packages to be put into the warehouse are collected through the feature collection terminal on the station side. A 512-dimensional query vector is generated according to the same algorithm process as the initial feature entry stage. Step 3: Vector index matching. The system uses Facebook's open-source Faiss library to construct an IndexIVFFlat type vector index from all feature fingerprint vectors in the comprehensive feature database. Using IndexFlatL2 as the quantizer, the vector space is divided into 100 units to achieve fast retrieval of near nearest neighbors. The function is called to perform vector indexing and return the 5 feature fingerprint vectors most similar to the query vector. Step 4: Matching and determination. Calculate the cosine similarity between the query vector and the 5 feature fingerprint vectors. Based on the similarity value, give the determination result: when the cosine similarity is ≥ 0.95, the identification is successful and the corresponding record ID is returned; when 0.85 ≤ cosine similarity < 0.95, trigger a second weight comparison for auxiliary verification; when the cosine similarity is < 0.85, the identification fails and the process is transferred to manual intervention or retry for exception handling. Step 5: Storage location allocation and shelving. After successful identification, the system requests an available storage location from the storage location management module and assigns a unique physical storage location number; the system controls the actuator to transport the package to the designated storage location. Step 6: Update the database, change the package status from "in transit" to "pending pickup", record its storage location number, and complete the warehousing and shelving process.
[0034] Step 2: User Notification Only after a package is successfully received and assigned a storage location will the user notification module send a pickup notification to the corresponding user. The notification may take the form of an SMS or an app push, and the content will clearly state the station name, storage location number, and pickup method.
[0035] Step 3: Customer picks up the package Step 1: After arriving at the station, the user completes identity verification by using the facial recognition or QR code scanning function of the pickup terminal; Step 2: Based on the verified user identity, the system queries the comprehensive feature database to obtain all packages under that user name that are in the "pending pickup" status and their storage location information; Step 3: The control system guides the actuator or prompts staff to retrieve the package from the designated storage location and deliver it to the user; Step 4: After the user confirms the pickup, the system will update the package status to "Picked up" and mark the corresponding storage location as available, thus completing the entire pickup process.
[0036] This invention is a comprehensive sorting and identification method for express delivery items based on multi-feature fusion, and also includes a return and exchange process and an unmanned pickup process; The return and exchange process is as follows: 1. Front-end trigger: User application When a user submits a return or exchange request on an e-commerce platform, a return order is generated after the system approves it. This order is uniquely linked to the original order and its corresponding "feature fingerprint" record.
[0037] 2. End-point verification: Package identity authentication Users bring the goods back to the rural unmanned station in their original packaging (or any packaging) and place them in the identification area.
[0038] The recognition and matching engine is activated, collecting real-time features of returned goods and generating query vectors.
[0039] Instead of blindly searching the global database, the engine directly compares the data one-to-one with the original "feature fingerprint" associated with the returned order.
[0040] Verification logic: Match successful (similarity ≥ 0.95): The system confirms that "this item is the same as that item," and verification passes. The product information and status are updated to "pending return for quality inspection" in the database, triggering the subsequent refund or exchange process. No proof is required from the user throughout the entire process.
[0041] Matching failure (similarity < 0.85): The system determines this as "product mismatch," generates an exception record, and issues an alert. This step effectively prevents malicious acts such as "replacing genuine goods with counterfeit goods" or "switching products."
[0042] 3. Backend processing: Asset status management and data iteration After verification: The system marks the status of this "feature fingerprint" record as "returned".
[0043] If the product is in good condition and can be resold, the "feature fingerprint" can be inherited and reused. When a new user purchases this product, the system can re-bind the new user information with this feature fingerprint to create a new logistics record.
[0044] Verification failed / manual intervention: If the matching fails due to severe wear and tear of the product or packaging replacement, but the product is confirmed to be genuine by manual inspection, the administrator can force the verification to pass.
[0045] The system will record the reason for the manual intervention and can optionally update the product's "feature fingerprint" to learn the feature changes that occur during the product's circulation process, thereby enabling the model to learn and adapt.
[0046] The unmanned package collection process is as follows: 1. User front-end operations: Initiating a shipment: Users can fill in the recipient's information, select a courier service, and pay the fee in advance via the touchscreen at the station or, more preferably, via a mobile app / mini-program. The system then generates a shipment order and a corresponding QR code.
[0047] 2. Automated operation within the station: Identity verification and order triggering: Upon arrival at the station, the user verifies their identity by facial recognition or scanning the QR code on their mobile phone at the identification area. The system then retrieves the created shipping order and prepares to proceed to the physical handling stage.
[0048] Item feature collection and "feature fingerprint" creation: Users place unpackaged items (or simply packaged items whose core features are visible) into the recognition area.
[0049] The system controls sensors to collect the multi-dimensional physical characteristics (visual, 3D, weight) of the item, and the process is completely consistent with the delivery and warehousing process.
[0050] The feature extraction and fusion module binds these features to the sender's user ID, generating a new feature fingerprint representing "this item to be sent belongs to this sender".
[0051] The system creates a new record in the cloud database, with the status "pending dispatch".
[0052] Standardized packaging and labeling: Based on the identified item's size and weight, the system automatically assigns a suitable standardized empty cardboard box (pre-placed within the station) or prompts the user to use packaging of a specific size.
[0053] After packaging, the printer prints out a simple shipping label containing only a QR code for the shipping order and key text information. The user or a robotic arm affixes this label to the package. Please note: This QR code is only for manual assistance in tracking and sorting; the core identifier remains the "feature fingerprint."
[0054] Outbound verification and logistics triggering: The package with the label affixed is placed back into the identification area (or automatically transported by the system to the outbound verification station).
[0055] The system performs a second identification before shipment to confirm that the package's characteristic fingerprint matches the pending shipment record in the database.
[0056] Once a match is successful, the system updates the package status to "in transit" and notifies the courier to pick it up, or automatically places it in the corresponding logistics compartment. The package then officially enters the logistics network.
[0057] To achieve the above method, this invention further provides a comprehensive sorting and identification system for express delivery items based on multi-feature fusion. Through the collaborative work of multiple modules, it achieves the fusion and identification of item features and user information, specifically including the following components: Feature acquisition terminals: Deployed in centralized distribution warehouses and regional unmanned stations, core equipment includes high-definition cameras, 3D structured light sensors, and weight sensors, used to comprehensively collect multi-dimensional feature information of items, covering color, texture, shape, size, weight, and special markings. The feature acquisition terminals on the centralized distribution warehouse side are more sophisticatedly configured, including at least two ≥5-megapixel global shutter industrial cameras (equipped with polarizing filters to eliminate reflections, shooting from top and 45-degree side angles respectively), a 3D camera based on structured light or time-of-flight principles, a dynamic weighing module with an error ≤±10g, and an industrial control computer equipped with a high-performance CPU and GPU. All sensors are connected to the industrial control computer via gigabit Ethernet or USB 3.0 interfaces. The regional unmanned station side can have a simplified configuration based on cost requirements, needing only one high-performance 3D camera with both RGB and depth information acquisition capabilities and one dynamic scale.
[0058] User Management Module: Used to store and manage core user information, including user ID, facial information, mobile phone number, frequently used delivery address, etc., providing data support for the association and binding of items with users.
[0059] Feature extraction and fusion module: As the core algorithm module of the system, it is responsible for associating and fusing the visual features of the items, 3D point cloud features, weight features, etc. acquired by the feature acquisition terminal with the corresponding user identity information in the user management module to generate a unique "item-user" combined feature fingerprint, and storing it in the comprehensive feature database.
[0060] Comprehensive Feature Database: Employs a cloud-native distributed database (such as Alibaba Cloud PolarDB or MongoDBAtlas) to store the feature fingerprint data of all items already in the database. It supports high-concurrency read and write operations and the storage and fast retrieval of massive vector data, ensuring efficient and stable data access.
[0061] Identification and matching engine: mainly deployed in regional unmanned stations, its core function is to collect multi-dimensional features of items in real time when they enter the identification area, calculate the similarity with the feature fingerprints in the comprehensive feature database, and quickly find the record with the highest matching degree, thereby accurately determining the item information and its user.
[0062] The specific implementation process of the comprehensive sorting and identification method and system for express delivery items based on multi-feature fusion of the present invention is as follows: Villager Zhang San (user ID: U_Rural_1001) purchased a 10kg bag of rice on an e-commerce platform and completed the sorting, identification, and pickup process using the method and system of this invention.
[0063] After Zhang San placed the order, the package (order number: ORD_20251126xxxx) entered the distribution warehouse. The system scanned the waybill to obtain the order number and queried the user ID bound to the user management module: U_Rural_xxxx; The feature acquisition terminal on the central distribution warehouse side is activated. The top-view and side-view industrial cameras capture images of the package. The 3D sensor collects point cloud data and calculates the dimensions as [300, 200, 500] mm and the volume as 30,000,000 mm³. The dynamic weighing module measures the weight as 10,000 g. The feature extraction and fusion module extracts image feature vectors through the ResNet-50 model, splices the top view and side view features to obtain a 4096-dimensional visual feature vector, and performs Z-score normalization on the visual features, size and volume features, point cloud distribution histogram, and weight features. After splicing, the dimensionality is reduced to 512-dimensional feature fingerprints through an autoencoder. The system creates records in the comprehensive feature database as follows: json { "RecordID":"FP_8a3b7c2e", "FeatureVector":[0.12,-0.45,0.88,...], / / 512-dimensional feature vector "UserID":"U_Rural_xxxx", "OrderID":"ORD_20251126xxxx", "Timestamp":"2025-xx-xx 10:30:25", "Status":"in_transit", "StoredLocation":null } Inbound Identification and Shelving: When the unmanned delivery vehicle arrives at the station, the package is sent to the identification area. The feature acquisition terminal on the station side collects real-time features and generates a query vector. The identification and matching engine searches through the Faiss index and finds that the cosine similarity between the query vector and FP_8a3b7c2e is 0.98 (≥0.95), and the identification is successful. The system applies for a storage location from the storage location management module and is assigned A-01-02. The control system drives the AGV to transport the package to the storage location, the database updates the package status to "ready_for_pickup", and the storage location is updated to "A-01-02". User notification: The user notification module sends an SMS to Zhang San's mobile phone number: "Your package has arrived at XX Station and is stored in locker A01, compartment 02. Please pick it up using facial recognition." User pickup: After arriving at the station, Zhang San scans his face at the pickup terminal. The system confirms his identity as U_Rural_xxxx through facial recognition, queries the database to obtain the package location information, and controls the AGV to retrieve the package from location A-01-02 and deliver it to Zhang San. After Zhang San confirms pickup, the system updates the package status to "picked_up" and marks location A-01-02 as available.
[0064] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A comprehensive sorting and identification method for express delivery items based on multi-feature fusion, characterized in that, This includes initial feature entry, end-point identification and return / exchange processing procedures, and unmanned pickup procedures; The initial feature entry is completed in the centralized distribution warehouse. The visual features of the items are collected from multiple angles through the feature acquisition terminal, and after being bound to the user ID, a unique feature fingerprint is generated and stored in the database. The terminal identification is completed at the regional unmanned station, including three steps: warehouse identification and shelving, user notification, and user pickup. The return and exchange process revolves around fingerprint verification to verify the identity of returned goods. The unmanned package collection process creates a unique fingerprint for new packages and binds the sender's information, thereby automating the collection process.
2. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 1, characterized in that: In the initial feature input, the visual features of the item are extracted using a ResNet-50 or EfficientNet-B4 model trained on the package image. After removing the last classification layer, the output of the second-to-last layer is taken as a single-view feature vector. The top view and side view features are then fused by concatenating the vectors to obtain a 4096-dimensional visual feature vector.
3. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 2, characterized in that: The process of generating the feature fingerprint is as follows: The visual feature vector, the size and volume features extracted from the 3D point cloud, the point cloud distribution histogram, and the weight features are Z-score standardized. After being stitched together, the dimensions are reduced to a 512-dimensional dense vector by principal component analysis or autoencoder, which is the feature fingerprint. The feature fingerprint is then bound to the user ID corresponding to the logistics order. In the unmanned package collection process, the multi-dimensional physical characteristics of the items to be sent are collected and bound to the sender's user ID to generate a new feature fingerprint representing that the items to be sent belong to the sender. In the return and exchange process, the return order is uniquely linked to the original order and the corresponding feature fingerprint record.
4. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 3, characterized in that: In the end-point identification process of warehouse entry and shelving: All the feature fingerprint vectors that have been entered in the cloud feature fingerprint database are used to construct an IndexIVFFlat index type vector index using the Faiss library, and the vector space is divided into 100 units using IndexFlatL2 as the quantizer. After collecting the features of the package to be identified, a 512-dimensional query vector is generated; By using vector indexing to perform an approximate nearest neighbor search on the query vector generated from the package to be identified, the 5 most similar result vectors are returned. Calculate the cosine similarity between the query vector and these 5 result vectors; Different matching results are given based on the cosine similarity value; In the outbound verification stage of the unmanned package collection process, the characteristics of the package with the label attached are collected to generate a query vector, which is then matched with the feature fingerprint vector of the pending shipment status in the database. In the final verification stage of the return and exchange process, real-time features of returned goods are collected to generate query vectors, which are then compared one-to-one with the original feature fingerprints associated with the return order.
5. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 4, characterized in that: The matching result is as follows: The recognition is considered successful when the cosine similarity is ≥0.95; When 0.85 ≤ cosine similarity < 0.95, a second weight comparison is triggered. When the cosine similarity is less than 0.85, the exception handling process is initiated. In the verification logic of the return and exchange process, the verification is considered successful when the cosine similarity is ≥0.95, confirming that this item is the same as that item. If the cosine similarity is less than 0.85, the match is deemed to have failed, indicating that the goods do not match, an abnormal record is generated, and an alarm is issued. In the outbound verification of the unmanned package collection process, a successful match is determined when the cosine similarity is ≥0.95, and the package status is updated to "in transit".
6. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 5, characterized in that: During the user pickup process, the user completes identity verification by facial recognition or QR code scanning. The system automatically queries all packages under the user's name with pending pickup status and storage location information, retrieves the package, and updates the package status to "picked up" after the user confirms, and empties the storage space. In the scenario of manual intervention in the return and exchange process, if the matching fails due to severe wear and tear of the product or replacement of the packaging, but the manual inspection confirms that it is genuine, the administrator can force the verification to pass, the system records the reason for the manual intervention, and can choose to update the feature fingerprint of the product.
7. The comprehensive sorting and identification method for express delivery items based on multi-feature fusion according to claim 6, characterized in that: In the last-mile identification process, only after the package is successfully stored and assigned a storage location will a text message or app push notification be sent to the corresponding user, which includes the station name, storage location number and pickup method. In the unmanned package collection process, users fill in the recipient information in advance through the touch screen in the station or through a mobile app / app, select the express delivery service and pay the fee, and the system generates a shipping order and a corresponding QR code. In the return and exchange process, users submit a return and exchange application on the e-commerce platform. After the system approves the application, a return order is generated. After verification, the product information and status are updated to "pending return for quality inspection," and the subsequent refund or exchange process is triggered. If the verification fails, the refund / exchange is rejected and the user is notified.
8. A comprehensive sorting and identification system for express delivery items based on multi-feature fusion, characterized in that: It includes a feature acquisition terminal, a user management module, a feature extraction and fusion module, a comprehensive feature database, and a recognition and matching engine; The feature acquisition terminal is deployed in centralized distribution warehouses and regional unmanned stations. It includes a high-definition camera, a 3D structured light sensor and a weight sensor, and is used to collect information on the color, texture, shape, size and weight of items. The feature extraction and fusion module associates and fuses the visual features of the item with the user identity information in the user management module to generate a feature fingerprint of the item and user combination and stores it in the comprehensive feature database. The identification and matching engine collects the features of the package to be identified in real time and performs similarity calculations with a comprehensive feature database to complete the identification of the item and its owner.
9. The comprehensive sorting and identification system for express delivery items based on multi-feature fusion according to claim 8, characterized in that: The feature acquisition terminal of the centralized distribution warehouse includes at least two global shutter industrial cameras, a 3D camera, a dynamic weighing module, and an industrial control computer. The industrial cameras are equipped with polarizing filters to eliminate reflections.
10. The comprehensive sorting and identification system for express delivery items based on multi-feature fusion according to claim 9, characterized in that: The comprehensive feature database adopts a cloud-native distributed database, which is capable of reading, writing, and storing and retrieving vector data.