Safety management system and method for self-service storage cabinet

Abstract

The invention relates to the field of self-service storage cabinet safety management, and discloses a self-service storage cabinet safety management system and method, and the method comprises the steps: collecting the appearance, geometry, material and initial posture of an article, estimating light field distribution, an inertia response threshold value and a perturbation sensitive factor through a lightweight model, and generating an initial state feature vector of the article; periodically recording the position, the posture and the light receiving change of the article to form a preliminary path timeline, and generating an optimized path timeline according to the security level of the article and the current calculation load; continuity verification and cross-modal offset analysis are carried out, abnormity is identified, and abnormity marks and path offset information are output; generating a disturbance domain in combination with the abnormal mark, the user authorization and the cabinet door state, and comparing the disturbance domain with the path offset to form an update path record; and performing short-term anomaly prediction by using a historical trend, and adjusting a final output strategy according to the security level and the calculated load. The method has the advantage of improving the safety management capability.

Description

Technical Field

[0001] This invention relates to the field of self-service storage locker security management, specifically a self-service storage locker security management system and method. Background Technology

[0002] Self-service storage locker systems are widely used for parcel storage and item safekeeping in public areas. Existing technologies mainly rely on RFID, QR codes, facial recognition, locker door detection, weight sensing, and video surveillance to prevent tampering and misplacement by identifying users and comparing the appearance of items. However, these methods have significant shortcomings. First, they only focus on appearance consistency and cannot identify high-concealment risks of items being replaced internally despite having the same appearance. For example, dangerous goods or empty substitutes may be placed in the original packaging, and it is impossible to determine whether the internal structure of the item has changed. Second, they assume that items remain stationary during storage and only compare them at the insertion and retrieval points. They cannot identify instances where items are taken away, replaced, or their contents are removed and then returned when the locker door is briefly opened. Third, they ignore changes in the relative positions and contact relationships between multiple items and cannot determine whether items have been moved, inserted into other items, disassembled and hidden, or rearranged. Existing systems have significant gaps in internal consistency verification, continuous monitoring of the storage process, and identification of changes in relationships between items, making it difficult to effectively prevent risks such as tampering, replacement, content theft, and the disguise of dangerous goods. Therefore, it is essential to design a security management system and method for self-service storage lockers to improve security management capabilities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a self-service storage locker security management system and method, which has the advantage of improving security management capabilities and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving security management capabilities, this invention provides the following technical solution: a self-service storage cabinet security management method, comprising the following steps:

[0005] The object's appearance, geometry, material, and initial posture are collected. The light field distribution, inertial response threshold, and perturbation sensitivity factor are estimated through a lightweight model. Combined with security level weighting and historical data, the object's initial state feature vector is generated.

[0006] Based on the initial state feature vector, the position, posture and light exposure changes of the object are recorded periodically to form a preliminary path timeline. Then, according to the object's safety level and the current computing load, the acquisition frequency, sensor priority and path accuracy are dynamically adjusted to generate an optimized path timeline.

[0007] Based on the optimized path timeline, continuity verification and cross-modal migration analysis are performed to identify anomalies and output anomaly markers and path migration information.

[0008] The disturbance domain is generated based on anomaly markers, user authorization, and cabinet door status. The matching strictness of the disturbance domain is adjusted by combining security level and dynamic modal weight. The strategy is adjusted with reference to the computational load and compared with the path offset to form an updated path record.

[0009] Based on the updated path records, the integrity of the entire path link is verified. Dynamic weighting is performed by combining cross-modal offset, disturbance matching and self-healing segments. Short-term anomaly prediction is performed using historical trends. The final output strategy is adjusted according to the security level and computational load.

[0010] Preferably, the process of generating the initial state feature vector of an item is as follows:

[0011] Multi-angle RGB and depth images of the object are acquired, and appearance texture features, volume contours and geometric boundary information are extracted through computer vision algorithms;

[0012] Based on the material type, density, and surface reflectivity, the inertial response threshold and perturbation sensitivity factor are calculated and stored as vectorized parameters.

[0013] A lightweight convolutional neural network is used to estimate the light field distribution on the surface of an object and in local space, and output a numerical representation.

[0014] The appearance, geometry, material, attitude, light field, inertia, and sensitivity factors are uniformly encoded to form an initial state feature vector.

[0015] Preferably, the process of forming a preliminary path timeline is as follows:

[0016] The monitoring module is initialized based on the initial state feature vector, and the three-dimensional spatial position of the item in the storage cabinet is collected periodically.

[0017] Simultaneously record the attitude angle and surface illumination status, and use RGB image analysis to calculate the illumination value;

[0018] The collected time-series data are sorted by timestamps to form a preliminary path timeline.

[0019] Preferably, the process of generating the optimized path timeline is as follows:

[0020] Based on the preliminary path timeline and the security level of the items, the collection priority is determined, and the sampling frequency of the path timeline of high-security items is increased accordingly.

[0021] By combining real-time computing load, the number of sensors and the acquisition cycle participating in the path timeline update are dynamically adjusted;

[0022] The spatial resolution and sampling interval of each recording point in the path timeline are dynamically optimized, and the optimized path timeline is output.

[0023] Preferably, the process of identifying anomalies and outputting anomaly markers and path offset information is as follows:

[0024] The continuity of data at adjacent time points in the optimized path timeline is verified by using a weighted difference method to detect abnormal changes in position, attitude, and lighting conditions, and to calculate a continuity score.

[0025] By using multimodal data for joint migration analysis, the deviations of each mode are fused according to weights to generate a comprehensive path offset vector;

[0026] An anomaly marker is generated when the overall path offset vector exceeds a preset threshold or the continuity score is lower than a safety threshold.

[0027] Preferably, the process of generating the disturbance domain based on anomaly markers, user authorization, and cabinet door status is as follows:

[0028] Based on anomaly markers, combined with cabinet door status and user authorization information, the three-dimensional spatial range in which items may move within the storage cabinet is determined.

[0029] Discretize the spatial range to generate a preliminary three-dimensional disturbance domain. Each grid point records the spatial coordinates, time node, and anomaly risk weight.

[0030] The anomaly risk weight is dynamically calculated based on the severity of the anomaly marker, the security level of the item, and the potential for disturbance, and a complete three-dimensional disturbance domain is generated through voxel segmentation.

[0031] Preferably, the process of adjusting the strategy based on the computing load is as follows:

[0032] Based on the three-dimensional perturbation domain, the matching threshold of the perturbation domain is adjusted in combination with the security level of the item, and items with higher security levels correspond to stricter matching standards;

[0033] Weights are assigned to different modal data in the perturbation domain, and weighted matching scores are calculated.

[0034] The matching strategy is dynamically adjusted based on the current computational load, including the priority of perturbation domain processing and computational accuracy.

[0035] The weighted matching score is compared with the threshold to generate a matching result set.

[0036] Preferably, the process of forming an updated path record is as follows:

[0037] Based on the matching result set, spatial intersection calculation is performed on the abnormal offset information recorded in the path timeline to identify the overlapping areas of items at each time node.

[0038] The offset vectors of the overlapping areas are weighted and combined with the perturbation domain matching score, abnormal offset magnitude, item safety level and historical offset trend to generate time series updated path records.

[0039] The offset changes of consecutive time nodes are smoothed and interpolated, and the abnormal states and matching confidence of each time node are marked. The final update path record is then output.

[0040] Preferably, the process of adjusting the final output strategy according to the security level and computational load is as follows:

[0041] Perform end-to-end integrity verification on the updated path records to detect missing data, outliers, and offset anomalies.

[0042] Dynamic weighting is performed based on cross-modal migration and disturbance matching results to output a comprehensive anomaly index;

[0043] Using historical path data and offset trends, short-term anomaly prediction is performed through time series models, generating predicted values ​​and updating path status.

[0044] The final output strategy is dynamically adjusted based on the security level of the items and the system's computational load.

[0045] A self-service storage locker security management system includes:

[0046] Feature acquisition module: Acquires the appearance, geometry, material and initial posture of the object, and estimates the light field distribution, inertial response threshold and perturbation sensitivity factor through a lightweight model to generate an initial state feature vector;

[0047] Path management module: Based on the initial state feature vector, it periodically records the position, posture and light exposure changes of items to form a preliminary path timeline, and generates an optimized path timeline by combining the safety level and computational load;

[0048] Anomaly detection module: Performs continuity verification and cross-modal offset analysis on the optimized path timeline, identifies anomalies, and outputs anomaly markers and path offset information;

[0049] Disturbance matching module: Combines anomaly markers, user authorization, and cabinet door status to generate a disturbance domain, and adjusts the matching strictness according to the security level and dynamic modal weights, while referring to the computational load adjustment strategy to form an update path record;

[0050] Strategy control module: Based on the updated path records, the integrity of the overall path link is verified. It combines cross-modal offset, disturbance matching and dynamic weighting of self-healing segments, and uses historical trends to predict short-term anomalies and adjust the final output strategy.

[0051] Compared with the prior art, the present invention provides a self-service storage locker security management system and method, which has the following beneficial effects:

[0052] This invention constructs a complete management chain from initial state acquisition, path timeline generation, anomaly identification to dynamic repair and prediction, enabling more refined, continuous, and reliable state monitoring of items throughout the storage process. Based on fundamental information such as appearance, geometry, material, and attitude, and combined with light field distribution, inertial threshold, and perturbation sensitivity factors, an initial feature vector is formed, providing a reliable benchmark for subsequent monitoring. Furthermore, by periodically recording changes in attitude, position, and light exposure, and considering the item's security level and real-time computational load, the acquisition frequency, sensor priority, and path accuracy are dynamically adjusted to effectively avoid monitoring degradation under high load conditions. Subsequently, continuous verification and cross-modal analysis are conducted based on the optimized path timeline. The dynamic offset analysis can detect anomalies and output offset information at an early stage. It further combines anomaly markers, user authorization, and cabinet door status to construct a disturbance domain. By strengthening the matching rigor through dynamic modal weights and security level constraints, it can achieve in-depth identification of suspicious behaviors. On this basis, it forms an updated path record, performs integrity verification and weighted repair on the overall link, and uses historical trends to carry out short-term anomaly prediction. This makes the path record not only correctable but also forward-looking. Through the above mechanisms, this method can still ensure the stability, accuracy, and timeliness of monitoring under the conditions of complex changes in multi-source data, different security level requirements, and diverse computing loads, thereby significantly improving the security management capabilities of the storage process and the overall reliability of the system. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the method of the present invention;

[0054] Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are 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.

[0056] Example 1: Please refer to Figure 1 As shown in the figure, a self-service storage locker security management method according to an embodiment of the present invention includes the following steps:

[0057] S1: Collect the appearance, geometry, material and initial posture of the object, estimate the light field distribution, inertial response threshold and perturbation sensitivity factor through a lightweight model, and generate the initial state feature vector of the object by combining security level weighting and historical data.

[0058] The process of generating the initial state feature vector of the item in S1 is as follows:

[0059] Multi-angle RGB and depth images of the object are acquired, and appearance texture features, volume contours and geometric boundary information are extracted through computer vision algorithms;

[0060] During the acquisition phase, the object is placed on a fixed acquisition platform. A combined acquisition device, including an RGB camera and a depth camera, is used to acquire multi-angle image sequences at a set rotation angle. The resolution of the acquired RGB images is maintained at 1280×720. The depth data is in the form of a synchronized depth frame matrix, with each frame containing distance values. After acquisition, edge detection and contour extraction algorithms are performed on each frame to generate the two-dimensional boundary curve of the object. Then, three-dimensional point cloud reconstruction is performed based on the depth data. A volume approximation model is obtained through voxelization. Texture information is extracted through a local feature algorithm to generate a descriptor subset. The geometric boundary is structured and stored using the minimum bounding volume, principal direction vector, and bounding box parameters of the point cloud.

[0061] Based on the material type, density, and surface reflectivity, the inertial response threshold and perturbation sensitivity factor are calculated and stored as vectorized parameters.

[0062] Material information is determined by mapping from a material database or comparing spectral reflectance, while parameters such as material density, hardness, and light absorption characteristics are recorded. The surface of the object is subjected to light source illumination tests, and the reflection intensity at different incident angles is recorded. A surface reflectivity model is fitted using the reflection intensity curve. The density, volume, and shape distribution are substituted into the inertial model to calculate the inertial response threshold. The perturbation sensitivity factor is calculated by applying a small perturbation vector to the point cloud model and recording the change in reconstruction error. The error is expressed as an Euclidean distance increment or surface offset, ultimately forming a set of numerical sensitivity factors associated with the material and structure.

[0063] A lightweight convolutional neural network is used to estimate the light field distribution on the surface of an object and in local space, and output a numerical representation.

[0064] In the light field estimation stage, the acquired multi-angle RGB images are used as input. After being normalized according to a fixed size, they are input into a lightweight convolutional neural network. The model structure consists of several convolutional layers, depthwise separable convolutional layers, and an upsampling module. The network performs spatial illumination estimation on the outer surface area of ​​the object. The output is a light intensity matrix corresponding to the input image and a local light source direction estimation vector. During training or inference, the model processes each viewpoint independently. Finally, the light field outputs from multiple viewpoints are combined according to a preset viewpoint index to form a light field parameter matrix with uniform resolution.

[0065] The appearance, geometry, material, attitude, light field, inertia, and sensitivity factors are uniformly encoded to form an initial state feature vector;

[0066] All features are numerically standardized or dimensionally compressed to compress high-dimensional texture descriptors into fixed-dimensional vectors. Geometric boundary parameters are encoded into structured vectors containing length, width, height, principal direction, and point cloud statistical features. Light field information is transformed into a set of average illumination feature vectors according to pixel block division. Inertia and sensitivity factors are kept as scalar or small-dimensional parameter sets. Finally, they are concatenated according to field order to generate a single initial state feature vector.

[0067] S2: Based on the initial state feature vector, the position, posture and light exposure changes of the object are recorded periodically to form a preliminary path timeline. The acquisition frequency, sensor priority and path accuracy are dynamically adjusted according to the object's safety level and the current computing load to generate an optimized path timeline.

[0068] The process of forming the initial path timeline in S2 is as follows:

[0069] The monitoring module is initialized based on the initial state feature vector, and the three-dimensional spatial position of the item in the storage cabinet is collected periodically.

[0070] Before forming the path timeline, the geometric parameters, attitude matrix, and light field reference data in the initial state feature vector are imported into the monitoring module as the initial reference. The monitoring module includes a depth camera, an RGB camera, and a coordinate analysis unit fixed inside the cabinet. The position reference is aligned with the cabinet coordinate system through the intrinsic parameter calibration matrix. The spatial positioning process is triggered according to the set acquisition cycle. The depth camera outputs the three-dimensional point cloud frame of the object. After spatial filtering and farthest point correction, the current three-dimensional position coordinates are calculated through centroid extraction and attitude fitting algorithms.

[0071] Simultaneously record the attitude angle and surface illumination status, and use RGB image analysis to calculate the illumination value;

[0072] PCA principal component analysis is performed on the current point cloud frame to extract the principal axis direction vector, and it is converted into three sets of angle values: pitch angle, roll angle, and yaw angle according to the cabinet coordinate system. The surface illumination state is calculated from the current image frame acquired by the RGB camera. After the image is grayscaled and local brightness is sampled, the average light intensity of the surface area is calculated according to a fixed grid to generate an illumination array containing the brightness values ​​of each grid.

[0073] The collected time-series data is sorted by timestamp to form a preliminary path timeline;

[0074] After obtaining time series entries such as position, attitude, and illumination, all records are sorted in ascending order of timestamp. The sorting uses a stable sorting algorithm to maintain the original relative order of multi-source data within the same period. The sorted sequence is stored in the form of a path event list. Each record contains the acquisition time, three-dimensional position coordinates, attitude angle group, and illumination array. Finally, the sequence is arranged in a linear structure to form a preliminary path timeline.

[0075] The process of generating the optimized path timeline in S2 is as follows:

[0076] Based on the preliminary path timeline and the security level of the items, the collection priority is determined, and the sampling frequency of the path timeline of high-security items is increased accordingly.

[0077] Based on the item identifiers in the preliminary path timeline, the corresponding safety level parameters are read from the item parameter table, for example, represented by numerical levels 1 to 5. The safety level is mapped to the sampling frequency adjustment coefficient, such as level 5 corresponding to coefficient 1.5, level 3 corresponding to coefficient 1.0, and level 1 corresponding to coefficient 0.7. The acquisition scheduling module performs periodic updates based on the current default sampling period and coefficients. For example, the periodic update for high-level items is 500ms ÷ 1.5, and the updated periodic value is recorded.

[0078] By combining real-time computing load, the number of sensors and the acquisition cycle participating in the path timeline update are dynamically adjusted;

[0079] The current CPU utilization, memory utilization, and sensor I / O queue length are obtained from the processor load monitoring module. Sensor scheduling strategies are invoked based on threshold rules. For example, when CPU utilization exceeds 80%, some non-critical sensors are paused, and only the depth camera and the main RGB camera are kept running. When utilization is below 40%, all sensors are resumed. The acquisition cycle is adjusted a second time according to the computing load. For example, when the CPU load is high, the cycle is extended to 1.2 times the original cycle to reduce real-time computing pressure.

[0080] The spatial resolution and sampling interval of each recording point in the path timeline are dynamically optimized, and the optimized path timeline is output.

[0081] The spatial rate of change is calculated for each recorded point in the initial path timeline by dividing the Euclidean distance between adjacent points by the time interval, and an array of rates of change is created. When the rate of change is high, a high-resolution sampling strategy is enabled for the segment, which calls the high-density point cloud mode of the depth camera and shortens the sampling interval. When the rate of change is low, a low-resolution strategy is adopted, which records only key nodes and samples them at longer time intervals. All updated recorded points are sorted by timestamp and written into a new path data structure to form the optimized path timeline.

[0082] S3: Based on the optimized path timeline, perform continuity verification and cross-modal offset analysis, identify anomalies, and output anomaly markers and path offset information.

[0083] The process of identifying anomalies and outputting anomaly markers and path offset information in S3 is as follows:

[0084] The continuity of data at adjacent time points in the optimized path timeline is verified by using a weighted difference method to detect abnormal changes in position, attitude, and lighting conditions, and to calculate a continuity score.

[0085] By using multimodal data for joint migration analysis, the deviations of each mode are fused according to weights to generate a comprehensive path offset vector;

[0086] An anomaly marker is generated when the overall path offset vector exceeds a preset threshold or the continuity score is lower than a safety threshold.

[0087] S4: Generate a disturbance domain based on anomaly markers, user authorization, and cabinet door status. Adjust the matching strictness of the disturbance domain by combining security level and dynamic modal weights. Adjust the strategy with reference to the computational load and compare it with the path offset to form an updated path record.

[0088] The process of generating the disturbance domain in S4 based on anomaly markers, user authorization, and cabinet door status is as follows:

[0089] Based on anomaly markers, combined with cabinet door status and user authorization information, the three-dimensional spatial range in which items may move within the storage cabinet is determined.

[0090] The module reads two adjacent records in the time stamp order from the optimized path timeline and extracts the three-dimensional position coordinates, attitude angles, and illumination values ​​respectively. The continuity verification module performs weighted difference calculation on the three types of data. The position difference is obtained by calculating the Euclidean distance between two points, the attitude difference is obtained by summing the absolute differences of the three-axis angles, and the illumination difference is obtained by calculating the difference in brightness values. The three differences are weighted and summed according to preset weights to generate a continuity difference value. The difference value is then mapped to the 0-1 interval according to the normalization rule as a continuity score.

[0091] Discretize the spatial range to generate a preliminary three-dimensional disturbance domain. Each grid point records the spatial coordinates, time node, and anomaly risk weight.

[0092] Multimodal data, including 3D position points, attitude angles, displacement of key points in depth point cloud, and illumination gradient, are read from the path timeline records. The offset of each mode is calculated separately according to its data format. For example, the offset of the position mode is obtained by Euclidean distance from the position at the previous moment, the offset of the attitude mode is obtained by summing the three-axis angle changes, the offset of the point cloud mode is calculated by key point matching to determine local deformation, and the offset of the illumination mode is obtained by the absolute value of the brightness gradient change. The offsets of each mode are weighted and summed according to the mode weights stored in the configuration table to generate a comprehensive path offset vector in the form of a 3D vector.

[0093] The anomaly risk weight is dynamically calculated based on the severity of the anomaly marker, the security level of the item, and the potential for disturbance, and a complete three-dimensional disturbance domain is generated through voxel segmentation.

[0094] The system reads the comprehensive path offset vector from the offset analysis cache and the continuity score at the current moment from the continuity record table. The magnitude of the offset vector is compared with a preset offset threshold. For example, if the offset threshold is 2cm, the offset status is set to offset overlimit when the magnitude is greater than the threshold. At the same time, the continuity score is compared with a safety threshold. For example, if the continuity threshold is set to 0.4, the continuity status is set to continuity anomaly when the score is lower than the value. When any condition is triggered, an anomaly marker is generated at the corresponding time point. The anomaly type, timestamp, and offset vector are written into the anomaly event table. Then, a three-dimensional perturbation domain is generated based on the offset of the offset vector in the x, y, and z directions, with the current reference point coordinates as the center.

[0095] The process of adjusting the strategy based on the reference computing load in S4 is as follows:

[0096] Based on the three-dimensional perturbation domain, the matching threshold of the perturbation domain is adjusted in combination with the security level of the item, and items with higher security levels correspond to stricter matching standards;

[0097] The system reads the 3D disturbance domain data structure corresponding to the current step from the abnormal event table and parses the minimum point coordinates, maximum point coordinates, and reference point coordinates. It reads the item safety level parameters from the item information table and maps the safety level to a numerical level. For example, it maps levels A, B, and C to 3, 2, and 1, respectively. Based on the safety level values, it calculates the matching threshold through a predefined threshold mapping table. For example, the matching threshold = basic threshold × 1 + safety level coefficient. The calculated matching threshold is written to the disturbance matching parameter area.

[0098] Weights are assigned to different modal data in the perturbation domain, and weighted matching scores are calculated.

[0099] Modal data, including 3D spatial offset, attitude angle change, illumination change, and other sensor sampling data, are read from the perturbation domain. Modal weight parameters, such as position data weight, attitude data weight, and illumination weight, are read from the modal weight table. The deviation values ​​of each mode are normalized to unify their numerical range to the 0–1 interval. Then, weighted calculations are performed sequentially: weighted offset = normalized offset value × position data weight, weighted attitude = normalized attitude value × attitude data weight, and weighted illumination = normalized illumination value × illumination weight. All weighted modal deviations are summed to obtain the weighted matching score.

[0100] The matching strategy is dynamically adjusted based on the current computational load, including the priority of perturbation domain processing and computational accuracy.

[0101] The system reads the current CPU utilization, memory utilization, and sensor data throughput from the system status monitoring table, and calculates the computational load value at the current moment. For example, when the CPU utilization exceeds 70% or the data throughput exceeds the preset limit, the load status is set to high load. The matching strategy is adjusted according to the load status. When in a high load state, the number of sampling points that need to be processed in the disturbance domain is reduced, the spatial resolution is reduced, and the computational accuracy of the modal fusion operation is reduced, such as reducing floating-point precision or adopting a simplified computation mode. When the load status is low load, the full resolution processing mode is restored, and the complete modal fusion process is enabled.

[0102] The weighted matching score is compared with a threshold to generate a matching result set;

[0103] The weighted matching score is read from the matching calculation cache, and the current matching threshold is read from the perturbation matching parameter area. The two are compared. When the weighted score is greater than the threshold, the matching status is set to "matching deviation exceeds the limit". When the weighted score is less than or equal to the threshold, the matching status is set to "matching normally". After the status judgment is completed, a matching result entry is created. The matching status, weighted score, threshold value and the timestamp of the current comparison time are written into the entry. The entry is added to the matching result set data table to form the matching result set at the current time.

[0104] The process of creating an update path record in S4 is as follows:

[0105] Based on the matching result set, spatial intersection calculation is performed on the abnormal offset information recorded in the path timeline to identify the overlapping areas of items at each time node.

[0106] The system reads the set of records to be processed from the matching result set table and the path timeline table. For each abnormal offset record, it extracts its timestamp, reference point coordinates, and offset vector. The 3D perturbation domain is discretized into a voxel grid. For each time node, the set of voxels occupied by the abnormal offset is marked on the voxel grid layer. The system performs a set intersection operation on the set of voxels in the perturbation domain marked as exceeding the matching deviation limit in the matching result set. The intersection operation uses a bitmap or sparse matrix representation to save memory and returns the index set of the intersection voxels. Each intersection voxel is mapped back to the 3D coordinate range. Adjacent voxels are merged to generate several overlapping subdomains. An entry is generated for each subdomain, which includes: timestamp, minimum / maximum coordinates of the subdomain, covering voxel index, and the corresponding set of original offset vectors. The entries are written to the overlapping region cache table.

[0107] The offset vectors of the overlapping areas are weighted and combined with the perturbation domain matching score, abnormal offset magnitude, item safety level and historical offset trend to generate time series updated path records.

[0108] The system reads each entry in the overlapping area cache table. For each subdomain, it reads the perturbation domain matching score from the matching result set, the abnormal offset amplitude from the path timeline, the security level value from the item information table, and calculates the historical offset trend factor from the historical offset record table. It then normalizes the above factors according to the weight parameters defined in the configuration file and calculates the comprehensive weight value. Subsequently, it iterates through the set of original offset vectors in the subdomain, multiplies each offset vector by the comprehensive weight, performs vector summation, and calculates the weighted average based on the total weight to obtain the synthetic offset vector of the subdomain. The synthetic offset vector is written into the record field of the corresponding time point in the path timeline, along with the participating factor value, normalized weight, matching score, and historical trend factor, generating time-series updated path record entries. The entries are then appended to the updated path record table in timestamp order to generate time-series updated path records.

[0109] The offset changes of consecutive time nodes are smoothed and interpolated, and the abnormal state and matching confidence of each time node are marked, and the final update path record is output.

[0110] The time series segments in the update path record table are segmented by timestamp. First, the time series is resampled to a uniform time interval. Linear interpolation is used to fill in the missing position and attitude data at time points. For the resampled numerical series, the system applies a sliding window to smooth the position and attitude series. For segments that require higher continuity, cubic spline interpolation can be performed on the position series to generate a continuous curve and output intermediate sampling points. Then, the matching confidence is calculated based on the matching score, the magnitude of the synthetic offset vector, and the variance of the historical residuals. The smoothed position / attitude, interpolation identifier, synthetic offset vector, matching score, and calculated matching confidence are written into the update record at each time node. The final set of these records is used as the final update path record.

[0111] S5: Based on the updated path records, the integrity of the entire path link is checked. It combines cross-modal offset, disturbance matching and self-healing segments for dynamic weighting, uses historical trends to predict short-term anomalies, and adjusts the final output strategy according to the security level and computational load.

[0112] The process of adjusting the final output strategy based on the security level and computational load in S5 is as follows:

[0113] Perform end-to-end integrity verification on the updated path records to detect missing data, outliers, and offset anomalies.

[0114] The system reads each record in the updated path record table, checks the sequence continuity using timestamps as indexes, calculates linear interpolation from adjacent records for missing time nodes and fills in the missing entries, performs threshold detection on the offset vector of each record, uses the offset limit table in the configuration file to determine whether the entry exceeds the preset range, writes the corresponding identifier in the anomaly mark field for records that exceed the limit, compares the numerical range of the original position field and the synthetic offset vector field to check for unresolved values, and if an anomaly is found, adds a data anomaly label next to the record and writes it to the verification log file. After completing all verifications, the processing results are updated back to the path record table.

[0115] Dynamic weighting is performed based on cross-modal migration and disturbance matching results to output a comprehensive anomaly index;

[0116] The multimodal migration vector list is read from the cross-modal migration analysis module, and the matching score for each time node is read from the perturbation matching module. For each node, the three fields of offset, rate of change of offset direction and perturbation matching score are normalized according to the weighting rules in the configuration file, and then weighted and accumulated according to the corresponding weights to calculate the comprehensive anomaly value. If the node contains multiple modal data, a weighted average is performed on the comprehensive anomaly values ​​generated by each modality.

[0117] Using historical path data and offset trends, short-term anomaly prediction is performed through time series models, generating predicted values ​​and updating path status.

[0118] Extract data segments of continuous time windows from the historical path record table, read the offset vector, offset direction change and comprehensive anomaly index of each node, and model the data using the time series model specified in the configuration file. The model first performs difference or smoothing on the sequence, and then trains and predicts the predicted values ​​of the offset vector and comprehensive anomaly index of the next time window. Write the predicted values ​​into the predicted offset field and the predicted anomaly field, and mark the predicted status of the time node in the path status table.

[0119] The final output strategy is dynamically adjusted based on the security level of the items and the computational load of the system.

[0120] The system reads the security level value corresponding to the current item from the item information table, and reads the real-time utilization of CPU, memory and I / O from the resource monitoring module to form the system load value. It then looks up the output policy corresponding to the security level-system load combination in the policy configuration table, such as output frequency, whether to attach prediction data, and whether to generate a detailed offset report. Based on the policy, it operates on the final output content. If the policy requires simplified output, it deletes non-critical fields and retains only the core offset indicators. If a strict mode is required, it adds complete fields and historical prediction chains. After processing, the final output is written to the final output record table and pushed to the specified interface according to the policy requirements.

[0121] Example 2: As Figure 2 As shown, a self-service storage locker security management system includes:

[0122] Feature acquisition module: Acquires the appearance, geometry, material and initial posture of the object, and estimates the light field distribution, inertial response threshold and perturbation sensitivity factor through a lightweight model to generate an initial state feature vector;

[0123] Path management module: Based on the initial state feature vector, it periodically records the position, posture and light exposure changes of items to form a preliminary path timeline, and generates an optimized path timeline by combining the safety level and computational load;

[0124] Anomaly detection module: Performs continuity verification and cross-modal offset analysis on the optimized path timeline, identifies anomalies, and outputs anomaly markers and path offset information;

[0125] Disturbance matching module: Combines anomaly markers, user authorization, and cabinet door status to generate a disturbance domain, and adjusts the matching strictness according to the security level and dynamic modal weights, while referring to the computational load adjustment strategy to form an update path record;

[0126] Strategy control module: Based on the updated path records, the integrity of the overall path link is verified. It combines cross-modal offset, disturbance matching and dynamic weighting of self-healing segments, and uses historical trends to predict short-term anomalies and adjust the final output strategy.

[0127] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0128] 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. A method for secure management of self-service storage lockers, characterized in that, Includes the following steps: The object's appearance, geometry, material, and initial posture are collected. The light field distribution, inertial response threshold, and perturbation sensitivity factor are estimated through a lightweight model. Combined with security level weighting and historical data, the object's initial state feature vector is generated. Based on the initial state feature vector, the position, posture and light exposure changes of the object are recorded periodically to form a preliminary path timeline. Then, according to the object's safety level and the current computing load, the acquisition frequency, sensor priority and path accuracy are dynamically adjusted to generate an optimized path timeline. Based on the optimized path timeline, continuity verification and cross-modal migration analysis are performed to identify anomalies and output anomaly markers and path migration information. The disturbance domain is generated based on anomaly markers, user authorization, and cabinet door status. The matching strictness of the disturbance domain is adjusted by combining security level and dynamic modal weight. The strategy is adjusted with reference to the computational load and compared with the path offset to form an updated path record. Based on the updated path records, the integrity of the entire path link is verified. Dynamic weighting is performed by combining cross-modal offset, disturbance matching and self-healing segments. Short-term anomaly prediction is performed using historical trends. The final output strategy is adjusted according to the security level and computational load.

2. The self-service storage locker security management method according to claim 1, characterized in that, The process of generating the initial state feature vector of an item is as follows: Multi-angle RGB and depth images of the object are acquired, and appearance texture features, volume contours and geometric boundary information are extracted through computer vision algorithms; Based on the material type, density, and surface reflectivity, the inertial response threshold and perturbation sensitivity factor are calculated and stored as vectorized parameters. A lightweight convolutional neural network is used to estimate the light field distribution on the surface of an object and in local space, and output a numerical representation. The appearance, geometry, material, attitude, light field, inertia, and sensitivity factors are uniformly encoded to form an initial state feature vector.

3. The self-service storage locker security management method according to claim 2, characterized in that, The process of forming the initial path timeline is as follows: The monitoring module is initialized based on the initial state feature vector, and the three-dimensional spatial position of the item in the storage cabinet is collected periodically. Simultaneously record the attitude angle and surface illumination status, and use RGB image analysis to calculate the illumination value; The collected time-series data are sorted by timestamps to form a preliminary path timeline.

4. The self-service storage locker security management method according to claim 3, characterized in that, The process of generating the optimized path timeline is as follows: Based on the preliminary path timeline and the security level of the items, the collection priority is determined, and the sampling frequency of the path timeline of high-security items is increased accordingly. By combining real-time computing load, the number of sensors and the acquisition cycle participating in the path timeline update are dynamically adjusted; The spatial resolution and sampling interval of each recording point in the path timeline are dynamically optimized, and the optimized path timeline is output.

5. A self-service storage locker security management method according to claim 4, characterized in that, The process of identifying anomalies and outputting anomaly markers and path offset information is as follows: The continuity of data at adjacent time points in the optimized path timeline is verified by using a weighted difference method to detect abnormal changes in position, attitude, and lighting conditions, and to calculate a continuity score. By using multimodal data for joint migration analysis, the deviations of each mode are fused according to weights to generate a comprehensive path offset vector; An anomaly marker is generated when the overall path offset vector exceeds a preset threshold or the continuity score is lower than a safety threshold.

6. The self-service storage locker security management method according to claim 5, characterized in that, The process of generating a disturbance domain based on anomaly markers, user authorization, and cabinet door status is as follows: Based on anomaly markers, combined with cabinet door status and user authorization information, the three-dimensional spatial range in which items may move within the storage cabinet is determined. Discretize the spatial range to generate a preliminary three-dimensional disturbance domain. Each grid point records the spatial coordinates, time node, and anomaly risk weight. The anomaly risk weight is dynamically calculated based on the severity of the anomaly marker, the security level of the item, and the potential for disturbance, and a complete three-dimensional disturbance domain is generated through voxel segmentation.

7. A self-service storage locker security management method according to claim 6, characterized in that, The process of adjusting strategies based on computational load is as follows: Based on the three-dimensional perturbation domain, the matching threshold of the perturbation domain is adjusted in combination with the security level of the item, and items with higher security levels correspond to stricter matching standards; Weights are assigned to different modal data in the perturbation domain, and weighted matching scores are calculated. The matching strategy is dynamically adjusted based on the current computational load, including the priority of perturbation domain processing and computational accuracy. The weighted matching score is compared with the threshold to generate a matching result set.

8. A self-service storage locker security management method according to claim 7, characterized in that, The process of creating an updated path record is as follows: Based on the matching result set, spatial intersection calculation is performed on the abnormal offset information recorded in the path timeline to identify the overlapping areas of items at each time node. The offset vectors of the overlapping areas are weighted and combined with the perturbation domain matching score, abnormal offset magnitude, item safety level and historical offset trend to generate time series updated path records. The offset changes of consecutive time nodes are smoothed and interpolated, and the abnormal state and matching confidence of each time node are marked, and the final update path record is output.

9. A self-service storage locker security management method according to claim 8, characterized in that, The process of adjusting the final output strategy based on the security level and computational load is as follows: Perform end-to-end integrity verification on the updated path records to detect missing data, outliers, and offset anomalies. Dynamic weighting is performed based on cross-modal migration and disturbance matching results to output a comprehensive anomaly index; Using historical path data and offset trends, short-term anomaly prediction is performed through time series models, generating predicted values ​​and updating path status. The final output strategy is dynamically adjusted based on the security level of the items and the system's computational load.

10. A self-service storage locker security management system, applied to the method described in any one of claims 1-9, characterized in that, include: Feature acquisition module: Acquires the appearance, geometry, material and initial posture of the object, and estimates the light field distribution, inertial response threshold and perturbation sensitivity factor through a lightweight model to generate an initial state feature vector; Path management module: Based on the initial state feature vector, it periodically records the position, posture and light exposure changes of items to form a preliminary path timeline, and generates an optimized path timeline by combining the safety level and computational load; Anomaly detection module: Performs continuity verification and cross-modal offset analysis on the optimized path timeline, identifies anomalies, and outputs anomaly markers and path offset information; Disturbance matching module: Combines anomaly markers, user authorization, and cabinet door status to generate a disturbance domain, and adjusts the matching strictness according to the security level and dynamic modal weights, while referring to the computational load adjustment strategy to form an update path record; Strategy control module: Based on the updated path records, the integrity of the overall path link is verified. It combines cross-modal offset, disturbance matching and dynamic weighting of self-healing segments, and uses historical trends to predict short-term anomalies and adjust the final output strategy.

Images