CN-122023260-A - Online safety recognition system and method for tobacco overhead warehouse based on image recognition

CN122023260ACN 122023260 ACN122023260 ACN 122023260ACN-122023260-A

Abstract

The invention discloses an online safety recognition system and method for a tobacco overhead warehouse based on image recognition, wherein the online safety recognition system comprises the steps of generating a second tray three-dimensional observation data set, obtaining a tray multi-view feature set, outputting a tray multi-view dense feature tensor, recording an optimized improved condition implicit nerve shape representation model, generating a tray three-dimensional reconstruction model in an reasoning stage, carrying out geometric registration on the tray three-dimensional reconstruction model and a preset tray reference model to obtain a registration difference model, calculating tray deformation parameters based on the registration difference model, comparing the tray deformation parameters with a preset tobacco logistics safety threshold to obtain a deformation judgment result, and generating a warehouse control system instruction according to the deformation judgment result. The invention can be suitable for trays with different sizes, categories and batches, and can be used for dynamic geometric registration and difference analysis with a standard tray reference model.

Inventors

LI WEI
CAO SISI
LIU YONG
PENG YUNJUN
Zhong Xiangqiong
ZHONG LIANG
HE WEIPING
HE JIAN
ZOU XIAOCHUN
DONG HUIMIN
LI XUEYONG

Assignees

湖南省烟草公司衡阳市公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. An online safety identification method of a tobacco overhead warehouse based on image identification is characterized by comprising the following steps: collecting a first tray three-dimensional observation data set and preprocessing the first tray three-dimensional observation data set to generate a second tray three-dimensional observation data set; Dividing the second tray three-dimensional observation data set into multi-view sparse observation subsets according to the acquisition view angles, and generating corresponding view angle feature vectors for each sparse observation subset to obtain a tray multi-view angle feature set; inputting the multi-view feature set of the tray into a view fusion network, and outputting a multi-view dense feature tensor of the tray; Constructing a tray condition vector, constructing an improved condition implicit nerve shape representation model in combination with a tray multi-view dense feature tensor, inputting the tray multi-view dense feature tensor into the improved condition implicit nerve shape representation model in a training stage to obtain a tray three-dimensional implicit field, applying structural geometric consistency constraint to the improved condition implicit nerve shape representation model, optimizing parameters of the improved condition implicit nerve shape representation model according to the structural geometric consistency constraint, and recording the optimized improved condition implicit nerve shape representation model; in the reasoning stage, inputting a new second tray three-dimensional observation data set into the optimized improved condition implicit nerve shape representation model, generating a corresponding tray three-dimensional implicit field, extracting a zero equivalent surface from the tray three-dimensional implicit field, and generating a tray three-dimensional reconstruction model; Performing geometric registration on the tray three-dimensional reconstruction model and a preset tray reference model to obtain a registration difference model, and calculating tray deformation parameters based on the registration difference model; and comparing the deformation parameters of the tray with preset tobacco logistics safety thresholds to obtain deformation judgment results, and generating a storage control system instruction according to the deformation judgment results to complete online safety identification of the tobacco overhead warehouse.
2. The method for online safety recognition of a tobacco overhead library based on image recognition according to claim 1, wherein the dividing the second tray three-dimensional observation data set into multi-view sparse observation subsets according to the acquisition view angles, generating a corresponding view angle feature vector for each sparse observation subset, comprises: for each tray three-dimensional observation data unit in the second tray three-dimensional observation data set, associating corresponding acquisition view angle identifiers thereof, and constructing an acquisition view angle identifier set; dividing the second tray three-dimensional observation data set according to the acquisition view angle identification set to obtain a plurality of tray multi-view angle sparse observation subsets corresponding to different acquisition view angles; Carrying out local geometric alignment on the tray multi-view sparse observation subset and a preset tray reference model under the same coordinate system, calculating the spatial deviation distribution of the tray multi-view sparse observation subset relative to the preset tray reference model according to an alignment result, and generating a view deviation feature vector corresponding to an acquisition view based on the spatial deviation distribution; And collecting view angle deviation feature vectors corresponding to all the collected view angles to construct a multi-view angle feature set of the tray.
3. The method for online safety identification of a tobacco overhead warehouse based on image identification according to claim 1, wherein the inputting the tray multi-view feature set into the view fusion network comprises: numbering and organizing the multi-view angle feature set of the tray to construct a view angle input sequence; Inputting the view angle input sequence into a view angle feature coding layer of a view angle fusion network, and performing feature coding on each view angle deviation feature vector in the view angle input sequence to obtain an intermediate view angle coding feature set; Based on the intermediate view coding feature set, respectively calculating a visual feature score and a geometric consistency score for each acquisition view, adding the visual feature score and the geometric consistency score to obtain a fusion score of the acquisition view, and based on the fusion scores of all the acquisition views, calculating a view weight coefficient corresponding to each acquisition view by adopting a normalization method; weighting and summing the view weight coefficients of all the acquired views to the corresponding intermediate view coding feature vectors to obtain tray multi-view fusion feature vectors; Inputting the multi-view fusion feature vector of the tray into a feature reconstruction layer of the view fusion network, and generating a multi-view dense feature tensor of the tray through feature mapping and dimension rearrangement operation.
4. The method for online safety recognition of a tobacco overhead library based on image recognition according to claim 1, wherein the constructing an improved conditional implicit neural shape representation model comprises: extracting tray size information, tray category information and tray batch information corresponding to a current tray to be detected, and uniformly coding to generate a tray condition vector; Combining the tray multi-view dense feature tensor and the tray condition vector at a feature splicing layer to construct an improved condition implicit nerve shape representation model; In the training stage, inputting the tray multi-view dense feature tensor into an improved condition implicit neural shape representation model to obtain a tray three-dimensional implicit field, and constructing an implicit field data consistency constraint item based on a second tray three-dimensional observation data set; Applying structural geometric consistency constraint to the improved implicit neural shape representation model, and constructing structural geometric consistency constraint items according to the structural geometric consistency constraint, wherein the geometric consistency constraint items are obtained by weighting a planeness constraint item, an orthogonality constraint item and a symmetry constraint item; And adding the implicit field data consistency constraint term and the structural geometric consistency constraint term, and optimizing model parameters of the improved condition implicit neural shape representation model as a training objective function to obtain the optimized improved condition implicit neural shape representation model.
5. The method for online safety identification of tobacco overhead warehouse based on image identification according to claim 4, wherein the calculation of the planarity constraint term comprises selecting a sampling point set positioned in an upper surface area of a tray in a three-dimensional implicit field of the tray to obtain an upper surface sampling point set, calculating an upper surface fitting plane which is most matched with the upper surface sampling point set by adopting a plane fitting method based on three-dimensional space coordinates of the upper surface sampling point set, determining an upper surface normal vector and an upper surface offset by the upper surface fitting plane, calculating distances from each sampling point in the upper surface sampling point set to the upper surface fitting plane based on the upper surface normal vector, the upper surface offset and the upper surface sampling point set, and averaging square values of all the distances; The calculation of the orthogonality constraint term comprises the steps of carrying out square operation on an inner product between an upper surface normal vector and a lateral structure normal vector; The symmetry constraint item is calculated by extracting symmetry sampling points from a tray three-dimensional implicit field according to preset symmetry plane parameters, determining the position of each symmetry sampling point under a three-dimensional space coordinate system, mapping the symmetry sampling points to symmetry point positions with respect to a symmetry plane by a symmetry transformation operator for each symmetry sampling point, respectively obtaining implicit field predicted values of an original symmetry sampling point and a corresponding symmetry mapping point on the tray three-dimensional implicit field, performing difference to obtain symmetry errors, squaring the symmetry errors of all the symmetry sampling points, and averaging after summation.
6. The method for online safety identification of a tobacco overhead warehouse based on image identification according to claim 1, wherein the step of extracting a zero isosurface from a tray three-dimensional implicit field to generate a tray three-dimensional reconstruction model comprises the following steps: In the reasoning stage, a new second tray three-dimensional observation data set is collected, and a new tray multi-view dense characteristic tensor is generated; Uniformly coding the tray size vector, the tray category identification and the tray batch identification to generate a new tray condition vector; Inputting the new multi-view dense feature tensor of the tray and the new condition vector of the tray into an optimized implicit nerve shape representation model of the improved condition to generate a corresponding three-dimensional implicit field of the tray; Constructing an inference sampling coordinate set under a unified space reference coordinate system, and calculating a predicted value of the three-dimensional implicit field of the tray on the inference sampling coordinate set to obtain a discrete sampling field of the three-dimensional implicit field of the tray; and performing zero equivalent surface extraction according to the discrete sampling field to generate a tray three-dimensional reconstruction model.
7. The method for online safety identification of tobacco overhead warehouse based on image identification according to claim 1, wherein the performing geometric registration of the three-dimensional reconstruction model of the tray with a preset reference model of the tray comprises: Performing geometric registration on the tray three-dimensional reconstruction model and a preset tray reference model to obtain optimal space transformation parameters, and transforming the preset tray reference model to a space position aligned with the tray three-dimensional reconstruction model according to the optimal space transformation parameters to obtain a registered preset tray reference model; calculating a space difference vector between the three-dimensional space coordinates of the current to-be-detected tray sampling point and the three-dimensional space coordinates of the registered standard tray sampling point according to each pair of corresponding sampling points between the three-dimensional reconstruction model of the tray and the registered preset tray reference model, and corresponding all the sampling points and the space difference vectors thereof to construct a registration difference model; based on the registration difference model, obtaining the central deflection of the tray through the average value of the projection lengths of the space difference vectors at all sampling points of the central area of the tray in the normal vector direction of the upper surface; Based on the registration difference model, obtaining the edge warping degree of the tray through the maximum value of the absolute value of the projection length of the spatial difference vector at all sampling points of the edge region of the tray in the direction of the upper surface normal vector; Constructing a neighborhood relation set based on a registration difference model, setting a fracture judgment threshold value, counting the occurrence times of a certain pair of adjacent sampling points when the space difference mutation value between the adjacent sampling points is larger than the fracture judgment threshold value, and taking the ratio of the number of all the super-threshold mutation point pairs to the total number of the neighborhood relation set as the structural fracture degree of the tray; and constructing the deformation parameters of the tray through the central deflection of the tray, the edge warpage of the tray and the structural fracture of the tray.
8. The method for online safety identification of a tobacco overhead warehouse based on image identification according to claim 1, wherein the obtaining the deformation determination result comprises: Comparing the deformation parameters of the tray with preset tobacco logistics safety thresholds item by item to generate a central area state of the tray, an edge area state of the tray and a continuity state of the overall structure of the tray; generating a tray deformation judgment result based on the central area state of the tray, the edge area state of the tray and the overall structure continuity state of the tray; and generating a warehouse control system instruction comprising a normal release instruction, a maintenance shunt instruction and a scrapping rejection instruction according to the tray deformation judging result.
9. The method for online safety identification of tobacco overhead warehouse based on image identification according to claim 8, wherein the tray deformation judgment result comprises a qualified state, a maintenance required state and a scrapped state; When the central area state of the tray, the edge area state of the tray and the continuity state of the overall structure of the tray are all safe states, judging that the deformation judgment result of the tray is a qualified state; When at least one tray deformation parameter exceeds a corresponding safety threshold and does not reach a scrapping judgment condition, judging that the tray deformation judgment result is in a maintenance-required state; When the structural fracture degree of the tray exceeds the structural fracture safety threshold or a plurality of tray deformation parameters simultaneously exceed the corresponding safety threshold, judging that the tray deformation judgment result is in a scrapped state.
10. An online safety recognition system for a tobacco overhead warehouse, for performing the online safety recognition method for the tobacco overhead warehouse according to any one of claims 1 to 9, comprising: The data acquisition module is used for acquiring a first tray three-dimensional observation data set of the tray and preprocessing the data set to generate a second tray three-dimensional observation data set; the multi-view feature construction module is used for dividing the second tray three-dimensional observation data set into multi-view sparse observation subsets according to the acquisition view angles to obtain a tray multi-view feature set; The view fusion module is used for inputting the multi-view feature set of the tray into the view fusion network and outputting a multi-view dense feature tensor of the tray; the condition implicit modeling and training module is used for forming a tray condition vector, constructing an improved condition implicit nerve shape representation model in combination with the tray multi-view dense feature tensor and training to obtain an optimized improved condition implicit nerve shape representation model; The implicit reasoning and reconstruction module is used for inputting a new second tray three-dimensional observation data set into the optimized improved condition implicit neural shape representation model in a reasoning stage to generate a tray three-dimensional reconstruction model; The deformation analysis module is used for performing geometric registration on the tray three-dimensional reconstruction model and a preset tray reference model to obtain a registration difference model and calculating tray deformation parameters; the judging and controlling module is used for comparing the tray deformation parameters with preset tobacco logistics safety thresholds, generating tray deformation judging results and outputting corresponding warehouse control system instructions according to the tray deformation judging results.

Description

Online safety recognition system and method for tobacco overhead warehouse based on image recognition Technical Field The invention relates to the technical field of tobacco overhead libraries, in particular to an online safety recognition system and method for a tobacco overhead library based on image recognition. Background Along with the continuous promotion of intelligent storage and commodity circulation automation level, the overhead three-dimensional warehouse of tobacco trade extensively adopts the tray to carry out the storage and the circulation of goods as basic carrier, and the tray takes place deformation easily under long-term high frequency use and complex environment, including local deflection, edge warpage or structure fracture, deformation not only influences the automation operation efficiency of warehouse system, still probably brings the stack potential safety hazard. The existing tray deformation detection technology mainly comprises two kinds of methods based on two-dimensional image processing and three-dimensional point cloud analysis, wherein the two-dimensional image processing method mainly relies on edge detection and profile extraction traditional visual algorithms, is greatly influenced by environmental illumination change, reflection interference and background, cannot accurately capture three-dimensional structure information and micro deformation characteristics of a tray, has the problem of insufficient detection precision, and adopts a laser scanning or depth camera to collect tray surface point cloud data in the three-dimensional point cloud analysis method, and performs deformation judgment in a point cloud registration, voxel modeling or grid reconstruction mode. However, due to dense shelves, cargo shielding and limited sensor acquisition view angles in the tobacco overhead warehouse environment, serious sparseness and local deletion of point cloud data are often caused, and based on the traditional point cloud registration or grid reconstruction method, surface reconstruction errors are easy to generate when the data are incomplete or the sampling quality is poor, real deformation and acquisition noise cannot be effectively distinguished, and even misjudgment or omission occurs. The current data driving model based on deep learning has certain feature extraction and morphological reduction capability, but generally lacks prior constraint on the physical structure of a tray, faces special interference factors in a tobacco warehouse, and is easily influenced by pseudo deformation and abnormal noise problems in training and reasoning stages by a model purely relying on data fitting, so that stability and reliability of detection results are difficult to guarantee, meanwhile, the adaptability and generalization capability of the existing method for different types of trays are limited, and high-precision and high-robustness deformation detection is difficult to realize in industrial sites where multiple types and batches of trays are mixed. Disclosure of Invention The invention aims to provide an online safety recognition system and method for a tobacco overhead warehouse based on image recognition, which can adapt to trays of different sizes, categories and batches, and is used for dynamic geometric registration and difference analysis with a standard tray reference model. According to the embodiment of the invention, the online safety identification method for the tobacco overhead warehouse based on image identification comprises the following steps: collecting a first tray three-dimensional observation data set and preprocessing the first tray three-dimensional observation data set to generate a second tray three-dimensional observation data set; Dividing the second tray three-dimensional observation data set into multi-view sparse observation subsets according to the acquisition view angles, and generating corresponding view angle feature vectors for each sparse observation subset to obtain a tray multi-view angle feature set; inputting the multi-view feature set of the tray into a view fusion network, and outputting a multi-view dense feature tensor of the tray; Constructing a tray condition vector, constructing an improved condition implicit nerve shape representation model in combination with a tray multi-view dense feature tensor, inputting the tray multi-view dense feature tensor into the improved condition implicit nerve shape representation model in a training stage to obtain a tray three-dimensional implicit field, applying structural geometric consistency constraint to the improved condition implicit nerve shape representation model, optimizing parameters of the improved condition implicit nerve shape representation model according to the structural geometric consistency constraint, and recording the optimized improved condition implicit nerve shape representation model; in the reasoning stage, inputting a new second tray three-dimensional observation data