Dressing compliance real-time detection method for advanced manufacturing industry

Abstract

The invention provides an advanced manufacturing industry-oriented dressing compliance real-time detection method, which comprises the following steps of: acquiring dense personnel images of different scenes and different postures, constructing an advanced manufacturing industry-oriented personnel dressing data set, marking a human body detection frame, human body key points and dressing on the data set, and determining the dressing compliance of the advanced manufacturing industry. Training a human body and key point model and a dressing detection model which can adapt to dense people; constructing a weighted bipartite graph by using the human body detection result and the dressing detection result, calculating an edge weight by using the human body detection frame, the human body key points and the dressing detection frame, and associating the personnel detection result and the dressing detection result in the image by calculating the optimal matching of the weighted bipartite graph; and filtering shielded persons in the image according to a human body detection result, judging whether dressing of non-shielded persons in the image meets requirements or not based on a dressing association result, and outputting an algorithm detection result. According to the invention, the detection precision and the detection speed of the dressing compliance detection method are improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of computer technology and image recognition technology, and in particular relates to a real-time detection method for dressing compliance for advanced manufacturing. Background Art

[0002] In China's advanced manufacturing industries such as pharmaceuticals, semiconductors, clean rooms for food production, laboratories, chemical production, etc., due to the need for standardized management and production safety, there are strict requirements for personnel's clothing in production environments such as pharmaceutical production workshops, laboratories, and food processing workshops. The main clothing types include protective clothing, masks, gloves, headgear, goggles, face shields, etc. In the production process, if the clothing is worn non-compliance and not corrected in time, it may pollute the production environment. However, it is difficult to achieve full-day coverage supervision by manually checking video surveillance and other methods, and it is difficult to achieve timely and accurate dressing compliance detection due to the influence of the subjective factors of the inspectors.

[0003] Therefore, the technology of personnel clothing detection has emerged accordingly. And with the continuous development of intelligent monitoring systems and the increase in labor costs, the demand of enterprises for accurate and efficient dressing compliance detection algorithms is increasing day by day. Currently, the existing dressing compliance detection algorithms mainly consist of a human body and key point detection network and a dressing detection network. The human body and key point network determines the area of the human body in the image and judges the human body posture, and the dressing detection network sequentially estimates the dressing types based on the local area of the image where the human body is located.

[0004] Specifically, since there are often multiple people in the actual application scenario, the existing dressing compliance detection methods have the following problems in actual application:

[0005] 1. The algorithm only uses the local area information of the image where the human body is located for compliance analysis, and does not consider the relationship between the human bodies in the image. When people are close and interacting, it is very easy to be affected by non-target people, that is, the local area of the image where the target is located contains part of the dressing information of adjacent people, resulting in false alarms or missed alarms.

[0006] 2. The algorithm sequentially performs dressing compliance detection on all people in the image, and the algorithm time consumption increases linearly with the increase in the number of people. When there are many people in the scene, the algorithm will be very slow and the real-time performance is poor.

[0007] Therefore, when the existing dressing compliance detection methods are actually applied, they will be affected by factors such as the number of people and the interaction between people, and have deficiencies such as poor real-time performance, low accuracy, and poor user experience. Summary of the Invention

[0008] In view of this, the present invention aims to overcome the deficiencies of the above problems in the prior art, and proposes a real-time detection method for dressing compliance in advanced manufacturing. By calculating the optimal matching of the weighted bipartite graph, the association between the personnel and dressing results in the image is realized, which greatly reduces the false alarms of the algorithm caused by personnel interaction, effectively improves the detection accuracy and speed of the algorithm, reduces the computing power consumption, and can perform higher-concurrency processing on edge computing devices, better adapting to the production environments of different advanced manufacturing industries.

[0009] To achieve the above object, the technical solution of the present invention is realized as follows:

[0010] The first aspect of the present invention provides a real-time detection method for dressing compliance in advanced manufacturing, including the following steps:

[0011] Step 1: Collect dense personnel images in different scenarios and poses, construct a personnel dressing dataset for advanced manufacturing, annotate the human detection frames, human key points, and dressings of the dataset, and train a human and key point model and a dressing detection model that can adapt to dense personnel.

[0012] Step 2: Use the human detection results and dressing detection results to construct a weighted bipartite graph, calculate the edge weights using the human detection frames, human key points, and dressing detection frames, and associate the human detection results and dressing detection results in the image by calculating the optimal matching of the weighted bipartite graph.

[0013] Step 3: Filter out the occluded personnel in the image according to the human detection results, judge whether the dressings of the non-occluded personnel in the image meet the requirements based on the dressing association results, and output the algorithm detection results.

[0014] Further, in the step 1, the human detection frames, human key points, and dressings of the dataset are annotated, low-quality data is filtered, the pre-annotation results are corrected, and the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0015] Further, in the step 1, the human and key point model uses the yolov10-pose model to output the human detection frame and 33 human key points, including the key points of the head, torso, hands, and feet, and uses it to calculate the centroid of the key points corresponding to the dressing, and further calculates the distance component in the edge weight of the bipartite graph.

[0016] Further, the model loss function uses Repulsion loss, and the Repulsion loss function is as follows:

[0017] L = L Attr +α*L RepGT +β*L RepBox ;;

[0018] Among them, L Attr is the attraction loss term, which makes the predicted bounding box closer to the ground truth bounding box. L Rep is the repulsion loss term, which makes the predicted bounding box away from the surrounding ground truth bounding boxes. α and β are weight balance terms, both with a value of 0.5.

[0019] Furthermore, the specific steps of step 2 include:

[0020] Filter the clothing detection results with low confidence;

[0021] Construct a bipartite graph of person-clothing rights, associate and process the clothing detection results by category, and calculate the edge weights of the bipartite graph with rights;

[0022] Use the Hungarian K-M algorithm to perform the optimal matching calculation of the bipartite graph with rights.

[0023] Furthermore, the specific steps of using the Hungarian K-M algorithm to perform the optimal matching calculation of the bipartite graph with rights include:

[0024] In the bipartite graph, the left nodes are the clothing detection results, and an integer value L xi is given to the i-th (1 ≤ i ≤ n) left node, and the right nodes are the person detection results, and an integer value L yj is given to the j-th (1 ≤ j ≤ n) right node, and at the same time, it satisfies: L xi +L yj ≥weight (i，j) where weight (i，j) is the edge weight between the i-th left node and the j-th right node. These integer values L xi and L yj are called the top labels of the nodes. First, on the premise of satisfying L xi +L yj ≥weight (i，j) , assign a random top label to each node to expand the scale of the equal subgraph until a perfect matching exists in the equal subgraph.

[0025] Furthermore, the specific steps of step 3 include:

[0026] Sort the algorithm results in descending order based on the sum of the confidence of the human body bounding box and the average confidence of the human body key points output by the human body and key point algorithm;

[0027] Check in turn whether the detection bounding box is close to the image edge. If the detection bounding box crosses the image edge, delete it;

[0028] For the remaining detection boxes, calculate the overlap ratio with other detection boxes in sequence. If the overlap ratio is greater than 0.2, delete the detection box with a low average confidence level; ensure that the final remaining human detection results are neither close to the image edge nor have obvious overlaps with each other

[0029] For the remaining human detection results, analyze in sequence whether the associated clothing detection results comply with the regulations, and output the algorithm detection results.

[0030] The second aspect of the present invention provides a real-time clothing compliance detection system for advanced manufacturing, including:

[0031] A data acquisition unit for collecting dense human images in different scenarios and postures, constructing a personnel clothing dataset for advanced manufacturing, annotating the human detection boxes, human key points, and clothing in the dataset, and training a human and key point model and a clothing detection model that can adapt to dense humans;

[0032] A data processing unit for constructing a weighted bipartite graph using the human detection results and clothing detection results, calculating the edge weights using the human detection boxes, human key points, and clothing detection boxes, and associating the human detection results and clothing detection results in the image by calculating the optimal matching of the weighted bipartite graph;

[0033] A result output unit for filtering occluded persons in the image according to the human detection results, judging whether the clothing of non-occluded persons in the image meets the requirements based on the clothing association results, and outputting the algorithm detection results.

[0034] The third aspect of the present invention provides an electronic device, including a processor and a memory communicatively connected to the processor and used for storing executable instructions of the processor. The processor is used to execute the above-mentioned real-time clothing compliance detection method for advanced manufacturing.

[0035] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, and the computer program is executed by the processor to perform the above-mentioned real-time clothing compliance detection method for advanced manufacturing.

[0036] Compared with the prior art, the real-time clothing compliance detection method for advanced manufacturing described in the present invention has the following advantages:

[0037] The present invention constructs a weighted bipartite graph using the human detection results and clothing detection results, and realizes the accurate association of the human and clothing detection results in the image by calculating the optimal matching of the weighted bipartite graph, avoiding false alarms caused by human interaction and improving the algorithm accuracy.

[0038] The present invention aims at the dressing compliance application scenario, improves the IOU calculation method, can effectively solve the problem of low IOU scores between detection frames of different scales, and accurately calculates the weights of the bipartite graph;

[0039] The present invention only performs dressing detection once, uses a weighted bipartite graph to associate the personnel detection results with the dressing detection results, avoids repeatedly calling the dressing detection algorithm when there are multiple people in the image, the algorithm processing speed no longer changes with the number of people, greatly improves the algorithm operation speed, and effectively improves the practicability and real-time performance of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0041] Figure 1 is the overall process schematic diagram of the method of the present invention;

[0042] Figure 2 is the algorithm result association flowchart of the present invention;

[0043] Figure 3 is the optimal matching calculation flowchart of the weighted bipartite graph of the present invention;

[0044] Figure 4 is the schematic diagram of the calculation process of the weighted bipartite graph of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0046] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation to the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "a plurality" is two or more.

[0047] In the description of the present invention, it should be noted that unless otherwise clearly specified and defined, the terms "installation", "connection", and "coupling" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0048] The present invention will be described in detail below with reference to the drawings and in conjunction with embodiments.

[0049] Embodiment 1:

[0050] The embodiment of the present invention provides a real-time detection method for dressing compliance in advanced manufacturing, and the overall process is as Figure 1 shown:

[0051] 1. Dataset construction and model training

[0052] To solve the above-mentioned problems, first, a dataset for training the human body and key point model and dressing detection model for dense personnel needs to be constructed. According to the work dressing specifications of current advanced manufacturing, in typical advanced manufacturing production scenarios such as pharmaceutical, semiconductor, food production clean rooms, laboratories, and chemical production, real video data in different scenarios is collected. The existing human body and key point model and dressing detection model are used to pre-annotate the dataset, filter low-quality data, and finally manually check and correct the pre-annotation results, and divide the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0053] There is no clear requirement for the human key point detection algorithm and dressing detection algorithm used in the implementation of the present invention, and any algorithm can be used. However, to ensure the implementation effect of the present invention, the human key point detection algorithm should be able to perform dense personnel detection. In this embodiment, the yolov10-pose model is used to output the human detection frame and 33-point human key points, including the key points of the head, torso, hands, and feet. The centroid of the key points corresponding to the dressing can be calculated by it, and then the distance component in the bipartite graph edge weight can be calculated. The model loss function uses Repulsion loss. This loss function can make the prediction box close to the real box while repelling the same type of detection boxes, so that the training model can effectively reduce the misdetection or missed detection of the detection box caused by personnel occlusion. The Repulsion loss function is as follows:

[0054] L = L Attr + α * L RepGT + β * L RepBox ;

[0055] where, LAttr is the attraction loss term, which makes the predicted bounding box closer to the ground truth box, L Rep is the repulsion loss term, which makes the predicted bounding box away from the surrounding ground truth boxes. α and β are weight balance terms, both with a value of 0.5.

[0056] In this embodiment, the yolov10 model is used for image clothing detection training. The input of the model is the original image, rather than the local area where the person in the image is located. The output of the model is all the clothing detection results in the image.

[0057] 2. Optimal matching of weighted bipartite graph and result association

[0058] After processing the image using the human detection algorithm and the clothing detection algorithm, the human key point detection results and clothing detection results of the image are obtained. The detection results of the people and the clothing detection results in the image are associated by calculating the optimal matching of the weighted bipartite graph.

[0059] A bipartite graph is a special graph structure in graph theory. Let G=(V, E) be an undirected graph. If the vertex V can be divided into two non-overlapping subsets (A, B), and the two vertices i, j associated with each edge in the graph belong to these two different vertex sets (A, B) respectively, then the graph G is called a bipartite graph.

[0060] In the given bipartite graph G, M is a subgraph of G. If any two edges in the edge set of M do not attach to the same vertex, then M is called a matching. Selecting such a subset with the largest number of edges becomes the maximum matching problem of the graph, that is, the optimal matching of the bipartite graph. If the edges have weights, it is the optimal matching of the weighted bipartite graph. In the present invention, we abstract the person detection results and the clothing detection results into two non-overlapping subsets A, B in the graph G, and use the positional relationship between the person detection results and the clothing detection results to abstract the edge weights, and then abstract the result association problem into the optimal matching problem of the weighted bipartite graph. Therefore, by calculating the optimal matching of the weighted bipartite graph, the optimal allocation of the person detection results and the clothing detection results in the graph can be obtained. As Figure 2 shown:

[0061] (1) First, filter out the clothing detection results with low confidence to avoid the influence of misdetected data on the matching results.

[0062] (2) Construct a person-clothing weighted bipartite graph, and perform association processing on the clothing detection results by category, such as protective clothing, gloves, hair covers, masks, etc. Gloves are distinguished by left and right hands to ensure that each category of clothing detection results only corresponds to one person. Calculate the edge weights of the weighted bipartite graph. The weight calculation formula is:

[0063] weight (i，j) =selfIOU (i，j) +λ·(200 - distance (i，j) );

[0064] The weight formula fuses the improved IOU result and the Euclidean distance result between the clothing detection box and the centroid of the key points of the corresponding parts of the human body. To balance the value ranges of the two parts, the adjustment variable λ = 0.005 is used here. In the formula, selfIOU (i，j) is the selfIOU result between the detection box of human body i and the detection box of clothing detection result j. The original IOU calculation method is as follows:

[0065]

[0066] where Rect i is the human body detection box, Rect j is the clothing detection result, and the calculation method of selfIOU (i，j) is as follows:

[0067]

[0068] Since the sizes of the detection boxes of actual clothing types vary greatly, the detection boxes of protective clothing are basically the same as those of the human body, while the detection boxes of gloves are often much smaller than those of the human body. Using the original IOU calculation method will cause the IOU values calculated for different types to vary too much and cannot reflect the real situation. The selfIOU calculation method can effectively reflect the overlapping relationship between the clothing detection result and the human body detection result and will not affect the calculation of the weight.

[0069] On a 1080P resolution image, the upper limit of the distance result is set to 200 pixels. If the resolutions are inconsistent, the threshold can be adaptively scaled. If the Euclidean distance distance (i，j) between the clothing detection box and the centroid of the key points of the corresponding parts of the human body is greater than 200 pixels, it is considered that it is not the clothing result of this person, and there is no connection between the human body detection result and the clothing detection result. If it is less than 200 pixels, the weight of the distance part is calculated based on the distance.

[0070] Fusing the above two parts to calculate the weight, the final weight result is obtained, and the value range is [0, 2]. Taking this threshold as the edge weight of the constructed bipartite graph G, the optimal matching result of this bipartite graph is calculated.

[0071] (3) Use the Hungarian K-M algorithm to calculate the optimal matching of the weighted bipartite graph. The K-M algorithm is a greedy extension of the Hungarian algorithm, and its idea is to give each vertex of the bipartite graph a label, that is, a top label, and then use the Hungarian algorithm to solve the optimal matching problem under the complete matching. By modifying the top labels of some points, the total number of feasible edges in the graph is continuously increased until there is a perfect matching composed only of feasible edges in the graph. At this time, this matching must be the best.

[0072] Specifically, in the bipartite graph, the left - hand nodes are the results of clothing detection, and an integer value \(L\) is assigned to the \(i\) - th (\(1\leq i\leq n\)) left - hand node. xi The right - hand nodes are the results of person detection, and an integer value \(L\) is assigned to the \(j\) - th (\(1\leq j\leq n\)) right - hand node. yj At the same time, it satisfies: \(L\) xi +\(L\) yj \(\geq\) weight (i，j) where weight (i，j) is the edge weight between the \(i\) - th left - hand node and the \(j\) - th right - hand node calculated in (2) (set to negative infinity when there is no edge weight). These integer values \(L\) xi and \(L\) yj are called the vertex labels of the nodes. First, on the premise of satisfying \(L\) xi +\(L\) yj \(\geq\) weight (i，j) , an arbitrary vertex label is assigned to each node, and then an appropriate strategy is adopted to continuously expand the scale of the equal - subgraph until a perfect matching exists in the equal - subgraph.

[0073] As Figure 3 shown, first check whether the number of current - type clothing - detection results is consistent with the number of person - detection results. If not, complete it by adding virtual nodes, and set the edge weights of all these nodes to negative infinity.

[0074] The matching process and results are as Figure 4 shown. Figure 4 (a) First, assign vertex - label values \(L\) xi =\(\max(\text{weight} (i，j) )\), \(L\) yj = 0 to the constructed bipartite graph. Figure 4 (b) The green results in (b) are the results of successful matching of clothing results \(x1\), \(x2\) in sequence. The red one is the matching result of \(x3\), and its right - hand matching result conflicts with \(x2\), so the matching fails. Calculate the adjustment amount \(\Delta\) according to the following formula. Each adjustment will make the equal - subgraph increase at least one edge without reducing the edges in the equal - subgraph. The calculation method is as follows:

[0075] \(\Delta=\min(L xi +L yj -\text{weight} (i，j) )\);

[0076] Calculate the \(\Delta\) values of the two conflicting points \(x2\), \(x3\) in sequence. \(\Delta\) x2 = 0.52, \(\Delta\) x3 = 0.32. Take the minimum value of \(\Delta\) as 0.32. Increase the right - hand vertex label \(y1\) by \(\Delta\), and decrease the left - hand vertex labels \(x2\), \(x3\) by \(\Delta\). Finally, \(x2\) matches \(y1\) and \(x3\) rematches. Figure 4(c) In this case, x3 continues to match y2, conflicting with x1 and resulting in a failed match. The value of Δ is recalculated as 0.35. After adjusting the top labels, x3 finally matches y3 successfully, as shown in Figure 4 (d), and the matching is completed.

[0077] According to the above steps (1), (2), and (3), the association matching between the dressing results of each type and the personnel detection results is completed in sequence.

[0078] 3. Dress code compliance check and output of the algorithm results

[0079] To minimize the impact of occlusion on the personnel detection algorithm as much as possible, the human detection model has been specifically optimized. The detected personnel will include human targets with partial occlusion. For such targets, due to occlusion, some dressing results are missing, which may lead to misjudgment of the algorithm. Therefore, before outputting the compliance results, it is necessary to remove the personnel with severe occlusion.

[0080] First, based on the sum of the confidence of the human body box and the average confidence of the human body key points output by the human body and key point algorithm, the algorithm results are sorted in descending order;

[0081] Check in sequence whether the detection box is close to the image edge. If the detection box crosses the image edge, it is deleted;

[0082] For the remaining detection boxes, calculate the intersection over union (IOU) with other detection boxes in sequence (i，j) . If IOU > 0.2, delete the detection box with a lower average confidence; ensure that the final remaining human detection results are neither close to the image edge nor have obvious overlap with each other.

[0083]

[0084] For the remaining personnel detection results, analyze in sequence whether the associated dressing detection results meet the regulations and output the algorithm detection results.

[0085] Example 2:

[0086] The embodiment of the present invention provides a real-time dressing compliance detection system for advanced manufacturing, including:

[0087] A data acquisition unit, which is used to collect dense personnel images in different scenarios and postures, construct a personnel dressing dataset for advanced manufacturing, annotate the human body detection box, human body key points, and dressing on the dataset, and train a human body and key point model and a dressing detection model that can adapt to dense personnel;

[0088] A data processing unit, configured to construct a weighted bipartite graph by using the human body detection result and the clothing detection result, calculate the edge weights by using the human body detection box, human body key points and clothing detection box, and associate the human detection result and the clothing detection result in the image by calculating the optimal matching of the weighted bipartite graph;

[0089] A result output unit, configured to filter out occluded persons in the image according to the human body detection result, judge whether the clothing of non-occluded persons in the image meets the requirements based on the clothing association result, and output the algorithm detection result.

[0090] Embodiment III:

[0091] Embodiment III of the present invention provides an electronic device, including a processor and a memory communicatively connected to the processor and used for storing executable instructions of the processor, and the processor is configured to execute the above-mentioned real-time clothing compliance detection method for advanced manufacturing.

[0092] Embodiment IV:

[0093] Embodiment IV of the present invention provides a computer-readable storage medium, storing a computer program, and the computer program is executed by a processor to perform the above-mentioned real-time clothing compliance detection method for advanced manufacturing.

[0094] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A real-time detection method for dressing compliance in advanced manufacturing, characterized in that: It includes the following steps: Step 1: Collect dense personnel images in different scenarios and poses, construct a personnel clothing dataset for advanced manufacturing, annotate the dataset with human detection boxes, human key points, and clothing, and train a human and key point model and a clothing detection model that can adapt to dense personnel; Step 2: Use the human detection results and clothing detection results to construct a weighted bipartite graph, calculate the edge weights using the human detection boxes, human key points, and clothing detection boxes, and associate the personnel detection results and clothing detection results in the image by calculating the optimal matching of the weighted bipartite graph; Step 3: Filter out occluded personnel in the image based on the human detection results, judge whether the clothing of non-occluded personnel in the image meets the requirements based on the clothing association results, and output the algorithm detection results.

2. The real-time detection method for dressing compliance oriented to advanced manufacturing according to claim 1, characterized in that: In the above Step 1, when annotating the dataset with human detection boxes, human key points, and clothing, filter out low-quality data, correct the pre-annotation results, and divide the dataset into a training set, a validation set, and a test set in a ratio of 8:1:

1.

3. The real-time detection method for dressing compliance for advanced manufacturing according to claim 1, characterized in that: In the above Step 1, the human and key point model uses the yolov10-pose model to output the human detection box and 33-point human key points, including the key points of the head, torso, hands, and feet, and uses them to calculate the centroid of the key points corresponding to the clothing parts, and then calculates the distance component in the edge weight of the bipartite graph.

4. The real-time detection method for dressing compliance in advanced manufacturing according to claim 3, characterized in that: The model loss function uses Repulsion loss, and the Repulsion loss function is as follows: L = L Attr + α * L RepGT + β * L RepBox ; Among them, L Attr is the attraction loss term, which makes the predicted bounding box closer to the ground truth bounding box. L Rep is the repulsion loss term, which makes the predicted bounding box away from the surrounding ground truth bounding boxes. α and β are weight balance terms, and their values are both 0.

5.

5. The real-time detection method for dressing compliance for advanced manufacturing according to claim 1, wherein: The specific content of the above Step 2 includes: Filter out low-confidence clothing detection results; Construct a personnel-clothing weighted bipartite graph, perform association processing on the clothing detection results by category, and calculate the edge weights of the weighted bipartite graph; Use the Hungarian K-M algorithm to calculate the optimal matching of the weighted bipartite graph.

6. The real-time detection method for dressing compliance oriented to advanced manufacturing according to claim 5, characterized in that: The specific content of using the Hungarian K-M algorithm to calculate the optimal matching of the weighted bipartite graph includes: In a bipartite graph, the left - hand nodes are the results of clothing detection. An integer value L is assigned to the i - th (1 ≤ i ≤ n) left - hand node xi , and the right - hand nodes are the results of person detection. An integer value L is assigned to the j - th (1 ≤ j ≤ n) right - hand node yj . At the same time, the following conditions are satisfied: L xi +L yj ≥weight (i，j) , where weight (i，j) is the edge weight between the i - th left - hand node and the j - th right - hand node. These integer values L xi 、L yj are called the vertex labels of the nodes. First, on the premise of satisfying L xi +L yj ≥weight (i，j) , an arbitrary vertex label is assigned to each node to expand the scale of the equal - subgraph until a perfect matching exists in the equal - subgraph.

7. The real-time detection method for dressing compliance for advanced manufacturing according to claim 1, characterized in that: The specific content of the above Step 3 includes: Sort the algorithm results in descending order based on the sum of the human box confidence and the average confidence of human key points output by the human and key point algorithm; Check in turn whether the detection boxes are close to the image edge. If the detection box crosses the image edge, delete it; For the remaining detection boxes, calculate the overlap ratio with other detection boxes in sequence. If the overlap ratio is greater than 0.2, delete the detection box with a low average confidence; ensure that the finally remaining human detection results are neither close to the image edge nor have obvious overlap with each other For the remaining personnel detection results, analyze in sequence whether the associated clothing detection results meet the regulations, and output the algorithm detection results.

8. A real-time detection system for dressing compliance in advanced manufacturing, characterized in that: It includes: A data acquisition unit, which is used to collect dense personnel images in different scenarios and poses, construct a personnel clothing dataset for advanced manufacturing, annotate the dataset with human detection boxes, human key points, and clothing, and train a human and key point model and a clothing detection model that can adapt to dense personnel; A data processing unit, which is used to construct a weighted bipartite graph using the human detection results and clothing detection results, calculate the edge weights using the human detection boxes, human key points, and clothing detection boxes, and associate the personnel detection results and clothing detection results in the image by calculating the optimal matching of the weighted bipartite graph; A result output unit is configured to filter out occluded persons in the image according to the human detection result, determine whether the clothing of the non-occluded persons in the image meets the requirements based on the clothing association result, and output the algorithm detection result.

9. An electronic device, comprising a processor and a memory communicatively connected to the processor and configured to store executable instructions of the processor, wherein: The processor is configured to execute a real-time detection method for clothing compliance in advanced manufacturing as described in any one of claims 1-7 above.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements a real-time detection method for clothing compliance in advanced manufacturing as described in any one of claims 1-7.

Images