Accurate warehouse goods checking method based on cooperation of RFID and unmanned aerial vehicle

Abstract

The invention discloses a warehouse goods accurate inventory method based on RFID and unmanned aerial vehicle cooperation, and the method comprises the steps: S1, pasting an RFID tag on a warehouse-in goods, binding the warehouse-in goods, and putting the warehouse-in goods on a shelf, and storing the goods inventory information in a warehouse management system; s2, establishing a three-dimensional model and a space coordinate system of the warehouse, wherein each storage location corresponds to a unique three-dimensional space coordinate; s3, the unmanned aerial vehicle executes an inventory task according to a preset route; s4, the unmanned aerial vehicle binds the RFID tag with the strongest signal when hovering in front of each storage location, so that the storage locations are in one-to-one correspondence with the goods; s5, comparing the RFID signal value acquired when the unmanned aerial vehicle flies to the next storage location with the RFID signal value acquired when the unmanned aerial vehicle flies to the previous storage location, and binding the tag with the strongest signal to the current storage location; and S6, after the checking task is finished and the unmanned aerial vehicle returns, the system automatically generates a checking report, compares the RFID real-time data with the WMS inventory record and marks a difference item. According to the invention, efficient and accurate unmanned inventory checking can be realized, and the inventory quantity and inventory position accuracy of inventory checking can be ensured.

Description

Technical Field

[0001] This invention relates to a warehouse inventory counting method, and more particularly to a precise warehouse inventory counting method based on RFID (Radio Frequency Identification) and drone collaboration, which is applicable to automated inventory counting in scenarios such as large warehouses and logistics centers. Background Technology

[0002] In the context of Industry 4.0, high-precision, real-time sensing of the spatial location of physical entities in complex environments has become a core prerequisite for achieving intelligent decision-making. Radio Frequency Identification (RFID) technology, with its inherent characteristics of non-line-of-sight, non-contact, and multi-target concurrent identification, has shown great potential in asset tracking, warehousing and logistics, and intelligent manufacturing. However, traditional RFID systems primarily serve for identification, with positioning accuracy limited to meters or even tens of meters. This is fundamentally due to its single-dimensional, low-robust ranging model based on Received Signal Strength Indication (RSSI). This model simplifies the complex propagation of spatial electromagnetic waves into an ideal path loss model. However, the multipath effects, shadowing fading, reader and tag antenna polarization mismatch, and spatial anisotropic radiation of RF signals prevalent in real-world environments result in a highly nonlinear and uncertain relationship between the RSSI value and the actual distance, severely restricting its application effectiveness in precise positioning scenarios.

[0003] In warehouse management, inventory counting is a crucial step in ensuring that the physical inventory matches the records. "Inventory counting" refers to the operational process by which a company periodically or temporarily checks and counts the actual quantity of inventory, aiming to verify the consistency between the records and the actual inventory and ensure the accuracy of inventory management.

[0004] In recent years, with the development of the low-altitude economy, drone technology has been gradually applied to warehouse management, especially in the field of inventory counting. RFID technology, due to its non-contact and batch identification characteristics, combined with the flexibility, maneuverability, and wide coverage of drones, makes drone + RFID collaboration an ideal choice for automated inventory counting in the industry. However, existing technologies still have the following shortcomings: 1. Inaccurate warehouse location: During the flight of the drone, it continuously acquires signals emitted by RFID in batches, but it cannot accurately determine the specific warehouse location corresponding to each tag. It can only realize quantity inventory, which cannot meet the needs of warehouse location management.

[0005] 2. Complex data processing: How to efficiently integrate RFID data collected by drones and automatically compare it with the WMS inventory system still needs optimization.

[0006] Therefore, there is an urgent need for a precise inventory counting method for warehouse goods based on RFID and drone collaboration to improve the level of intelligent warehouse inventory counting. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a warehouse goods accurate inventory method based on RFID and drone collaboration, which can not only achieve very high-precision RFID "fingerprint" positioning, but also achieve efficient and low-risk unmanned inventory, and can guarantee the accuracy of both inventory quantity and inventory location.

[0008] To address the aforementioned technical problems, this invention provides a method for accurate warehouse inventory counting based on RFID and drone collaboration, comprising the following steps: S1, after receiving goods, RFID tags are affixed and bound to the shelves, and the inventory information is stored in the warehouse management system; S2, a three-dimensional model and spatial coordinate system of the warehouse are established, with each storage location corresponding to a unique three-dimensional spatial coordinate; S3, the drone performs the inventory counting task according to a preset flight path; S4, when the drone hovers in front of each storage location, it binds the RFID tag with the strongest signal, ensuring a one-to-one correspondence between storage locations and goods; S5, when the drone flies to the next storage location, the RFID signal value collected is compared with that of the previous storage location, and the tag with the strongest signal is bound to the current storage location; S6, after the inventory counting task is completed and the drone returns, the system automatically generates an inventory report, compares the real-time RFID data with the WMS inventory records, and marks the discrepancies for manual review.

[0009] Further, step S1 includes: after the goods arrive, randomly select 10% of the sample size for random inspection and after confirming that there are no errors, first affix RFID tags to complete the binding relationship between the tags and the goods, and then scan the goods and the storage location number after the goods are put on the shelf to complete the binding relationship between the goods and the storage location.

[0010] Furthermore, step S2 completes the point cloud map construction during the offline training phase: a drone equipped with a lidar is used to perform a three-dimensional scan of the entire warehouse scene to collect point clouds; the collected point cloud data is preprocessed to extract key geometric features; a mapping relationship between point cloud feature descriptors and spatial locations is established; and all feature-location pairs are stored in the point cloud feature database to form a point cloud feature library that can be used for positioning.

[0011] Furthermore, the detailed mapping process for step S2 is as follows: i. 3D spatial scanning: Plan the drone flight path to ensure coverage of all areas of the warehouse, and use high-precision LiDAR for multi-view scanning.

[0012] ii. Unified coordinate system: Arrange several reflective column calibration points with known coordinates in the warehouse to obtain a set of references with precise coordinates {P1(x1,y1,z1), P2(x2,y2,z2),...Pn(xn,yn,zn)}; take the projection center of the UAV nest on the ground as the origin of the world coordinate system (0,0,0) to ensure that all scanned point clouds are unified to the same global coordinate system.

[0013] iii. Point cloud data processing: Downsampling, noise reduction, and ground segmentation are performed on the original point cloud. Key feature points are extracted, the local coordinate system is initialized, and the local point cloud is transformed into the global coordinate system and fused with the existing map.

[0014] Further, step S2, point cloud feature extraction and database construction, is performed as follows: point cloud segmentation and keyframe extraction, assigning a unique identifier and precise pose information to each keyframe; for keyframe point cloud P1, feature descriptors are extracted to obtain the first feature vector: V1=[FPFH1, mean curvature, variance of curvature, plane normal vector, line segment direction, point density, height statistics]; The vector V1 is bound to its corresponding geographic location P1(x1,y1,z1) and stored as a record in the database; then the second keyframe point cloud P2 is recorded, and the whole process is repeated until all point clouds are collected; thus, a priori map feature database that supports fast retrieval based on feature similarity is obtained.

[0015] Furthermore, the warehouse 3D model established in step S2 is 1:1 in size with the actual warehouse. The warehouse 3D model includes a receiving and dispatching area and high-level shelves in the storage area. The height of each shelf corresponds to the actual items. The center position of each storage location corresponds to its own corresponding 3D spatial coordinate point (x, y, z).

[0016] Furthermore, in step S3, connecting two points in spatial coordinates forms a straight line. Connecting the spatial points of each storage location forms the preset flight path for the drone. The drone will perform different flight inventory tasks based on different flight paths. The drone briefly hovers at each storage location to collect the signals of all RFID tags at the current location. Based on the RFID signal strength, the drone matches the storage location and binds the tag with the strongest signal to the current storage location. If the strongest signal strength of a storage location is lower than a preset threshold, it is removed, ensuring that empty storage locations are not bound to goods.

[0017] Further, step S3, the online positioning stage, includes: i. Real-time point cloud acquisition: The UAV acquires local point clouds in real time during the inventory process; and extracts the real-time feature vector V current. ii. Feature matching and position estimation: The feature vector most similar to V current is searched in the prior map and compared to obtain the real-time position of P current.

[0018] Furthermore, steps S4 and S5, based on the real-time collected tag signals, use a combination of the following two algorithms to find the most matching location in the prior map feature database: Basic K-nearest neighbor localization algorithm: find the K reference points in the feature database that are most similar to the real-time vector, and then take the average of the position coordinates of these K points as the final localization result; Weighted K-nearest neighbor localization algorithm: perform a weighted average of the K nearest neighbor points according to their similarity, with points having higher similarity having greater weight.

[0019] Further, step S4 or S5 uses Euclidean distance or Manhattan distance to measure similarity. The specific matching and binding process is as follows: i. Obtain the current signal vector: Convert the currently read tag signal into a vector V_current; each dimension of the vector represents a known tag ID, and its value is the currently read RSSI value; ii. Obtain the reference feature vector: From the feature database, retrieve the reference feature vector V_ref of the current storage location L_current; iii. Calculate the similarity: Calculate the similarity between the current vector V_current and the reference feature vector V_ref using the reciprocal of the Euclidean distance or Manhattan distance, and introduce weights during the calculation, giving higher weights W_i to tags with strong signals in V_ref (representing that it is an inherent tag of this storage location), where the weight W_i is a function of the tag's RSSI value in V_ref; iv. Find the best matching tag: Calculate the signal change vector ΔV = V_current - V_ref; If the dimension corresponding to the label of the current storage location has a large positive change in ΔV, while the signal changes of other labels are small or negative, it is determined that new goods have been placed in the current storage location; v. Decision and binding: Traverse the ΔV vector and find the label T_best with the largest increase in signal strength; At the same time, set a minimum gain threshold. When the maximum gain exceeds the minimum gain threshold, bind the storage location L_current to the label T_best.

[0020] Furthermore, the inventory report automatically output in step S6 includes two types of discrepancies: incorrect inventory location and missing inventory in the storage location.

[0021] Compared with existing technologies, this invention offers the following advantages: The warehouse inventory management method based on RFID and drone collaboration provided by this invention utilizes drones equipped with RFID reading and writing devices, combined with precise spatial positioning technology, to achieve efficient and accurate inventory management, effectively solving the problems of low efficiency, large errors, and high risks associated with high-altitude operations in traditional inventory management methods. It possesses the following technical advantages: 1. Utilize information technology to overcome the inherent limitations of RFID in space environment applications, and achieve stable and accurate spatial coordinate positioning of RFID technology that relies on signal strength and phase change mechanisms in complex environments.

[0022] 2. By using the above methods and combining drones with RFID for precise coordinate positioning, inventory management of goods can be effectively achieved, realizing a one-to-one correspondence between goods and warehouse locations, thus solving the problem of traditional drone inventory counting being unable to locate goods.

[0023] 3. Location-level precision inventory provides a reliable inventory solution for large warehouses with strict requirements for location accuracy, achieving efficient, accurate, and low-cost unmanned inventory. Attached Figure Description

[0024] Figure 1 This is a flowchart of the precise inventory counting process for warehouse goods based on RFID and drone collaboration, as described in this invention. Figure 2 This is a schematic diagram illustrating the precise location of warehouse goods based on RFID according to the present invention. Detailed Implementation

[0025] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a flowchart illustrating the precise inventory counting process for warehouse goods based on RFID and drone collaboration, as described in this invention.

[0027] Please see Figure 1 The present invention provides a method for accurate inventory counting of warehouse goods based on RFID and drone collaboration, which includes the following steps: S1. When the goods arrive at the warehouse receiving area, after comparing them with the delivery list and confirming that they are correct, RFID tags are affixed, the goods are manually placed on high shelves, and the inventory information of the goods is maintained in the WMS (Inventory Management System).

[0028] S2. Establish a 3D model and precise spatial coordinate system for the warehouse. Set the center of the drone's nest on the ground as the origin of the spatial coordinate system (0,0,0). Each storage location on the high-level rack in the warehouse corresponds to a unique 3D spatial coordinate (x,y,z). The connection between each spatial point (i.e., storage location) is planned as the flight path of the drone.

[0029] S3. The drone performs inventory tasks according to the preset route.

[0030] S4. The drone hovers briefly at each storage location (approximately 0.5 seconds) to collect all RFID tag signals at the current location. Based on the RFID signal strength (RSSI, Received Signal Strength Indicator), the drone matches the storage location and binds the tag with the strongest signal to the current storage location, ensuring an accurate correspondence between the storage location and the goods. S5. When the drone flies to the next storage location, it will also collect all the RFID tag signals of this storage location and compare them with all the RFID signal values ​​of the previous storage location. Then, it will bind the tag with the strongest signal to the current storage location to ensure the accuracy of the goods bound to the storage location. This process will be repeated to cover the entire inventory task. S6. All collected information will be stored in the drone system. After the inventory task is completed and the drone returns, the system will automatically generate an inventory report, compare the RFID data with the WMS inventory records, and mark the differences for manual review.

[0031] In this embodiment, after the goods arrive in step S1, 10% of the sample size is randomly selected for inspection. After the inspection is confirmed to be correct, RFID tags are manually affixed to complete the binding relationship between the tags and the goods. After the goods are put on the shelf, the goods and the storage location number are scanned to complete the binding relationship between the goods and the storage location, thereby realizing the correspondence between the RFID tags, the goods and the storage location.

[0032] In this embodiment, the 3D warehouse model established in step S2 is 1:1 in size with the actual warehouse. The model layout completely restores the actual warehouse scene, including the receiving and dispatching area and the high-level shelves in the storage area. The height of each shelf corresponds one-to-one with the actual items. The established spatial rectangular coordinate system takes the center of the drone's nest projection on the ground as the origin (0,0,0). The center position of each storage location corresponds to its own corresponding 3D spatial coordinate point (x,y,z). Connecting two points forms a straight line. Connecting the spatial points of each storage location forms the preset flight path of the drone. The drone will perform different flight inventory tasks according to different flight paths.

[0033] In this embodiment, in step S4, the drone hovers at each storage location along a preset route. First, it compares the current real-time location with the prior map. Then, it collects all RFID tag signals at the current location. The radio frequency signal emitted by the RFID tag closest to the drone is the strongest when it is received by the drone. Therefore, the drone binds the tag with the strongest signal to the current storage location, thereby accurately realizing the precise correspondence between the storage location and the goods. The RFID tag will not be matched to a nearby storage location.

[0034] In this embodiment, each time the drone performs a task in step S5, it receives signals from different RFID tags while hovering and binds the goods with the strongest signal strength at the current location to that storage location. It also analyzes all signal strengths received in an inventory task in real time and calculates a threshold through an algorithm. If the strongest signal strength of a storage location is also lower than the threshold, it is removed to ensure that empty storage locations are not bound to goods.

[0035] After the drone finishes its inventory count in step S6 of this embodiment, it will compare the real-time inventory data with the WMS inventory records and automatically output an inventory report marking discrepancies, including two types of discrepancies: incorrect inventory location and missing inventory in the storage location. The operator can then recount the physical inventory based on the marked discrepancies.

[0036] To achieve accurate inventory counting in warehouses, a precise RFID positioning method is crucial. Therefore, this invention employs the following positioning method: First, point cloud map construction is completed during the offline training phase: a drone equipped with LiDAR is used to perform a 3D scan of the entire warehouse scene to collect point cloud data. The collected point cloud data is preprocessed to extract key geometric features. A mapping relationship between point cloud feature descriptors and spatial locations is established. All feature-location pairs are stored in the point cloud feature database to form a point cloud feature library that can be used for localization.

[0037] Secondly, the online positioning phase includes: i. Real-time point cloud acquisition: The UAV acquires local point clouds in real time during the inventory process; extracts the real-time feature vector Vcurrent. ii. Feature matching and position estimation: Search for the feature vector most similar to Vcurrent in the prior map and compare them to obtain the real-time position of Pcurrent.

[0038] The following examples illustrate the application of the present invention in accurate warehouse inventory counting based on RFID and drone collaboration.

[0039] Once the goods arrive at the warehouse receiving area, they are checked against the delivery list and RFID tags are affixed. The goods are then manually placed on high shelves, and the inventory information is maintained in the WMS (Inventory Management System). Establish a 3D model and precise spatial coordinate system for the warehouse. Set the center of the drone's nest on the ground as the origin of the spatial coordinate system (0,0,0). Each storage location on the high-level rack in the warehouse corresponds to a unique 3D spatial coordinate (x,y,z). The connection between each spatial point (i.e., storage location) is planned as the flight path of the drone. The drone performs inventory tasks along a preset route and hovers briefly at each storage location (about 0.5 seconds) to collect all RFID tag signals at the current location. Based on the RFID signal strength (RSSI, Received Signal Strength Indicator), the drone matches the storage location and binds the tag with the strongest signal to the current storage location to ensure the accurate correspondence between storage location and goods. When the drone flies to the next storage location, it will collect all the RFID tag signals of this storage location and compare them with all the RFID signal values ​​of the previous storage location. Then, it will bind the tag with the strongest signal to the current storage location to ensure the accuracy of the goods bound to the storage location. This process will be repeated to cover the entire inventory task. All collected information will be stored in the drone system. After the inventory task is completed and the drone returns, the system will automatically generate an inventory report, compare the RFID data with the WMS inventory records, and mark the differences for manual review.

[0040] As a preferred approach, point cloud map construction is completed during the offline training phase: a drone equipped with LiDAR is used to perform a 3D scan of the entire warehouse scene to collect point clouds, the collected point cloud data is preprocessed to extract key geometric features, a mapping relationship between point cloud feature descriptors and spatial locations is established, and all feature-location pairs are stored in the point cloud feature database to form a point cloud feature library that can be used for localization.

[0041] As a preferred method, the detailed process for mapping during the offline training phase is as follows: i. 3D spatial scanning: Plan the drone flight path to ensure coverage of all areas of the warehouse, and use high-precision LiDAR for multi-view scanning.

[0042] ii. Unified coordinate system: Arrange several reflective column calibration points with known coordinates in the warehouse to obtain a set of references with precise coordinates {P1(x1,y1,z1), P2(x2,y2,z2),...Pn(xn,yn,zn)}; take the projection center of the UAV nest on the ground as the origin of the world coordinate system (0,0,0) to ensure that all scanned point clouds are unified to the same global coordinate system.

[0043] iii. Point cloud data processing: Downsampling, noise reduction, and ground segmentation are performed on the original point cloud. Key feature points are extracted, the local coordinate system is initialized, and the local point cloud is transformed into the global coordinate system and fused with the existing map.

[0044] Preferably, during the offline training phase, point cloud feature extraction and database construction are performed as follows: point cloud segmentation and keyframe extraction, assigning a unique identifier and precise pose information to each keyframe; for keyframe point cloud P1, feature descriptors are extracted to obtain the first feature vector: V1=[FPFH1, mean curvature, variance of curvature, plane normal vector, line segment direction, point density, height statistics]; The vector V1 is bound to its corresponding geographic location P1(x1,y1,z1) and stored as a record in the database; then the second keyframe point cloud P2 is recorded, and the whole process is repeated until all point clouds are collected; thus, a priori map feature database that supports fast retrieval based on feature similarity is obtained.

[0045] Preferably, the online positioning stage includes: i. Real-time point cloud acquisition: The UAV acquires local point clouds in real time during the inventory process; and extracts the real-time feature vector V current. ii. Feature matching and position estimation: The feature vector most similar to V current is searched in the prior map and compared to obtain the real-time position of P current.

[0046] As a preferred method, during the online positioning phase, the best matching location is found in the prior map feature database based on the real-time collected tag signals. After online positioning begins, real-time signals are collected, followed by feature engineering and data standardization of the real-time data. A combination of KNN and WKNN matching positioning algorithms is used to obtain the K nearest neighbor values ​​and calculate the weighted position to evaluate the accuracy. The two algorithms are as follows: The basic K-Nearest Neighbors (KNN) localization algorithm finds the K reference points in a feature database that are most similar to the real-time vector, and then takes the average of the position coordinates of these K points as the final localization result. Similarity is usually measured by Euclidean distance or Manhattan distance.

[0047] Weighted K-Nearest Neighbor (WKNN) localization algorithm: It calculates a weighted average of K nearest neighbor points based on their similarity (the inverse of the distance). Points with higher similarity have greater weights, resulting in more accurate results.

[0048] As a preferred method, after the goods arrive, a random sample of 10% is selected for inspection. If no errors are found, RFID tags are manually affixed to complete the binding relationship between the tags and the goods. After the goods are put on the shelves, the RFID tags and storage location numbers are scanned to complete the binding relationship between the tags and the storage locations, thereby realizing the correspondence between RFID tags, goods and storage locations.

[0049] Ideally, establishing a 3D model and precise spatial coordinate system for the warehouse requires a systematic solution combining hardware, software, and computing. By registering and mapping collected real-world data with the virtual warehouse model, each physical storage location is assigned a precise and unique virtual coordinate. This process includes the following key stages: 1. Data Acquisition and Model Building: Using an existing CAD model of the building design as an initial reference, a 3D laser scanner (LiDAR) mounted on a drone is dynamically moved to scan, enabling rapid and high-precision acquisition of point cloud data for the entire warehouse. This combination significantly reduces data processing volume, outputting a dataset containing millions of point clouds, each with its own (x, y, z) coordinates. This raw data contains noise and a large amount of redundant information, thus requiring preprocessing. This invention employs an iterative nearest-point algorithm and its variants for point cloud registration. The point clouds scanned from multiple angles during the drone's dynamic movement are iteratively analyzed to find the closest point pairs between two point clouds, and the optimal rigid body transformation (rotation matrix and translation vector) is calculated. This is combined with industrial inspection (comparison with the CAD model) to achieve optimal alignment. Finally, multiple point cloud fragments are merged into a complete warehouse model.

[0050] 2. Storage location identification and coordinate calculation: Since the warehouse shelves are arranged in a regular pattern, the RANSAC algorithm is used to extract large planes (such as the ground and shelf panels) from the point cloud. Then, the Euclidean clustering algorithm is used to separate different shelves and goods. After identifying the shelf height and columns, the boundaries of each storage location are virtually defined based on prior knowledge, generating an Axis-Aligned Bounding Box (AABB) for each location. The coordinate axes of the AABB are parallel to the world coordinate system, so this cubic space can be perfectly defined using the coordinates of two diagonal points (min_x, min_y, min_z) and (max_x, max_y, max_z). The three-dimensional coordinates (x, y, z) of the center point of each storage location are then calculated. x = (min_x + max_x) / 2; y = (min_y + max_y) / 2; z = (min_z + max_z) / 2; the coordinates (x, y, z) of the center point are the "unique three-dimensional spatial coordinates" of the storage location.

[0051] 3. Data association and storage: Data Structure: Create an SQL database table, where each entry represents a storage location. For example: location_ID:"B01-01-01" (Storage location number, B01 row 1 column 1 level); coordinate_x: 1310.0 (unit: millimeters); coordinate_y: 850.0 (unit: millimeters); coordinate_z: 1100.0 (unit: millimeters); bbox_min_x: 1260.0, bbox_max_x: 1360.0 (stores AABB information, which can be used for path planning and obstacle avoidance); calculate the center coordinates and dimensions of all AABBs, bind them with the storage location ID, and store them in the database.

[0052] As a preferred method, the ability of a drone to perform different flight inventory tasks based on different flight routes depends on the countless point clouds in its map and the SLAM algorithm used to construct the map. A point cloud is a sparse sample of the physical world's surface. It is not a continuous surface, but rather uses a large number of discrete points to infinitely approximate real objects. To improve map accuracy, a 3D LiDAR SLAM (Simultaneous Localization and Mapping) algorithm is introduced, the core of which simultaneously solves the two tasks of localization and mapping. The LiDAR on the drone continuously analyzes the matching relationships between the scanned point clouds during dynamic movement, estimating its complete trajectory. With precise pose, each frame of point cloud is transformed to the same coordinate system. Thousands of single-frame point clouds are precisely stitched together to form a global, consistent, and seamless point cloud map. Simultaneously, the SLAM algorithm also has "loop closure detection" capabilities. When the drone returns to its original location, the algorithm can still identify and correct the small errors accumulated throughout the trajectory, thus greatly improving the global consistency of the map. For warehouse mapping, 16-line or 32-line mechanical rotating radar is usually chosen, which can ensure accuracy while offering good cost performance and a 360° field of view.

[0053] Preferably, the drone hovers at each storage location along a preset route, collecting signals from all RFID tags at that location. Based on prior knowledge, the RFID tag closest to the drone will be bound to that storage location. To improve the accuracy of the matching, before formal production, the drone can collect a set of RFID signal data at the precise center point of each storage location, recording the average signal strength (RSSI_mean) and the number of times each tag appears (Count) for each route, forming a "signal fingerprint" database. Establishing the "background signal characteristics" of each storage location helps eliminate errors caused by environmental reflections and multipath interference in subsequent matching.

[0054] As a preferred approach, the prior map established in the early stages is utilized, and a location algorithm based on weighted nearest neighbor signal fingerprint matching (more robust and accurate) is employed to enhance anti-interference capabilities. This process can be divided into the following key stages: 1. Obtain the current signal vector: Convert the currently read tag signal into a vector V_current. Each dimension of the vector represents a known tag ID, and its value is the currently read RSSI value (if not read, fill with a very small value, such as -100 dBm).

[0055] 2. Obtain the reference feature vector: Retrieve the reference feature vector V_ref for the current location L_current from the signal fingerprint database. This vector is the average RSSI value of each tag at this location recorded during the database construction phase.

[0056] 3. Calculate similarity (or distance): Calculate the similarity between the current signal vector V_current and the reference feature vector V_ref. This uses the reciprocal of the Euclidean distance mentioned earlier. During the calculation, weights are introduced to assign higher weights to labels with strong signals in V_ref (representing that they are inherent labels for this location). For example, the weight W_i can be a function of the RSSI value of that label in V_ref.

[0057] 4. Finding the best-matching label: Identify which label makes V_current and V_ref most similar. A clever approach is to calculate the signal change vector ΔV = V_current - V_ref. Ideally, if a new item (label T_new) is placed in the current storage location, the dimension corresponding to T_new will show a large positive change in ΔV (signal enhancement). Other labels will show small or negative signal changes (because the new item may slightly obscure them).

[0058] 5. Decision and Binding: Traverse the ΔV vector to find the tag T_best with the largest increase in signal strength; at the same time, a minimum gain threshold can be set (e.g., +5 dBm). Binding operation is only performed when the maximum gain exceeds the threshold; otherwise, it is considered that there are no new goods in the storage location or the goods have been taken away; bind the storage location L_current with the tag T_best.

[0059] This algorithm utilizes prior knowledge to effectively suppress environmental noise and interference from nearby storage locations, making it highly reliable.

[0060] As a preferred approach, RFID signals are processed in real time during the continuous flight of the drone, and the most likely tags are intelligently bound to the current storage location, thus optimizing the dynamic storage location-tag binding. When the drone arrives at the current storage location, it collects all tag signals, performs signal preprocessing to filter candidate tags, and then executes a binding decision after time-series context analysis. At the start of the decision-making process, a confidence check is performed. Signals with high confidence can be bound immediately, and the binding relationship is recorded. For signals with low confidence, the decision is delayed, and more evidence is awaited. The drone then flies to the next storage location, and so on, to complete the current inventory task. This algorithm optimizes decisions based on historical data and provides confidence assessment and decision traceability, which can significantly improve the accuracy and efficiency of warehouse inventory counting.

[0061] Even better, each time the drone performs a mission, it analyzes the signal strength of all signals received during the current inventory task in real time. A dynamic threshold calculation algorithm is used to evaluate the warehouse location signals. If the strongest signal strength of a warehouse location is also below the threshold, it is marked as an empty warehouse location. The threshold statistics are continuously updated during flight to ensure that empty warehouse locations are not bound to goods. This dynamic threshold calculation algorithm uses statistical methods based on normal distribution and quartiles to ensure robustness and has good tolerance to outliers and noisy signals.

[0062] As a preferred method, after the drone completes the inventory task, it will compare the real-time inventory data with the WMS inventory records and automatically output an inventory report marking the differences, including two types of differences: incorrect inventory location and missing inventory in the storage location. The operator can then recount the physical inventory based on the marked differences.

[0063] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A warehouse goods precise inventory method based on RFID and UAV cooperation, characterized in that, Includes the following steps: S1. After receiving goods, RFID tags are affixed and the goods are put on the shelves, and the inventory information is stored in the warehouse management system. S2. Establish a three-dimensional model and spatial coordinate system for the warehouse, with each storage location corresponding to a unique three-dimensional spatial coordinate. S3. The drone performs the inventory task according to the preset route; S4. When the drone hovers in front of each storage location, it attaches the RFID tag with the strongest signal, so that the storage location corresponds to the goods one by one. S5. When the drone flies to the next storage location, the RFID signal value collected is compared with the RFID signal value of the previous storage location, and then the tag with the strongest signal is bound to the current storage location. S6. After the inventory task is completed and the drone returns, an inventory report is automatically generated, comparing the real-time RFID data with the WMS inventory record and marking the differences for manual review.

2. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 1, characterized in that, Step S1 includes: after the goods arrive, randomly select 10% of the sample size for inspection and after confirming that there are no errors, first affix RFID tags to complete the binding relationship between the tags and the goods, and then scan the goods and the storage location number after the goods are put on the shelf to complete the binding relationship between the goods and the storage location.

3. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 1, characterized in that, Step S2 completes the point cloud map construction during the offline training phase: a drone equipped with a LiDAR is used to perform a 3D scan of the entire warehouse scene to collect point clouds, the collected point cloud data is preprocessed to extract key geometric features; a mapping relationship between point cloud feature descriptors and spatial locations is established; all feature-location pairs are stored in the point cloud feature database to form a point cloud feature library that can be used for positioning.

4. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 3, characterized in that, The detailed mapping process for step S2 is as follows: i. 3D spatial scanning: Plan the drone flight path to ensure coverage of all areas of the warehouse, and use high-precision lidar for multi-view scanning; ii. Unified coordinate system: Arrange several reflective column calibration points with known coordinates in the warehouse to obtain a set of references with precise coordinates {P1(x1,y1,z1), P2(x2,y2,z2),...Pn(xn,yn,zn)}; take the projection center of the UAV nest on the ground as the origin of the world coordinate system (0,0,0) to ensure that all scanned point clouds are unified to the same global coordinate system; iii. Point cloud data processing: Downsampling, noise reduction, and ground segmentation are performed on the original point cloud. Key feature points are extracted, the local coordinate system is initialized, and the local point cloud is transformed into the global coordinate system and fused with the existing map.

5. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 4, characterized in that, Step S2, point cloud feature extraction and database construction, is performed as follows: point cloud segmentation and keyframe extraction, assigning a unique identifier and precise pose information to each keyframe; for keyframe point cloud P1, feature descriptors are extracted to obtain the first feature vector: V1 =[FPFH1, Curvature Mean, Curvature Variance, Plane Normal Vector, Line Segment Direction, Point Density, Height Statistics]; The vector V1 is bound to its corresponding geographic location P1(x1,y1,z1) and stored as a record in the database; then the second keyframe point cloud P2 is recorded, and the whole process is repeated until all point clouds are collected; thus, a priori map feature database that supports fast retrieval based on feature similarity is obtained.

6. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 1, characterized in that, The warehouse 3D model established in step S2 is 1:1 in size with the actual warehouse. The warehouse 3D model includes a receiving and dispatching area and high-level shelves in the storage area. The height of each shelf corresponds to the actual items. The center position of each storage location corresponds to its own corresponding 3D spatial coordinate point (x, y, z).

7. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 1, characterized in that, In step S3, connecting two points in spatial coordinates forms a straight line. Connecting the spatial points of each storage location forms the preset flight path of the UAV. The UAV performs different flight inventory tasks according to different flight paths. The UAV briefly hovers at each storage location to collect all RFID tag signals at the current location. Based on the RFID signal strength, the storage location is matched, and the tag with the strongest signal is bound to the current storage location. If the strongest signal strength of a certain storage location is lower than a preset threshold, it will be removed, so that empty storage locations will not be bound to goods.

8. The method for accurate inventory counting of warehouse goods based on RFID and UAV collaboration as described in any one of claims 3-5, characterized in that, The online positioning stage in step S3 includes: i. Real-time point cloud acquisition: The UAV acquires local point clouds in real time during the inventory process; and extracts the real-time feature vector V_current. ii. Feature matching and position estimation: The feature vector most similar to V_current is searched in the prior map and compared to obtain the real-time position of P_current.

9. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 8, characterized in that, Steps S4 and S5 use a combination of the following two algorithms to find the most matching position in the feature database based on the real-time collected tag signals: Basic K-nearest neighbor localization algorithm: find the K reference points in the feature database that are most similar to the real-time vector, and then take the average of the position coordinates of these K points as the final localization result; Weighted K-Nearest Neighbor Algorithm: The algorithm calculates a weighted average of K nearest neighbor points based on their similarity, with points that are more similar having a higher weight.

10. The method for accurate warehouse inventory counting based on RFID and UAV collaboration as described in claim 9, characterized in that, Step S4 or S5 uses Euclidean distance or Manhattan distance to measure similarity. The specific matching and binding process is as follows: i. Obtain the current signal vector: Convert the currently read tag signal into a vector V_current; each dimension of the vector represents a known tag ID, and its value is the currently read RSSI value; ii. Obtain the reference feature vector: Retrieve the reference feature vector V_ref of the current storage location L_current from the feature database; iii. Calculate similarity: Calculate the similarity between the current vector V_current and the reference feature vector V_ref using the reciprocal of the Euclidean distance or Manhattan distance, and introduce weights during the calculation, giving higher weights W_i to those labels with strong signals in V_ref (representing that it is an inherent label of this location). The weights W_i are a function of the RSSI value of the label in V_ref. iv. Find the best matching label: Calculate the signal change vector ΔV = V_current - V_ref; if the dimension corresponding to the label of the current storage location has a large positive change in ΔV, while the signal changes of other labels are small or negative, then it is determined that new goods have been put into the current storage location; v. Decision and Binding: Traverse the ΔV vector to find the label T_best with the largest increase in signal strength; at the same time, set a minimum gain threshold, and when the maximum gain exceeds the minimum gain threshold, bind the storage location L_current to the label T_best.

11. The warehouse cargo accurate inventory method based on RFID and UAV collaboration as described in claim 1, characterized in that, The inventory report automatically output in step S6 marks discrepancies, including two types of discrepancies: incorrect inventory location and missing inventory in the storage location.

Images