A vending machine settlement system and method supporting multi-modal payment

Abstract

The application relates to the technical field of automatic vending equipment, and discloses an automatic vending cabinet settlement system and method supporting multi-modal payment, which comprises a voucher buffer module, a topological mapping module, a reverse check module and an arbitration settlement module. The voucher buffer module listens to a payment authentication request and presses a candidate voucher state vector into a buffer pool. The topological mapping module extracts a three-dimensional skeleton key point sequence, assigns a skeleton anchor point target, and generates a virtual anchor point three-dimensional coordinate and a confidence coefficient through skeleton kinematics graph backtracking when physical occlusion occurs. The reverse check module combines layer frame weight mutation absolute height and human body envelope height to calculate a relative taking height index, compares the Z-axis displacement of the key point sequence with a mechanical constraint condition, and removes invalid vectors. The arbitration settlement module calculates a kinematics correlation index based on a space attenuation term, a time attenuation term and a confidence coefficient, and executes a deduction or a block according to the comparison result of the maximum value and a safety isolation threshold. The application solves the occlusion problem by using physical law cross verification visual tracking to prevent mistaken deduction and theft.

Description

Technical Field

[0001] This invention relates to the field of automated vending equipment technology, specifically to an automated vending machine settlement system and method that supports multimodal payment. Background Technology

[0002] With the development of unmanned retail technology, automated vending machines with visual recognition and gravity sensing functions have been widely used, and their payment authentication methods are gradually developing towards multimodal directions such as facial recognition, palm vein recognition, and near-field communication.

[0003] In real-world applications, there are often multiple concurrent interactions or dense queues. Most existing payment systems rely on simple timing-based judgments or static visual features to link payment accounts with item retrieval actions. However, due to cabinet structure limitations and the complexity of user retrieval movements, external visual tracking devices frequently encounter physical obstructions when capturing key points of the user's hands or body. Once obstruction occurs, target tracking is easily lost, making it impossible for the system to continuously and accurately determine the actual person retrieving the item.

[0004] Meanwhile, although existing vending machines are equipped with gravity sensors, current technologies typically only use them to calculate the increase or decrease in the quantity or type of goods. In complex concurrent scenarios, existing systems lack deep cross-modal correlation between the spatial location of internal gravity events and the human posture captured by external vision. Because they fail to cross-validate the kinematic changes during item retrieval with physical spatial constraints, existing systems often cannot effectively distinguish between situations where one person pays and another retrieves the item, or where users interfere with the visual system's judgment through abnormal movements. This disconnect between internal and external sensory data leads to a decrease in the settlement accuracy of vending machines in open environments, easily resulting in incorrect order deductions and fraudulent transactions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automated vending machine settlement system and method that supports multimodal payment. This solves the problems of incorrect deductions and fraudulent transactions caused by visual tracking loss due to occlusion and lack of cross-validation of internal and external cross-modal physical constraints in existing automated vending machines during multi-user concurrent interaction scenarios.

[0006] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides an automated vending machine checkout system supporting multimodal payment, comprising: The voucher buffer module is used to listen for concurrent payment authentication requests, instantiate candidate voucher state vectors containing multi-dimensional attribute parameters and push them into the circular buffer pool, and at the same time issue cabinet door unlocking instructions. The topology mapping module is used to extract the three-dimensional skeleton key point sequence of the human target, assign skeleton anchor point targets to the candidate voucher according to the payment modality attribute parameters in the candidate voucher state vector, and when physical occlusion occurs, backtrack the parent node through the skeleton directed kinematic graph constructed by the key point sequence, infer and calculate the three-dimensional coordinates of the virtual anchor point, and generate the inference confidence coefficient. The reverse verification module is used to extract the absolute Z-axis height of the physical shelf and the absolute timestamp of the physical retrieval event of the shelf weight mutation event, calculate the relative retrieval height index by combining the static envelope height of the human target, extract the Z-axis displacement of the key point sequence of the three-dimensional skeleton and compare it with the mechanical constraint conditions corresponding to the relative retrieval height index, and eliminate candidate voucher state vectors that do not meet the constraint conditions. The arbitration settlement module is used to calculate the kinematic correlation index of the remaining candidate voucher state vector in the annular buffer pool based on the spatial Euclidean distance attenuation term, the time difference attenuation term, and the inference confidence coefficient. The maximum value is compared with the preset security isolation threshold. If the condition is met, a deduction request is issued; if the condition is not met, the deduction instruction is blocked.

[0007] Furthermore, the credential buffer module includes a vector encapsulation subunit, used to define the candidate credential state vector as a matrix containing a user's unique identifier, sensor modality type parameters, three-dimensional physical space absolute coordinate vectors, and the candidate credential's absolute timestamp. The vector encapsulation subunit executes a dual-queue update rule: calculating the difference between the candidate credential's absolute timestamp and the current system time, and removing candidate credential state vectors whose difference is greater than or equal to a preset time window length from the circular buffer pool; when the total number of valid vectors in the circular buffer pool reaches the maximum space length, subsequent vector pushes are blocked. This mechanism avoids system blocking between payment authentication and physical door opening actions.

[0008] Furthermore, the topology mapping module includes a modal anchor mapping subunit, which is used to assign the preferred skeleton anchor target for external visual tracking based on the payment modality attribute parameters in the candidate credential state vector; when the payment modality attribute parameter is a face recognition modality, the cervical spine node is specified as the preferred skeleton anchor target; when the payment modality attribute parameter is a near-field communication modality or a palm vein recognition modality, the wrist node is specified as the preferred skeleton anchor target.

[0009] Furthermore, the topology mapping module also includes a kinematic compensation inference subunit. When the 2D feature confidence of the preferred skeleton anchor target is lower than a preset visibility threshold, it backtracks along the directed kinematic graph of the skeleton towards the root node to find the nearest valid parent node with a feature confidence greater than or equal to the preset visibility threshold. Using the 3D coordinates of the nearest valid parent node, the unit direction vector calculated based on the motion trajectory of the historical frame buffer, the static envelope height, and the standard scaling constant, the 3D coordinates of the virtual anchor point are calculated by superimposing them. And based on the backtracking level depth, an inference confidence coefficient is generated using a linear decay formula. Thus, the spatial position of the target is reconstructed within the visual occlusion area, and the computational error introduced by the spatial inference is evaluated.

[0010] Furthermore, the reverse verification module includes a height index mapping subunit, which is used to extract the static envelope height of the human target and divide the absolute height of the physical shelf Z-axis by the static envelope height to calculate the dimensionless relative object retrieval height index; according to the preset low-domain division threshold and high-domain division threshold, the relative object retrieval height index is divided into low-domain, middle-domain or high-domain spatial states.

[0011] Furthermore, the reverse verification module also includes a cross-domain displacement verification subunit, used to extract the skeleton root node from the 3D skeleton key point sequence, calculate the two-dimensional Euclidean distance between the skeleton root node and the vending machine on the horizontal plane, and when the two-dimensional Euclidean distance is less than a preset operating radius threshold, extract the difference in Z-axis coordinates between the skeleton root node at the time of item retrieval and the time of authentication as the Z-axis displacement. The mechanical constraints specifically include: when in a low-domain space state, the Z-axis displacement must be less than or equal to the negative value of a preset center of gravity sinking scalar threshold; when in a high-domain space state, the Z-axis displacement must be greater than or equal to a preset center of gravity stretching scalar threshold. This verification method utilizes the physical laws of human center of gravity changes, eliminating the possibility of posture misjudgment during external visual tracking by using the gravity event height of the shelf.

[0012] Furthermore, the arbitration settlement module includes an index fusion calculation subunit, which calculates the kinematic correlation index through the following logic: obtaining the three-dimensional absolute spatial distance between the skeleton anchor point target corresponding to the candidate voucher and the geometric center point of the front outer edge of the physical shelf where a weight change occurs, and calculating the spatial Euclidean distance attenuation term in combination with the spatial attenuation constant; calculating the absolute difference between the authentication timestamp and the absolute timestamp of the physical item retrieval event, and calculating the time difference attenuation term in combination with the time attenuation constant; multiplying the spatial Euclidean distance attenuation term by the spatial dimension weight term, multiplying the time difference attenuation term by the time dimension weight term, and multiplying the weighted sum of the two by the inference confidence coefficient to obtain the final kinematic correlation index.

[0013] Furthermore, the arbitration settlement module includes a settlement boundary determination subunit, which is used to sort the kinematic correlation indices of all valid vouchers in descending order, obtain the maximum and second largest values ​​of the kinematic correlation indices, and calculate the difference margin between the maximum and second largest values; the conditions to be met specifically include that the maximum value is greater than or equal to the safety isolation threshold and the difference margin is greater than or equal to the preset difference margin threshold.

[0014] Furthermore, the credential buffer module includes a concurrent listening subunit. The concurrent listening subunit configures network time protocol services at the system bus level to synchronize and align the internal hardware clock of the multimodal payment sensor array with the clock of the internal gravity sensing matrix, so that the absolute timestamp of the candidate credential and the absolute timestamp of the physical retrieval event are on the same absolute time axis.

[0015] A second aspect of the present invention provides a method for settling accounts in an automated vending machine that supports multimodal payments, applied to the aforementioned automated vending machine settlement system that supports multimodal payments, comprising the following steps: S1 listens for concurrent payment authentication requests, instantiates a candidate credential state vector containing multi-dimensional attribute parameters and pushes it into the circular buffer pool, and simultaneously issues a cabinet door unlocking command. S2, extract the three-dimensional skeleton key point sequence of the human target, assign skeleton anchor point targets to the candidate voucher according to the payment modal attribute parameters in the candidate voucher state vector, when physical occlusion occurs, backtrack the parent node through the skeleton directed kinematic graph constructed by the key point sequence, infer and calculate the three-dimensional coordinates of the virtual anchor point, and generate the inference confidence coefficient. S3, extract the absolute height of the physical shelf Z-axis and the absolute timestamp of the physical retrieval event of the shelf weight mutation event, calculate the relative retrieval height index by combining the static envelope height of the human target, extract the Z-axis displacement of the three-dimensional skeleton key point sequence and compare it with the mechanical constraint conditions of the corresponding relative retrieval height index, and remove the candidate voucher state vectors that do not meet the constraint conditions. S4 calculates the kinematic correlation index of the remaining candidate voucher state vectors in the annular buffer pool based on the spatial Euclidean distance attenuation term, the time difference attenuation term, and the deduction confidence coefficient. The maximum value is compared with the preset security isolation threshold. If the condition is met, a deduction request is issued; if the condition is not met, the deduction instruction is blocked.

[0016] This invention provides an automated vending machine checkout system and method supporting multimodal payment. It has the following beneficial effects: 1. This invention utilizes a topology mapping module to backtrack parent nodes using a directed kinematic graph of the skeleton constructed from a sequence of key points in the 3D skeleton when a human target is physically occluded. This allows for the calculation of the 3D coordinates of virtual anchor points and the generation of a projection confidence coefficient. This feature enables the reconstruction of the target's spatial position in multimodal payment scenarios, addressing blind spots caused by concurrent interactions among multiple users, thus reducing the target tracking loss rate due to external visual occlusion.

[0017] 2. This invention extracts the absolute Z-axis height of the physical shelf from the sudden weight change event using a reverse verification module. It then calculates the relative object retrieval height index by combining this with the static envelope height of the human target. Finally, it compares the Z-axis displacement of the 3D skeleton keypoint sequence with the corresponding mechanical constraints. This design utilizes the physical laws governing the change in the center of gravity when a person retrieves an object. By cross-validating the objective height events from the internal gravity sensor with the behavior capture of the external vision system, it effectively eliminates potential posture misjudgments during visual tracking, preventing erroneous charges and fraudulent transactions.

[0018] 3. This invention, through an arbitration settlement module, integrates spatial Euclidean distance attenuation, time difference attenuation, and inference confidence coefficients to calculate the kinematic correlation index of candidate voucher state vectors. The maximum value is then compared with a security isolation threshold to determine whether to issue a deduction instruction. This mechanism combines the weights of three dimensions—spatial span, time delay, and location inference error—to quantitatively match and filter concurrent authentication requests, improving the accuracy of vending machine order settlement in complex open scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system block diagram of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the voucher buffer module of the present invention; Figure 4 This is a logical schematic diagram of the topology mapping module of the present invention; Figure 5 This is a schematic diagram of the reverse verification module of the present invention; Figure 6 This is a schematic diagram of the arbitration settlement module of the present invention; Figure 7 This is a bar chart comparing the performance of the present invention in complex concurrent scenarios.

[0020] Among them, 10 is the voucher buffer module; 20 is the topology mapping module; 30 is the reverse verification module; and 40 is the arbitration settlement module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 This invention provides an automated vending machine settlement system supporting multimodal payment, comprising: The voucher buffer module 10 is connected to the multimodal payment sensor array for listening to concurrent payment authentication requests. After any sensor completes feature comparison and account authentication, it instantiates a candidate voucher state vector containing multidimensional attribute parameters, pushes the candidate voucher state vector into a ring buffer pool with a preset time length, and simultaneously sends a cabinet door unlocking command to the vending machine execution mechanism. The ring buffer pool releases the linear blocking control logic between the payment authentication event and the physical door opening action. The topology mapping module 20 is connected to the external visual capture module. It is used to receive the two-dimensional video stream output by the external visual capture module and extract the three-dimensional skeleton key point sequence of the human target. According to the payment modality attribute parameters in the candidate credential state vector, it assigns the skeleton anchor point target of external visual tracking to each candidate credential. When the skeleton key point of the human target is physically occluded, it calculates the three-dimensional coordinates of the virtual anchor point by tracing back the parent node through the skeleton directed kinematic graph and generates the inference confidence coefficient for the candidate credential. The reverse verification module 30 is connected to the internal gravity sensing matrix and is used to extract the shelf weight mutation event data output by the internal gravity sensing matrix during the period from the opening to the closing of the cabinet door. The event data includes the absolute height of the physical shelf Z-axis when the weight mutation occurs and the absolute timestamp of the physical retrieval event. The relative retrieval height index is calculated based on the static envelope height of the human target and the absolute height of the physical shelf Z-axis. The Z-axis displacement of the skeleton key point sequence within the retrieval time window is extracted and compared with the segmented ergonomic mechanical constraints corresponding to the relative retrieval height index. Candidate credential state vectors that do not meet the mechanical constraints are removed from the circular buffer pool. The arbitration settlement module 40 is connected to the topology mapping module 20 and the reverse verification module 30 respectively. It is used to perform multi-dimensional parameter fusion calculation on the remaining candidate voucher state vectors in the ring buffer pool. It calculates the kinematic correlation index of each candidate voucher state vector based on the spatial Euclidean distance attenuation term, the time difference attenuation term, and the inference confidence coefficient. It compares all the calculated kinematic correlation indices, selects the maximum value, and compares it with the preset security isolation threshold. When the security isolation threshold condition is met, it binds the item collection settlement flow to the target user identifier corresponding to the maximum value, issues an asynchronous deduction network request, and clears the ring buffer pool. When the security isolation threshold condition is not met, it blocks the deduction instruction and generates an pending order status.

[0023] Please see the appendix Figure 2 This invention provides a method for settling accounts at automated vending machines that supports multimodal payments, comprising the following steps: S10: Listen for concurrent payment authentication requests. After any sensor completes feature comparison and account authentication, instantiate a candidate voucher state vector containing multi-dimensional attribute parameters, push the candidate voucher state vector into a circular buffer pool with a preset time length, and simultaneously issue a door unlocking command to the vending machine execution mechanism. S20: Receive the two-dimensional video stream output by the external visual capture module and extract the three-dimensional skeleton key point sequence of the human target. Assign the skeleton anchor point target of external visual tracking to each candidate credential according to the payment modal attribute parameter in the candidate credential state vector. When the skeleton key point of the human target is physically occluded, calculate the three-dimensional coordinates of the virtual anchor point by backtracking the parent node through the skeleton directed kinematic graph and generate the inference confidence coefficient for the candidate credential. S30: Extract shelf weight mutation event data from the internal gravity sensing matrix during the period from cabinet door opening to closing. The event data includes the absolute height of the physical shelf Z-axis where the weight mutation occurs and the absolute timestamp of the physical retrieval event. Calculate the relative retrieval height index based on the static envelope height of the human target and the absolute height of the physical shelf Z-axis. Extract the Z-axis displacement of the skeleton key point sequence within the retrieval time window and compare it with the segmented ergonomic mechanical constraints corresponding to the relative retrieval height index. Remove candidate credential state vectors that do not meet the mechanical constraints from the circular buffer pool. S40: Perform multi-dimensional parameter fusion calculation on the remaining candidate voucher state vectors in the circular buffer pool. Calculate the kinematic correlation index of each candidate voucher state vector based on the spatial Euclidean distance attenuation term, the time difference attenuation term, and the inferred confidence coefficient. Compare all calculated kinematic correlation indices to select the maximum value and compare it with the preset security isolation threshold. If the security isolation threshold condition is met, bind the item collection settlement transaction to the target user identifier corresponding to the maximum value, issue an asynchronous deduction network request, and clear the circular buffer pool. If the security isolation threshold condition is not met, block the deduction instruction and generate an pending order status.

[0024] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0025] See attached document Figure 3 , Figure 3 This is a logical schematic diagram of a credential buffer module according to an embodiment of the present invention.

[0026] In this embodiment, the credential buffer module 10 is mainly used to execute a multimodal concurrent asynchronous buffer decoupling mechanism. To implement concurrent request capture and linear state machine decoupling logic, the credential buffer module 10 is further divided into a concurrent monitoring subunit, a vector encapsulation subunit, and an instruction decoupling subunit.

[0027] The concurrent monitoring subunit communicates with the multimodal payment sensor array. The multimodal payment sensor array includes independent external hardware devices such as a face recognition camera, an RFID card reader, and a palm vein scanner. The concurrent monitoring subunit is configured with parallel polling tasks in the underlying operating system of the edge computing control unit, allocating independent monitoring threads for sensor devices with different communication bus interfaces.

[0028] To eliminate the risk of out-of-order issues caused by internal clock skew in heterogeneous sensors, as a preferred approach, the concurrent monitoring subunit is configured with a Network Time Protocol (NTP) service at the system bus level, forcing microsecond-level alignment of the hardware clocks of all external hardware devices. Based on this unified clock reference, when any external hardware device captures a user's payment characteristics and completes account authentication, that external hardware device sends a hardware interrupt request to the edge computing control unit.

[0029] The concurrent listening subunit takes over the authentication event through an interrupt handler and does not trigger exclusive mutex locks on other communication bus interfaces. The specific configuration process for the edge computing control unit's bus polling and allocation of independent listening threads can be written conventionally by those skilled in the art based on the operating system and hardware platform. The underlying communication interface protocol configuration is well-known in the field and will not be elaborated upon here.

[0030] After completing the non-blocking takeover of the underlying signals, in order to achieve a standardized mathematical expression for heterogeneous authentication data, the vector encapsulation subunit instantiates a candidate credential state vector containing multi-dimensional attribute parameters. Candidate credential state vector The mathematical expression for it is defined as: ; In the formula, The unique identifier of the associated user is used to complete the authentication comparison. The sensor modality type parameter that triggers the authentication event. The absolute coordinate vector of the external hardware device that triggers the authentication event in the three-dimensional physical space in the three-dimensional global world coordinate system is determined by the three spatial components measured offline in advance. constitute, The absolute timestamp for external hardware devices to complete feature comparison and account authentication operations. Indicates matrix transpose. This is the index number of the current candidate credential in the buffer pool.

[0031] Generate candidate voucher state vector Then, the vector encapsulation subunit pushes it into a pre-allocated circular buffer pool in system memory. Considering the memory limits for edge computing in practical engineering, the circular buffer pool has a fixed maximum space length. and preset time window length Preset time window length The specific value range is set to 1.5 to 2.0 times the mechanical delay of the vending machine door's magnetic closing action, typically ranging from 3000 milliseconds to 5000 milliseconds, to ensure that this time envelope can fully cover a single physical item retrieval action. Each time a new vector is pushed in, the system executes a dual-queue update rule.

[0032] On one hand, the vector encapsulation subunit traverses the existing candidate credential state vectors in the circular buffer pool and extracts the absolute timestamp of each vector. Calculate the difference between this time and the current system time, and set the difference to be greater than or equal to the current system time. The candidate credential state vector is dequeued and destroyed from memory; On the other hand, when concurrent push requests within a unit of time cause the total number of valid vectors in the buffer pool to reach the maximum space length... When this happens, the system blocks the subsequent push of vectors and throws a congestion exception status quantity to the upper-layer bus, thereby preventing core memory overflow.

[0033] The instruction decoupling subunit is connected to the vector encapsulation subunit. The instruction decoupling subunit detects a new candidate credential state vector. At the synchronization moment of pushing the data into the circular buffer pool, a non-blocking door unlocking command is issued to the vending machine actuator via the general input / output interface. The vending machine actuator is typically an electromagnetic door lock relay. Based on the aforementioned data residency mechanism of the circular buffer pool, the command decoupling subunit does not perform exclusive binding of the currently triggering user identifier at the global variable level during the issuance of the unlocking command. The system maintains a parallel listening state, allowing authentication requests from other users to continue generating new candidate credential state vectors and pushing them into the circular buffer pool. Through this mechanism, the credential buffer module 10 eliminates the linear blocking control logic established between the payment authentication event and the physical door opening action in traditional vending systems, transforming the single-threaded queuing settlement mode into a multi-request asynchronous parallel buffer mode.

[0034] See attached document Figure 4 , Figure 4 This is a logical schematic diagram of a topology mapping module according to an embodiment of the present invention.

[0035] In this embodiment, the topology mapping module 20 is used to perform cross-domain modality-driven dynamic node mapping and directed layer-level compensation mechanism. To achieve cross-domain fusion and spatial fault-tolerant inference of external machine vision tracking data and multimodal payment attribute parameters, the topology mapping module 20 is further divided into a skeleton extraction and envelope initialization subunit, a modal anchor point mapping subunit, and a kinematic compensation inference subunit.

[0036] The skeleton extraction and envelope initialization subunit communicates with the external vision capture module. Due to complex lighting and background interference in the physical environment, direct geometric calculations are insufficient to obtain stable poses. Therefore, the skeleton extraction and envelope initialization subunit receives a continuous two-dimensional video stream from the external vision capture module and feeds it as an input tensor into a pre-built three-dimensional human pose estimation neural network model. As a preferred approach, this model employs a high-resolution network architecture. During the model's forward inference phase, the input data is preprocessed and uniformly scaled to the required size dimension. The standard tensor, in which =3 corresponds to RGB three-channel image data. Internally, the model uses parallel high- and low-resolution convolutional subnetworks for feature extraction and multi-scale fusion. The output is a three-dimensional heatmap containing multiple skeletal key points with anatomical and physical significance, which is then converted into a local three-dimensional coordinate sequence after the expectation is calculated.

[0037] To ensure that those skilled in the art can reproduce the model, its offline training steps include: constructing a sample library using publicly available large-scale 3D human pose datasets such as Human3.6M; defining training samples as real-world RGB images and labels as 3D spatial coordinates of skeletons manually annotated by motion capture equipment; and using mean squared error (MSE) as the loss function for model training, iteratively updating network weights through backpropagation until the loss value converges to below a set threshold.

[0038] After acquiring the local 3D coordinate sequence, the skeleton extraction and envelope initialization subunit calls the camera intrinsic and extrinsic parameter matrices, which have been pre-calibrated offline by the external vision capture module. Through coordinate system rigid body transformation operations, it maps each local skeleton keypoint relative to the camera into an absolute 3D coordinate sequence in the 3D global world coordinate system. During the coordinate system transformation operation of matrix inversion, to prevent matrix singularity issues caused by extreme camera distortion or coplanar lattices, the system introduces damped least squares in the inverse matrix calculation. This is achieved by superimposing small regularization parameters on the diagonal to ensure the stability of the calculation results.

[0039] Based on this absolute three-dimensional coordinate sequence, the skeleton extraction and envelope initialization subunit further establishes the static envelope height of the target individual. The system calculates the principal axis direction vector of the target skeleton's torso in real time. When the angle between this principal axis direction vector and the vertical line of the world coordinate system is determined to be less than a preset upright tilt angle threshold (usually ranging from 5 to 10 degrees to filter out interference from non-upright postures such as bending over), the maximum spatial span value from the anatomical head vertex node to the foot end node in the global absolute Z-axis direction is extracted as the static envelope height. This height parameter objectively describes the longitudinal physical dimensions of an individual user, establishing a basic conversion benchmark for eliminating errors in sports school testing caused by differences in height among different users.

[0040] The modal anchor mapping subunit is connected to the skeleton extraction and envelope initialization subunit. To establish a rigid coupling between discrete financial payment data and continuous visual tracking data, the modal anchor mapping subunit uses payment modal attribute parameters from the candidate voucher state vectors within the circular buffer pool. Dynamically assign the preferred skeleton anchor point target for external visual tracking to each candidate credential. .

[0041] The system maintains a hard-coded association mapping table internally. When the payment modality attribute parameters... When extracting facial features for face recognition, the system forcibly designates the cervical vertebrae as the primary skeletal anchor point target from the key points of human skeletal anatomy. The technical reason for this mapping strategy is that the cervical vertebrae possess stable positional output in visual pose estimation algorithms and are always physically adjacent to the facial payment sensing area. When the payment modality attribute parameters... When using near-field communication or palm vein recognition modes, the system designates the wrist node on the user interaction side as the preferred skeleton anchor point target. This setup is based on physical causality, where the wrist is the final stable kinematic transmission node that drives the palm or NFC device closer to the payment sensing area. Through this cross-domain parameter mapping mechanism, the system transforms the abstract payment credential into a continuously trackable biomechanical entity in physical space.

[0042] The kinematic compensation deduction subunit is connected to the two subunits mentioned above, and is used to maintain the continuity of spatial tracking in situations where visual occlusion is caused by dense crowds. In actual sales scenarios, the preferred skeleton anchor point target extracted by the external visual capture module is... The confidence level of the two-dimensional features is lower than the preset visibility threshold. The risk of this visibility threshold. The value range is typically set to 0.4 to 0.6. When the confidence level falls below this range, continued acceptance will lead to significant jumps and drifts in the 3D coordinate calculation. To address the aforementioned feature loss defect, the kinematic compensation inference subunit defines the extracted human skeleton as a directed kinematic graph with hierarchical topological relationships. In this directed kinematic diagram, Represents a set of anatomical joint nodes. It represents the set of rigid skeletal lines connecting adjacent joint nodes. The topological hierarchy of the graph is distributed in a directed divergent pattern from the central node of the trunk to the end nodes of the limbs.

[0043] When the preferred skeleton anchor point target When physical occlusion occurs, the kinematic compensation deduction subunits are set along the edges of the directed kinematic graph. Perform a hierarchical reverse backtracking traversal towards the root node to find features with confidence scores greater than or equal to the visibility threshold. The nearest valid parent node Locking the nearest valid parent node. Subsequently, the kinematic compensation deduction subunit, combined with prior ergonomic data, calculates the reconstructed three-dimensional coordinates of the virtual anchor point using the following deduction algorithm formula. : ; In the formula, This refers to the absolute 3D coordinates of the nearest valid parent node in the 3D global world coordinate system. To calculate the unit direction vector from the parent node to the occluded child node, which is obtained from the motion trajectory cached in the skeleton's historical frames, the system simultaneously performs vector magnitude verification when calculating this unit direction vector. Based on multi-dimensional fault-tolerant logic, when it detects that the denominator magnitude approaches zero (i.e., the historical motion amount is extremely small, constituting a singular state in spatial calculation), the system abandons pure mathematical operations and forcibly calls the default unit vector of the limb's natural drooping posture to replace it, in order to avoid system crashes caused by division operations. The target static envelope height established in the aforementioned calculation steps, Using a standard proportional constant of the length of a specific skeletal segment relative to the total height of an individual, based on an anthropometric database, the coordinates of the upper limb joints that are not physically obscured are used as the origin, and the theoretical link length is superimposed on the individual's actual height and scaled proportionally. This allows for the reconstruction of the spatial position of the target anchor point with high certainty within the visual blind spot.

[0044] After completing the compensatory reconstruction of the spatial coordinates, the kinematic compensation deduction subunit simultaneously generates the corresponding deduction confidence coefficient for the candidate evidence. Deducing the confidence coefficient The computational logic is based on the backtracking hierarchy depth of the directed kinematics graph. This depth value is defined as the number of graph line segments traversed from the preferred anchor point to the selected parent node, and its specific linear decay expression is as follows: ; In the formula, The preset confidence penalty step size parameter is set to a value ranging from 0.15 to 0.25. When the preferred skeleton anchor point target is not occluded, the backtracking depth... The confidence coefficient is zero. The confidence coefficient is maintained at its maximum value of 1.0. When physical occlusion occurs in the system and the backtracking level deepens, the extrapolation confidence coefficient is adjusted. The corresponding weighting decreases accordingly. This mechanism ensures that the degradation of the system's underlying visual features will not trigger a deadlock or crash in the upper-level control logic. At the same time, by actively reducing the association weight of the fuzzy certificate in the subsequent arbitration and settlement stage, it objectively reflects the risk of distance measurement errors brought about by deep spatial deduction.

[0045] See attached document Figure 5 , Figure 5 This is a schematic diagram of the logic of a reverse verification module according to an embodiment of the present invention.

[0046] In this embodiment, the reverse verification module 30 is used to execute a Z-axis kinematics reverse rigidity verification mechanism based on the relative reachability domain, thereby constructing a conditional constraint closed loop between the internal hardware physical state and the external human biomechanical characteristics. To achieve the above-mentioned cross-domain physical quantity comparison and abnormal behavior filtering, the reverse verification module 30 is further divided into a gravity event extraction subunit, a height index mapping subunit, and a cross-domain displacement verification subunit.

[0047] The gravity event extraction subunit is communicatively connected to the internal gravity sensing matrix. The gravity sensing matrix deployed inside the vending machine consists of an array of pressure sensors distributed at the bottom of each shelf. Each pressure sensor has a pre-calibrated absolute Z-axis physical height in a three-dimensional global coordinate system. During the operating cycle from door opening to closing, the gravity event extraction subunit continuously reads the weight readings from the gravity sensing matrix at a fixed sampling rate. In actual physical environments, compressor vibration or door opening and closing can introduce low-frequency mechanical interference. Based on these operating conditions, as a preferred approach, the gravity event extraction subunit 31 calculates the weight change gradient using a first-order difference algorithm combined with a sliding time window.

[0048] When the system detects a negative gradient in weight change on a local shelf that exceeds a preset minimum weight threshold for the item (typically 10 to 20 grams, determined based on the weight of the lightest item in the shelf), it determines that a valid physical item retrieval event has occurred. The gravity event extraction subunit extracts the absolute Z-axis height of the physical shelf corresponding to this sharp weight reduction event. And record the absolute timestamp of the event. This absolute timestamp The authentication events in the credential buffer module 10 are clock-aligned using the same network time protocol to ensure that internal hardware physical events and external visual tracking states are on the same absolute time axis. For the underlying analog-to-digital conversion and filtering / denoising algorithms of the gravity sensor, those skilled in the art can use conventional moving average filtering or Kalman filtering for code writing; the underlying signal processing is well-known in the field and will not be elaborated upon here.

[0049] The height index mapping subunit is connected to both the gravity event extraction subunit and the topology mapping module 20. Considering the height differences among users, the biomechanical actions corresponding to the same physical shelf height vary significantly. Before proceeding to specific mechanical judgments, the system needs to establish a unified dimensionless evaluation benchmark. For each candidate credential state vector in the annular buffer pool, the height index mapping subunit extracts the static envelope height of its associated target individual. And based on the absolute height of the physical shelf's Z-axis Calculate the relative height index Its computational model is defined as follows: ; In the formula, It is a dimensionless relative retrieval height index, whose physical meaning lies in objectively representing the relative reachable spatial position of the physical shelf relative to the individual user. The absolute height of the physical shelf that experienced the weight mutation event along the Z-axis in the global world coordinate system. The static envelope height of the target individual. This is the index number of the current candidate credential. Before performing this division operation, the system is configured with denominator zero-tolerance logic. When the static envelope height of the target individual is detected... If the height is less than the preset lower limit of the effective height (e.g., 0.5 meters), it is determined to be a distortion of the visual tracking data. The system will forcibly call the local standard adult average height as the denominator to avoid the overflow anomaly caused by division by zero.

[0050] The relative retrieval height index was calculated. Subsequently, the height index mapping subunit, based on ergonomic prior distribution patterns, divides the three-dimensional object retrieval space into three discrete execution domains. The system presets two dimensionless partitioning thresholds: a low-domain partitioning threshold... High-domain partitioning threshold . The value range is set to 0.3 to 0.45, representing the shelf height being in the area from the user's knees to below the waist; The value range is set to 0.75 to 0.85, representing the area where the shelf height is above the user's shoulders. Based on these two thresholds, the system divides the space into a low domain ( ), Zhongyu ( ) and high domain ( This provides state guidance for subsequent segmented biomechanical verification.

[0051] The cross-domain displacement verification subunit is connected to the height index mapping subunit to perform the final candidate credential elimination judgment. In human biomechanics, the spatial displacement of the pelvic center node can most stably reflect the changing state of the body's center of gravity. Based on this physical causal relationship, the cross-domain displacement verification subunit extracts the root node (anatomical pelvic center) of the corresponding skeleton in the global world coordinate system as the centroid reference.

[0052] To eliminate the high-frequency jitter noise introduced by single-frame visual pose estimation, the system uses authentication timestamps. Absolute timestamp of physical item retrieval The width of each nearby opening is A time smoothing window (typically 3 to 5 frames) is used. After calculating the average coordinates within two consecutive time windows, the system subtracts the average Z-axis coordinate at the authentication time from the average Z-axis coordinate at the physical object retrieval time. This signed difference operation logic is then used to calculate the actual relative displacement of the root node in the global Z-axis direction. ,in This represents the root node index in the skeleton graph topology.

[0053] To avoid misjudgments caused by relying on extreme value determination in a single spatial dimension, the cross-domain displacement verification subunit adopts a multi-dimensional weighted verification logic that integrates horizontal plane proximity and vertical center of gravity change. The system first calculates the two-dimensional Euclidean distance between the root node of the target skeleton and the vending machine in the global XY plane. Only when The system will only continue to process the actual displacement when the actual displacement is less than the preset operating radius threshold (e.g., 0.8 meters). Substitute the relative object retrieval height index The corresponding piecewise ergonomics and biomechanics constraint function outputs a Boolean filtering function. The result.

[0054] The specific constraint determination logic is as follows: when in the lower domain ( When this happens, it means the user must perform a deep squat or bend over significantly, and the system mandates that the actual displacement must meet certain requirements. ,in The preset scalar threshold for center of gravity subsidence (an empirical value is usually taken as 0.25 meters to 0.35 meters); when in the middle region ( When the user's arm is naturally extended, they can reach the product without significant changes in their center of gravity. In this range, the system is exempt from strong Z-axis displacement verification constraints. However, in the higher domain (…),… When the user simultaneously rises to their feet or stretches their torso upwards, the system mandates that the actual displacement must meet certain conditions. ,in The preset scalar threshold for center of gravity stretching (an empirical value is usually taken as 0.04 meters to 0.08 meters).

[0055] The cross-domain displacement verification subunit 33 outputs a final Boolean value based on the above judgment results. If the skeleton kinematic features (including horizontal distance constraints and Z-axis displacement constraints) corresponding to the candidate credential do not meet any of the above piecewise mechanical constraint conditions, then the filtering function... If the output is 0, the cross-domain displacement verification subunit triggers a hardware memory clearing instruction, hard-removing the candidate credential state vector from the circular buffer. If the physical constraints are met, the filtering function outputs 1, and the credential state vector is retained and sent to the next-level arbitration module. The technical purpose of this reverse verification mechanism is to directly use the deterministic events of the internal gravity sensor to reverse-negate fraudulent behavior that may exist in the external vision system due to excessive distance or incorrect posture by introducing the rigid change law of the center of gravity in human anatomy, thereby significantly improving the system's damage prevention accuracy in complex scenarios.

[0056] See attached document Figure 6 , Figure 6 This is a logical diagram of an arbitration settlement module according to an embodiment of the present invention.

[0057] In this embodiment, the arbitration settlement module 40 is used to execute the kinematic correlation index arbitration algorithm that integrates spatiotemporal parameters. To realize the quantitative mathematical calculation of collision avoidance decision and the system state machine closing logic, the arbitration settlement module 40 is further divided into an index fusion calculation subunit, a correlation sorting and filtering subunit, and a settlement boundary determination subunit.

[0058] The index fusion calculation subunit is communicatively connected to the topology mapping module 20 and the reverse verification module 30, respectively. In complex contactless payment scenarios, when multiple users are simultaneously in front of the vending machine, multiple candidate credential state vectors, all of which have passed the basic physical condition constraints, are often concurrently stored in the annular buffer pool. To accurately locate the unique target that actually performs the physical retrieval behavior, the system needs to introduce multi-dimensional quantitative evaluation benchmarks to avoid the random errors caused by single-dimensional calculations. Based on the above logic, the index fusion calculation subunit extracts the associated multi-dimensional parameters for each remaining candidate credential state vector in the annular buffer pool, performs fusion calculations, and outputs a single dimensionless kinematic correlation index. .

[0059] Specifically, the kinematic correlation index The mathematical formula for calculation is defined as follows: ; In the formula, For the first The final kinematic correlation index corresponding to each candidate credential state vector represents, in physical terms, the comprehensive matching probability between a specific biometric authenticator and the current physical retrieval event. The inference confidence coefficients, passed from the topology mapping module 20, are used as a global compensation term to directly apply a scaling penalty to the final result. This is a spatial Euclidean distance parameter. To avoid the dimensionality reduction calculation error caused by treating the entire physical shelving as the measurement object, this spatial Euclidean distance parameter... It is explicitly defined as the three-dimensional absolute spatial distance between the skeleton anchor point target corresponding to the candidate credential and the geometric center point of the front outer edge of the physical shelf where a weight change occurs at the absolute timestamp of the physical retrieval event. The preset spatial decay constant, typically ranging from 1.5 to 2.5, is calibrated offline based on variance data of a standard adult's outstretched arm. It is used to control the score decay curvature when spatial distance diverges. This is the time difference parameter, which is equal to the absolute difference between the authentication timestamp and the absolute timestamp of item retrieval. The preset time decay constant, typically ranging from 0.2 to 0.5, is determined by fitting the average pause time distribution of a large number of users selecting goods. and These are the spatial dimension weights and the time dimension weights, respectively.

[0060] In the above formula, as a preferred approach, the system introduces a normalization constraint on the weight terms, which is to force them to be satisfied. =1.0. In actual physical conditions, after users complete authentication, unpredictable pauses often occur in front of the counter during product selection, leading to a time difference parameter. There are significant discrete fluctuations. Therefore, the spatial dimension weighting term... The value range is rigidly set to 0.65 to 0.75, while the time dimension weighting item... The value is correspondingly limited to between 0.25 and 0.35. This formula constructs a nonlinear decay model by introducing the natural base exp. Its technical purpose is to provide an exponential competitive advantage when user anchor points are spatially close and actions are coherent; simultaneously, it derives a confidence compensation term. The multiplication intervention ensures that when visual data encounters severe occlusion and multi-level parent node backtracking is performed, the priority of the credential in collision avoidance arbitration is objectively downgraded.

[0061] The correlation sorting and filtering subunit is connected to the index fusion calculation subunit. After completing the individual calculation, the correlation sorting and filtering subunit traverses the entire annular buffer pool and reads the set of kinematic correlation indices for all valid vouchers that have not been eliminated. The system calls the quicksort algorithm to sort the one-dimensional array in descending order and extract the maximum value. and the second largest value (when array length) (Time) and its corresponding target candidate credentials. This traversal and extreme value extraction process converges the underlying concurrent asynchronous data into a definite business processing object. To avoid one-sided judgments caused by relying solely on a single extreme value, the system introduces a multi-dimensional judgment logic based on the degree of competition difference, rather than simply accepting the highest score.

[0062] The settlement boundary determination subunit is connected to the associated sorting and filtering subunit and is used to perform exit control of the system's closed loop. The settlement boundary determination subunit extracts the maximum value. With preset security isolation threshold Perform a comparison and simultaneously calculate the margin of difference between the maximum and second largest values. Security isolation threshold The core parameters for the system to intercept false deductions are usually derived from the false positive rate statistics of historical test datasets, with values ​​ranging from 0.60 to 0.70; the difference margin threshold. It is typically set to 0.15 to measure the lead of the highest-scoring candidate over other interfering players.

[0063] When the maximum value meets the security isolation threshold condition (i.e.) And when multiple valid vouchers exist, the difference margin condition is met simultaneously (i.e. When the collision arbitration is successful and the target confidence level is high, the system determines that the collision avoidance arbitration is successful. The settlement boundary determination subunit rigidly binds the item collection settlement flow output by the gravity matrix to the target user identifier corresponding to the maximum value. To prevent network input / output from blocking the real-time monitoring process of the edge computing control unit, the system encapsulates the target user's consumption data and sends an asynchronous deduction network request to the cloud payment gateway. At the same time as the network socket throws a success signal, the system executes a memory pointer reset instruction, clears all residual state vectors in the current circular buffer pool, and resets the cabinet door electromagnetic lock to the closed armed state, forming a normal automatic vending closed loop.

[0064] For the construction of encapsulation protocols for asynchronous network requests and the processing of asymmetric encrypted transmission, those skilled in the art can use the standard HTTPS protocol for interface calls, which is a well-known technology in the field and will not be elaborated here.

[0065] When the maximum value does not meet the security isolation threshold condition (i.e.) ), or the highest and second-highest scores are too close, resulting in insufficient margin for difference (i.e. When this occurs, it means that the current candidate voucher cannot meet the certainty requirements for financial-grade deduction due to its distance, long time interval, or visual degradation.

[0066] This scenario typically corresponds to situations where the authenticator leaves and a bystander steals the goods, or complex fraud and confusion scenarios involving two users physically intercourse. Based on this, the settlement boundary determination subunit triggers the system boundary circuit breaker strategy. The system directly blocks the deduction request instruction sent to the cloud at the software level to prevent erroneous deductions. Simultaneously, the system packages the current item pickup transaction history, visual snapshots of multiple people in the vicinity, and gravity event sequences, generating an pending order status and suspending it for upload to the cloud audit server. This circuit breaker mechanism, while ensuring the reliability of the consumer's financial account, preserves complete on-site data for subsequent human customer service intervention or system self-learning feature recall.

[0067] To further aid in understanding the core technical solution of this invention and enhance the authenticity and engineering persuasiveness of the solution, the following are specific application examples, namely complex anti-collision and anti-theft scenarios, as well as experimental verification and effect comparison.

[0068] Please see the appendix Figure 7 Specific application examples: Scenario: Suppose the following concurrent interaction occurs in front of a vending machine in an office building: User A (height) Approach the device from 1.80 meters away and use facial recognition to complete the authentication.

[0069] Meanwhile, User B (height) =1.60 meters, standing behind side A) Use your mobile phone to bring it close to the card reader to complete NFC payment authentication.

[0070] User A opened the cabinet door but did not immediately take any items, instead hesitating and selecting. At this moment, User B suddenly reached under A's armpit and reached into the bottom shelf of the vending machine (at an absolute height of...). A (0.35 meters) took a bottle of Coke. The cabinet door was then closed by A.

[0071] Processing flow description S10 (credential buffer): The system concurrently detects two authentication attempts and instantiates the vector. (Modality = Face) and (Modality = NFC), push-in time window =90 seconds of circular buffer pool, and issue an unlock command.

[0072] S20 (Topology Mapping): For Assign cervical vertebrae as anchor points; Assign the right wrist node as the anchor point. Because B reaches into the lower level to retrieve goods, his right wrist is physically obscured by A's body and the cabinet door (confidence drops to 0.3, below the threshold of 0.4). The system triggers directed graph backtracking, finding B's right elbow as the nearest valid parent node, and deducing the virtual wrist anchor point. Since one level of backtracking was performed, the confidence coefficient is calculated. =1.0 - 0.2 × 1 = 0.8. User A is not obscured. =1.0.

[0073] S30 (Reverse Verification - Core Interception Point): Gravity matrix detects the underlying layer ( =0.35m) The weight suddenly decreased.

[0074] For User A: Relative Retrieval Height Index =0.35 / 1.80=0.19 (less than the lower threshold) =0.3), the system determined that the behavior was in the low domain. The system forcibly checked the Z-axis displacement of A's pelvis and found that A was only tilting his head down, and the actual displacement of the center of mass was 0.3. =-0.05m. Since -0.05>-0.25 (the preset deep squatting and sinking threshold has not been reached). The system directly determines that A's action does not conform to the biomechanical laws of bottom object retrieval, and vectors it. Forcefully removed from the buffer pool.

[0075] For User B: Relative Retrieval Height Index =0.35 / 1.60=0.21 (in the low range). The system extracts the Z-axis displacement of B's ​​pelvis and finds that B performed a squatting motion. =-0.38m. Since -0.38 ≤ -0.25, it conforms to the laws of mechanics, so the filter function outputs 1, retaining... .

[0076] S40 (Arbitration Settlement): Since A has been removed, only B remains in the buffer pool. Even with exponential fusion calculation, This also naturally became the maximum value, and the system ultimately deducted the Coca-Cola bill accurately to user B's NFC account, preventing a case where A opened the door and B took the goods, resulting in A being maliciously charged.

[0077] Experimental verification and effect comparison: To verify the effectiveness of this invention, a test dataset containing 1000 real physical item retrieval events was constructed. This dataset includes standard single-person item retrieval (500 events), two-person densely interleaved item retrieval (300 events), and malicious proxy retrieval / theft (200 events).

[0078] Comparison with benchmark: Comparative Example 1: FIFO queue logic based on first-to-first-to-rights settlement.

[0079] Comparative Example 2: Hand skeleton tracking relies solely on an external camera, without incorporating gravity Z-axis and human height index.

[0080] The present invention provides a solution that includes multimodal decoupling, reverse verification of the Z-axis centroid, and topology fusion arbitration.

[0081] Evaluation indicators: Settlement accuracy rate: The percentage of goods that are correctly linked to the actual recipients.

[0082] False positive rate: The proportion of serious customer complaints where users were charged even though they did not receive the goods.

[0083] System circuit breaker rate: The percentage of pending orders that the system cannot confirm and that are manually transferred to the system.

[0084] The experimental data are shown in Table 1: Table 1: Performance Comparison in Complex Concurrency Scenarios ; in conclusion: Experiments show that the multimodal fusion settlement scheme proposed in this invention has absolute superiority in complex concurrent interaction scenarios involving dense interleaving and malicious fraudulent transactions. Compared with traditional time-series queuing methods and pure visual tracking methods, this scheme not only improves the system settlement accuracy to 98.7%, but also successfully reduces the false positive and incorrect deduction rate, which is prone to causing serious customer complaints, to 0.0%, while significantly reducing the system's circuit breaker-to-manual transfer rate.

[0085] The fundamental reason for this breakthrough lies in the fact that this invention overcomes the technical bottleneck of single visual tracking being susceptible to occlusion interference, and innovatively introduces a reverse biomechanical verification mechanism based on the Z-axis physical shelving and the center of gravity of the human skeleton. This mechanism, starting from the underlying logic of physical and mechanical laws, can hard-verify and eliminate abnormal object-taking behaviors, thereby preventing malicious proxy taking and collision-induced deductions in complex scenarios, providing unmanned retail terminals with an extremely high level of financial security and system reliability.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated vending machine settlement system supporting multimodal payment, characterized in that, include: The voucher buffer module is used to listen for concurrent payment authentication requests, instantiate candidate voucher state vectors containing multi-dimensional attribute parameters and push them into the circular buffer pool, and at the same time issue cabinet door unlocking instructions. The topology mapping module is used to extract the three-dimensional skeleton key point sequence of the human target, assign skeleton anchor point targets to the candidate voucher according to the payment modality attribute parameters in the candidate voucher state vector, and when physical occlusion occurs, backtrack the parent node through the skeleton directed kinematic graph constructed by the key point sequence, infer and calculate the three-dimensional coordinates of the virtual anchor point, and generate the inference confidence coefficient. The reverse verification module is used to extract the absolute Z-axis height of the physical shelf and the absolute timestamp of the physical retrieval event of the shelf weight mutation event, calculate the relative retrieval height index by combining the static envelope height of the human target, extract the Z-axis displacement of the key point sequence of the three-dimensional skeleton and compare it with the mechanical constraint conditions corresponding to the relative retrieval height index, and eliminate candidate voucher state vectors that do not meet the constraint conditions. The arbitration settlement module is used to calculate the kinematic correlation index of the remaining candidate voucher state vector in the annular buffer pool based on the spatial Euclidean distance attenuation term, the time difference attenuation term and the inference confidence coefficient, compare the maximum value with the preset security isolation threshold, issue a deduction request when the condition is met, and block the deduction instruction when the condition is not met. The reverse verification module includes a height index mapping subunit, which is used to extract the static envelope height of the human target and divide the absolute height of the physical shelf Z-axis by the static envelope height to calculate the dimensionless relative object retrieval height index. Based on the preset low-domain division threshold and high-domain division threshold, the relative object retrieval height index is divided into low-domain, medium-domain, or high-domain spatial states. The reverse verification module also includes a cross-domain displacement verification subunit, which is used to extract the skeleton root node from the three-dimensional skeleton key point sequence, calculate the two-dimensional Euclidean distance between the skeleton root node and the vending machine on the horizontal plane, and when the two-dimensional Euclidean distance is less than a preset operation radius threshold, extract the Z-axis coordinate difference of the skeleton root node at the time of item retrieval and the time of authentication as the Z-axis displacement. The mechanical constraints specifically include: when in a low-domain space state, the Z-axis displacement must be less than or equal to the negative value of a preset scalar threshold value for center of gravity sinking. When in a high-domain space state, the Z-axis displacement must be greater than or equal to a preset centroid stretching scalar threshold. The arbitration settlement module includes an index fusion calculation subunit, used to calculate the kinematic correlation index through the following logic: Obtain the three-dimensional absolute spatial distance between the skeleton anchor point target corresponding to the candidate credential and the geometric center point of the front outer edge of the physical shelf where a weight change occurs, and calculate the spatial Euclidean distance attenuation term in combination with the spatial attenuation constant; Calculate the absolute difference between the authentication timestamp and the absolute timestamp of the physical item retrieval event, and calculate the time difference decay term in combination with the time decay constant; The kinematic correlation index is obtained by multiplying the spatial Euclidean distance attenuation term by the spatial dimension weight term, multiplying the time difference attenuation term by the time dimension weight term, and multiplying the weighted sum of the two by the inference confidence coefficient.

2. The automated vending machine settlement system supporting multimodal payment according to claim 1, characterized in that, The credential buffer module includes a vector encapsulation subunit, which defines the candidate credential status vector as a matrix containing a user unique identifier, sensor modality type parameters, three-dimensional physical space absolute coordinate vector, and candidate credential absolute timestamp. The vector encapsulation subunit executes a dual queue update rule: calculates the difference between the absolute timestamp of the candidate voucher and the current system time, and removes the candidate voucher state vectors whose difference is greater than or equal to the preset time window length from the circular buffer pool; When the total number of valid vectors in the annular buffer pool reaches the maximum spatial length, the subsequent vector push is blocked.

3. The automated vending machine settlement system supporting multimodal payment according to claim 1, characterized in that, The topology mapping module includes a modal anchor mapping subunit, which is used to allocate the preferred skeleton anchor target for external visual tracking based on the payment modal attribute parameters in the candidate credential state vector. When the payment modality attribute parameter is a face recognition modality, the cervical spine node is specified as the preferred skeleton anchor point target; When the payment modality attribute parameter is a near-field communication modality or a palm vein recognition modality, the wrist node is specified as the preferred skeleton anchor point target.

4. The automated vending machine settlement system supporting multimodal payment according to claim 3, characterized in that, The topology mapping module also includes a kinematic compensation inference subunit. When the two-dimensional feature confidence of the preferred skeleton anchor point target is lower than the preset visibility threshold, it backtracks along the directed kinematic graph of the skeleton towards the root node to find the nearest effective parent node whose feature confidence is greater than or equal to the preset visibility threshold. The three-dimensional coordinates of the virtual anchor point are calculated by superimposing the three-dimensional coordinates of the nearest effective parent node, the unit direction vector calculated based on the motion trajectory of the historical frame buffer, the static envelope height, and the standard scaling constant. The inference confidence coefficient is generated by calculating the backtracking level depth using a linear decay formula.

5. An automated vending machine settlement system supporting multimodal payment according to claim 1, characterized in that, The arbitration settlement module includes a settlement boundary determination subunit, which is used to sort the kinematic correlation indices of all valid vouchers in descending order, obtain the maximum and second largest values ​​of the kinematic correlation indices, and calculate the margin of difference between the maximum and second largest values. The conditions for satisfying the security isolation threshold specifically include: the maximum value is greater than or equal to the security isolation threshold, and the difference margin is greater than or equal to the preset difference margin threshold.

6. The automated vending machine settlement system supporting multimodal payment according to claim 1, characterized in that, The credential buffer module includes a concurrent monitoring subunit. The concurrent monitoring subunit configures a network time protocol service at the system bus level to synchronize and align the internal hardware clock of the multimodal payment sensor array with the clock of the internal gravity sensing matrix, so that the absolute timestamp in the candidate credential state vector and the absolute timestamp of the physical retrieval event are on the same absolute time axis.

7. A method for settling accounts for automated vending machines supporting multimodal payments, applied to an automated vending machine settlement system supporting multimodal payments as described in any one of claims 1-6, characterized in that, Includes the following steps: S1, listen for concurrent payment authentication requests, instantiate a candidate voucher state vector containing multi-dimensional attribute parameters after authentication is completed, push the candidate voucher state vector into the circular buffer pool, and issue a door unlocking command to the vending machine execution mechanism. S2, receive the two-dimensional video stream and extract the three-dimensional skeleton key point sequence of the human target, assign skeleton anchor point targets to the candidate voucher according to the payment modal attribute parameters in the candidate voucher state vector, when the skeleton anchor point target is physically occluded, calculate the three-dimensional coordinates of the virtual anchor point by backtracking the parent node through the skeleton directed kinematic graph, and generate the inference confidence coefficient. S3, extract the absolute Z-axis height of the physical shelf where the weight mutation event occurs and the absolute timestamp of the physical retrieval event, calculate the relative retrieval height index based on the static envelope height of the human target and the absolute Z-axis height of the physical shelf, extract the Z-axis displacement of the key point sequence of the three-dimensional skeleton and compare it with the mechanical constraint conditions corresponding to the relative retrieval height index, and remove the candidate credential state vectors that do not meet the mechanical constraint conditions from the annular buffer pool. S4. Based on the spatial Euclidean distance attenuation term, the time difference attenuation term, and the inference confidence coefficient, calculate the kinematic correlation index of the remaining candidate voucher state vectors in the annular buffer pool, filter out the maximum value and compare it with the preset security isolation threshold. If the security isolation threshold is met, issue a deduction request and clear the annular buffer pool. If the security isolation threshold is not met, block the deduction instruction and generate an pending order status.

Images