An e-commerce intelligent warehouse cargo management system

Abstract

This invention discloses an intelligent e-commerce warehouse goods management system, belonging to the field of intelligent warehousing and logistics management technology. The system adopts a multi-source heterogeneous data fusion architecture, including core components such as an RFID tag module, RFID reader / writer, visual recognition module, and data fusion processing module. The data fusion processing module employs spatiotemporal alignment units, adaptive Kalman filtering algorithms, and hybrid positioning algorithms to achieve high-precision fusion of RFID data and visual data. The method includes steps such as system initialization, goods entry identification, data preprocessing, spatiotemporal alignment, data fusion, location calculation, data storage, real-time tracking, outbound verification, and report generation. Through fusion algorithms and architecture design, this invention achieves high-precision positioning, millisecond-level real-time response, and system availability, significantly outperforming existing technologies and providing a new technical solution for intelligent warehousing.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing and logistics management technology, specifically to a warehouse cargo tracking system and method. Background Technology

[0002] In recent years, with the rapid development of the e-commerce industry, the challenges facing warehouse management have become increasingly prominent. The wide variety of goods, frequent inbound and outbound operations, and complex inventory management have made traditional warehouse management methods unsustainable. Currently, mainstream warehouse management relies mainly on manual record-keeping and barcode scanning. This method is not only inefficient but also prone to human error, failing to meet the accuracy and real-time requirements of modern logistics.

[0003] RFID technology enables contactless batch identification and is widely used in warehouse management, but it still has the following technical shortcomings: Limited positioning accuracy: Existing RFID positioning systems typically only provide location information with meter-level accuracy, which cannot meet the needs of refined management. For example, the patent publication number CN119378871A applied for by Guangzhou Fengyijie Electronic Technology Co., Ltd. improves warehouse management efficiency through RFID technology, but its positioning accuracy is still limited, making it difficult to achieve precise tracking of goods.

[0004] Signal obstruction problem: In complex warehouse environments, factors such as metal shelves and stacked goods can cause RFID signal attenuation or obstruction, affecting the reliability of identification. For example, patent publication number CN118798243B uses a combination of high-frequency and ultra-high-frequency readers, but it still does not effectively solve the identification blind spot problem caused by signal obstruction.

[0005] Insufficient real-time tracking capability: Existing RFID systems mainly focus on identification during the entry and exit of goods, lacking the ability to track goods in real time within the warehouse. For example, patent publication number CN119100040A mainly addresses the problem of entry and exit detection, but cannot achieve continuous tracking of goods during their movement within the warehouse.

[0006] Visual recognition technology can provide accurate location and status information, but it faces the following technical bottlenecks in large-scale warehousing scenarios: Processing speed limitations: In large-scale warehouse environments, a large amount of video data from cameras needs to be processed, and the processing speed of existing vision systems is insufficient to meet real-time requirements. For example, patent publication number CN118658129B, even when focusing on computer vision-based cargo recognition, exhibits performance bottlenecks when processing large-scale video data.

[0007] Poor environmental adaptability: The visual recognition system is sensitive to environmental factors such as lighting conditions and obstructions, and its performance is unstable in complex warehouse environments.

[0008] High computational resource consumption: High-precision visual recognition requires a large amount of computational resources, which increases the system cost and complexity.

[0009] While some technologies have attempted to combine RFID and visual recognition—for example, patent publication number CN208061320U discloses a smart vending machine based on a hybrid RFID and visual recognition technology—this combination is limited to the specific application scenario of smart vending machines. The technical solutions are relatively simple and cannot solve the complex tracking problems in large-scale warehousing environments. The TagVision system researched by Beijing Institute of Technology, while achieving the fusion of RFID and vision, primarily targets fine-grained tracking in indoor scenarios and is not optimized for the characteristics of warehousing environments, making it difficult to directly apply to warehouse management.

[0010] Therefore, there is a lack of existing technologies that can effectively integrate the advantages of RFID and visual recognition to solve the problem of high-precision, real-time tracking of goods in large-scale warehousing environments. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide an e-commerce intelligent warehouse goods management system that achieves high-precision, real-time, and reliable tracking of warehouse goods through a multi-source heterogeneous data fusion architecture and algorithm.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: An e-commerce intelligent warehouse goods management system includes: The RFID tag module uses ultra-high frequency passive RFID tags to store the identification information and attribute data of goods; the tags are designed to be anti-metal and can be adapted to metal shelf environments. RFID reading and writing equipment includes fixed readers and handheld readers. Fixed readers use a multi-antenna array design, with a reading distance of approximately 6-8 meters, and support reading 150-200 tags per second. They are installed in key locations such as warehouse entrances and exits and main passageways. Handheld readers are for use by operators during mobile operations. The visual recognition module includes multiple high-definition network cameras and an image processing unit. The cameras employ pixel sensors, and the image processing unit adopts an accelerated architecture. An auxiliary ranging module, including a TOF sensor, is installed in key areas of the warehouse to provide a reference distance for the visual algorithm. The data fusion processing module includes an RFID data processing unit, a visual data processing unit, a spatiotemporal alignment unit, a data fusion unit, a location calculation unit, and an anomaly detection unit. It is used to receive multi-source heterogeneous data from RFID reading and writing devices and visual recognition modules, and to process and analyze the data through fusion algorithms. The central control module is used to coordinate the work of various modules and achieve unified management of cargo tracking; The data storage module is used to store cargo information, tracking records, and system operation data; The display interaction module is used to show cargo tracking information and system status.

[0013] The image processing unit of this system adopts a GPU-accelerated architecture, and the camera adopts a pixel CMOS sensor; The central control module of this system uses an industrial-grade computer to run warehouse management software and a database management system. The system is used to coordinate the work of various modules and achieve unified management and scheduling of cargo tracking; The data storage module of this system adopts a relational database architecture, supports distributed storage and parallel querying, and is used to store cargo information, tracking records and system operation data; The spatiotemporal alignment unit of the system uses a timestamp synchronization algorithm and a spatial coordinate transformation algorithm to synchronize the RFID reading time and visual acquisition time to milliseconds, and converts data from different coordinate systems into a unified warehouse coordinate system.

[0014] The data fusion unit of this system uses an adaptive Kalman filter algorithm to dynamically adjust the fusion weights based on the confidence levels of RFID data and visual data, thereby achieving the fusion of multi-source data.

[0015] The system's location calculation unit employs a hybrid positioning algorithm that combines triangulation and visual ranging. Based on fused data, it calculates the three-dimensional coordinates of goods in the warehouse, achieving a positioning accuracy of up to centimeters.

[0016] The anomaly detection unit of the system adopts an anomaly detection algorithm based on machine learning. It establishes a normal behavior model by analyzing historical data and detects abnormal cargo status and system anomalies in real time.

[0017] The RFID data processing unit of the system performs multi-level filtering, verification and parsing on the tag information read by the RFID reader, including signal strength threshold filtering, removal of duplicate data, data integrity verification and EPC code parsing.

[0018] The system's visual data processing unit performs preprocessing, feature extraction, and target recognition on images captured by the camera, including image grayscale conversion, Gaussian filtering, edge detection, feature point extraction, and target detection.

[0019] An e-commerce intelligent warehouse goods management system includes the following steps: S1: System initialization, including RFID device frequency calibration and power adjustment, camera intrinsic and extrinsic parameter calibration, time synchronization via Network Time Protocol (NTP), system parameter configuration and database initialization; S2: When goods enter the warehouse, the label information of the goods is automatically read by a fixed RFID reader, and high-resolution images of the goods are captured by a high-definition camera at the same time. S3: Preprocess RFID data and visual data. RFID data preprocessing includes signal strength threshold filtering, duplicate data removal, data integrity verification and EPC code parsing. Visual data preprocessing includes image grayscale conversion, Gaussian filtering, edge detection, feature point extraction and deep learning target detection. S4: RFID data and visual data are associated and matched through a spatiotemporal alignment algorithm. Time alignment uses the NTP protocol to achieve millisecond-level time synchronization. Spatial alignment uses camera calibration parameters to convert image coordinates into world coordinates. Feature matching uses a feature matching algorithm. S5: Based on the fused data, calculate the precise location and status information of goods in the warehouse, use an adaptive Kalman filter algorithm to fuse RFID positioning information and visual positioning information, dynamically adjust the fusion weight according to the data confidence level, and use a hybrid positioning algorithm that combines triangulation and visual ranging. S6: Store the cargo's three-dimensional coordinate location information, status information, timestamp, and other data in the database, and update the inventory records in real time; S7: When goods move within the warehouse, the location changes of goods are tracked in real time through a distributed RFID reader and camera network, and the goods trajectory records are updated. S8: When goods leave the warehouse, the goods information is verified by both RFID reading and visual recognition; S9: The system automatically generates various statistical reports based on the tracking data.

[0020] The spatiotemporal alignment algorithm described in this method includes the following steps: S1 timestamp synchronization uses the Network Time Protocol (NTP) to achieve time synchronization of all devices in the system, with a synchronization accuracy of milliseconds. S2 time alignment establishes a mapping relationship between RFID reading time series and visual acquisition time series, and uses a linear interpolation algorithm to process data with mismatched timestamps; S3 spatial calibration uses Zhang Zhengyou's calibration method to calibrate the camera's intrinsic parameters and a checkerboard calibration board to calibrate the extrinsic parameters. S4 coordinate transformation uses camera calibration parameters to convert image coordinates into world coordinates, establishing a mapping relationship between the image coordinate system and the warehouse coordinate system; S5 data association uses a feature matching algorithm to associate the EPC code of the RFID tag with the visually recognized characteristics of the goods.

[0021] The adaptive Kalman filter algorithm described in this method includes the following steps: S1 state prediction: Predicts the current state based on the state estimate of the previous time step and the motion model; S2 covariance prediction: Predicting state covariance based on process noise covariance; S3 Kalman gain calculation: The Kalman gain is calculated based on the measurement noise covariance and the prediction covariance. S4 State Update: Updates the state estimate based on Kalman gain and measurements; S5 covariance update: Update the state covariance based on the Kalman gain; S6 adaptive weight adjustment dynamically adjusts the fusion weights based on the confidence levels of RFID and visual data.

[0022] The present invention has the following advantages over existing data: By organically combining RFID coarse positioning and visual fine positioning, and using an adaptive Kalman filter algorithm, centimeter-level positioning accuracy is achieved, which is an order of magnitude higher than the meter-level positioning accuracy of existing RFID systems.

[0023] By leveraging GPU-accelerated image processing and a distributed architecture design, it supports real-time tracking of hundreds of items in large-scale warehouse environments, reducing real-time tracking latency.

[0024] By integrating multi-source data and adaptive weight adjustment, the system can maintain stable recognition performance even in complex warehousing environments, thus improving system availability.

[0025] By establishing a normal behavior model through machine learning algorithms, abnormal cargo status and system anomalies can be automatically detected.

[0026] It adopts a modular design and standardized interfaces, which facilitates system upgrades and function expansion, and supports seamless integration with other warehouse management systems. Attached Figure Description

[0027] 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 the structures shown in these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a detailed structural block diagram of the data fusion processing module of the present invention; Figure 3 This is an overall flowchart of the method of the present invention; Figure 4 This is a detailed flowchart of the spatiotemporal alignment algorithm of the present invention; Figure 5 This is a schematic diagram of the adaptive Kalman filter algorithm of the present invention; Figure 6 This is a schematic diagram illustrating the deployment of the system of the present invention in a typical warehouse environment; Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1: System architecture as follows Figure 1 and Figure 6 As shown, the intelligent warehouse cargo tracking system based on the fusion of RFID and visual recognition provided by the present invention adopts a layered modular architecture, which is divided into a perception layer, a processing layer, an application layer and a display layer from bottom to top.

[0031] The perception layer includes RFID tag modules, RFID readers / writers, and a visual recognition module, responsible for collecting raw data. The RFID tag modules use ultra-high frequency passive RFID tags, operating at 860-960MHz. Each tag stores a 96-bit EPC code and a 128-bit user storage area to record information such as product identification, name, specifications, batch number, and production date. The tags feature an anti-metal design, an IP67-rated protective shell, and an added humidity compensation algorithm. It incorporates a humidity compensation algorithm that includes ambient humidity (RH) and a calibration coefficient (k_h), and deploys dual-frequency RFID readers: adding an 866MHz band to address signal attenuation in high humidity. Built-in absorbing material allows it to adapt to metal shelving environments, with a reading distance of 6-8 meters.

[0032] RFID reading and writing equipment includes fixed RFID readers and handheld RFID readers. Fixed RFID readers use the Impinj R420 reader, employing a multi-antenna array, with a reading distance of several meters, supporting the reading of 200 tags per second. They connect to the network via an RJ45 interface and support PoE power supply. Fixed readers are installed at warehouse entrances, main aisles, and key areas to form a comprehensive RFID identification network. Handheld RFID readers use the Zebra RFD8500, supporting Bluetooth and USB connectivity, and are provided to warehouse operators for goods identification and retrieval during mobile operations.

[0033] The visual recognition module includes a webcam and an image processing unit. The webcam is mounted on the warehouse ceiling to ensure all storage areas are within visual coverage. The image processing unit uses a GPU-accelerated architecture to support real-time image processing and deep learning inference.

[0034] In this embodiment, the visual recognition module uses a Hikvision DS-2CD2T45D-I5 high-definition network camera and an NVIDIA Jetson Xavier NX image processing unit. The recommended camera uses a 2-megapixel CMOS sensor, supports H.264 video encoding at 25fps, has a 3.6mm focal length, and a 90-degree field of view. The cameras are mounted on the warehouse ceiling, each covering an area of ​​approximately 100 square meters, ensuring all storage areas are within visual coverage. The image processing unit uses a GPU-accelerated architecture, equipped with a 32-core CPU, a 384-core NVIDIA GPU, 8GB of RAM, and 16GB of storage, supporting real-time image processing and deep learning inference. It also allows for manual storage replacement and automatic cloud uploads.

[0035] Example 2: Data fusion processing module as follows Figure 2 As shown, the data fusion processing module is the core of this invention. It adopts a pipeline architecture to refine multi-source heterogeneous data step by step, specifically including six processing units: The RFID data processing unit is responsible for multi-level processing of the tag data uploaded by the RFID reader / writer. First, signal strength threshold filtering is performed, retaining only valid read records with RSSI values ​​greater than -80dBm; then, duplicate data removal is performed, retaining only the strongest signal read record of the same tag within a short time window; next, data integrity verification is performed, checking the integrity of fields such as EPC code, timestamp, and antenna ID; finally, EPC code parsing is performed to extract the identification information of the goods.

[0036] The visual data processing unit is responsible for image feature reduction and target capture. Limited by the computing power of edge computing devices, this unit extracts frames from the raw video stream captured by the camera as needed, and then sequentially performs grayscale conversion and Gaussian filtering (σ=1.5) to suppress noise caused by the complex lighting conditions in the warehouse. Based on this, it calls the lightweight YOLOv5 network to directly output structured target data containing category, bounding box, and confidence score, skipping the traditional cumbersome feature point matching and ensuring processing latency is controlled within the hundreds of milliseconds.

[0037] The spatiotemporal alignment unit is a key component of this invention, responsible for resolving the spatiotemporal alignment issues between RFID data and visual data. Time alignment employs Network Time Protocol (NTP) to achieve time synchronization across all devices within the system; PTP (Precise Time Protocol) can also be added, achieving millisecond-level synchronization accuracy. Spatial alignment is achieved through camera calibration. The Zhang Zhengyou calibration method is used to calibrate the camera's intrinsic parameters, employing 20 chessboard images in different poses. Each chessboard is 30mm × 30mm in size, with internal corner points of 10 × 7, to obtain the camera's intrinsic parameter matrix and distortion coefficients. An extrinsic parameter calibration is performed using a chessboard calibration board, establishing a mapping relationship between the image coordinate system and the warehouse coordinate system.

[0038] Specifically, the coordinate transformation is achieved using a pinhole camera model and homogeneous coordinate transformation. Let the pixel coordinates of the goods in the image be (u, v), and their corresponding three-dimensional coordinates in the warehouse world coordinate system be (X_w, Y_w, Z_w). The transformation relationship satisfies the following matrix equation:

[0039] Where s is the scale factor, f_x and f_y are the equivalent focal lengths of the camera on the x and y axes, respectively, and c_x and c_y are the coordinates of the principal point of the image. These parameters constitute the intrinsic parameter matrix, obtained through the Zhang Zhengyou calibration method. R is a 3x3 rotation matrix, and T is a 3x1 translation vector. R and T constitute the extrinsic parameter matrix, obtained through the known position of the checkerboard calibration board in the warehouse. Given the known pixel coordinates (u, v) of the goods, to address the error in visual distance judgment caused by the shape of the goods, a laser-assisted ranging module, such as a TOF sensor, is added to key areas to provide a reference distance for the visual algorithm. By combining depth information or solving for the scale factor s through multi-view geometric constraints, the world coordinates (X_w, Y_w, Z_w) of the goods can be inversely solved.

[0040] The data fusion unit employs an adaptive Kalman filter algorithm, which is the core algorithm of this invention. Specifically, the adaptive Kalman filter algorithm constructs a state vector X_k=[x_k,y_k,z_k,\dot{x}_k,\dot{y}_k,\dot{z}_k]^T based on the three-dimensional position and velocity of the cargo. Combined with... Figure 5As shown, the traditional Kalman filter algorithm uses fixed fusion weights, while this invention adopts an adaptive weight adjustment mechanism, specifically performing the following steps: Predict the current state based on the uniform motion model: , where F is the state transition matrix, B is the control matrix, and U_k is the control vector; Covariance prediction: , where P is the state covariance matrix and Q is the process noise covariance; Adaptive Fusion of Measurement Vectors: Constructing Fusion Measurement Vectors Where α_k is the adaptive weight. This invention defines a confidence mapping function: Let the confidence level of the RFID signal be C_r, and the confidence level of visual detection be C_v, then the formula for calculating the adaptive weight α_k is: The corresponding fusion measurement noise covariance matrix R_k is updated as follows: Where R_{rfid} is the RFID positioning variance and R_{vis} is the visual positioning variance; then, the Kalman gain is used to calculate: Where H is the observation matrix; State update: Covariance update: , where I is the identity matrix.

[0041] Through the above steps, when the visual recognition confidence level C_v is high, α_k automatically decreases, and the system trusts the visual positioning results more. When the RFID signal is extremely strong, α_k automatically increases, thereby achieving the optimal fusion positioning output X_{k|k}. For example, when RSSI > -60dBm, the weight of RFID data is increased to 0.7; when the visual recognition confidence level is greater than 0.8, the weight of visual data is increased to 0.7; under normal circumstances, the weights of RFID data and visual data are each 0.5.

[0042] The location calculation unit employs a hybrid positioning algorithm, combining RFID triangulation and visual ranging. RFID triangulation utilizes the signal strength of RFID antennas at at least three known locations to calculate the approximate tag location using a path loss model. Specifically, this RFID triangulation method is based on a logarithmic distance path loss model. Let the signal strength received by the fixed RFID reader's antenna from the tag be RSSI(d) (unit: dBm), and its value be related to the physical distance from the reader to the tag. d The following relationship must be satisfied:

[0043] Where d_0 is the reference distance (usually taken as 1 meter), RSSI(d_0) is the signal strength at the reference distance, and n is the environmental path loss index (taken as 3.5~4.5 in areas with dense metal shelving). εThe noise is Gaussian random noise with a mean of 0. The distance estimate from the tag to the i-th antenna is obtained by inverse solving the above equation. di :

[0044] Furthermore, using the RFID antenna coordinates (x_i, y_i, z_i) at at least three known locations, the following system of nonlinear equations is established: The equations were solved iteratively using the least squares method to obtain the approximate three-dimensional coordinates (x_{rfid}, y_{rfid}, z_{rfid}) of the RFID tag.

[0045] Visual ranging utilizes the geometric relationships of monocular vision to calculate distance using the known actual size of the object and the pixel size in the image; specifically, this visual ranging method leverages the pinhole imaging principle and the geometric relationships of similar triangles. Let the actual physical width of the cargo target be W (unit: meters), and the pixel width occupied by this target in the image be... w (Unit: pixels), the horizontal focal length of the camera is... fx (Unit: pixels), then the vertical distance from the optical center of the camera to the surface of the goods. Z Calculated using the following formula:

[0046] By combining the pixel coordinate offset of the target in the image, the three-dimensional coordinates of the goods in the camera coordinate system can be calculated. Then, through the extrinsic parameter matrix of the aforementioned spatiotemporal alignment unit, the coordinates are transformed to the warehouse world coordinate system to obtain the visual positioning coordinates (x_{vis}, y_{vis}, z_{vis}). The two positioning results are then weighted and fused to achieve centimeter-level positioning accuracy.

[0047] The anomaly detection unit serves as a safety fallback, employing the Isolation Forest algorithm to perform unsupervised learning on the trajectory sequence output from the fusion process. By extracting speed, acceleration, and path features during normal cargo relocation to construct a normal behavior profile, this unit can accurately detect abnormal events such as sudden position changes, excessive speed, or illegal paths, significantly improving stability.

[0048] Example 3: Tracking Method Flow like Figure 3 As shown, the present invention provides a smart warehouse cargo tracking method based on the fusion of RFID and visual recognition. System initialization. First, the RFID device is calibrated to precisely adjust the operating frequency to 915MHz and the power to approximately 30dBm. Then, the camera is calibrated, including internal and external parameter calibration. Next, the NTP time server is configured to achieve time synchronization of all devices in the system. Finally, system parameters are configured, including database connection parameters, network parameters, algorithm parameters, etc., and the database table structure is initialized.

[0049] Taking a specific warehousing operation as an example: When goods pass through the warehousing channel, a fixed RFID reader automatically reads the goods' tag information, while a camera simultaneously captures a high-resolution image of the goods. The RFID data includes the tag's EPC code 3034F5B90000000000001234, the reading timestamp 2024-01-15 10:30:25, the signal strength -65dBm, and the antenna ID Antenna1. The visual data includes the image frame (1920×1080 pixels), the acquisition timestamp 2024-01-15 10:30:25, the camera ID Camera05, and the target detection result Box: [100,200,300,400], Class: "Carton", Confidence: 0.95.

[0050] The RFID data was filtered using a signal strength threshold, retaining 6 valid read records with RSSI values ​​greater than -80dBm; duplicate data removal was performed, merging read records of the same tag within a 1-second time window to obtain 3 records; data integrity verification was performed to ensure all fields were complete and valid; EPC code parsing was performed, such as extracting the cargo ID "1234". The visual image was converted to grayscale; noise removal was performed; edge detection was performed to extract edge information; multiple feature points were extracted using the ORB algorithm; and a YOLOv5 model was used to detect one cargo target with a confidence level of 0.95.

[0051] Time synchronization is achieved using the NTP protocol with a synchronization accuracy of ±5ms. A mapping relationship is established between the RFID reading time series and the visual acquisition time series, and a linear interpolation algorithm is used to handle data with mismatched timestamps. The image coordinates [100, 200] are converted to world coordinates [5.2, 3.8, 1.5] (unit: meters) using camera calibration parameters. A feature matching algorithm based on deep learning is used to associate the RFID tag EPC code 1234 with the visually recognized cargo features, with an association confidence level of 0.92, thereby achieving temporal and spatial alignment.

[0052] An adaptive Kalman filter algorithm is used to fuse RFID and visual positioning information. The RFID triangulation method calculates the tag position as [5.1, 3.9, 1.4] with a confidence level of 0.85; the visual ranging method calculates the cargo position as [5.3, 3.7, 1.6] with a confidence level of 0.95. In this embodiment, assuming the current RFID signal confidence level C_r = 0.85 and the visual detection confidence level C_v = 0.95, substituting these into the adaptive weight formula, the system calculates an RFID weight of 0.44 and a visual weight of 0.56. For ease of understanding, this can be approximated as RFID: 0.3 and visual: 0.7. Substituting the coarse RFID positioning result and the fine visual positioning result into the state update equation, the fused position is calculated to be [5.24, 3.76, 1.54]. Due to the higher confidence level of visual recognition, the positioning error is 5cm.

[0053] The data such as the three-dimensional coordinates of the goods [5.24, 3.76, 1.54], the status information "normal", and the timestamp "2024-01-15 10:30:25" are stored in a distributed database, and the inventory records are updated in real time. The status of goods "1234" is updated from "in transit" to "in stock", and the location is updated to [5.24, 3.76, 1.54], thereby updating the warehouse data in real time.

[0054] When goods move within the warehouse, their location is tracked in real time via a distributed RFID reader and camera network. The system updates the goods' location every second, continuously recording their trajectory. When goods move from location [5.24, 3.76, 1.54] to [6.8, 4.2, 1.5], the system detects a movement distance of 1.8 meters, a movement time of 2 seconds, and a movement speed of 0.9 m / s, all within the normal range for real-time tracking.

[0055] When goods leave the warehouse, the system reads the EPC code 1234 of the tag through RFID, confirms that the appearance of the goods is normal through visual recognition, and allows the goods to leave the warehouse after passing the double verification to prevent incorrect shipment.

[0056] The system automatically generates cargo flow reports based on tracking data, displaying the complete trajectory of goods from warehousing to outbound; it also generates inventory analysis reports, statistically analyzing the inventory levels and turnover rates of various goods; and it generates anomaly reports, recording detected anomalies. Reports are output in PDF format and can be automatically sent via email.

[0057] In summary, the system also includes an edge module, where edge devices upload 10% of their sample data to the cloud daily, and the cloud model is updated and distributed to the edge devices monthly. A digital twin module is also added: a 3D warehouse model is built in the cloud, enabling offline simulation and path optimization, which allows for better model training.

[0058] The above-described values ​​are merely examples and are not intended to limit the scope of this invention. They are only preferred embodiments of this patent, but the scope of protection of this patent is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in this patent, based on the technical solution and inventive concept of this patent, shall fall within the scope of protection of this patent.

Claims

1. An e-commerce intelligent warehouse goods management system, characterized in that, include: The RFID tag module uses ultra-high frequency passive RFID tags to store the identification information and attribute data of goods. RFID reading and writing equipment, including fixed RFID readers and handheld RFID readers, wherein the fixed RFID readers adopt a multi-antenna array design; The visual recognition module includes multiple high-definition network cameras and an image processing unit. The cameras use pixel-level sensors, and the image processing unit uses an accelerated architecture. An auxiliary ranging module, including a TOF sensor, is installed in key areas of the warehouse to provide a reference distance for the visual algorithm. The data fusion processing module includes an RFID data processing unit, a visual data processing unit, a spatiotemporal alignment unit, a data fusion unit, a location calculation unit, and an anomaly detection unit. It is used to receive multi-source heterogeneous data from RFID reading and writing devices and visual recognition modules, and to process and analyze the data through fusion algorithms. The central control module is used to coordinate the work of various modules and achieve unified management of cargo tracking; The data storage module is used to store cargo information, tracking records, and system operation data; The display interaction module is used to show cargo tracking information and system status.

2. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The spatiotemporal alignment unit uses a timestamp synchronization algorithm and a spatial coordinate transformation algorithm to synchronize the RFID reading time and visual acquisition time to milliseconds, and converts data from different coordinate systems into a unified warehouse coordinate system.

3. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The data fusion unit employs an adaptive Kalman filter algorithm to dynamically adjust the fusion weights based on the confidence levels of RFID data and visual data, thereby achieving the fusion of multi-source data.

4. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The location calculation unit uses a hybrid positioning algorithm that combines triangulation and visual ranging to calculate the three-dimensional coordinates of goods in the warehouse based on fused data.

5. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The anomaly detection unit employs a machine learning-based anomaly detection algorithm. By analyzing historical data, it establishes a normal behavior model and detects abnormal cargo status and system anomalies in real time.

6. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The RFID data processing unit performs multi-level filtering, verification, and parsing on the tag information read by the RFID reader, including signal strength threshold filtering, removal of duplicate data, data integrity verification, and EPC code parsing.

7. The e-commerce intelligent warehouse goods management system according to claim 1, characterized in that, The visual data processing unit performs preprocessing, feature extraction, and target recognition on the images captured by the camera, including image grayscale conversion, Gaussian filtering, edge detection, feature point extraction, and target detection.

8. An e-commerce intelligent warehouse goods management system, characterized in that, Includes the following steps: Step S1: System initialization, including RFID device frequency calibration and power adjustment, camera intrinsic and extrinsic parameter calibration, time synchronization via Network Time Protocol (NTP), system parameter configuration, and database initialization; Step S2: When goods enter the warehouse, the goods tag information is automatically read by a fixed RFID reader, and high-resolution images of the goods are captured by a high-definition camera at the same time. Step S3: Preprocess the RFID data and visual data. RFID data preprocessing includes signal strength threshold filtering, duplicate data removal, data integrity verification and EPC code parsing. Visual data preprocessing includes image grayscale conversion, Gaussian filtering, edge detection, feature point extraction and deep learning target detection. Step S4: The RFID data and visual data are associated and matched through a spatiotemporal alignment algorithm. The time alignment uses the NTP protocol to achieve millisecond-level time synchronization, the spatial alignment uses camera calibration parameters to convert image coordinates into world coordinates, and the feature matching algorithm is used for feature matching. Step S5: Calculate the precise location and status information of goods in the warehouse based on the fused data. Use an adaptive Kalman filter algorithm to fuse RFID positioning information and visual positioning information. Dynamically adjust the fusion weights according to the data confidence level. Use a hybrid positioning algorithm that combines triangulation and visual ranging. Step S6: Store the cargo's three-dimensional coordinate location information, status information, timestamp, and other data in the database, and update the inventory records in real time; Step S7: When goods move within the warehouse, the location changes of the goods are tracked in real time through a distributed RFID reader and camera network, and the goods trajectory records are updated. Step S8: When goods leave the warehouse, the goods information is verified by both RFID reading and visual recognition; Step S9: The system automatically generates various statistical reports based on the tracking data.

9. The method according to claim 8, characterized in that, The spatiotemporal alignment algorithm includes the following steps: S1 timestamp synchronization uses the Network Time Protocol (NTP) to achieve time synchronization of all devices in the system, with a synchronization accuracy of milliseconds. S2 time alignment establishes a mapping relationship between RFID reading time series and visual acquisition time series, and uses a linear interpolation algorithm to process data with mismatched timestamps; S3 spatial calibration uses Zhang Zhengyou's calibration method to calibrate the camera's intrinsic parameters and a checkerboard calibration board to calibrate the extrinsic parameters. S4 coordinate transformation uses camera calibration parameters to convert image coordinates into world coordinates, establishing a mapping relationship between the image coordinate system and the warehouse coordinate system; S5 data association uses a feature matching algorithm to associate the EPC code of the RFID tag with the visually recognized characteristics of the goods.

10. The e-commerce intelligent warehouse goods management system according to claim 8, characterized in that, The adaptive Kalman filter algorithm includes the following steps: S1 state prediction: Predicts the current state based on the state estimate of the previous time step and the motion model; S2 covariance prediction: Predicting state covariance based on process noise covariance; S3 Kalman gain calculation: The Kalman gain is calculated based on the measurement noise covariance and the prediction covariance. S4 State Update: Updates the state estimate based on Kalman gain and measurements; S5 covariance update: Update the state covariance based on the Kalman gain; S6 adaptive weight adjustment dynamically adjusts the fusion weights based on the confidence levels of RFID and visual data.

Images