A process image recognition method based on a YOLO model

Abstract

The application discloses a process image recognition method based on a YOLO model, relates to the technical field of process image recognition, and is used for solving the problems of redundant detection range and low response efficiency. After the warehousing of goods is completed, the goods are marked and compared with a target warehousing quantity, whether the warehousing quantity is abnormal is judged, if the warehousing quantity is abnormal, the stopping frequency of a transport vehicle at each goods shelf is acquired, the reuse frequency of the goods shelf is calculated, the height data of the goods are collected, the stacking state is calculated, the goods shelf review index is generated by fusion, and whether the visual recognition mechanism is entered is judged. In the recognition mechanism, the detection area is determined based on the marked goods shelf, the image is input into the YOLO model for recognition, the data of the goods location occupation and the channel gap are output, the area correction coefficient is calculated and the detection area is adjusted, the YOLO model is input again for accurate recognition, and the final result is generated, so that the layered detection and review of the abnormal goods shelf are realized, the high-frequency global call of YOLO is avoided, and the real-time performance and accuracy of recognition are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of process image recognition, and more particularly, to a process image recognition method based on a YOLO model. BACKGROUND

[0002] With the wide application of intelligent warehousing and logistics systems, automatic identification and dynamic monitoring of warehouse goods have gradually become the key link to improve warehouse efficiency and reduce labor costs. Currently, the identification of warehouse goods mainly relies on RFID, barcode scanning and image-based target detection technology. Among them, the image recognition method based on deep learning is widely used due to its good feature extraction and generalization ability. The YOLO series model, as a kind of efficient end-to-end target detection algorithm, can realize the positioning and classification of multiple targets in an image in a single forward propagation, has high real-time detection performance, and is suitable for goods and shelf identification tasks in dynamic warehouse environment.

[0003] The prior art has the following disadvantages:

[0004] Currently, traditional YOLO recognition usually performs global scanning on the entire warehouse area. When the warehousing exception occurs only in part of the shelves or storage locations, full-field detection still needs to be performed, causing waste of computing resources and increase of identification time delay. There is a lack of adaptive triggering and region screening mechanism based on warehousing exception, which leads to the inability of the identification process to accurately locate the abnormal area, resulting in redundant detection range and low response efficiency. Therefore, a process image recognition method based on a YOLO model is proposed.

[0005] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a process image recognition method based on a YOLO model, which uses an abnormal trigger driven layered detection and region adaptive correction mechanism to solve the problems raised in the above background.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical scheme, a process image recognition method based on a YOLO model, comprising the following steps:

[0008] Step S1: After the goods are warehoused, the goods are marked, the current warehousing quantity of the marked goods is obtained from the inventory database, and the target warehousing quantity of the marked goods is compared with the current warehousing quantity to determine whether the warehousing quantity of the marked goods is abnormal.

[0009] Step S2: When the storage quantity of the marked goods is abnormal, the stop frequency of the transport vehicle at the same shelf is obtained through the vehicle scheduling database, and the shelf reuse frequency is calculated. The height difference reference of the goods height of each shelf is detected, and the stacking state of the shelf is analyzed. The shelf is marked in combination with the shelf reuse frequency, and whether to enter the visual identification mechanism is judged.

[0010] Step S3: In the visual identification mechanism, a detection area is set based on the marked shelf, and a YOLO model is input. The YOLO model is used to identify the goods location occupancy data and the channel gap data of the detection area.

[0011] Step S4: Based on the goods location occupancy data, the occupancy distribution state is analyzed. The occupancy distribution state and the channel gap data are fused to generate a region correction coefficient, and the detection area is adjusted. The YOLO model is used to identify the adjusted detection area and generate an identification result.

[0012] In a preferred embodiment, in step S1, when the goods are stored in the warehouse, the warehouse management unit generates a corresponding storage number for the goods according to the storage order of the goods. After the goods are stored in the warehouse, the goods are marked based on the storage number of each goods;

[0013] Access the current storage quantity of the marked goods in the inventory database, and call the target storage quantity of the marked goods through the storage plan unit;

[0014] If the current storage quantity is inconsistent with the target storage quantity, it is judged that the storage quantity of the marked goods is abnormal;

[0015] Otherwise, it is judged that the storage quantity of the marked goods is normal.

[0016] In a preferred embodiment, in step S2, when the storage quantity of the marked goods is abnormal, the shelf stop record of the transport vehicle in the storage process of the marked goods is accessed through the vehicle scheduling database, including the shelf number stopped by the transport vehicle and the stop operation event;

[0017] The storage duration of the marked goods is evenly divided into multiple statistical periods. In each statistical period, the stop operation events under the same shelf number are accumulated to obtain the stop frequency in the statistical period;

[0018] The stop frequencies of the same shelf in multiple statistical periods are statistically analyzed, and the maximum value and the minimum value of the stop frequency are processed to obtain the stop frequency fluctuation value;

[0019] The ratio of the stop frequency fluctuation value to the preset frequency fluctuation threshold value is used as the shelf reuse frequency.

[0020] In a preferred embodiment, in step S2, the goods height of each shelf is obtained by the warehouse monitoring unit, and the goods height of each shelf is combined into a height data set;

[0021] The median of the height data set is taken as a goods height median, the absolute value of the difference between each goods height and the goods height median is taken as a goods height deviation, and the median of the goods height deviation is taken as a goods deviation median;

[0022] The height lower limit reference value and the height upper limit reference value are calculated based on the goods height median and the goods deviation median;

[0023] The absolute value of the difference between the height lower limit reference value and the height upper limit reference value is taken as a height difference reference.

[0024] In a preferred embodiment, in step S2, the ratio of the goods height deviation of the shelf to the height difference reference is taken as the stacking state of the shelf;

[0025] The stacking state of the shelf and the shelf reuse frequency are standardized respectively, and a shelf review index is calculated;

[0026] If the shelf review index is greater than a preset review index threshold, it is determined to mark the shelf;

[0027] Otherwise, it is determined not to mark the shelf;

[0028] When there is a marked shelf, a visual recognition mechanism is entered.

[0029] In a preferred embodiment, in step S3, in the visual recognition mechanism, the spatial information of the marked shelf is called from the warehouse database, including the spatial position coordinates, the level number and the boundary coordinates of the shelf;

[0030] Three-dimensional positioning data of the shelf is generated according to the spatial position coordinates of the shelf, and the initial detection range of the shelf in the video image is determined with the boundary coordinates of the shelf, and the initial detection range is defined as a detection area;

[0031] The image frame corresponding to the detection area is intercepted from the global video monitoring stream and input to the YOLO model for recognition processing, and based on the pixel data in the detection area, a target detection frame set is generated, and the YOLO model outputs the goods location occupancy data and the channel gap data based on the target detection frame set.

[0032] In a preferred embodiment, in step S3, the class label corresponding to each target detection frame in the target detection frame set is recognized by the YOLO model to obtain the goods location occupancy data:

[0033] If the identified category label is a cargo, the cargo occupancy identifier of the corresponding target detection frame is set to 1; if the identified category is empty or background, the cargo occupancy identifier is set to 0, and the cargo occupancy identifiers of the target detection frames are combined into a storage position occupancy data set;

[0034] The channel gap width between adjacent target detection frames is calculated based on the spatial position relationship of the target detection frames with the identified category label as cargo in the target detection frame set.

[0035] The average value of the channel gap widths between all adjacent target detection frames with channel gaps is calculated as channel gap data.

[0036] In a preferred embodiment, in step S4, the occupancy distribution state including the storage position occupancy rate and the occupancy distribution uniformity index is obtained by statistical calculation of the cargo occupancy identifiers in the storage position occupancy data set.

[0037] The storage position occupancy rate represents the ratio of the number of target detection frames with the identified category label as cargo to the total number of target detection frames in the detection area.

[0038] The occupancy distribution uniformity index is the spatial variance of the target detection frames with the identified category label as cargo in the detection area, and the calculation formula is:

[0039] ;

[0040] Wherein, is the occupancy distribution uniformity index, is the storage position occupancy rate, is the cargo occupancy identifier, is the total number of target detection frames, is the index value of the target detection frame.

[0041] In a preferred embodiment, in step S4, the ratio of the occupancy distribution uniformity index and the channel gap data is taken as the area correction coefficient.

[0042] The area correction coefficient is multiplied by the area of the detection area to obtain an adjusted detection area.

[0043] The image frame corresponding to the adjusted detection area is re-input into the YOLO model for identification to generate a new target detection frame set.

[0044] The technical effects and advantages of the present application are:

[0045] The application judges whether the marked goods storage quantity is abnormal by marking the goods after the goods storage is completed and comparing with the target storage quantity, when the marked goods storage quantity is abnormal, the stopping frequency of the transport vehicle at each goods shelf is obtained, the goods shelf reuse frequency is calculated, the goods height data is collected through the warehouse monitoring unit, the goods shelf stacking state is calculated, the two are fused to generate a goods shelf review index, and it is judged whether to enter the visual recognition mechanism, in the visual recognition mechanism, the detection area is determined based on the spatial position information of the marked goods shelf, and the corresponding image frame is input into the YOLO model for recognition, the goods location occupation data and the channel gap data are output, the regional correction coefficient is calculated based on the output result, the detection area is adjusted and then input into the YOLO model for accurate recognition, and the final recognition result is generated, realizing the hierarchical detection and review of the abnormal storage goods shelf, avoiding the high-frequency global call of the YOLO model, and improving the real-time performance and accuracy of the warehouse storage recognition. BRIEF DESCRIPTION OF DRAWINGS

[0046] Fig. 1 The implementation flowchart of the process image recognition method based on the YOLO model.

[0047] Fig. 2 The step schematic diagram of the process image recognition method based on the YOLO model. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0049] The application judges whether the marked goods storage quantity is abnormal by marking the goods after the goods storage is completed and comparing with the target storage quantity, when the marked goods storage quantity is abnormal, the stopping frequency of the transport vehicle at each goods shelf is obtained, the goods shelf reuse frequency is calculated, the goods height data is collected through the warehouse monitoring unit, the goods shelf stacking state is calculated, the two are fused to generate a goods shelf review index, and it is judged whether to enter the visual recognition mechanism, in the visual recognition mechanism, the detection area is determined based on the spatial position information of the marked goods shelf, and the corresponding image frame is input into the YOLO model for recognition, the goods location occupation data and the channel gap data are output, the regional correction coefficient is calculated based on the output result, the detection area is adjusted and then input into the YOLO model for accurate recognition, and the final recognition result is generated, realizing the hierarchical detection and review of the abnormal storage goods shelf, avoiding the high-frequency global call of the YOLO model.

[0050] Example 1, as Figs. 1-2As shown, a process image recognition method based on a YOLO model includes the following steps:

[0051] Step S1: After the goods are stored, the goods are marked, the current storage quantity of the marked goods is obtained from the inventory database, and the target storage quantity of the marked goods is called to compare with the current storage quantity to determine whether the storage quantity of the marked goods is abnormal;

[0052] Step S2: When the storage quantity of the marked goods is abnormal, the stopping frequency of the transport vehicle at the same shelf is obtained from the vehicle scheduling database, and the shelf reuse frequency is calculated, the height difference reference of the goods height of each shelf is detected, and the stacking state of the shelf is analyzed, the shelf is marked in combination with the shelf reuse frequency, and it is determined whether to enter the visual recognition mechanism;

[0053] Step S3: In the visual recognition mechanism, a detection area is set based on the marked shelf, and the YOLO model is input, and the YOLO model is used to identify the storage position occupancy data and the channel gap data of the detection area;

[0054] Step S4: Based on the storage position occupancy data, the occupancy distribution state is analyzed, the occupancy distribution state and the channel gap data are fused to generate a region correction coefficient, and the detection area is adjusted, the YOLO model is used to identify the adjusted detection area, and an identification result is generated.

[0055] The specific implementation is as follows:

[0056] In step S1, when the goods are stored, the storage management unit generates a corresponding storage number for the goods according to the storage order of the goods, and after the goods are stored, the goods are marked based on the storage number of each goods;

[0057] The current storage quantity of the marked goods is obtained from the inventory database, wherein the current storage quantity refers to the total quantity of the marked goods;

[0058] The target storage quantity of the marked goods is called through the storage plan unit, and the target storage quantity refers to the storage quantity set according to the storage plan;

[0059] The current storage quantity and the target storage quantity are compared to determine whether the storage quantity of the marked goods is abnormal:

[0060] If the current storage quantity is inconsistent with the target storage quantity, it is determined that the storage quantity of the marked goods is abnormal;

[0061] Otherwise, it is determined that the storage quantity of the marked goods is normal;

[0062] When the quantity of the marked goods entering the warehouse is abnormal, a hierarchical judgment and triggering mechanism is entered to analyze the data of the shelves storing the marked goods to determine whether the shelves have quantity storage errors of the marked goods; when the quantity of the marked goods entering the warehouse is normal, the current quantity entering the warehouse is stored;

[0063] It should be explained that the warehouse management unit is an execution unit for controlling the warehouse entry process, including warehouse task scheduling and generating a warehouse entry number; the inventory database is a data resource library for storing and managing the state information of goods in the warehouse system, storing the warehouse entry number obtained by the marker; the warehouse planning unit is a business unit for managing and issuing warehouse task plans, which can be used to set the target warehouse quantity.

[0064] This step takes the quantity consistency verification after the goods enter the warehouse as the detection logic. When the goods enter the warehouse, the current warehouse quantity of the marked goods is obtained from the inventory database, and the target warehouse quantity is called. The two are compared to determine whether the warehouse quantity of the marked goods is abnormal. If the warehouse quantity of the marked goods is abnormal, the data of the shelves storing the marked goods is analyzed to further determine whether the shelves have quantity record errors or storage abnormalities, so as to determine whether to enter the subsequent visual recognition link, reduce the high-frequency calling of the YOLO model, reduce the system computing power consumption, and ensure the subsequent recognition accuracy while having higher execution efficiency.

[0065] In step S2, when the quantity of the marked goods entering the warehouse is abnormal, the shelf parking record of the transport vehicle in the storage process of the marked goods is obtained from the vehicle scheduling database, including the shelf number parked by the transport vehicle and the parking operation event;

[0066] The shelf parking record refers to the historical behavior data of the loading and unloading operation of the transport vehicle in the process of storing the marked goods into the shelves, and the parking operation event refers to a record unit for completing a loading and unloading operation process of the transport vehicle at the shelf position;

[0067] The warehouse duration of the marked goods entering the warehouse is evenly divided into multiple statistical periods, and the parking operation events under the same shelf number are accumulated in each statistical period to obtain the parking frequency in the statistical period;

[0068] The warehouse duration refers to the time interval experienced from the beginning to the completion of the whole process of storing the marked goods.

[0069] The parking frequencies of the same shelf in multiple statistical periods are statistically analyzed, and the maximum value and the minimum value of the parking frequency are subtracted to obtain the parking frequency fluctuation value;

[0070] The ratio of the parking frequency fluctuation value to the preset frequency fluctuation threshold value is used as the shelf reuse frequency;

[0071] The shelf reuse frequency refers to the dynamic change degree of the stop frequency, reflects the use activity and operation fluctuation of the shelf in the warehouse entry operation process, and the higher the shelf reuse frequency, the more the shelf is accessed or operated in a short time, or the state frequently switches, or there are cross-layer mixed storage or quantity recording errors, and the warehouse entry state of the shelf needs to enter the visual identification mechanism for further confirmation.

[0072] It needs to be explained that the vehicle scheduling database is a data management database for recording the operation behavior of the transport vehicle on the marked goods storage, and in the embodiment, it is used to obtain the shelf stop record of the transport vehicle between the shelves; the preset frequency fluctuation threshold value can be set according to the operation characteristics of the stored marked goods, the average loading and unloading cycle and the shelf scheduling density, when the average loading and unloading cycle of the marked goods is short, the preset frequency fluctuation threshold value should be set to a higher value accordingly, to avoid misjudgment of normal high-frequency operation as abnormal.

[0073] The goods height of each shelf is obtained by the warehouse monitoring unit, and the goods height refers to the vertical distance from the goods location bearing reference surface to the highest point of the goods;

[0074] The goods height of each shelf is combined into a height data set, the median of the height data set is taken as the goods height median, the absolute value of the difference between each goods height and the goods height median is taken as the goods height deviation, and the median of the goods height deviation is taken as the goods deviation median;

[0075] The first deviation correction amount and the second deviation correction amount are obtained by multiplying the preset first configuration coefficient and the second configuration coefficient with the goods deviation median respectively;

[0076] The difference between the goods height median and the first deviation correction amount is taken as the height lower limit reference value, and the sum of the goods height median and the second deviation correction amount is taken as the height upper limit reference value, and the absolute value of the difference between the height lower limit reference value and the height upper limit reference value is taken as the height difference reference;

[0077] The stacking state of the shelf is analyzed based on the height difference reference, and the ratio of the goods height deviation corresponding to the shelf to the height difference reference is taken as the stacking state of the shelf;

[0078] The stacking state of the shelf refers to the deviation degree of the goods height from the normal height interval, the larger the stacking state of the shelf, the more obvious the deviation of the goods height from the normal interval, the worse the loading stability of the shelf, and the more likely to appear the abnormal conditions such as goods quantity error, interlayer mixed storage, and the like, which need to enter the visual identification mechanism for further confirmation.

[0079] The shelf review index is calculated after the stacking state of the shelf and the shelf reuse frequency are standardized respectively: , wherein, is the stacking state after standardization, a standardization-processed shelf reuse frequency, 、 and a preset adjustment factor, a shelf review index;

[0080] The shelf review index reflects the deviation of the shelf in the two dimensions of spatial stacking and operation frequency. The larger the shelf review index is, the greater the height deviation of the shelf in the vertical stacking structure is, and the more frequently or repeatedly the shelf is used in the time dimension. The loading state of the shelf is unstable, and the quantity record is more prone to errors.

[0081] The shelf review index is compared with a preset review index threshold to determine whether to mark the shelf:

[0082] If the shelf review index is greater than the preset review index threshold, it is determined to mark the shelf.

[0083] On the contrary, it is determined not to mark the shelf.

[0084] When there is a marked shelf, enter the visual recognition mechanism.

[0085] When there is a marked shelf, enter the visual recognition mechanism to further detect and identify the marked shelf.

[0086] When there is no marked shelf, do not perform local detection on individual shelves, but identify each shelf through a YOLO model and generate an identification result.

[0087] It should be explained that the warehouse monitoring unit is a module for collecting the state of the marked shelf and the height of the goods during the storage process. The preset first and second configuration coefficients are used to adjust the sensitivity of the height deviation of the goods, which can be set according to the volume category of the marked goods and the shelf layer density. The standardization processing method includes but is not limited to standard linear transformation based on interval scaling, Z-Score standardization method based on statistics, or normalization method based on nonlinear mapping function. Here, the application method of standardization processing is not described. The preset adjustment factor can be set according to the shelf usage frequency, detection period, and visual recognition trigger rate. The preset review index threshold is a critical value for determining whether to mark the shelf, which can be set according to the historical identification accuracy and the proportion of abnormal samples.

[0088] By calculating the shelf review index in two dimensions of shelf reuse frequency and stacking state, a hierarchical screening mechanism from quantity detection to shelf structure review to visual verification is realized, and the stability, real-time performance, and data reliability of the warehouse storage process are improved.

[0089] In step S3, in the visual recognition mechanism, a detection region is established according to the marked shelf, the spatial information of the marked shelf is called from the warehouse database, including the spatial position coordinates, the level number and the boundary coordinates of the shelf;

[0090] The spatial position coordinates are three-dimensional coordinate data of each reference point of the shelf in the horizontal (X-axis), vertical (Y-axis) and height (Z-axis) directions in the warehouse coordinate system, which are used to determine the actual position and level distribution of the shelf in space;

[0091] The boundary coordinates are the upper left and lower right pixel coordinates of the shelf in the video image coordinate system, which are used to limit the projection range of the shelf in the image and serve as the boundary parameters of the detection region;

[0092] The three-dimensional positioning data of the shelf is generated according to the spatial position coordinates of the shelf, and the initial detection range of the shelf in the video image is determined according to the boundary coordinates of the shelf, and the initial detection range is defined as the detection region;

[0093] It should be noted that the warehouse database is a comprehensive information database for storing and managing various structured data in the warehouse system, and its data content includes shelf information, storage location number information, warehouse entry and exit records, spatial coordinate data, level number information and video monitoring mapping parameters, etc.

[0094] After determining the detection region, the corresponding image frame is intercepted from the global video monitoring stream as the input image, which is input into the YOLO model for recognition processing. The YOLO model is based on a convolutional neural network structure, which sequentially extracts target features and outputs detection results through feature extraction layers, feature fusion layers and prediction output layers;

[0095] In the recognition process, the YOLO model generates a set of target detection boxes based on the pixel data in the detection region, where each target detection box represents the position coordinates and size information of the recognized goods or channel target in the detection region, expressed as the center coordinates, width and height of the target detection box;

[0096] The YOLO model outputs the storage location occupancy data and the channel gap data based on the set of target detection boxes;

[0097] The YOLO model identifies the class labels corresponding to each target detection box in the set of target detection boxes to obtain the storage location occupancy data. If the identified class label is a good, the good occupancy identifier of the corresponding target detection box is set to 1; if the identified class label is empty or background, the good occupancy identifier is set to 0, and the good occupancy identifiers of each target detection box are combined into a storage location occupancy data set;

[0098] The channel gap data is calculated based on the spatial position relationship of the target bounding boxes whose class labels are identified as goods in the target bounding box set, and is used to describe the channel distance and stacking density between goods. The target bounding box set is arranged in the horizontal order of the center coordinates of the target bounding boxes, and the horizontal coordinate of the right boundary of the i-th target bounding box is:

[0099]

[0100] wherein, is the horizontal coordinate of the right boundary of the i-th target bounding box, is the horizontal coordinate of the center of the i-th target bounding box, is the width of the i-th target bounding box.

[0101] The horizontal coordinate of the left boundary of the i+1-th target bounding box is:

[0102]

[0103] wherein, is the horizontal coordinate of the left boundary of the i+1-th target bounding box, is the horizontal coordinate of the center of the i+1-th target bounding box, is the width of the i+1-th target bounding box.

[0104] The channel gap width between adjacent target bounding boxes is defined as:

[0105]

[0106] wherein, is the channel gap width between adjacent target bounding boxes, is the horizontal coordinate of the right boundary of the i-th target bounding box, is the horizontal coordinate of the left boundary of the i+1-th target bounding box.

[0107] When , it indicates that there is a channel gap between the goods; when , it indicates that the adjacent target bounding boxes overlap or adhere, indicating the state of stacking or goods close to each other.

[0108] The average value of the channel gap widths between all adjacent target bounding boxes with channel gaps is calculated as the channel gap data.

[0109] ​​​It should be noted that the YOLO model is a real-time target detection model based on deep learning, and its core structure is a convolutional neural network, which can complete the positioning and classification of targets in a single forward propagation. The convolutional neural network structure is composed of a plurality of convolutional layers, pooling layers and activation functions, which are used to extract multi-scale and multi-channel feature information from input images; the feature extraction layer is responsible for local convolution operation and nonlinear transformation of the pixel data of the input image, and generates a preliminary feature map to represent the local features of each region in the image; the feature fusion layer integrates and enhances the feature maps of different scales through cross-layer connection, feature splicing or upsampling operation, so as to balance the detection ability of large and small targets in the image; the prediction output layer generates the final target detection frame set and its corresponding class confidence based on the fused feature map, and outputs the center coordinates, width, height and class probability value of each detection frame through regression calculation, so as to complete the accurate identification and positioning of the goods and channel targets in the detection area.

[0110] In step S4, the occupancy distribution state in the detection area is analyzed based on the goods location occupancy data, and the occupancy distribution state is obtained by statistical calculation of the goods occupancy identifier in the goods location occupancy data set, including the goods location occupancy rate and the occupancy distribution uniformity index. Specifically, the goods location occupancy rate represents the ratio of the number of target detection frames with the identification class label as goods to the total number of target detection frames in the detection area, and its calculation formula is:

[0111] ;

[0112] Wherein, is the goods location occupancy rate, is the goods occupancy identifier, is the total number of target detection frames, is the index value of the target detection frame.

[0113] The occupancy distribution uniformity index is the spatial variance of the target detection frame with the identification class label as goods in the detection area, and its calculation formula is:

[0114] ;

[0115] Wherein, is the occupancy distribution uniformity index, is the goods location occupancy rate, is the goods occupancy identifier.

[0116] Based on the occupancy distribution uniformity index, it is fused with the channel gap data to calculate the area correction coefficient of the detection area, which is used for spatial adjustment of the detection area to optimize the input range of the YOLO model and improve the recognition accuracy. The calculation formula of the area correction coefficient is:

[0117] ;

[0118] wherein, is a region correction coefficient, is an occupancy distribution uniformity index, is a channel gap data.

[0119] When , it indicates that the detection region has concentrated occupancy or insufficient channel gap, and the detection region needs to be enlarged to cover the possible missed storage locations;

[0120] When , it indicates that the detection region has low occupancy or large channel gap, and the detection region needs to be reduced to reduce the invalid calculation amount.

[0121] The region correction coefficient is multiplied by the area of the detection region to obtain the adjusted detection region, which is used to optimize the input of the YOLO model.

[0122] The image frame corresponding to the adjusted detection region is re-input into the YOLO model for recognition to generate a new target detection frame set, and the new storage occupancy data and channel gap data are output accordingly. The output result obtained in this process is the accurate recognition result of the adjusted detection region, reflecting the real occupancy state and spatial distribution characteristics of the goods in the detection region, and providing reliable quantitative basis for subsequent abnormality judgment, shelf state review and warehouse operation optimization.

[0123] Finally, it should be noted that in this paper, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0124] Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0125] In this document, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise. It will be further understood that the terms "comprises," "comprising," "includes," and "including," or the like, when used in this specification, specify the presence of stated features, integers, steps, operations, components, parts, or the like, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or the like.

[0126] Various embodiments described in this specification are described with reference to particular implementations. Embodiments can be practiced with other systems, components, materials, and techniques than those described without departing from the spirit and scope of the present disclosure. The terms "comprise," "comprising," "include," "including," "contain," "containing," "have," "having," "composed of," and the like can have the meaning ascribed to them under

[0127] The above description of disclosed embodiments is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. While specific embodiments of, and examples for, the application are described herein for illustrative purposes, various equivalent modifications are possible within the scope of the application, as those skilled in the relevant art will recognize. The teachings of the application provided herein can be applied to other embodiments and examples without departing from the spirit and scope of the description after reading and understanding the previous description. Accordingly, the application is not limited to that precisely as shown and described. Since many embodiments of the application can be made without departing from the spirit and scope of the application, the application resident in the claims along with the specification, including the drawings, is to be accorded the widest scope under the law.

Claims

1. A process image recognition method based on the YOLO model, characterized in that: Includes the following steps: Step S1: After the goods are received into the warehouse, mark the goods, access the inventory database to get the current quantity of marked goods received into the warehouse, retrieve the target quantity of marked goods received into the warehouse and compare it with the current quantity of marked goods received into the warehouse to determine whether the quantity of marked goods received into the warehouse is abnormal. Step S2: When the quantity of goods entering the warehouse is abnormal, obtain the frequency of the transport vehicle stopping at the same shelf through the vehicle scheduling database and calculate the shelf reuse frequency. Detect the height of goods on each shelf, set a height difference benchmark, and analyze the stacking status of the shelves. Combine the shelf reuse frequency to mark the shelf and determine whether to enter the visual recognition mechanism. Step S3: In the visual recognition mechanism, a detection area is set based on the marked shelves and the YOLO model is input. The YOLO model is used to identify the location occupancy data and aisle gap data of the detection area. Step S4: Analyze the occupancy distribution based on the occupancy data, integrate the occupancy distribution with the channel gap data to generate a region correction coefficient and adjust the detection area, use the YOLO model to identify the adjusted detection area and generate the identification result; In step S4, the occupancy distribution status, including the occupancy rate and the occupancy distribution uniformity index, is obtained by statistically calculating the cargo occupancy identifiers in the cargo occupancy data set. The cargo occupancy rate represents the ratio of the number of target detection boxes in the detection area whose category label is cargo to the total number of target detection boxes; The occupancy distribution uniformity index is the spatial variance of the bounding boxes of targets labeled as goods within the detection area, and its calculation formula is as follows: ； in, The occupancy distribution uniformity index, For the occupancy rate of storage space, For cargo occupancy identification, This represents the total number of target detection boxes. This is the index value of the target detection box; In step S4, the ratio of the occupancy distribution uniformity index to the channel gap data is used as the regional correction coefficient; The adjusted detection area is obtained by multiplying the area correction factor by the area of ​​the detection area. The image frames corresponding to the adjusted detection areas are re-input into the YOLO model for recognition, generating a new set of target detection boxes.

2. The process image recognition method based on the YOLO model according to claim 1, characterized in that: In step S1, when goods are stored in the warehouse, the warehouse management unit generates corresponding warehouse entry numbers for the goods according to the order of their entry. After the goods are stored, the goods are marked based on their warehouse entry numbers. Access the inventory database to obtain the current inbound quantity of marked goods, and retrieve the target inbound quantity of marked goods through the inbound planning unit; If the current quantity of goods received is inconsistent with the target quantity of goods received, then the quantity of goods received is marked as abnormal. Conversely, if the quantity of goods marked as having entered the warehouse is normal, it is determined that the quantity is normal.

3. The process image recognition method based on the YOLO model according to claim 1, characterized in that: In step S2, when the quantity of goods entering the warehouse is abnormal, the vehicle scheduling database is accessed to obtain the rack docking records of the transport vehicles during the storage of the marked goods, including the rack number where the transport vehicles are docked and the docking operation event. The storage duration of marked goods is divided into multiple statistical periods. Within each statistical period, the stopping operation events under the same shelf number are accumulated to obtain the stopping frequency within the statistical period. Statistical analysis was performed on the frequency of stops at the same shelf in multiple statistical periods, and the difference between the maximum and minimum frequency of stops was used to obtain the frequency fluctuation value. The ratio of the frequency fluctuation value to the preset frequency fluctuation threshold is used as the shelf reuse frequency.

4. The process image recognition method based on the YOLO model according to claim 1, characterized in that: In step S2, the height of goods on each shelf is obtained through the warehouse monitoring unit, and the height of goods on each shelf is combined into a height dataset; The median of the height dataset is taken as the median of the cargo height. The absolute value of the difference between each cargo height and the median cargo height is taken as the cargo height deviation. The median of the cargo height deviations is taken as the median of the cargo deviations. The lower limit reference value and the upper limit reference value of the height are calculated based on the median height of the cargo and the median deviation of the cargo. The height difference benchmark is obtained by taking the absolute value of the difference between the lower and upper height benchmark values.

5. The process image recognition method based on the YOLO model according to claim 4, characterized in that: In step S2, the ratio of the goods height deviation on the shelf to the height difference benchmark is used as the stacking state of the shelf. The shelf verification index was calculated after standardizing the stacking status and shelf reuse frequency of the shelves respectively. If the shelf verification index is greater than the preset verification index threshold, then the shelf will be marked. Conversely, it is determined that the shelves should not be marked. When a marked shelf is present, the visual recognition mechanism is activated.

6. The process image recognition method based on the YOLO model according to claim 1, characterized in that: In step S3, the visual recognition mechanism retrieves the spatial information of the marked shelves from the warehouse database, including the spatial location coordinates, layer number, and boundary coordinates of the shelves. The three-dimensional positioning data of the shelf is generated based on the spatial coordinates of the shelf, and the initial detection range of the shelf in the video image is determined by the boundary coordinates of the shelf. The initial detection range is defined as the detection area. Image frames corresponding to the detection area are extracted from the global video surveillance stream and input into the YOLO model for recognition processing. Based on the pixel data within the detection area, a set of target detection boxes is generated. Based on the set of target detection boxes, the YOLO model outputs the cargo space occupancy data and aisle gap data.

7. The process image recognition method based on the YOLO model according to claim 6, characterized in that: In step S3, the YOLO model is used to identify the category labels corresponding to each target detection box in the target detection box set to obtain the storage location occupancy data: If the identified category label is goods, the goods occupancy flag of the corresponding target detection box is set to 1; if the identified category is empty space or background, the goods occupancy flag is set to 0, and the goods occupancy flags of each target detection box are combined into a cargo space occupancy dataset. The width of the channel gap between adjacent target detection boxes is calculated based on the spatial positional relationship of the target detection boxes in the target detection box set that identify the category label as goods. The average width of the channel gap between all adjacent target detection boxes with channel gaps is calculated as the channel gap data.

Images