Intelligent inventory warehouse management system and method for black light warehouse

Abstract

The present application relates to the technical field of intelligent warehousing, in particular to a kind of intelligent inventory warehousing management system and method for black light warehouse, including in the process of transportation AGV, dispatching inventory robot enters wake zone and works along, obtains continuous track point set;Based on multi-view point cloud data, probability occupation fusion is carried out, three-dimensional voxel model is constructed to remove occlusion and object geometric center and bounding box are extracted;Combined with geometric center directional emission sweep frequency sound wave, extract acoustic impedance frequency spectrum features to determine content material category;Synchronous acquisition RFID carrier wave phase data, combined with trajectory information realizes label three-dimensional space positioning;Through spatial inclusion relationship and SKU material information cross validation, output high reliability automatic inventory result.The present application is suitable for unlighted, unmanned warehousing scene, can improve inventory accuracy and abnormal detection ability without affecting logistics operation.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing technology, specifically to an intelligent inventory management and warehousing system and method for dark warehouses. Background Technology

[0002] With the explosive growth of e-commerce, fully automated, unmanned "lights-out warehouses" have become a key model for improving logistics efficiency and reducing energy consumption. In such scenarios, massive SKUs are typically placed randomly to maximize space utilization, and transport AGVs operate at high speeds and high frequencies.

[0003] Existing warehouse inventory management technologies primarily rely on machine vision (such as OCR barcode recognition) or radio frequency identification (RFID). However, in the specific scenarios described above, these technologies face significant technical bottlenecks: First, machine vision solutions heavily depend on lighting conditions, requiring frequent high-energy supplemental lighting in dark environments. Furthermore, when goods are densely stacked, line-of-sight is obstructed, limiting access to surface information and failing to reach deeper blind spots. Second, while RFID technology does not require illumination, it is highly susceptible to multipath effects and signal reflections in the complex electromagnetic environment of high-density metal shelving, leading to cross-reading between adjacent storage compartments and hindering accurate spatial binding of labels to physical locations. Additionally, existing inventory operations are typically performed by robots with independent right-of-way, easily causing path conflicts with AGVs in high-speed workflows, resulting in decreased overall scheduling efficiency. More critically, conventional methods can only verify the existence of outer packaging or labels, failing to effectively verify the physical material and condition of the contents without contact, thus failing to address the management pain point of "matching records with actual contents but with abnormal contents." Therefore, there is an urgent need for a technical solution that can utilize the spatiotemporal correlation during the operation process to solve the problems of occlusion, positioning drift, and perspective verification.

[0004] To address this, an intelligent inventory and storage management system and method for dark warehouses are proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent inventory and storage management system and method for dark warehouses. By scheduling an inventory robot to follow the transport AGV, the system performs multi-view scanning of the storage compartments and performs probabilistic voxel reconstruction to obtain a three-dimensional model of the object. The system combines acoustic impedance detection to determine the material of the contents and uses RFID carrier phase to achieve three-dimensional positioning of the tag. The system outputs the inventory results through spatial matching and material cross-verification, thus achieving highly reliable automatic inventory.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An intelligent inventory and warehouse management system for dark warehouses includes:

[0008] The trajectory planning module is used to obtain the task path of the transport AGV, filter the transport AGVs whose paths pass laterally through the inventory location, schedule the inventory robot to enter the wake zone and perform lateral undulation following motion, and output a continuous trajectory point set with spatiotemporal labels.

[0009] The temporal voxel reconstruction module is used to continuously scan the target compartment from multiple perspectives during the movement of the inventory robot along the trajectory, perform probabilistic occupancy fusion on the point cloud data at different times, and output the de-occluded 3D voxel model and the geometric center coordinates of the object.

[0010] The acoustic impedance detection module is used to control the aiming direction based on the geometric center coordinates, emit sweep frequency acoustic wave signals, collect vibration response signals on the packaging surface, extract acoustic impedance spectrum characteristics, and determine the material type of the contents.

[0011] The phase holographic positioning module is used to continuously collect carrier phase data of RFID tags during movement, and combine it with a continuous trajectory point set to perform phase unwrapping and synthetic aperture back projection calculation to obtain the three-dimensional spatial coordinates of the tag.

[0012] The comprehensive judgment module is used to determine whether the three-dimensional spatial coordinates fall within the bounding box of the object detected in the three-dimensional voxel model, and to cross-validate the SKU information corresponding to the label with the material category, and output the inventory results.

[0013] Preferably, the probability occupancy fusion includes:

[0014] The spatial range of the target storage compartment is discretized into voxel units, and the occupancy probability value of each voxel unit is initialized. Based on the robot pose corresponding to the acquisition time, the point cloud of each frame is transformed from the sensor coordinate system to the global coordinate system. For each point in the transformed point cloud, a ray is projected from the sensor position. The occupancy probability of the voxel unit through which the ray passes is updated and shifted towards the idle direction. The occupancy probability of the voxel unit where the ray terminates is updated and shifted towards the occupancy direction. The probability update is performed by traversing multiple consecutive frames of point cloud. Based on the comparison result between the final occupancy probability value and the preset threshold, the occupancy status of each voxel unit is determined.

[0015] Preferably, the three-dimensional voxel model adopts a layered architecture, including: a probabilistic grid layer, which stores the position index of each voxel unit and the occupancy probability value updated cumulatively over multiple frames; a state label layer, which stores the occupancy state label of each voxel unit determined by comparing the occupancy probability value with a threshold; a connected component analysis layer, which performs three-dimensional connected component labeling on the voxel units marked as occupied, identifies independent object clusters, and stores the index set of voxel units contained in each cluster; and an object attribute layer, which calculates the geometric center coordinates, bounding box range, and volume estimate of the object based on the coordinates of the voxel units contained in each object cluster.

[0016] Preferably, the determination of the material category of the contents includes:

[0017] The aiming angle of the directional sound source is calculated based on the geometric center coordinates of the object, and the direction of the sound source is controlled. A sweeping sound wave signal with a frequency that varies with time is generated and emitted towards the target object. The response signal is collected through a microphone array, and the response signal is preprocessed by filtering and normalization. The transfer function between the excitation signal and the response signal is calculated, and spectral feature parameters including the main resonant frequency, resonant peak width, high-frequency attenuation rate, and low-frequency energy proportion are extracted from the transfer function. The spectral feature parameters are input into a pre-trained classification model, and the material category label and confidence score are output.

[0018] Preferably, the steps for obtaining the three-dimensional spatial coordinates are as follows: during the movement of the inventory robot, record the tag identifier, carrier phase value, and reading time of each RFID tag reading; query the corresponding robot pose from the continuous trajectory point set according to the reading time, and construct the phase history sequence of each tag by grouping according to the tag identifier; perform unwrapping processing on the phase history sequence, detect the jump of adjacent phase values ​​and compensate for periodic folding, and restore the continuous phase change curve; establish a candidate position grid in the target area, calculate its theoretical phase for each candidate position, and calculate the correlation between the theoretical phase and the measured phase; select the candidate position with the highest correlation as the three-dimensional spatial coordinate estimation result of the tag.

[0019] Preferably, the judgment process of the comprehensive judgment module is as follows:

[0020] Obtain the 3D spatial coordinates of each label and the bounding box range of each object; for the 3D spatial coordinates of each label, check whether all three components are located between the minimum and maximum corner points of the bounding box. If so, determine that the label falls within the bounding box of the object and establish the association between the label and the object; for labels whose coordinates do not fall within any bounding box, mark them as label drift anomalies; for bounding boxes in which no associated labels are detected, mark them as label missing anomalies.

[0021] Preferably, the material cross-validation process includes:

[0022] Based on the label identifier, query the associated SKU information and expected material category; based on the association between the label and the object, compare the expected material of each label's SKU with the actual measured material category of the associated object; for cases where the expected material and the actual material are inconsistent, mark it as a material anomaly and record the anomaly details;

[0023] The inventory results are compiled and generated, including the inventory grid identification, the number and list of detected objects, the number and list of identification tags, the spatial matching status, the material verification status, the overall inventory status, and the list of anomaly details.

[0024] A smart inventory and warehouse management method for dark warehouses includes:

[0025] The system acquires the task path of the transport AGV, filters AGVs whose paths laterally pass through the inventory location, schedules the inventory robot to enter the wake zone and executes lateral undulation following motion, and outputs a continuous trajectory point set with spatiotemporal tags. During the movement of the inventory robot along the trajectory, it continuously scans the target storage compartment from multiple perspectives, performs probabilistic occupancy fusion on the point cloud data at different times, and outputs an unoccluded 3D voxel model and the geometric center coordinates of the object. Based on the geometric center coordinates, it controls the aiming direction, emits a sweeping acoustic wave signal, collects the vibration response signal of the packaging surface, extracts the acoustic impedance spectrum characteristics, and determines the material category of the contents. During the movement, it continuously collects the carrier phase data of the RFID tag, performs phase unwrapping and synthetic aperture back projection calculation in combination with the continuous trajectory point set, and obtains the 3D spatial coordinates of the tag. It determines whether the 3D spatial coordinates fall within the bounding box of the object detected in the 3D voxel model, and cross-validates the SKU information corresponding to the tag with the material category, outputting the inventory results.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. This invention uses a trajectory planning module to select transport AGVs that pass laterally through the storage location and schedules an inventory robot to enter the wake zone to perform lateral undulating following motion, completing the inventory task without interfering with the existing logistics operation rhythm. This invention fully utilizes the spatiotemporal resources during AGV operation to achieve "inventory while moving". Combined with a continuous trajectory point set with spatiotemporal tags, it provides a unified benchmark for the temporal-spatial alignment of multi-sensor data, enabling the system to maintain stable, continuous, and high-frequency inventory operation capabilities even in a dark warehouse environment, significantly improving the overall throughput efficiency and automation level of the warehousing system.

[0028] 2. This invention introduces a temporal voxel reconstruction mechanism based on probabilistic occupancy fusion, which uniformly maps point cloud data acquired at multiple time points and from multiple viewpoints to a 3D voxel space. Through ray projection and probabilistic update strategies, it effectively eliminates the modeling incompleteness caused by mutual occlusion of packaging within the cargo compartment and limited viewpoints. This invention can progressively approximate the true spatial occupancy state, stably outputting an unoccluded 3D voxel model, and extracting reliable geometric center and bounding box information based on this. This mechanism significantly improves the robustness of object spatial localization and scale estimation, providing accurate spatial constraints for subsequent acoustic detection and RFID positioning.

[0029] 3. This invention integrates three-dimensional spatial positioning, acoustic impedance material identification, and RFID tag information for unified judgment. By determining the inclusion relationship between the tag's spatial coordinates and the voxel bounding box, it achieves a one-to-one association between the tag and the physical goods. Furthermore, it cross-validates the expected material of the tag SKU with the acoustically measured material. This invention can effectively identify hidden anomalies such as tag drift, missing tags, and misplaced goods, significantly reducing the risk of "tagd but no goods" or "goods with mislabeled tags," and significantly improving the reliability, interpretability, and anomaly detection capability of inventory results. Attached Figure Description

[0030] Figure 1 A structural diagram of an intelligent inventory and storage management system for a dark warehouse provided by the present invention;

[0031] Figure 2 A flowchart of an intelligent inventory and storage management method for a dark warehouse is provided by the present invention;

[0032] Figure 3 A flowchart for determining the material category of the contents provided by this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0034] like Figure 1 and Figure 2 As shown, the intelligent inventory and storage management method for a dark warehouse provided by this invention is applied to an intelligent inventory and storage management system for a dark warehouse; it includes a trajectory planning module, a temporal voxel reconstruction module, an acoustic impedance detection module, a phase holographic positioning module, and a comprehensive judgment module; the data flow relationship between the modules is as follows: the continuous trajectory point set output by the trajectory planning module is simultaneously transmitted to the temporal voxel reconstruction module and the phase holographic positioning module; the geometric center coordinates output by the temporal voxel reconstruction module are transmitted to the acoustic impedance detection module, and the bounding box output by the temporal voxel reconstruction module is transmitted to the comprehensive judgment module; the material category output by the acoustic impedance detection module is transmitted to the comprehensive judgment module; the three-dimensional spatial coordinates of the label output by the phase holographic positioning module are transmitted to the comprehensive judgment module; the comprehensive judgment module summarizes the above information and combines it with the SKU data in the storage management system to output the final inventory result.

[0035] The specific technical solutions are as follows: A trajectory planning module is used to acquire the task path of the transport AGV, filter transport AGVs whose paths laterally pass through the inventory location, schedule the inventory robot to enter the wake zone and perform lateral undulation following motion, and output a continuous trajectory point set with spatiotemporal tags; a temporal voxel reconstruction module is used to continuously scan the target compartment from multiple perspectives during the movement of the inventory robot along the trajectory, perform probabilistic occupancy fusion of point cloud data at different times, and output an unoccluded 3D voxel model and the geometric center coordinates of the object; an acoustic impedance detection module is used to control the aiming direction based on the geometric center coordinates, emit sweep frequency acoustic signals, collect vibration response signals from the packaging surface, extract acoustic impedance spectrum features, and determine the material category of the contents; a phase holographic positioning module is used to continuously collect carrier phase data of RFID tags during movement, combine the continuous trajectory point set to perform phase unwrapping and synthetic aperture back projection calculations to obtain the 3D spatial coordinates of the tag; a comprehensive judgment module is used to determine whether the 3D spatial coordinates fall within the bounding box of the object detected in the 3D voxel model, and cross-validate the SKU information corresponding to the tag with the material category, outputting the inventory results.

[0036] Example 1:

[0037] In this embodiment, the inventory robot is a mobile platform equipped with lidar, RFID reader, directional sound source, microphone array and high-precision positioning system; the transport AGV is an automated guided vehicle that performs cargo handling tasks; the dark warehouse is a fully automated warehousing facility that can operate without artificial lighting, and is equipped with high-density shelves, automated storage and retrieval equipment and warehouse control system.

[0038] The trajectory planning module obtains the assigned task paths of all transport AGVs within the current time window from the warehouse control system. The task paths are stored in the form of a sequence of path points, with each path point containing location coordinates and estimated arrival time.

[0039] The trajectory planning module calculates the spatial relationship between each task path and the inventory location, models the shelf where the inventory location is located as a plane, and extracts the normal vector of the plane; it traverses each path segment of each task path and calculates the angle between the direction vector of the path segment and the normal vector of the shelf; when the angle is less than the preset observation angle threshold, it determines that the path segment has lateral observation conditions; it accumulates the total length of the path segments that meet the lateral observation conditions in each path, and selects the transportation AGV with the largest total length as the following AGV.

[0040] The inventory robot navigates to the area behind the selected AGV and enters the wake zone. The front boundary distance of the wake zone is determined based on the braking safety distance calculated from the maximum operating speed of the AGV and the ground friction coefficient. The rear boundary distance is the front boundary distance plus the preset follow zone length. The inventory robot maintains its relative position with the AGV in the wake zone and moves synchronously with the AGV.

[0041] When the inventory robot enters the shelf area to be inventoried along with the AGV, it starts the lateral undulation following mode. In this mode, the inventory robot performs periodic lateral reciprocating motion relative to the AGV, and the motion trajectory is a sine wave. The amplitude of the lateral motion is determined according to the width of the aisle and the size of the robot to ensure that the motion range does not exceed the aisle boundary. The frequency of the lateral motion is determined according to the speed of the AGV and the target scanning resolution.

[0042] The positioning system in the trajectory planning module continuously records the pose information of the inventory robot. The positioning system uses laser synchronous positioning and mapping technology and inertial measurement unit fusion to achieve centimeter-level positioning accuracy. Each pose record includes the timestamp of the acquisition time, the three-dimensional position coordinates of the inventory robot in the global coordinate system, the three-axis attitude angles, and the current velocity vector. Continuous pose records constitute a continuous trajectory point set with spatiotemporal labels. This continuous trajectory point set is stored in the system cache for subsequent modules to call.

[0043] Furthermore, the probability occupancy fusion includes:

[0044] The spatial range of the target storage compartment is discretized into voxel units, and the occupancy probability value of each voxel unit is initialized. Based on the robot pose corresponding to the acquisition time, the point cloud of each frame is transformed from the sensor coordinate system to the global coordinate system. For each point in the transformed point cloud, a ray is projected from the sensor position. The occupancy probability of the voxel unit through which the ray passes is updated and shifted towards the idle direction. The occupancy probability of the voxel unit where the ray terminates is updated and shifted towards the occupancy direction. The probability update is performed by traversing multiple consecutive frames of point cloud. Based on the comparison result between the final occupancy probability value and the preset threshold, the occupancy status of each voxel unit is determined.

[0045] Specifically, the temporal voxel reconstruction module performs the probability occupancy fusion process as follows:

[0046] The spatial range of the voxel grid is determined based on the spatial boundary information of the target grid registered in the warehouse management system. The spatial boundary information includes the minimum and maximum corner coordinates of the grid in the global coordinate system. The spatial range is uniformly divided into a voxel cell array according to the preset voxel resolution. A three-dimensional integer index is assigned to each voxel cell, and the global coordinates of the center point of the voxel are calculated based on the index and resolution. During initialization, the occupancy probability values ​​of all voxel cells are set to the intermediate values ​​representing the unknown state.

[0047] For each frame of laser point cloud acquired by the inventory robot during its movement, a coordinate transformation is performed. Based on the acquisition timestamp of that frame of point cloud, the robot's pose at the corresponding moment is retrieved from the continuous trajectory point set. If there is a discrepancy between the acquisition time and the most recently recorded timestamp in the continuous trajectory point set, the pose estimate at that moment is obtained by linear interpolation between two adjacent pose records. Using the pose information obtained from the query or interpolation, the coordinates of each point in the point cloud are transformed from the local coordinate system of the lidar to the global coordinate system through rotation and translation. The transformation process is as follows: first, a rotation matrix is ​​constructed based on the attitude angle; the local coordinates are multiplied by the rotation matrix to obtain the rotated coordinates; and then the position coordinates are added to obtain the global coordinates.

[0048] For each point in the transformed point cloud, a probability update is performed: The voxel cell into which the point falls is determined. The global coordinates of the point are subtracted from the raster origin coordinates, divided by the voxel resolution, and rounded down to obtain the voxel index. An occupancy direction probability update is performed on this voxel cell, increasing its occupancy probability value by a preset occupancy increment. Simultaneously, all voxel cells traversed along a straight line from the LiDAR position to the point are iterated. This traversal uses a 3D linear rasterization method, progressively advancing along the three coordinate axes, sequentially visiting each voxel traversed by the ray. For each voxel cell traversed but not terminated, an idle direction probability update is performed, decreasing its occupancy probability value by a preset idle increment. The probability update uses logarithmic probability representation to avoid numerical underflow caused by probability multiplication. The occupancy probability is converted to logarithmic probability form for storage and updating. Addition and subtraction operations are performed directly during updates, and the output is converted back to standard probability form.

[0049] As the inventory robot moves along the trajectory, the point cloud in multiple consecutive frames sequentially performs the above coordinate transformation and probability update process. Due to the continuous change in the observation angle, some voxels are not effectively observed in the early frames because they are occluded by objects in front, but they are exposed in the later frames due to the change in the viewpoint, thus obtaining effective observation data. The accumulation of multiple frames causes the occupancy probability value of each voxel unit to gradually converge to a stable state.

[0050] After point cloud acquisition is completed, all voxel units are traversed and their status is determined based on their final occupancy probability values. Voxel units with occupancy probability values ​​higher than a preset occupancy threshold are marked as occupied, voxel units with occupancy probability values ​​lower than a preset idle threshold are marked as idle, and voxel units with occupancy probability values ​​between the two thresholds are marked as unknown.

[0051] Furthermore, the three-dimensional voxel model adopts a layered architecture, including: a probability grid layer, which stores the position index of each voxel unit and the occupancy probability value updated cumulatively over multiple frames; a state label layer, which stores the occupancy state label of each voxel unit determined by comparing the occupancy probability value with a threshold; a connected component analysis layer, which performs three-dimensional connected component labeling on the voxel units marked as occupied, identifies independent object clusters, and stores the index set of voxel units contained in each cluster; and an object attribute layer, which calculates the geometric center coordinates, bounding box range, and volume estimate of the object based on the coordinates of the voxel units contained in each object cluster, wherein the geometric center coordinates are passed to the acoustic impedance detection module as the aiming target, and the bounding box is passed to the comprehensive judgment module as a spatial constraint condition.

[0052] Specifically, the 3D voxel model uses a hierarchical architecture to organize data, with the following layers from bottom to top: probability grid layer, state label layer, connected component analysis layer, and object attribute layer.

[0053] The probabilistic raster layer is the lowest-level data structure, storing basic information about the voxel raster. This layer maintains the following data fields for each voxel unit: a 3D integer index to uniquely identify the voxel's position within the raster; global coordinates of the center point for subsequent geometric calculations; and an occupancy probability value, reflecting the likelihood of the voxel being occupied after cumulative updates over multiple frames. The data in the probabilistic raster layer is continuously updated during the probabilistic occupancy fusion process.

[0054] The state label layer is built on top of the probability grid layer and stores the discretized state labels of each voxel unit. Based on the comparison between the occupancy probability value of each voxel in the probability grid layer and a preset threshold, this layer assigns one of three state labels to each voxel: occupied, idle, or unknown. The data of the state label layer is generated all at once after probability fusion is completed, providing binary input for subsequent connected component analysis.

[0055] The connected component analysis layer is responsible for identifying and segmenting independent objects. This layer takes all voxel units marked as occupied in the state label layer as input and performs 3D connected component marking. The connected component marking process starts from any unvisited occupied voxel and searches for its neighboring occupied voxels in 3D space. The adjacency relationship is defined using the six-neighbor rule, that is, six voxels directly adjacent along the positive and negative directions of the three coordinate axes are considered neighbors. The searched neighboring occupied voxels are marked as belonging to the same object cluster as the starting voxel, and the search continues to expand outward from these newly marked voxels. The process is iterated until all neighboring occupied voxels of the object cluster have been visited. Then, the next unvisited occupied voxel is selected, and a new object cluster marking process begins. The above steps are repeated until all occupied voxels are assigned to an object cluster. The connected component analysis layer assigns a unique identifier to each object cluster and stores the set of indices of all voxel units contained in the cluster.

[0056] The object attribute layer is the top-level data structure, storing the object-level geometric features extracted from the connected component analysis results. For each object cluster, the object attribute layer calculates and stores the geometric center coordinates, bounding box, and volume estimate. The geometric center coordinates are passed to the acoustic impedance detection module as the target position for the sound source to aim at. The bounding box is passed to the comprehensive judgment module as a spatial constraint condition.

[0057] The following steps are performed to obtain the 3D voxel model and the coordinates of the object's geometric center:

[0058] In the voxel grid construction step, the spatial boundary information of the target cell is queried from the warehouse management system, including the minimum and maximum boundary values ​​of the cell along the three coordinate axes in the global coordinate system. The length, width, and height of the cell are calculated based on the boundary information. The number of voxels in each direction is calculated according to the preset voxel resolution; the number of voxels in each direction is the result of dividing the dimension of that direction by the resolution and rounding up. A three-dimensional voxel array is created, with the three dimensions of the array corresponding to the number of voxels in the three coordinate axes. Each voxel cell in the array is assigned an index consisting of three integers, with the index value starting from 0 and increasing along each direction. The coordinates of the center point of each voxel are calculated based on the index by adding 0.5 to the index value, multiplying by the resolution, and then adding the coordinates of the grid origin. The occupancy probability of all voxel cells is initialized to 0, representing an initial unknown state.

[0059] In the multi-frame point cloud registration step, for each frame of laser point cloud acquired by the inventory robot, the acquisition timestamp of that frame is first obtained; the pose record closest to that timestamp is searched in the continuous trajectory point set; if a record with a perfectly matching timestamp exists, it is used directly; otherwise, two adjacent records with timestamps before and after that moment are selected, and the position and attitude are linearly interpolated according to the time difference ratio; the obtained pose information is used to construct a transformation relationship from the local coordinate system to the global coordinate system. For each point in the point cloud, its local coordinates are first multiplied by the attitude rotation matrix to complete the rotation transformation, and then added to the position vector to complete the translation transformation, obtaining the coordinates of the point in the global coordinate system. After the transformation, the point cloud data from different times and positions are aligned to a unified spatial reference system.

[0060] In the probability accumulation update step, for each registered point cloud point, it is first determined whether it falls within the spatial range of the voxel grid. If it exceeds the range, the point is skipped. For points within the range, their coordinates are converted into voxel indices by subtracting the grid origin from the coordinates, dividing by the resolution, and rounding down. The corresponding voxel unit is located according to the index, and the occupancy probability value of the unit is increased by a preset occupancy update amount. Then, ray traversal update is performed: starting from the global position at the acquisition time of the current frame, and ending at the global position of the current point cloud point, all voxels are traversed along the connecting line direction. The traversal adopts a progressive method, starting from the starting point and gradually advancing along the connecting line direction. At each step, it is determined whether the voxel index of the current position has changed. If it has changed, the probability update of the idle direction is performed for the newly entered voxel. For each voxel unit traversed by the ray, its occupancy probability value is reduced by a preset idle update amount. The voxel at the endpoint does not perform idle updates, only occupancy updates. With the input of multiple consecutive frames of point cloud, the occupancy probability value of each voxel unit gradually converges during repeated updates.

[0061] In the occupancy state binarization step, after all point cloud frames have been processed, each voxel unit in the voxel grid is traversed. For each unit, its occupancy probability value is compared with a preset occupancy threshold and an idle threshold. If the occupancy probability value is greater than the occupancy threshold, the state label of the unit is set to occupied; if the occupancy probability value is less than the idle threshold, it is set to idle; if it is between the two thresholds, it is set to unknown. The state label is stored in the state label layer in the form of an integer encoding.

[0062] In the connected component clustering step, an access marker array and a cluster identifier array of the same size as the voxel grid are created, and all elements are initialized to an unvisited state and without a cluster identifier. The current cluster identifier counter is initialized to 1. The voxel grid is traversed, and when a voxel in the occupied state that has not yet been visited is encountered, the region growth process is started with that voxel as the seed. The seed voxel is marked as visited and assigned the current cluster identifier, and then added to the waiting queue. A voxel is taken out from the queue, and its six neighboring voxels are checked. For each neighboring voxel, if its state is occupied and has not yet been visited, it is marked as visited, assigned the same cluster identifier, and added to the waiting queue. This process is repeated until the queue is empty, indicating that all voxels of the current object cluster have been marked. The cluster identifier counter is incremented by 1, and the traversal continues to find the next unvisited occupied voxel as a new seed. After the traversal is completed, all occupied voxels are assigned to an object cluster, and the final value of the cluster identifier counter is decremented by 1 to represent the number of detected objects.

[0063] In the geometric center calculation step, for each object cluster, all voxel indices belonging to that cluster are extracted from the cluster identifier array based on its cluster identifier; the global coordinates of the center point of the corresponding voxel are obtained from the probabilistic raster layer based on the index; the three components of all center point coordinates are summed, and then divided by the total number of voxels contained in the cluster to obtain the average value in the three directions. The combination of the three average values ​​constitutes the geometric center coordinates of the object and is stored in the object attribute layer.

[0064] In the bounding box generation step, for each object cluster, the center point coordinates of all voxels it contains are traversed. During the traversal, the minimum and maximum values ​​in three directions are maintained respectively. Initially, the minimum value is set to the maximum number and the maximum value is set to the minimum number. For each voxel coordinate, if a component is less than the current minimum value, the minimum value is updated; if it is greater than the current maximum value, the maximum value is updated. After the traversal is completed, the coordinates formed by the minimum values ​​in the three directions are the minimum corner coordinates of the bounding box, and the coordinates formed by the maximum values ​​in the three directions are the maximum corner coordinates of the bounding box. The minimum and maximum corners are stored in the object attribute layer and together define the axis-aligned bounding box of the object.

[0065] Furthermore, the determination of the material type of the contents is based on... Figure 3 ,include:

[0066] The aiming angle of the directional sound source is calculated based on the geometric center coordinates of the object, and the direction of the sound source is controlled. A sweeping sound wave signal with a frequency that varies with time is generated and emitted towards the target object. The response signal is collected through a microphone array, and the response signal is preprocessed by filtering and normalization. The transfer function between the excitation signal and the response signal is calculated, and spectral feature parameters including the main resonant frequency, resonant peak width, high-frequency attenuation rate, and low-frequency energy proportion are extracted from the transfer function. The spectral feature parameters are input into a pre-trained classification model, and the material category label and confidence score are output.

[0067] Specifically, the process by which the acoustic impedance detection module determines the material type of the contents is as follows:

[0068] The acoustic impedance detection module receives the geometric center coordinates of the target from the time-series voxel reconstruction module. Based on the geometric center coordinates and the current installation position of the sound source, it calculates the azimuth and pitch angles required for the sound source to point towards the target. The azimuth angle is the angle between the line connecting the geometric center and the sound source on the horizontal plane and the reference direction, and the pitch angle is the angle between this line and the horizontal plane. The calculated angle values ​​are converted into gimbal control commands to drive the gimbal to adjust the orientation of the sound source so that the main axis of the sound source points to the geometric center position of the target object.

[0069] The sound source uses a directional loudspeaker, whose sound beam has a certain directionality, enabling it to concentrate sound energy and project it towards the target. The sound source controller generates a swept-frequency sound wave signal, the frequency of which changes linearly with time, gradually increasing from the starting frequency to the ending frequency. The swept-frequency range covers the characteristic response frequency bands of packaging materials and common contents. After being amplified, the swept-frequency signal drives the loudspeaker to emit sound waves toward the target object.

[0070] The microphone array is deployed on the inventory robot and consists of multiple microphone units arranged in a preset geometric configuration to collect sound wave responses from the target direction. When the sound wave excitation signal acts on the target object and its packaging, the packaging surface and contents will generate forced vibrations, and the vibrations will radiate sound waves to the surrounding air. The microphone array synchronously collects the response signals of each channel at a preset sampling rate, and the collection time covers the complete sweep frequency cycle and the response decay time.

[0071] The acquired response signals are preprocessed as follows: First, bandpass filtering is performed to retain the signal components within the swept frequency range and filter out environmental noise and interference outside this range. Then, normalization processing is performed to adjust the amplitude of each channel signal to a uniform reference level, eliminating the effects of sound source power fluctuations and propagation distance differences. For multi-channel signals, beamforming processing can be selectively performed to weight and superimpose the signals of each channel according to preset delays and weights, enhancing the signal from the target direction and suppressing interference from other directions.

[0072] Spectral analysis is performed on the preprocessed response signal: the excitation signal and response signal are transformed by time and frequency respectively to obtain their representations in the frequency domain; the ratio of the spectrum of the response signal to the spectrum of the excitation signal is calculated to obtain the system transfer function; the transfer function reflects the frequency response characteristics of the entire acoustic path from the sound source through air propagation, acting on the packaging and contents, and then propagating through the air to the microphone; since the acoustic response signal includes the contribution of the packaging material, the classification model uses labeled samples with packaging conditions during the training phase, so that the model learns the comprehensive acoustic response feature patterns of different contents under packaging conditions, thereby achieving effective determination of the contents material inside the packaging.

[0073] Feature parameters are extracted from the transfer function: the main resonance frequency is the frequency corresponding to the maximum peak value in the amplitude spectrum of the transfer function, reflecting the inherent vibration characteristics of the target object and the packaging system; the resonance peak width is the frequency range width corresponding to the moment when the amplitude of the main peak drops to half of the peak value, reflecting the damping characteristics of the system; the high-frequency attenuation rate is the rate at which the amplitude of the high-frequency band decreases as the frequency increases, reflecting the material's ability to absorb high-frequency sound waves; the low-frequency energy ratio is the ratio of the low-frequency signal energy to the total energy of the entire frequency band, reflecting the relative intensity of the low-frequency vibration component; the extracted feature parameters are combined to form a feature vector.

[0074] The feature vectors are input into a pre-trained classification model to determine the material category: the classification model is trained offline using a large number of labeled samples; a support vector machine is selected for classification.

[0075] Training samples were obtained in the following ways: Sample source: typical goods in a dark warehouse, with no less than 30 varieties in each category; Sample collection conditions: actual packaging condition, 3m away from the sound source; Number of repetitions: at least 10 independent collections for each variety.

[0076] Specifically, representative products from each material category are selected and placed in their actual packaging state at the same detection location as the real-world application scenario. Acoustic stimulation and response acquisition are then performed. The acquired feature vectors are paired with known material category labels to form training samples. The training samples should cover multiple typical products and common packaging forms for each material category to enhance the model's generalization ability. The output of the classification model includes material category labels and confidence scores. The material category labels indicate which predefined category the contents of the target object belong to, such as liquid, solid fiber, solid rigid material, powder / particle, or cavity. The confidence score reflects the reliability of the classification results.

[0077] Further, the steps for obtaining the three-dimensional spatial coordinates are as follows: during the movement of the inventory robot, record the tag identifier, carrier phase value, and reading time of each RFID tag reading; query the corresponding robot pose from the continuous trajectory point set according to the reading time, and construct the phase history sequence of each tag by grouping according to the tag identifier; perform unwrapping processing on the phase history sequence, detect the jump of adjacent phase values ​​and compensate for periodic folding, and restore the continuous phase change curve; establish a candidate position grid in the target area, calculate its theoretical phase for each candidate position, and calculate the correlation between the theoretical phase and the measured phase; select the candidate position with the highest correlation as the three-dimensional spatial coordinate estimation result of the tag.

[0078] Specifically, the process of obtaining three-dimensional spatial coordinates is as follows:

[0079] The RFID reader mounted on the inventory robot continuously performs tag scanning operations during movement. When the reader successfully reads a tag, it records the relevant information of the read, including the tag's unique identifier, the carrier phase value at the time of reading, and the timestamp of the reading time. The carrier phase value reflects the phase difference between the radio frequency signal emitted by the reader and the signal reflected by the tag. This phase difference is related to the distance between the reader and the tag. Since the phase value is periodic, its value range is limited to zero to a complete cycle.

[0080] For each read record, the corresponding pose of the inventory robot at the corresponding time is queried from the continuous trajectory point set based on its timestamp. The query method is the same as the point cloud registration step. If the timestamp matches precisely, the corresponding pose is used directly; otherwise, it is obtained by interpolation from adjacent records. The queried pose information is associated with the read record to form a read record with pose annotation.

[0081] All read records are grouped according to the tag identifier: for the same tag identifier, all its read records are arranged in chronological order to form the phase history sequence of that tag; the phase history sequence includes all the times when the tag was read during the passage of the inventory robot, the corresponding phase values, and the position coordinates of the reader at each time.

[0082] Phase unwrapping is performed on the phase history sequence of each tag: Since the original phase values ​​are limited to the range of zero to one period, folding transitions occur when the actual phase change exceeds one period. The purpose of unwrapping is to eliminate folding transitions and restore a continuous phase change curve. The specific method is as follows: the difference between two adjacent phase values ​​in the phase history sequence is checked sequentially. When the absolute value of the difference exceeds half a period, it is determined that a folding transition has occurred, and subsequent phase values ​​are compensated for a full period. The compensation direction is determined by the sign of the difference: a positive transition is subtracted by one period, and a negative transition is added by one period. After traversing the entire sequence to complete unwrapping, a continuous phase change curve is obtained. A three-dimensional candidate location grid is established in the target detection area: the spatial range of the grid covers the target compartment and its adjacent areas, and the grid resolution determines the positioning accuracy; each node in the grid represents a candidate location, and the tag may be located at any of these locations.

[0083] For each candidate location, a back-projection calculation is performed: First, based on the geometric relationship between the candidate location and the reader positions at each read time in the phase history sequence, the theoretical phase that should occur if the tag were located at that candidate location is calculated. The calculation of the theoretical phase is based on the propagation principle of radio frequency signals: the signal is emitted from the reader, reaches the tag, and is reflected back to the reader; the round-trip path length is twice the distance from the reader to the tag, and the phase change is directly proportional to the path length and inversely proportional to the signal wavelength. For each read time in the sequence, the corresponding theoretical phase value is calculated based on the distance between the reader position and the candidate location at that time. In the established candidate location grid, a candidate point... The three-dimensional spatial coordinates are At the time of reading The spatial coordinates of the RFID reader antenna are: Then the Euclidean distance between the reader and the candidate point at that moment is... and the corresponding theoretical phase The calculation formula is as follows:

[0084] ;

[0085] in, The wavelength of the RFID radio frequency signal. This represents the phase change caused by the double path of the signal round trip. The periodic folding characteristics of the phase value are represented and correlated with the measured unwrapped phase.

[0086] Then, the correlation between the theoretical phase and the unwrapped measured phase is calculated. The correlation calculation measures the degree of agreement between the two phase sequences. The higher the degree of agreement, the more likely the label is to be located at the candidate position. After traversing all candidate positions and completing the back projection calculation, the candidate position with the highest correlation score is selected as the estimated three-dimensional spatial coordinate result of the label. At the same time, the location confidence of the estimation result is recorded. The confidence is positively correlated with the highest correlation score.

[0087] Furthermore, the judgment process of the comprehensive judgment module is as follows:

[0088] Obtain the 3D spatial coordinates of each label and the bounding box range of each object; for the 3D spatial coordinates of each label, check whether all three components are located between the minimum and maximum corner points of the bounding box. If so, determine that the label falls within the bounding box of the object and establish the association between the label and the object; for labels whose coordinates do not fall within any bounding box, mark them as label drift anomalies; for bounding boxes in which no associated labels are detected, mark them as label missing anomalies.

[0089] Specifically, the process of spatial matching judgment performed by the comprehensive judgment module is as follows:

[0090] The comprehensive judgment module obtains a list of three-dimensional spatial coordinates of each tag from the phase holographic positioning module and a list of bounding box parameters of each object from the temporal voxel reconstruction module. Each record in the tag coordinate list contains the tag identifier and three-dimensional coordinate values. Each record in the bounding box parameter list contains the object identifier, minimum corner coordinates, and maximum corner coordinates.

[0091] For each label coordinate, perform bounding box containment relationship detection: traverse the bounding boxes of all objects and check whether the label coordinate is inside a certain bounding box. The detection method is as follows: determine whether each of the three components of the label coordinate satisfies the condition that it is greater than or equal to the smallest corner component and less than or equal to the largest corner component of the corresponding bounding box; if all three components satisfy the condition, it is determined that the label falls inside the bounding box, and the association relationship between the label and the object is established; if a label coordinate satisfies the containment conditions of multiple bounding boxes at the same time, select the object whose geometric center of the bounding box is closest to the label coordinate to establish the association.

[0092] After traversal, identify spatial matching anomalies: for tags whose coordinates do not fall within any bounding box, determine them as tag drift anomalies. Tag drift anomalies may be caused by the following reasons: the tag detaches and hangs in the air, the tag is moved to a location outside the registration cell, or there is a large error in the positioning calculation. Record the tag identifier and the calculated coordinate position of the drift anomaly.

[0093] For object bounding boxes where no associated tags are detected internally, a tag missing anomaly is identified. Tag missing anomalies may be caused by the following reasons: damaged tags that cannot respond, tags that are obstructed causing excessive signal attenuation, or tags not being properly affixed when goods are received; record the identifier of the object with the missing anomaly and the location of its bounding box.

[0094] Output spatial matching results, including a list of normally matched label-object pairs and a list of abnormal records.

[0095] Furthermore, the material cross-validation process includes:

[0096] Based on the label identifier, query the associated SKU information and expected material category; based on the association between the label and the object, compare the expected material of each label's SKU with the actual measured material category of the associated object; for cases where the expected material and the actual material are inconsistent, mark it as a material anomaly and record the anomaly details;

[0097] The inventory results are compiled and generated, including the inventory grid identification, the number and list of detected objects, the number and list of identification tags, the spatial matching status, the material verification status, the overall inventory status, and the list of anomaly details.

[0098] Specifically, the process of the comprehensive judgment module performing material cross-validation and inventory result output is as follows:

[0099] For each tag with an established object association, query the SKU master data in the warehouse management system based on its tag identifier. The SKU master data includes the SKU code, product name, product category, and expected material category of the product associated with the tag. The expected material category is the expected physical material attribute of the contents of the product corresponding to the SKU, which is maintained by the system administrator when the product is put into storage.

[0100] Based on the association between tags and objects established in the spatial matching step, the expected material of the SKU associated with the tag is compared with the actual material category of the object; the actual material category comes from the output of the acoustic impedance detection module; the comparison adopts the category matching rule: if the expected material and the actual material belong to the same category, it is determined to be a match; if they belong to different categories, it is determined to be a mismatch.

[0101] For cases where the comparison results are mismatched, mark them as material anomalies and analyze possible causes of the anomalies; if the expected state is liquid but the actual state is an empty cavity, possible causes include an empty bottle or liquid leakage followed by evaporation; if the expected state is liquid but the actual state is solid, possible causes include incorrect shipment or liquid solidification; if the expected state is solid but the actual state is an empty cavity, possible causes include an empty box or goods being removed; record detailed information about the material anomalies, including the relevant labels, SKU information, expected material, actual material, and inferred possible causes.

[0102] Summarize all detection and verification results to generate a complete inventory report. The data structure of the inventory report includes the following fields: Cell ID, a unique code for the target cell in the warehousing system; Inventory Time, the start and end timestamps of this inventory; Number of Detected Objects, the number of independent object clusters identified by the temporal voxel reconstruction module; List of Detected Objects, containing the attribute information of each object, including object identifier, geometric center coordinates, bounding box parameters, estimated volume, and measured material type; Number of Identified Tags, the number of RFID tags successfully located by the phase holographic positioning module; List of Identified Tags, containing the location information of each tag. The system includes: label identification, 3D spatial coordinates, location reliability, and associated object identification; spatial matching status (normal if all labels fall within the bounding box of their corresponding objects and there are no missing anomalies, otherwise an anomaly exists); material verification status (pass if the expected material of all SKUs associated with the labels matches the actual measured material of the objects, otherwise an anomaly exists); comprehensive inventory status (normal if both spatial matching and material verification statuses are normal, otherwise an anomaly exists); and an anomaly details list, summarizing all detected anomalies, with each anomaly record including the anomaly type, involved label identification, involved object identification, anomaly description, and suggested handling measures.

[0103] This invention proposes an intelligent inventory and storage management system suitable for dark warehouses. By collaboratively utilizing AGV trajectories, 3D perception, acoustic detection, and RFID positioning, it achieves unmanned and highly reliable inventory counting. The system centers on trajectory planning, reusing existing task paths of transport AGVs to guide the inventory robot to perform lateral undulation following within the wake zone, enabling continuous multi-view observation of target cargo locations and completing high-density scanning without interfering with warehousing operations. Through temporal voxel reconstruction, multi-frame point clouds are probabilistically occupant-fused to effectively eliminate occlusion effects, accurately extract the geometric center and spatial bounding box of the goods, and form a stable 3D object model. Based on this, the system introduces acoustic impedance detection to perform non-contact identification of the material of the contents in the packaged state, overcoming the difficulty of detecting hidden anomalies such as misdeliveries and empty boxes relying solely on vision or tag information. Simultaneously, phase holographic positioning technology is used to perform high-precision 3D positioning of RFID tags and spatial matching with the voxel model to achieve reliable tag-object association. Finally, the comprehensive judgment module cross-validates spatial consistency and material consistency, and outputs complete inventory results and anomaly analysis based on SKU master data. This solution improves the accuracy, robustness, and intelligence of inventory management in dark warehouses, reduces reliance on manual labor, and enhances the ability to identify misplaced items, missing items, and abnormal contents.

[0104] Example 2:

[0105] This embodiment, based on the system architecture described in Embodiment 1, further describes the specific implementation methods of the multipath signal suppression mechanism, voxel confidence assessment mechanism, and multi-source confidence fusion mechanism. For parts not described in detail in this embodiment, please refer to the relevant content in Embodiment 1.

[0106] The intelligent inventory and warehouse management system in this embodiment also includes a trajectory planning module, a temporal voxel reconstruction module, an acoustic impedance detection module, a phase holographic positioning module, and a comprehensive judgment module. The trajectory planning module is responsible for selecting the following AGV and planning the following trajectory of the inventory robot, and outputting a continuous trajectory point set with spatiotemporal tags. The temporal voxel reconstruction module performs multi-view scanning and probabilistic occupancy fusion on the target compartment, and outputs a three-dimensional voxel model, the geometric center coordinates of the object, and the bounding box range. The acoustic impedance detection module aims at the target according to the geometric center coordinates and emits a frequency-sweeping sound wave, and determines the material type of the contents through spectrum analysis. The phase holographic positioning module collects the carrier phase data of the RFID tag, and obtains the three-dimensional spatial coordinates of the tag through phase unwrapping and back projection calculation. The comprehensive judgment module performs spatial matching judgment and material cross-verification, and outputs the inventory results.

[0107] In this embodiment, the phase holographic positioning module is equipped with a multipath signal suppression unit, the temporal voxel reconstruction module is equipped with a voxel confidence evaluation unit, and the comprehensive judgment module is equipped with a multi-source confidence fusion unit.

[0108] The multipath signal suppression unit in the phase holographic positioning module is used to identify and suppress the impact of multipath effects on RFID positioning accuracy in a metal shelf environment. The phase holographic positioning module also includes a multipath signal suppression unit, which performs the following processing:

[0109] Each time a tag is read, the RFID reader is controlled to transmit radio frequency signals sequentially at multiple preset operating frequencies and collect the carrier phase value at each frequency to form a multi-frequency phase measurement group for that reading.

[0110] For each time point in the phase history sequence of the same label, calculate the deviation of each phase value in the multi-frequency phase measurement group at that time point relative to the reference frequency phase value, and compare the deviation value with the theoretical deviation value calculated based on the direct path model.

[0111] When the difference between the measured deviation value and the theoretical deviation value exceeds the preset consistency threshold, it is determined that the measurement at that time point is significantly affected by the multipath effect, and that time point is marked as an abnormal measurement point.

[0112] When performing back projection calculations, phase data marked as abnormal measurement points are assigned lower weights or are removed, so that the positioning results are mainly based on measurement data that are less affected by multipath.

[0113] Each time a tag is read, the RFID reader transmits radio frequency signals sequentially on multiple preset operating frequencies. The reader supports multiple channels within the UHF RFID standard frequency band. After transmitting a signal on each frequency, the reader collects the carrier phase value of the signal returned by the tag at that frequency. The phase values ​​collected at each frequency during the same read are combined to form a multi-frequency phase measurement group, which is stored together with the read time and tag identification.

[0114] For the phase history sequence of the same tag, the multipath signal suppression unit performs a consistency check at each time point; selects a certain frequency in the measurement group as the reference frequency, calculates the deviation of the phase values ​​of other frequencies relative to the phase values ​​of the reference frequency, and calculates the theoretical deviation value according to the direct path model. This model is based on the principle of radio frequency signal propagation: the phase of the direct path changes linearly with the frequency, and the rate of change is proportional to the distance from the reader to the tag; compares the measured deviation value with the theoretical deviation value, and when the difference exceeds the preset consistency threshold, it is determined that the time point is significantly affected by the multipath effect and marked as an abnormal measurement point.

[0115] When performing back projection calculations, phase data marked as abnormal measurement points are assigned lower weights or are removed, so that the positioning results are mainly based on measurement data that have passed the consistency test.

[0116] The voxel confidence assessment unit in the temporal voxel reconstruction module is used to quantitatively evaluate the reliability of each voxel reconstruction result. The temporal voxel reconstruction module also includes a voxel confidence assessment unit, which performs the following processing:

[0117] During the probability cumulative update process, observation statistics are maintained for each voxel unit. The observation statistics include the cumulative number of times the voxel is directly hit by the laser point cloud, the cumulative number of times it is determined to be idle by light passing through, and the incident angle of the lidar relative to the voxel during each effective observation.

[0118] After probability fusion is completed, the observation sufficiency index is calculated based on the observation statistics of each voxel. The observation sufficiency index comprehensively considers the total number of effective observations and the distribution range of incident angles.

[0119] Based on the comparison results between the observation sufficiency index and the preset grading threshold, a confidence level label is assigned to each voxel score. The confidence level label includes three levels: high confidence, medium confidence, and low confidence.

[0120] The voxel confidence level distribution information of each object cluster is summarized to generate an object-level confidence score, and the object-level confidence score is passed to the comprehensive judgment module, so that the comprehensive judgment module can adjust the judgment strategy according to the reliability of the object reconstruction results.

[0121] During the probability accumulation update process, observation statistics are maintained for each voxel unit, including: hit count, which records the cumulative number of times the voxel is directly hit by the laser point cloud; crossing count, which records the cumulative number of times the voxel is determined to be idle by light crossing; and incident angle record, which records the incident direction vector of the lidar relative to the voxel during each effective observation.

[0122] After probabilistic fusion is completed, the observation sufficiency index for each voxel is calculated based on observation statistics. This index comprehensively considers the total number of effective observations and the distribution range of incident angles; the total number of effective observations is the sum of the hit count and the crossing count; the distribution range of incident angles is quantified by calculating the dispersion of all incident vectors, with a greater dispersion indicating that the voxel has been observed from more different directions.

[0123] Based on the comparison results between the observation sufficiency index and the preset grading threshold, a confidence level label is assigned to each voxel, including three levels: high confidence, medium confidence, and low confidence. After the connected component clustering is completed, the confidence level distribution of voxels contained in each object cluster is statistically analyzed, the object-level confidence score is calculated, and the score is passed to the comprehensive judgment module.

[0124] The multi-source confidence fusion unit in the comprehensive judgment module is used to integrate confidence information from various detection modules.

[0125] The comprehensive judgment module also includes a multi-source confidence fusion unit, which performs the following processing:

[0126] Object-level confidence scores are obtained from the temporal voxel reconstruction module, material classification confidence scores are obtained from the acoustic impedance detection module, and positional confidence scores for each label are obtained from the phase holographic localization module.

[0127] For each pair of labels and objects that have established a relationship, the confidence scores of the relevant items of the pair will be weighted and fused. The weights are determined based on the expected reliability of each detection module under the current scene conditions.

[0128] Based on the comparison between the fused overall confidence level and the preset confidence threshold, each association pair is divided into two categories: high-confidence results and results to be verified.

[0129] For the related pairs classified as pending verification results, mark them in the inventory report as requiring further verification, and record the specific factors that led to insufficient confidence for subsequent processing reference.

[0130] The multi-source confidence fusion unit obtains object-level confidence from the temporal voxel reconstruction module, material classification confidence scores from the acoustic impedance detection module, and location confidence of each tag from the phase holographic positioning module.

[0131] For each pair of tags and objects with established relationships, the relevant confidence scores are extracted, weighted, and fused to calculate the overall confidence score. The weights are determined based on the expected reliability of each detection module under the current scene conditions.

[0132] The overall confidence level is compared with a preset confidence threshold. Association pairs exceeding the threshold are classified as high-confidence results, while those below the threshold are classified as results requiring verification. For results requiring verification, the specific factors leading to insufficient confidence levels are marked and recorded in the inventory report, including insufficient reconstruction observations, uncertain material determination, or limited tag positioning accuracy, for reference in subsequent processing.

[0133] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent inventory and warehouse management system for darkened warehouses, characterized in that, include: The trajectory planning module is used to obtain the task path of the transport AGV, filter the transport AGVs whose paths pass laterally through the inventory location, schedule the inventory robot to enter the wake zone and perform lateral undulation following motion, and output a continuous trajectory point set with spatiotemporal labels. The temporal voxel reconstruction module is used to continuously scan the target compartment from multiple perspectives during the movement of the inventory robot along the trajectory, perform probabilistic occupancy fusion on the point cloud data at different times, and output the de-occluded 3D voxel model and the geometric center coordinates of the object. The probability occupancy fusion includes: The spatial range of the target storage compartment is discretized into voxel units, and the occupancy probability value of each voxel unit is initialized. According to the robot pose corresponding to the acquisition time, the point cloud of each frame is transformed from the sensor coordinate system to the global coordinate system. For each point in the transformed point cloud, a ray is projected from the sensor position. The occupancy probability of the voxel unit through which the ray passes is updated and shifted towards the idle direction. The occupancy probability of the voxel unit where the ray terminates is updated and shifted towards the occupancy direction. The probability update is performed by traversing multiple consecutive frames of point cloud. The occupancy status of each voxel unit is determined based on the comparison result between the final occupancy probability value and the preset threshold. The acoustic impedance detection module is used to control the aiming direction based on the geometric center coordinates, emit sweep frequency acoustic wave signals, collect vibration response signals on the packaging surface, extract acoustic impedance spectrum characteristics, and determine the material type of the contents. The phase holographic positioning module is used to continuously collect carrier phase data of RFID tags during movement, and combine it with a continuous trajectory point set to perform phase unwrapping and synthetic aperture back projection calculation to obtain the three-dimensional spatial coordinates of the tag. The comprehensive judgment module is used to determine whether the three-dimensional spatial coordinates fall within the bounding box of the object detected in the three-dimensional voxel model, and to cross-validate the SKU information corresponding to the label with the material category, and output the inventory results.

2. The intelligent inventory and storage management system for a dark warehouse according to claim 1, characterized in that, The 3D voxel model adopts a layered architecture, including: a probabilistic grid layer, which stores the position index of each voxel unit and the occupancy probability value updated cumulatively over multiple frames; a state label layer, which stores the occupancy state label of each voxel unit determined by comparing the occupancy probability value with a threshold; a connected component analysis layer, which performs 3D connected component labeling on the voxel units marked as occupied, identifies independent object clusters, and stores the index set of voxel units contained in each cluster; and an object attribute layer, which calculates the geometric center coordinates, bounding box range, and volume estimate of the object based on the coordinates of the voxel units contained in each object cluster.

3. The intelligent inventory and warehouse management system for a dark warehouse according to claim 1, characterized in that, The determination of the material category of the contents includes: The aiming angle of the directional sound source is calculated based on the geometric center coordinates of the object, and the direction of the sound source is controlled. A sweeping sound wave signal with a frequency that varies with time is generated and emitted towards the target object. The response signal is collected through a microphone array, and the response signal is preprocessed by filtering and normalization. The transfer function between the excitation signal and the response signal is calculated, and spectral feature parameters including the main resonant frequency, resonant peak width, high-frequency attenuation rate, and low-frequency energy proportion are extracted from the transfer function. The spectral feature parameters are input into a pre-trained classification model, and the material category label and confidence score are output.

4. The intelligent inventory and storage management system for a dark warehouse according to claim 1, characterized in that, The steps for obtaining the three-dimensional spatial coordinates are as follows: during the movement of the inventory robot, record the tag identifier, carrier phase value, and reading time of each RFID tag reading; query the corresponding robot pose from the continuous trajectory point set according to the reading time, and construct the phase history sequence of each tag by grouping according to the tag identifier; perform unwrapping processing on the phase history sequence, detect the jump of adjacent phase values ​​and compensate for periodic folding, and restore the continuous phase change curve; establish a candidate position grid in the target area, calculate its theoretical phase for each candidate position, and calculate the correlation between the theoretical phase and the measured phase; select the candidate position with the highest correlation as the three-dimensional spatial coordinate estimation result of the tag.

5. The intelligent inventory and storage management system for a dark warehouse according to claim 1, characterized in that, The judgment process of the comprehensive judgment module is as follows: Obtain the 3D spatial coordinates of each label and the bounding box range of each object; for the 3D spatial coordinates of each label, check whether all three components are located between the minimum and maximum corner points of the bounding box. If so, determine that the label falls within the bounding box of the object and establish the association between the label and the object; for labels whose coordinates do not fall within any bounding box, mark them as label drift anomalies; for bounding boxes in which no associated labels are detected, mark them as label missing anomalies.

6. The intelligent inventory and storage management system for a dark warehouse according to claim 1, characterized in that: The material cross-validation process includes: Based on the label identifier, query the associated SKU information and expected material category; based on the association between the label and the object, compare the expected material of each label's SKU with the actual measured material category of the associated object; for cases where the expected material and the actual material are inconsistent, mark it as a material anomaly and record the anomaly details; The inventory results are compiled and generated, including the inventory grid identification, the number and list of detected objects, the number and list of identification tags, the spatial matching status, the material verification status, the overall inventory status, and the list of anomaly details.

7. A smart inventory and warehouse management method for darkened warehouses, characterized in that, include: The task path of the transport AGV is obtained, the transport AGVs that pass laterally through the inventory location are selected, the inventory robot is scheduled to enter the wake zone and perform lateral undulation following motion, and the continuous trajectory point set with spatiotemporal labels is output; during the movement of the inventory robot along the trajectory, the target grid is continuously scanned from multiple perspectives, the point cloud data at different times are probabilistically occupied and fused, and the occlusion-free 3D voxel model and the geometric center coordinates of the object are output. The probability occupancy fusion includes: discretizing the spatial range of the target compartment into voxel units and initializing the occupancy probability value of each voxel unit; Based on the robot pose at the time of acquisition, the point cloud of each frame is transformed from the sensor coordinate system to the global coordinate system. For each point in the transformed point cloud, light is projected from the sensor position. The occupancy probability of the voxel unit through which the light passes is updated and shifted towards the idle direction. The occupancy probability of the voxel unit where the light terminates is updated and shifted towards the occupancy direction. The probability update is performed by traversing multiple consecutive point clouds. The occupancy status of each voxel unit is determined based on the comparison between the final occupancy probability value and the preset threshold. The aiming direction is controlled according to the geometric center coordinates, and a sweep frequency acoustic signal is emitted to collect the vibration response signal of the packaging surface. The acoustic impedance spectrum features are extracted and the material category of the contents is determined. During the movement, the carrier phase data of the RFID tag is continuously collected. The phase unwrapping and synthetic aperture back projection calculation are performed in combination with the continuous trajectory point set to obtain the three-dimensional spatial coordinates of the tag. It is determined whether the three-dimensional spatial coordinates fall within the bounding box of the object detected in the three-dimensional voxel model. The SKU information corresponding to the tag is cross-validated with the material category, and the inventory results are output.

Images